WO2024259429A2 - Methods and compositions for pharmaceutically relevant interactions - Google Patents

Methods and compositions for pharmaceutically relevant interactions Download PDF

Info

Publication number
WO2024259429A2
WO2024259429A2 PCT/US2024/034369 US2024034369W WO2024259429A2 WO 2024259429 A2 WO2024259429 A2 WO 2024259429A2 US 2024034369 W US2024034369 W US 2024034369W WO 2024259429 A2 WO2024259429 A2 WO 2024259429A2
Authority
WO
WIPO (PCT)
Prior art keywords
pme
cell
compound
compounds
formula
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/US2024/034369
Other languages
French (fr)
Other versions
WO2024259429A3 (en
Inventor
Aaron Esser-Kahn
Yifeng TANG
Andrew L. FERGUSON
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Chicago
Original Assignee
University of Chicago
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Chicago filed Critical University of Chicago
Publication of WO2024259429A2 publication Critical patent/WO2024259429A2/en
Publication of WO2024259429A3 publication Critical patent/WO2024259429A3/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61K—PREPARATIONS FOR MEDICAL, DENTAL OR TOILETRY PURPOSES
    • A61K45/00—Medicinal preparations containing active ingredients not provided for in groups A61K31/00 - A61K41/00
    • A61K45/06—Mixtures of active ingredients without chemical characterisation, e.g. antiphlogistics and cardiaca
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61K—PREPARATIONS FOR MEDICAL, DENTAL OR TOILETRY PURPOSES
    • A61K31/00—Medicinal preparations containing organic active ingredients
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07C—ACYCLIC OR CARBOCYCLIC COMPOUNDS
    • C07C243/00—Compounds containing chains of nitrogen atoms singly-bound to each other, e.g. hydrazines, triazanes
    • C07C243/24—Hydrazines having nitrogen atoms of hydrazine groups acylated by carboxylic acids
    • C07C243/38—Hydrazines having nitrogen atoms of hydrazine groups acylated by carboxylic acids with acylating carboxyl groups bound to carbon atoms of six-membered aromatic rings
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07C—ACYCLIC OR CARBOCYCLIC COMPOUNDS
    • C07C49/00—Ketones; Ketenes; Dimeric ketenes; Ketonic chelates
    • C07C49/76—Ketones containing a keto group bound to a six-membered aromatic ring
    • C07C49/82—Ketones containing a keto group bound to a six-membered aromatic ring containing hydroxy groups
    • C07C49/835—Ketones containing a keto group bound to a six-membered aromatic ring containing hydroxy groups having unsaturation outside an aromatic ring
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07C—ACYCLIC OR CARBOCYCLIC COMPOUNDS
    • C07C57/00—Unsaturated compounds having carboxyl groups bound to acyclic carbon atoms
    • C07C57/02—Unsaturated compounds having carboxyl groups bound to acyclic carbon atoms with only carbon-to-carbon double bonds as unsaturation
    • C07C57/13—Dicarboxylic acids
    • C07C57/145—Maleic acid
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07C—ACYCLIC OR CARBOCYCLIC COMPOUNDS
    • C07C59/00—Compounds having carboxyl groups bound to acyclic carbon atoms and containing any of the groups OH, O—metal, —CHO, keto, ether, groups, groups, or groups
    • C07C59/40—Unsaturated compounds
    • C07C59/58—Unsaturated compounds containing ether groups, groups, groups, or groups
    • C07C59/64—Unsaturated compounds containing ether groups, groups, groups, or groups containing six-membered aromatic rings
    • C07C59/66—Unsaturated compounds containing ether groups, groups, groups, or groups containing six-membered aromatic rings the non-carboxylic part of the ether containing six-membered aromatic rings
    • C07C59/68—Unsaturated compounds containing ether groups, groups, groups, or groups containing six-membered aromatic rings the non-carboxylic part of the ether containing six-membered aromatic rings the oxygen atom of the ether group being bound to a non-condensed six-membered aromatic ring
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D207/00—Heterocyclic compounds containing five-membered rings not condensed with other rings, with one nitrogen atom as the only ring hetero atom
    • C07D207/02—Heterocyclic compounds containing five-membered rings not condensed with other rings, with one nitrogen atom as the only ring hetero atom with only hydrogen or carbon atoms directly attached to the ring nitrogen atom
    • C07D207/18—Heterocyclic compounds containing five-membered rings not condensed with other rings, with one nitrogen atom as the only ring hetero atom with only hydrogen or carbon atoms directly attached to the ring nitrogen atom having one double bond between ring members or between a ring member and a non-ring member
    • C07D207/22—Heterocyclic compounds containing five-membered rings not condensed with other rings, with one nitrogen atom as the only ring hetero atom with only hydrogen or carbon atoms directly attached to the ring nitrogen atom having one double bond between ring members or between a ring member and a non-ring member with hetero atoms or with carbon atoms having three bonds to hetero atoms with at the most one bond to halogen, e.g. ester or nitrile radicals, directly attached to ring carbon atoms
    • C07D207/24—Oxygen or sulfur atoms
    • C07D207/26—2-Pyrrolidones
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D209/00—Heterocyclic compounds containing five-membered rings, condensed with other rings, with one nitrogen atom as the only ring hetero atom
    • C07D209/02—Heterocyclic compounds containing five-membered rings, condensed with other rings, with one nitrogen atom as the only ring hetero atom condensed with one carbocyclic ring
    • C07D209/44—Iso-indoles; Hydrogenated iso-indoles
    • C07D209/48—Iso-indoles; Hydrogenated iso-indoles with oxygen atoms in positions 1 and 3, e.g. phthalimide
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D213/00—Heterocyclic compounds containing six-membered rings, not condensed with other rings, with one nitrogen atom as the only ring hetero atom and three or more double bonds between ring members or between ring members and non-ring members
    • C07D213/02—Heterocyclic compounds containing six-membered rings, not condensed with other rings, with one nitrogen atom as the only ring hetero atom and three or more double bonds between ring members or between ring members and non-ring members having three double bonds between ring members or between ring members and non-ring members
    • C07D213/04—Heterocyclic compounds containing six-membered rings, not condensed with other rings, with one nitrogen atom as the only ring hetero atom and three or more double bonds between ring members or between ring members and non-ring members having three double bonds between ring members or between ring members and non-ring members having no bond between the ring nitrogen atom and a non-ring member or having only hydrogen or carbon atoms directly attached to the ring nitrogen atom
    • C07D213/60—Heterocyclic compounds containing six-membered rings, not condensed with other rings, with one nitrogen atom as the only ring hetero atom and three or more double bonds between ring members or between ring members and non-ring members having three double bonds between ring members or between ring members and non-ring members having no bond between the ring nitrogen atom and a non-ring member or having only hydrogen or carbon atoms directly attached to the ring nitrogen atom with hetero atoms or with carbon atoms having three bonds to hetero atoms with at the most one bond to halogen, e.g. ester or nitrile radicals, directly attached to ring carbon atoms
    • C07D213/62—Oxygen or sulfur atoms
    • C07D213/63—One oxygen atom
    • C07D213/64—One oxygen atom attached in position 2 or 6
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D215/00—Heterocyclic compounds containing quinoline or hydrogenated quinoline ring systems
    • C07D215/02—Heterocyclic compounds containing quinoline or hydrogenated quinoline ring systems having no bond between the ring nitrogen atom and a non-ring member or having only hydrogen atoms or carbon atoms directly attached to the ring nitrogen atom
    • C07D215/16—Heterocyclic compounds containing quinoline or hydrogenated quinoline ring systems having no bond between the ring nitrogen atom and a non-ring member or having only hydrogen atoms or carbon atoms directly attached to the ring nitrogen atom with hetero atoms or with carbon atoms having three bonds to hetero atoms with at the most one bond to halogen, e.g. ester or nitrile radicals, directly attached to ring carbon atoms
    • C07D215/38—Nitrogen atoms
    • C07D215/40—Nitrogen atoms attached in position 8
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D215/00—Heterocyclic compounds containing quinoline or hydrogenated quinoline ring systems
    • C07D215/02—Heterocyclic compounds containing quinoline or hydrogenated quinoline ring systems having no bond between the ring nitrogen atom and a non-ring member or having only hydrogen atoms or carbon atoms directly attached to the ring nitrogen atom
    • C07D215/16—Heterocyclic compounds containing quinoline or hydrogenated quinoline ring systems having no bond between the ring nitrogen atom and a non-ring member or having only hydrogen atoms or carbon atoms directly attached to the ring nitrogen atom with hetero atoms or with carbon atoms having three bonds to hetero atoms with at the most one bond to halogen, e.g. ester or nitrile radicals, directly attached to ring carbon atoms
    • C07D215/38—Nitrogen atoms
    • C07D215/42—Nitrogen atoms attached in position 4
    • C07D215/46—Nitrogen atoms attached in position 4 with hydrocarbon radicals, substituted by nitrogen atoms, attached to said nitrogen atoms
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D219/00—Heterocyclic compounds containing acridine or hydrogenated acridine ring systems
    • C07D219/04—Heterocyclic compounds containing acridine or hydrogenated acridine ring systems with hetero atoms or with carbon atoms having three bonds to hetero atoms with at the most one bond to halogen, e.g. ester or nitrile radicals, directly attached to carbon atoms of the ring system
    • C07D219/08—Nitrogen atoms
    • C07D219/10—Nitrogen atoms attached in position 9
    • C07D219/12—Amino-alkylamino radicals attached in position 9
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D231/00—Heterocyclic compounds containing 1,2-diazole or hydrogenated 1,2-diazole rings
    • C07D231/02—Heterocyclic compounds containing 1,2-diazole or hydrogenated 1,2-diazole rings not condensed with other rings
    • C07D231/10—Heterocyclic compounds containing 1,2-diazole or hydrogenated 1,2-diazole rings not condensed with other rings having two or three double bonds between ring members or between ring members and non-ring members
    • C07D231/14—Heterocyclic compounds containing 1,2-diazole or hydrogenated 1,2-diazole rings not condensed with other rings having two or three double bonds between ring members or between ring members and non-ring members with hetero atoms or with carbon atoms having three bonds to hetero atoms with at the most one bond to halogen, e.g. ester or nitrile radicals, directly attached to ring carbon atoms
    • C07D231/38—Nitrogen atoms
    • C07D231/42—Benzene-sulfonamido pyrazoles
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D239/00—Heterocyclic compounds containing 1,3-diazine or hydrogenated 1,3-diazine rings
    • C07D239/70—Heterocyclic compounds containing 1,3-diazine or hydrogenated 1,3-diazine rings condensed with carbocyclic rings or ring systems
    • C07D239/72—Quinazolines; Hydrogenated quinazolines
    • C07D239/86—Quinazolines; Hydrogenated quinazolines with hetero atoms directly attached in position 4
    • C07D239/94—Nitrogen atoms
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D251/00—Heterocyclic compounds containing 1,3,5-triazine rings
    • C07D251/02—Heterocyclic compounds containing 1,3,5-triazine rings not condensed with other rings
    • C07D251/12—Heterocyclic compounds containing 1,3,5-triazine rings not condensed with other rings having three double bonds between ring members or between ring members and non-ring members
    • C07D251/14—Heterocyclic compounds containing 1,3,5-triazine rings not condensed with other rings having three double bonds between ring members or between ring members and non-ring members with hydrogen or carbon atoms directly attached to at least one ring carbon atom
    • C07D251/16—Heterocyclic compounds containing 1,3,5-triazine rings not condensed with other rings having three double bonds between ring members or between ring members and non-ring members with hydrogen or carbon atoms directly attached to at least one ring carbon atom to only one ring carbon atom
    • C07D251/18—Heterocyclic compounds containing 1,3,5-triazine rings not condensed with other rings having three double bonds between ring members or between ring members and non-ring members with hydrogen or carbon atoms directly attached to at least one ring carbon atom to only one ring carbon atom with nitrogen atoms directly attached to the two other ring carbon atoms, e.g. guanamines
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D267/00—Heterocyclic compounds containing rings of more than six members having one nitrogen atom and one oxygen atom as the only ring hetero atoms
    • C07D267/02—Seven-membered rings
    • C07D267/08—Seven-membered rings having the hetero atoms in positions 1 and 4
    • C07D267/12—Seven-membered rings having the hetero atoms in positions 1 and 4 condensed with carbocyclic rings or ring systems
    • C07D267/16—Seven-membered rings having the hetero atoms in positions 1 and 4 condensed with carbocyclic rings or ring systems condensed with two six-membered rings
    • C07D267/20—[b, f]-condensed
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D279/00—Heterocyclic compounds containing six-membered rings having one nitrogen atom and one sulfur atom as the only ring hetero atoms
    • C07D279/10—1,4-Thiazines; Hydrogenated 1,4-thiazines
    • C07D279/14—1,4-Thiazines; Hydrogenated 1,4-thiazines condensed with carbocyclic rings or ring systems
    • C07D279/18—[b, e]-condensed with two six-membered rings
    • C07D279/22—[b, e]-condensed with two six-membered rings with carbon atoms directly attached to the ring nitrogen atom
    • C07D279/24—[b, e]-condensed with two six-membered rings with carbon atoms directly attached to the ring nitrogen atom with hydrocarbon radicals, substituted by amino radicals, attached to the ring nitrogen atom
    • C07D279/28—[b, e]-condensed with two six-membered rings with carbon atoms directly attached to the ring nitrogen atom with hydrocarbon radicals, substituted by amino radicals, attached to the ring nitrogen atom with other substituents attached to the ring system
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D307/00—Heterocyclic compounds containing five-membered rings having one oxygen atom as the only ring hetero atom
    • C07D307/02—Heterocyclic compounds containing five-membered rings having one oxygen atom as the only ring hetero atom not condensed with other rings
    • C07D307/34—Heterocyclic compounds containing five-membered rings having one oxygen atom as the only ring hetero atom not condensed with other rings having two or three double bonds between ring members or between ring members and non-ring members
    • C07D307/56—Heterocyclic compounds containing five-membered rings having one oxygen atom as the only ring hetero atom not condensed with other rings having two or three double bonds between ring members or between ring members and non-ring members with hetero atoms or with carbon atoms having three bonds to hetero atoms with at the most one bond to halogen, e.g. ester or nitrile radicals, directly attached to ring carbon atoms
    • C07D307/70—Nitro radicals
    • C07D307/71—Nitro radicals attached in position 5
    • C07D307/72—Nitro radicals attached in position 5 with hydrocarbon radicals, substituted by nitrogen-containing radicals, attached in position 2
    • C07D307/74—Nitro radicals attached in position 5 with hydrocarbon radicals, substituted by nitrogen-containing radicals, attached in position 2 by hydrazino or hydrazono or such substituted radicals
    • C07D307/76—Nitro radicals attached in position 5 with hydrocarbon radicals, substituted by nitrogen-containing radicals, attached in position 2 by hydrazino or hydrazono or such substituted radicals having carbonic acyl radicals or their thio or nitrogen analogues directly attached to the hydrazino or hydrazono radical, e.g. semicarbazides
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D309/00—Heterocyclic compounds containing six-membered rings having one oxygen atom as the only ring hetero atom, not condensed with other rings
    • C07D309/02—Heterocyclic compounds containing six-membered rings having one oxygen atom as the only ring hetero atom, not condensed with other rings having no double bonds between ring members or between ring members and non-ring members
    • C07D309/08—Heterocyclic compounds containing six-membered rings having one oxygen atom as the only ring hetero atom, not condensed with other rings having no double bonds between ring members or between ring members and non-ring members with hetero atoms or with carbon atoms having three bonds to hetero atoms with at the most one bond to halogen, e.g. ester or nitrile radicals, directly attached to ring carbon atoms
    • C07D309/14—Nitrogen atoms not forming part of a nitro radical
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D401/00—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, at least one ring being a six-membered ring with only one nitrogen atom
    • C07D401/02—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, at least one ring being a six-membered ring with only one nitrogen atom containing two hetero rings
    • C07D401/04—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, at least one ring being a six-membered ring with only one nitrogen atom containing two hetero rings directly linked by a ring-member-to-ring-member bond
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D401/00—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, at least one ring being a six-membered ring with only one nitrogen atom
    • C07D401/02—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, at least one ring being a six-membered ring with only one nitrogen atom containing two hetero rings
    • C07D401/06—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, at least one ring being a six-membered ring with only one nitrogen atom containing two hetero rings linked by a carbon chain containing only aliphatic carbon atoms
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D401/00—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, at least one ring being a six-membered ring with only one nitrogen atom
    • C07D401/02—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, at least one ring being a six-membered ring with only one nitrogen atom containing two hetero rings
    • C07D401/10—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, at least one ring being a six-membered ring with only one nitrogen atom containing two hetero rings linked by a carbon chain containing aromatic rings
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D403/00—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, not provided for by group C07D401/00
    • C07D403/02—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, not provided for by group C07D401/00 containing two hetero rings
    • C07D403/12—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, not provided for by group C07D401/00 containing two hetero rings linked by a chain containing hetero atoms as chain links
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D403/00—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, not provided for by group C07D401/00
    • C07D403/14—Heterocyclic compounds containing two or more hetero rings, having nitrogen atoms as the only ring hetero atoms, not provided for by group C07D401/00 containing three or more hetero rings
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D473/00—Heterocyclic compounds containing purine ring systems
    • C07D473/26—Heterocyclic compounds containing purine ring systems with an oxygen, sulphur, or nitrogen atom directly attached in position 2 or 6, but not in both
    • C07D473/32—Nitrogen atom
    • C07D473/34—Nitrogen atom attached in position 6, e.g. adenine
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D487/00—Heterocyclic compounds containing nitrogen atoms as the only ring hetero atoms in the condensed system, not provided for by groups C07D451/00 - C07D477/00
    • C07D487/02—Heterocyclic compounds containing nitrogen atoms as the only ring hetero atoms in the condensed system, not provided for by groups C07D451/00 - C07D477/00 in which the condensed system contains two hetero rings
    • C07D487/04—Ortho-condensed systems
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D491/00—Heterocyclic compounds containing in the condensed ring system both one or more rings having oxygen atoms as the only ring hetero atoms and one or more rings having nitrogen atoms as the only ring hetero atoms, not provided for by groups C07D451/00 - C07D459/00, C07D463/00, C07D477/00 or C07D489/00
    • C07D491/12—Heterocyclic compounds containing in the condensed ring system both one or more rings having oxygen atoms as the only ring hetero atoms and one or more rings having nitrogen atoms as the only ring hetero atoms, not provided for by groups C07D451/00 - C07D459/00, C07D463/00, C07D477/00 or C07D489/00 in which the condensed system contains three hetero rings
    • C07D491/14—Ortho-condensed systems
    • C07D491/147—Ortho-condensed systems the condensed system containing one ring with oxygen as ring hetero atom and two rings with nitrogen as ring hetero atom
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D495/00—Heterocyclic compounds containing in the condensed system at least one hetero ring having sulfur atoms as the only ring hetero atoms
    • C07D495/02—Heterocyclic compounds containing in the condensed system at least one hetero ring having sulfur atoms as the only ring hetero atoms in which the condensed system contains two hetero rings
    • C07D495/04—Ortho-condensed systems
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07D—HETEROCYCLIC COMPOUNDS
    • C07D513/00—Heterocyclic compounds containing in the condensed system at least one hetero ring having nitrogen and sulfur atoms as the only ring hetero atoms, not provided for in groups C07D463/00, C07D477/00 or C07D499/00 - C07D507/00
    • C07D513/02—Heterocyclic compounds containing in the condensed system at least one hetero ring having nitrogen and sulfur atoms as the only ring hetero atoms, not provided for in groups C07D463/00, C07D477/00 or C07D499/00 - C07D507/00 in which the condensed system contains two hetero rings
    • C07D513/04—Ortho-condensed systems
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07F—ACYCLIC, CARBOCYCLIC OR HETEROCYCLIC COMPOUNDS CONTAINING ELEMENTS OTHER THAN CARBON, HYDROGEN, HALOGEN, OXYGEN, NITROGEN, SULFUR, SELENIUM OR TELLURIUM
    • C07F9/00—Compounds containing elements of Groups 5 or 15 of the Periodic Table
    • C07F9/02—Phosphorus compounds
    • C07F9/547—Heterocyclic compounds, e.g. containing phosphorus as a ring hetero atom
    • C07F9/6561—Heterocyclic compounds, e.g. containing phosphorus as a ring hetero atom containing systems of two or more relevant hetero rings condensed among themselves or condensed with a common carbocyclic ring or ring system, with or without other non-condensed hetero rings
    • C07F9/65616—Heterocyclic compounds, e.g. containing phosphorus as a ring hetero atom containing systems of two or more relevant hetero rings condensed among themselves or condensed with a common carbocyclic ring or ring system, with or without other non-condensed hetero rings containing the ring system having three or more than three double bonds between ring members or between ring members and non-ring members, e.g. purine or analogs
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07H—SUGARS; DERIVATIVES THEREOF; NUCLEOSIDES; NUCLEOTIDES; NUCLEIC ACIDS
    • C07H15/00—Compounds containing hydrocarbon or substituted hydrocarbon radicals directly attached to hetero atoms of saccharide radicals
    • C07H15/20—Carbocyclic rings
    • C07H15/24—Condensed ring systems having three or more rings
    • C07H15/256—Polyterpene radicals
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07H—SUGARS; DERIVATIVES THEREOF; NUCLEOSIDES; NUCLEOTIDES; NUCLEIC ACIDS
    • C07H19/00—Compounds containing a hetero ring sharing one ring hetero atom with a saccharide radical; Nucleosides; Mononucleotides; Anhydro-derivatives thereof
    • C07H19/02—Compounds containing a hetero ring sharing one ring hetero atom with a saccharide radical; Nucleosides; Mononucleotides; Anhydro-derivatives thereof sharing nitrogen
    • C07H19/04—Heterocyclic radicals containing only nitrogen atoms as ring hetero atom
    • C07H19/06—Pyrimidine radicals
    • C07H19/067—Pyrimidine radicals with ribosyl as the saccharide radical
    • C—CHEMISTRY; METALLURGY
    • C07—ORGANIC CHEMISTRY
    • C07H—SUGARS; DERIVATIVES THEREOF; NUCLEOSIDES; NUCLEOTIDES; NUCLEIC ACIDS
    • C07H19/00—Compounds containing a hetero ring sharing one ring hetero atom with a saccharide radical; Nucleosides; Mononucleotides; Anhydro-derivatives thereof
    • C07H19/02—Compounds containing a hetero ring sharing one ring hetero atom with a saccharide radical; Nucleosides; Mononucleotides; Anhydro-derivatives thereof sharing nitrogen
    • C07H19/04—Heterocyclic radicals containing only nitrogen atoms as ring hetero atom
    • C07H19/12—Triazine radicals
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/48—Biological material, e.g. blood, urine; Haemocytometers
    • G01N33/50—Chemical analysis of biological material, e.g. blood, urine; Testing involving biospecific ligand binding methods; Immunological testing
    • G01N33/5005—Chemical analysis of biological material, e.g. blood, urine; Testing involving biospecific ligand binding methods; Immunological testing involving human or animal cells
    • G01N33/5008—Chemical analysis of biological material, e.g. blood, urine; Testing involving biospecific ligand binding methods; Immunological testing involving human or animal cells for testing or evaluating the effect of chemical or biological compounds, e.g. drugs, cosmetics
    • G01N33/5044—Chemical analysis of biological material, e.g. blood, urine; Testing involving biospecific ligand binding methods; Immunological testing involving human or animal cells for testing or evaluating the effect of chemical or biological compounds, e.g. drugs, cosmetics involving specific cell types
    • G01N33/5047—Cells of the immune system
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]

Definitions

  • This invention relates to the field of chemical biology, medicinal chemistry, and machine learning.
  • helper molecules known as adjuvants are often required to stimulate innate pathways involved in antigen presentation and processing that are critical in invoking a productive adaptive immune response.
  • adjuvants are often required to stimulate innate pathways involved in antigen presentation and processing that are critical in invoking a productive adaptive immune response.
  • signaling events Despite the necessity of such signaling events to maximize the potency, excessive activation of signaling pathways by adjuvants can cause undesirable systemic inflammation, and limit tolerability and dosage in a clinical setting.
  • NF-KB nuclear factor K-light- chainenhancer of activated B-cells
  • IRF interferon regulatory factors
  • the NF-KB pathway plays an essential role in inflammation as well as immune activation, while the IRF pathway produces type-I interferons that are essential for a productive antiviral response.
  • PRRs pattern recognition receptors
  • PAMPs pathogen-associated molecular patterns
  • PRRs are necessary for the activation of antigen-presenting cells (APCs) that act as a link between the innate and adaptive immune responses and play a critical role in detecting and responding to pathogens.
  • APCs antigen-presenting cells
  • 10 PRR agonists are molecules that bind to PRRs, mimicking the effects of pathogenic molecules and triggering an immune response.
  • PRR agonists have recently been used as adjuvants to activate both NF-KB and IRF pathways and are the most common targets for manipulating the innate immune response.i l
  • a breadth of PRR agonist-based adjuvants comprised of pathogenic motifs have been used as adjuvants in vaccines and immunotherapies, such as lipopolysaccharides (LPS) 12 and Monophosphoryl Lipid A (MPLA) 13 that target toll-like receptor (TLR) 4, synthetic oligodeoxynucleotides that contain unmethylated cytosine- phosphate-guanine dinucleotide motifs (CpG-ODN) targeting TLR 9 14 cyclic guanosine monophosphate- adenosine monophosphate (cyclic GMPAMP, cGAMP) that binds and activates the stimulator of interferon genes (STING), 15 polyriboisosinic:polyribocytidylic acid [poly(LC)] that are
  • SN50 is a cell permeable peptide that consists of nuclear localization sequence of the NF-KB subunits p50 and blocks the import of p50- containing dimers into the nucleus.
  • SN50 can reduce the levels of inflammatory cytokines TNF-a and IL-6 while enhancing antigen- specific antibody titers when delivered with the TLR9 agonist CpG. 20
  • small molecules present attractive candidates for immunomodulators due to their better synthetic accessibility and reduced potential for immunogenicity.
  • Moser et al. 21 also demonstrated that a small molecule NF-KB inhibitor honokiol and its derivatives can be used as immunomodulators with similar functions to SN50.
  • aspects of the disclosure relate to the discovery that the integration of laboratory screening data with machine learning methods can lead to an efficient drug identifying algorithm. Further aspects of the disclosure relate to compositions identified using the integrated algorithm, and methods of using the compositions for treating a disease.
  • compositions comprising one or more of the compounds disclosed herein.
  • the pharmaceutical composition further comprises one or more pattern recognition receptor agonists.
  • the pattern recognition receptor agonists may target one or more of TLR4, TLR9, and STING.
  • the pattern recognition receptor agonists may be one or more of LPS, MPLA, CpG, and cGAMP.
  • the method comprises delivering to the cell an effective amount of at least one compound disclosed herein, including one or more of PME-4855, PME-4426, PME-4119, PME-3974, and PME- 5149, or any combination thereof.
  • the effective amount results in a concentration of about, at least, or at most, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 pM.
  • the concentration may be a concentration in the cell and/or a concentration in an environment surrounding the cell.
  • the method further comprises delivering to the cell an effective amount of LPS, MPLA, and/or CpG.
  • the method further comprises delivering to the cell a pattern recognition agonist.
  • the cell is an immune cell.
  • the method comprises delivering to the cell an effective amount of one or more compounds disclosed herein including one or more of PME-4637, PME-4800, PME-5839, PME-5084, PME-4974, PME-4873, PME-5246, PME-4695, PME-3465, PME-4633, PME-4392, and PME-5071, or any combination thereof.
  • the effective amount results in a concentration of about, at least, or at most, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 pM.
  • the concentration may be a concentration in the cell and/or a concentration in an environment surrounding the cell.
  • the method further comprises delivering to the cell an effective amount of LPS, MPLA, and/or CpG.
  • the method further comprises delivering to the cell a pattern recognition agonist.
  • the cell is an immune cell.
  • the method comprises delivering to the cell an effective amount of at least one compound disclosed herein, including one or more of PME-4855, PME-4426, PME-4119, PME-3974, PME-5149, PME-4637, PME-4800, PME-5839, PME-5084, PME-4974, PME-4873, PME-5246, PME- 4695, PME-3465, PME-4633, PME-4392, and PME-5071, or any combination thereof.
  • the effective amount results in a concentration of about, at least, or at most, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 pM.
  • the concentration may be a concentration in the cell and/or a concentration in an environment surrounding the cell.
  • the method further comprises delivering to the cell an effective amount of LPS, MPLA, and/or CpG.
  • the method further comprises delivering to the cell a pattern recognition agonist.
  • the cell is an immune cell.
  • Also disclosed are methods of modulating one or more immune-related pathways in a cell comprising delivering to the cell an effective amount of one or more of PME-4855, PME-4426, PME-4119, PME-3974, PME-5149, PME-4637, PME-4800, PME- 5839, PME-5084, PME-4974, PME-4873, PME-5246, PME-4695, PME-3465, PME-4633, PME-4392, and PME-5071, or any combination thereof.
  • the effective amount results in a concentration of about, at least, or at most, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 pM.
  • the concentration may be a concentration in the cell and/or a concentration in an environment surrounding the cell.
  • the method further comprises delivering to the cell an effective amount of LPS, MPLA, and/or CpG.
  • the method further comprises delivering to the cell a pattern recognition agonist.
  • the cell is an immune cell.
  • the method can comprise one or more steps including administering to the patient at least one of the compounds described herein and/or at least one of the pharmaceutical compositions described herein, administering a vaccine to the patient, administering an additional therapeutic composition to the patient, and/or administering an anti-viral composition and/or an anti-cancer composition.
  • pro-inflammatory cytokines act antagonistically to the additional therapeutic composition.
  • the patient has, has been diagnosed with, or is suspected of having cancer, an infection, and/or an autoimmune disease.
  • Also disclosed are methods comprising administering to a human patient at least one of the compounds described herein, at least one of the pharmaceutical composition described herein, or any combination thereof.
  • the method comprises 1, 2, 3, 4, 5, or more steps including: generating a fixed-length input feature vector for each compound in a plurality compounds of interest; compiling each fixed-length input feature vectors into a latent space; receiving high-throughput screening data for each compound of the plurality of compounds of interest; associating the fixed-length input feature vectors in the latent space with the high- throughput screening data to generate a set of association information; and training a machine learning model by iteratively minimizing error to within a predetermined threshold using the set of association information.
  • the method further comprises identifying a second plurality of compounds of interest that is refined from the plurality of compounds of interest based on the trained machine learning model. In some aspects, the method further comprises predicting drug-target interactions for one or more compounds that are not in the plurality of compounds. In some aspects, the method further comprises predicting activity against the target for one or more compounds that are not in the plurality of compounds. In some aspects, the activity is agonistic activity. In some aspects, the activity is antagonistic activity. In some aspects, the activity comprises affinity to the target.
  • the plurality of compounds comprises a library of drug candidates. In some aspects, the plurality of compounds are used in an assay to generate the high- throughput screening data. In some aspects, only a subset of the plurality of compounds are used to generate the high-throughput training data.
  • the method further comprises performing one or more iterations of the method, wherein a subsequent iteration uses the second plurality of compounds of interest as the plurality of compounds of interest.
  • a subsequent iteration uses the second plurality of compounds of interest as the plurality of compounds of interest.
  • the method may be performed once to identify a set of compounds, that set is then used in a second iteration of the method to identify another set of compounds. This process can be repeated multiple times in order to identify candidate molecules.
  • the set of compounds and/or the candidate molecules are then used in a high-throughput screen, which generates data that is then used in another iteration of the method.
  • the high-throughput screening data comprises a signal from a plurality of vessels, wherein each vessel of the plurality of vessels comprises one compound of the compounds of interest and a target.
  • the plurality of vessel may be a well plate, such as a 384-well plate.
  • the target is a protein of interest.
  • the target is in a cell.
  • the cell comprises a reporter system capable of producing the signal.
  • the signal is fluorescence produced by the cell.
  • the high-throughput screening data does not comprise data from compounds of interest that are cytostatic or cytotoxic to the cell.
  • the generating a fixed-length input feature vector is performed by a variational autoencoder.
  • the variational autoencoder is trained by feedforward network architecture.
  • the compiling is performed via a feed forward architecture to generate a defined-node layer defining the latent space.
  • the feed forward architecture is a 500-200-100 fully-connected feed forward architecture.
  • the defined-node layer is a 100-node layer.
  • the variational autoencoder is trained with the plurality of compounds of interest and an additional set of compounds. The additional set of compounds may be compounds from a known library, such as a commercial library including the ZINC library.
  • the additional set may comprise at least, at most, or approximately 1,000, 2,000, 3,000, 4,000, 5,000, 6,000, 7,000, 8,000, 9,000, 10,000, 50,000, 100,000, 200,000, 300,000, 400,000, 500,000, 600,000, 700,000, 800,000, 900,000, 1,00,000 known compounds (or any range derivable therein).
  • the associating comprises an empirical mapping, which may be an automated empirical mapping, of each coordinate of each input vector to at least one biological response measured in the high-throughput screening data.
  • the empirical mapping comprises an amount of calculations that cannot be done by hand.
  • the training the machine learning model comprises training at least one Gaussian procession regression models for each measured biological response in the high-throughput screening data.
  • the method further comprises decoding the latent space into a plurality of interpretable compound vectors.
  • the decoding is performed by a variational autoencoder.
  • the decoding is performed by at least one gated recurrent units.
  • Also disclosed are methods for identifying a compound pharmacologically active against a target of interest comprising 1, 2, 3, 4, 5, 6, or more steps including: generating a fixed-length input feature vector for each compound in a plurality compounds of interest; compiling each fixed-length input feature vectors into a latent space; receiving high-throughput screening data for each compound in the plurality of compounds of interest; associating the high-throughput screening data with the latent space to generate a set of association information; applying, into a trained machine learning model, the latent space to generate an output feature vector predicting the pharmacological activity against a target of interest for each compound in the plurality of compounds of interest; and identifying one or more compounds with pharmacological activity over a determined threshold based on the output feature vector.
  • raw readings of a reporter of biological activity such as absorbance and/or luminescence from a reporter molecule, produced in each well are divided by the average reading of the positive controls (which may be the presence of agonists and absence of immunomodulators) which are on the same plate to define the fold change associated with each modulator relative to the baseline of the corresponding agonist.
  • A, B, and/or C includes: A alone, B alone, C alone, a combination of A and B, a combination of A and C, a combination of B and C, or a combination of A, B, and C. In other words, “and/or” operates as an inclusive or.
  • compositions and methods for their use can “comprise,” “consist essentially of,” or “consist of’ any of the ingredients or steps disclosed throughout the specification. Compositions and methods “consisting essentially of’ any of the ingredients or steps disclosed limits the scope of the claim to the specified materials or steps which do not materially affect the basic and novel characteristic of the claimed invention.
  • FIGs. 1A-1E show data-driven active learning framework for immunomodulator discovery.
  • Immunomodulator candidates are subjected to in vitro high throughput screening (HTS) using automated liquid handling platforms.
  • HTS high throughput screening
  • VAE Deep representational learning with a variational autoencoder
  • VAE Deep representational learning with a variational autoencoder
  • a supervised surrogate Gaussian process regression (GPR) model is trained to predict the immunomodulation of transcription factor levels in the NF-KB and IRF path- ways from all accumulated experimental measurements to date.
  • GPR Gaussian process regression
  • (D) The trained surrogate model is interrogated using Bayesian optimization (BO) to select the most promising batch of as-yet- untested immunomodulators for the next round of experimental screening. Experimental measurements are then used to retrain and update the GPR models and inform subsequent rounds of BO candidate selection within a virtuous cycle of QSAR model training and model- guided experimental screening.
  • NF-KB generalist capable of enhancement or suppression of the NF-K pathway when delivered in concert with any one of the LPS, MPLA, or CpG agonists
  • IRF generalist capable of enhancing the IRF pathway when delivered in concert with any one of the LPS, MPLA, or cGAMP agonists.
  • FIGs. 2A-2C show results of the data-driven immunomodulator discovery active learn- ing screen.
  • A Violin plots of the active learning screen over the 12 functional objectives over the eight agonist-pathway combinations of interest. Fold changes in the NF- KB and IRF responses are measured relative to delivery of the agonist alone.
  • Each violin for Round 1-4 contains the 720 molecules experimentally assayed in each round along the eight agonistpathway combinations, where as the violin for Round 0 denotes the 2674 molecules experimentally assayed prior to the active learning-assisted screen.
  • the orange and purple lines show, respectively, the top enhancer and inhibitor identified by the mean of the two independent experimental measurements, and the shading shows the range of the two measurements.
  • the NF-K generalist comprises the LPS, MPLA, and CpG agonists
  • the IRF generalist comprises the LPS, MPLA and cGAMP agonists
  • the MAE of the terminal model is reported on the right of last two panels with error bars estimated by 5-fold cross validation.
  • FIGs. 3A-3D show Measuring the progress of the active learning screen over 2D projections of the 100-dimensional latent space.
  • A Probability density function estimated by kernel density estimation of a projection of the 142,672 molecules, consisting of 139,998 candidate molecules and 2,674 molecules from previous study 24 into a 2D t-distributed Stochastic Neighbor Embedding (t-SNE) embedding of the 100D VAE latent space.
  • t-SNE 2D t-distributed Stochastic Neighbor Embedding
  • FIGs. 4A-4B show top-performing immunomodulator candidates.
  • A The two topperforming immunomodulator candidates in each of the 12 functional objectives. We present for each molecule its chemical structure along with their code names. A bar chart shows the ex- perimentally measured log2-fold change in the immunomodulatory profile along all eight agonist-pathway combinations of interest. We highlight the immunomodulatory property that makes the candidate highly ranked in terms of activity enhancement (red) or sup- pression (blue). The 17 candidates with names high-lighted in bold text were selected for additional characterization of their cytokine release profiles (note that PME-3465 and PME-5839 each appear top-ranked twice).
  • B An additional eight candidates with outstanding immunomodulatory profiles that were selected for additional cytokine characterization.
  • FIGs. 5A-5H show immunomodulator design rules for each of the eight agonistpathway combinations exposed by LASSO regression. Illustration of the up to six largest magnitude non-zero regression coefficients for LASSO linear regression models to predict the log2-fold change in immunomodulatory activity as a function of the presence or absence of particular molecular fragments or functional groups.
  • the features with positive weights are displayed in black text, while the ones with negative weights are displayed in red text. Positive weights imply that there is a positive correlation between the feature values and enhanced immunomodulation, while negative weights indicate a positive correlation between the feature values and inhibitory immunomodulation.
  • FIGs. 6A-6B show low-throughput measurement of cytokine release profiles within pri- mary cells of 17 top-performing candidates.
  • A Immunomodulation of 17 selected top-performing candidates over the release profiles of 13 cytokines activated by LPS, CpG and cGAMP, shown as suppression and enhancement, repectively. The extent of immunomodulation is visualized as color-maps depicting the log2-fold change values, in form of heat-maps contrasting corresponding cytokines and immunomodulator indices.
  • FIG. 7 shows architecture of VAE for deep representational learning of molecular latent space.
  • the input and output of the VAE model are 7535-element one-hot vector that are one-to-one with the SELFIES string representation of each molecule.
  • the encoder stack comprises three fully-connected feedforward layers with a 500-200-100 architecture.
  • a bottleneck layer comprising 100 neurons defines the 100D latent space embedding.
  • the decoder comprises three stacked GRUs each containing 200 neurons.
  • FIGs. 8A-8C shows VAE training curves over the augmented ZINC training library.
  • A VAE loss function (Equation SI).
  • B KLD component of VAE loss function (Equation S3, second term). The dashed vertical lines in each panel denote the epochs at which the KLD loss coefficient a was tuned from an initial value of 10-2 to 10-3 then 10-4.
  • C Exact reconstruction accuracy reporting the fraction of molecules in the training data whose SELFIES representations are exactly reconstructed by the trained network.
  • FIG. 9 shows correlation heatmap and cluster dendrogram of highly correlated descriptors.
  • Clustering analyses were conducted to organize highly correlated descriptors into larger groups/families. In this way, we identified six de- scriptor groups with high correlation (red blocks shown in the clustered heat map). For each of the six descriptor groups, we select one representative descriptor. Thus, we retained six descriptors, namely “fr_Al_OH”, “fr_Ar_OH”, “ ‘fr_phos_acid”, “fr_Ar_NH”, “fr_COO”, “fr_nitro”.
  • FIG. 10 shows the performance of LASSO regression models evaluated with respect to sparsity regularization parameter //,.
  • the plot presents the number of molecular descriptors that have non-zero learned coefficient values, which were identified by training the model using a range of //. values (displayed on the left y-axis in orange). Additionally, the plot shows the 5-fold cross-validation mean absolute error (MAE) score of the LASSO regression model at each corresponding //, value (displayed on the right y-axis in purple).
  • the optimal // is shown in red points where the corresponding MAE is the lowest.
  • the title of each plot denotes the corresponding immune signaling pathway and agonist used for stimulation.
  • the MAE of the model with the optimal // is reported in the upper-right comer of each panel with error bars estimated by 5-fold cross validation. In all cases, the MAE of the LASSO regression models are poorer than those for the corresponding GPR (c/. FIG. 2), but the performance differential is at most only 32.3%.
  • FIG. 11 shows full accounting of nonzero learned coefficient weights 0k associated with substructure features ranked by their magnitudes of weights in units of log2 fold change.
  • F n ,k a feature matrix represented as F n ,k.
  • F n ,k a feature matrix represented as F n ,k.
  • LASSO regression we apply LASSO regression to predict the calculated immunomodulation acquired from high throughput screening experiments and identifying the learned nonzero coefficient weights Ok with the highest magnitude from the reduced feature set retained by the LASSO model.
  • the features with positive weights are displayed in black text, while the ones with negative weights are displayed in red text. Positive weights imply that there is a positive correlation between the feature values and enhanced immunomodulation, while negative weights indicate a positive correlation between the feature values and inhibitory immunomodulation.
  • FIG. 12 shows results of active learning screen as raw data. NF-KB activity was measured as absorbance readings at 620 nm, and IRF activity was measured as luminescence readings in the units of relative light unit (RLU). The higher the absorbance or luminescence, the stronger the corresponding immune activity.
  • the CpG agonist shows minimal stimulation of the IRF pathway relative to no agonist present and was dropped from subsequent analyses.
  • FIG. 13 shows modulators (10/rM) alone minimally affects cytokine production 24 hours after addition. Each panel shows the secreted concentration for a specific cytokine stimulated by PBS (as negative control) and 17 selected top-performing immunomodulator candidates (without the addition of agonists), addressed for in total 13 cytokines.
  • the cytokine production result from the addition of modulators in the absence of agonists is within the same order of magnitude as that of the PBS negative control, showing that the addition of modulators alone minimally affect the production of cytokines.
  • the magnitude of the amount of IL-6 being released is still minimal immunologically.
  • the cytokine production is measured by LegendPlex.
  • FIG. 14 shows PME-4007 induces minimal cytotoxicity while enhancing the secretion of IL-6 and IL-27 stimulated by cGAMP.
  • A MTT Assay indicates that MSA-2, cGAMP, and the combination of cGAMP and PME-4007 induce minimal cytotoxicity to the cells as the cell viability is all higher than the cutoff 70%, and close to the level of negative control resting cells.
  • B PME-4007 slightly enhances the secretion of IL-6.
  • MSA-2 strongly stimulates the secretion of IL-6 and is much stronger than the stimulation induced by cGAMP.
  • C PME-4007 significantly enhances the secretion of IL-27 with cGAMP, and it is close to the level of stimulation induced by MSA-2.
  • FIG. 15 shows maximum pairwise Tanimoto similarities between the 2674 previously screened molecules used to train our initial GPR models and the top-performing candidates capable of >2-fold enhancement/inhibition identified by our active learning screen. Tanimoto similarities between all pairs associating one molecule in the previous screen and one molecule in the active learning screen were computed with ECFP4 molecular fingerprint using RDKit.31 The maximum Tanimoto similarities for each of the molecules identified in the active learning screen were determined by max; Tanimoto(xz, x;), where xz denoted the z lh molecule in the active learning screen and x; denoted the / lh molecule in the previous screen.
  • Tanimoto similarity quantifies the proportion of chem- ical substructures shared by a pair of molecules and it is a continuous number between 0 and 1.0, where 1.0 denotes complete identity.
  • the large support for the histogram in the 0-0.6 region indicates that the active learning screen is exploring molecules that are substantially diverse from the initial training data.
  • FIG. 16 shows maximum pairwise Tanimoto similarities between the 2674 previously screened molecules used to train our initial GPR models and the top-performing candidates capable of >2-fold enhancement/inhibition identified by our active learning screen. Tanimoto similarities between all pairs associating one molecule in the previous screen and one molecule in the active learning screen were computed with ECFP4 molecular fingerprint using RDKit.31 The maximum Tanimoto similarities for each of the molecules identified in the active learning screen were determined by max; Tanimoto(xz, x;), where xz denoted the z lh molecule in the active learning screen and x; denoted the molecule in the previous screen.
  • Tanimoto similarity quantifies the proportion of chem- ical substructures shared by a pair of molecules and it is a continuous number between 0 and 1.0, where 1.0 denotes complete identity.
  • the large support for the histogram in the 0-0.6 region indicates that the active learning screen is exploring molecules that are substantially diverse from the initial training data.
  • FIG. 17 shows 384-Well plate layout. One column and two rows on edges are not used to test and filled with cell culture media to avoid error induced by evaporation of water.
  • FIG. 18 shows immunomodulation of 17 selected top-performing candidates over cytokine release profiles activated by LPS, CpG and cGAMP.
  • the extent of immunomodulation is visualized as bar plots with error bars indicating standard errors associated with each experiment.
  • the bars are also color-coded by the log2 fold change values for better clarity in showing those immunomodulation with large extent, representing potent enhancers or suppressors.
  • FIG. 19 shows a block diagram of an example computer implemented method encompassing methods disclosed herein herein.
  • FIG. 20 shows a block diagram of an example drug- screening method encompassing methods disclosed herein
  • FIG. 21 shows analysis of the influence of featurization and regression model upon predictive performance of the surrogate model. All combinations of featurizations - VAE and Morgan fingerprints (FP) - and regression model - GPR and linear regression (LR) were compared: (1) VAE-GPR; (2) VAE-LR; (3) FP-GPR; and (4) FP-LR. Performance is compared based on the MAE in the log2-fold change in six immunomogical functional goals. The mean MAE is shown as the bars and the standard deviation of the MAE computed from 5-fold cross validation is shown as the error bars. The simplest model, FP-LR, has a poor MAE compared to the other three. The three remaining model combinations exhibit similar performance.
  • All combinations of featurizations - VAE and Morgan fingerprints (FP) - and regression model - GPR and linear regression (LR) were compared: (1) VAE-GPR; (2) VAE-LR; (3) FP-GPR; and (4) FP-LR. Performance is compared based on the MAE
  • FIG. 22 shows top performers identified in the active learning-assisted screen and their closest analog in the 2674 molecules from a prior screen.
  • the Tanimoto similarity quantifies the proportion of chemical substructures shared by a pair of molecules and it is a continuous number between 0 and 1.0, where 1.0 denotes complete topological identity.
  • FIG. 23 shows comparison of the number of immunomodulators with 1.5x, 2x, 5x, and lOx enhancement or suppression identified in the present work compared to a prior screen.
  • the number of compounds meeting the foldchange threshold (columns) is presented.
  • the left number represents the number of candidates meeting this criterion identified from our prior work
  • the right number represents the number identified in the present work.
  • the active learning-guided screen yielded a significant increase in the population of immunomodulator candidates, and in five instances identified candidates with activity levels that were not achieved in our prior screen.
  • the innate immune response is vital for the success of prophylactic vaccines and immunotherapies.
  • Control of signaling in innate immune pathways can improve prophylactic vaccines by inhibiting unfavorable systemic inflammation and immunotherapies by enhancing immune stimulation.
  • Aspects herein developed a machine learning-enabled active learning pipeline to guide in vitro experimental screening and discovery of small molecule immunomodulators that improve immune responses by altering the signaling activity of innate immune responses stimulated by traditional pattern recognition receptor agonists.
  • molecules were tested by in vitro high throughput screening (HTS) where modulation of the nuclear factor K-light-chain-enhancer of activated B-cells (NF-KB) and the interferon regulatory factors (IRF) pathways were measured.
  • HTS high throughput screening
  • these data were used to train data-driven predictive models linking molecular structure to modulation of the NF-KB and IRF responses using deep representational learning, Gaussian process regression, and Bayesian optimization.
  • an active learning-guided traversal of a 139,998 molecule library was performed.
  • a small portion of a library including less than or approximately 1, 2, or 3 percent (or any range derivable therein) of the library can be used to discover viable molecules with desired activity, including those capable of suppressing NF-KB activity by up to 15-fold, elevating NF-KB activity by up to 5-fold, and elevating IRF activity by up to 6-fold.
  • chemical design rules can be extracted to identify particular chemical fragments as principal drivers of activity, such as specific immunomodulation behaviors. Aspects herein validate the immunomodulatory effect of a subset of top candidates by measuring activity in vitro or in vivo, including measuring cytokine release profiles.
  • one molecule induced a 3-fold enhancement in IFN-/1 production when delivered with a cyclic di-nucleotide stimulator of interferon genes (STING) agonist.
  • STING interferon genes
  • machine learning-enabled screening approaches described herein lead to an efficient discovery pipeline that has furnished a library of novel small molecules with desired activity, including a strong capacity to enhance or suppress innate immune signaling pathways to shape and improve prophylactic vaccination and immunotherapies.
  • Input data for the methods described herein may comprise data collected from cell screening methods.
  • data is collected for performing methods disclosed herein.
  • high-throughput screening data is collected.
  • the data can be collected by incubating a target population, such as a population of cells, with a library of candidate compounds.
  • the data may be collected by any method disclosed herein.
  • the data is collected in a high-throughput manner, including using high-throughput culture plates and high-throughput analyzing apparatuses.
  • the method comprises contacting a population of cells and a library of candidate compounds.
  • the population of cells may comprise a reporting system, such as one or more reporter genes.
  • the reporting system may provide a signal when a target of interest, such as a drug target of interest, is affected by a candidate compound.
  • the contacting produces an interpretable signal.
  • the interpretable signal comprises luminescence produced by the population of cells.
  • the interpretable signal is detected by a plate reader or other apparatus capable of detecting the signal. The signal can be used in one or more of the methods described herein.
  • reporter cells such as RAW-DualTM macrophages as a reporter cell line
  • a reporter molecule such as secreted alkaline phosphatase (SEAP) and/or Lucia luciferase using substrates such as QUANTI-BlueTM and QUANTI-LucTM for absorbance and luminescence readings.
  • SEAP secreted alkaline phosphatase
  • Lucia luciferase using substrates such as QUANTI-BlueTM and QUANTI-LucTM for absorbance and luminescence readings.
  • cells can be seed into well-plates, such as 384-well plate, which are then contacted with the compounds of interest.
  • the activity is measured, such as by absorbance or luminescence readings.
  • the raw readings of absorbance and luminescence of each well are divided by the average reading of the positive controls (presence of agonists and absence of immunomodulators) which are on the same plate to define the fold change associated with each modulator relative to the baseline of the corresponding agonist.
  • This plate-based normalization can ensure measurement consistency by eliminating plate-to- plate and day-to-day variance.
  • Each immunomodulator can be incubated in two replicated plates and the results can be averaged. Experimental errors can be calculated from the standard deviation of the mean calculated from the two replicates with the same compound of interest using standard propagation of errors.
  • a schematic illustration of the experimental screening process is presented in FIG. 16.
  • An illustration of the plate layout used in the HTS experiments is presented in FIG. 17. [0061]
  • viability is monitored after overnight addition by monitoring confluency to insure the compounds of interest are not cytotoxic or cytostatic.
  • FIG. 19 is a block diagram illustrating a computer system 100 upon which aspects of the present teachings may be implemented.
  • computer system 100 can include a bus 102 or other communication mechanism for communicating information and a processor 104 coupled with bus 102 for processing information.
  • computer system 100 can also include a memory, which can be a random-access memory (RAM) 106 or other dynamic storage device, coupled to bus 102 for determining instructions to be executed by processor 104. Memory can also be used for storing temporary variables or other intermediate information during execution of instructions to be executed by processor 104.
  • computer system 100 can further include a read only memory (ROM) 108 or other static storage device coupled to bus 102 for storing static information and instructions for processor 104.
  • ROM read only memory
  • a storage device 110 such as a magnetic disk or optical disk, can be provided and coupled to bus 102 for storing information and instructions.
  • computer system 100 can be coupled via bus 102 to a display 112, such as a cathode ray tube (CRT) or liquid crystal display (LCD), for displaying information to a computer user.
  • a display 112 such as a cathode ray tube (CRT) or liquid crystal display (LCD)
  • An input device 114 can be coupled to bus 102 for communication of information and command selections to processor 104.
  • a cursor control 116 such as a mouse, a trackball or cursor direction keys for communicating direction information and command selections to processor 104 and for controlling cursor movement on display 112.
  • This input device 114 typically has two degrees of freedom in two axes, a first axis (i.e., x) and a second axis (i.e., y), that allows the device to specify positions in a plane.
  • a first axis i.e., x
  • a second axis i.e., y
  • input devices 114 allowing for 3-dimensional (x, y and z) cursor movement are also contemplated herein.
  • results can be provided by computer system 100 in response to processor 104 executing one or more sequences of one or more instructions contained in memory 106.
  • Such instructions can be read into memory 106 from another computer-readable medium or computer-readable storage medium, such as storage device 110.
  • Execution of the sequences of instructions contained in memory 106 can cause processor 104 to perform the processes described herein.
  • hard-wired circuitry can be used in place of or in combination with software instructions to implement the present teachings.
  • implementations of the present teachings are not limited to any specific combination of hardware circuitry and software.
  • FIG. 20 is a block diagram illustrating a workflow encompassing aspects of the present teaching, including aspects related to a drug-discovery process.
  • a library of compounds of interest 200 are manipulated into a latent space 204. The manipulation may be performed using aspects described herein including using a variational autoencoder.
  • the library of compounds of interest 200 may also be used in a high-throughput screen 202.
  • the high-throughput screen can include testing the compounds of interest 200 against one or more targets of interest to generate data.
  • the high-throughput screen 202 may be associated with the latent space 204 then used to train a model 206.
  • the model 206 can be used to identify or predict a set of active compounds 208.
  • the active compounds 208 may include compounds that were not present in the library of compounds of interest 200.
  • the active compounds 208 may then be fed back into the library of compounds 200 and the process may be repeated until a desired number or efficacy of active compounds 208 are identified.
  • computer-readable medium e.g., data store, data storage, etc.
  • computer-readable storage medium refers to any media that participates in providing instructions to processor 104 for execution.
  • Such a medium can take many forms, including but not limited to, non-volatile media, volatile media, and transmission media.
  • non-volatile media can include, but are not limited to, dynamic memory, such as memory 106.
  • transmission media can include, but are not limited to, coaxial cables, copper wire, and fiber optics, including the wires that comprise bus 102.
  • Computer-readable media include, for example, a floppy disk, a flexible disk, hard disk, magnetic tape, or any other magnetic medium, a CD-ROM, any other optical medium, punch cards, paper tape, any other physical medium with patterns of holes, a RAM, PROM, and EPROM, a FLASH-EPROM, another memory chip or cartridge, or any other tangible medium from which a computer can read.
  • instructions or data can be provided as signals on transmission media included in a communications apparatus or system to provide sequences of one or more instructions to processor 104 of computer system 100 for execution.
  • a communication apparatus may include a transceiver having signals indicative of instructions and data.
  • the instructions and data are configured to cause one or more processors to implement the functions outlined in the disclosure herein.
  • Representative examples of data communications transmission connections can include, but are not limited to, telephone modem connections, wide area networks (WAN), local area networks (LAN), infrared data connections, NFC connections, etc.
  • the methodologies described herein may be implemented by various means depending upon the application. For example, these methodologies may be implemented in hardware, firmware, software, or any combination thereof.
  • the processing unit may be implemented within one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, micro-controllers, microprocessors, electronic devices, other electronic units designed to perform the functions described herein, or a combination thereof.
  • ASICs application specific integrated circuits
  • DSPs digital signal processors
  • DSPDs digital signal processing devices
  • PLDs programmable logic devices
  • FPGAs field programmable gate arrays
  • processors controllers, micro-controllers, microprocessors, electronic devices, other electronic units designed to perform the functions described herein, or a combination thereof.
  • the methods of the present teachings may be implemented as firmware and/or a software program and applications written in conventional programming languages such as C, C++, Python, etc. If implemented as firmware and/or software, the aspects described herein can be implemented on a non-transitory computer-readable medium in which a program is stored for causing a computer to perform the methods described above. It should be understood that the various engines described herein can be provided on a computer system, such as computer system 100, whereby processor 104 would execute the analyses and determinations provided by these engines, subject to instructions provided by any one of, or a combination of, memory components 106/108/110 and user input provided via input device 114.
  • any of the various system aspects may have been presented as a group of particular components.
  • these systems should not be limited to the particular set of components, now their specific configuration, communication and physical orientation with respect to each other.
  • these components can have various configurations and physical orientations (e.g., wholly separate components, units and subunits of groups of components, different communication regimes between components).
  • the machine learning component of the framework comprises a deep representational learned projection of molecular candidates into a smooth and low-dimensional latent space suitable for regression and optimization, including by using a variational autoencoder (VAE; FIG. IB), the construction of learned QSAR models over this space, including by using Gaussian process regression (GPR, FIG. 1C), and the selection of new candidate molecules from the design space, including by using Bayesian optimization (BO, FIG. ID).
  • VAE variational autoencoder
  • GPR Gaussian process regression
  • BO Bayesian optimization
  • the selected candidate molecules can then be passed to experimental HTS by automated robotics (FIG. 1 A).
  • Experimental measurements can then be used to retrain and update the GPR models and inform subsequent rounds of BO candidate selection within a virtuous cycle of QSAR model training and model-guided experimental screening.
  • Multiple passes through the active learning loop can be conducted until desired performance metrics are reached or a performance ceiling is reached, which may be evinced by a lack of improvement over a particular number of cycles.
  • VAE variational auto-encoders
  • the VAE comprises two consecutive deep networks: an encoder converting the discrete chemical representations into fixed-length vectors defining an embedding into a latent space, and a decoder that inverts this operation to reconstruct the discrete chemical representations from their latent space projections (FIG. IB).
  • Chemical structures can be represented to the VAE as SELFIES strings flattened into one-hot vectors.
  • candidates that exist as salts or with other components are represented as a single SELFIES string in which the multiple molecules are concatenated using a special linking token.
  • one or more feature rich representations e.g, Smooth Overlap of Atomic Position (SOAP), 47 Many-body Tensor Representation (MBTR), 48 Coulomb matrices, 49 quantum mechanically calculated full and/or partial charges and optimized atomic coordinates
  • SOAP Smooth Overlap of Atomic Position
  • MBTR Many-body Tensor Representation
  • Coulomb matrices 48 quantum mechanically calculated full and/or partial charges and optimized atomic coordinates
  • the loss function for network training comprises two objectives that are simultaneously optimized: accurate reconstruction of the chemical representations by decoding from their latent space projections and preservation of a prior distribution - frequently a multi-dimensional Gaussian - over the latent space.
  • training of the network discovers a low-dimensional fixed- length vector representation of each molecule in the training data that preserves sufficient information to permit its accurate reconstruction within a compact distribution that promotes generalization to unseen data and from which it is easy to sample.
  • the latent space embedding furnishes a smooth, low-dimensional embedding in which chemical similarity of the training molecules is related to latent space proximity and that is well-suited to traversal and optimization by active learning.
  • the VAE model is constructed and trained in a defined algorithm such as PyTorch.
  • the parameters of the VAE model are optimized under 5-fold cross-validation (CV).
  • performance is achieved by employing a 500-200-100 fully-connected feedforward network architecture as the encoder, and a stack of two gated recurrent units (GRUs) as the decoder.
  • GRUs gated recurrent units
  • an additional 100-node layer following the encoder network serves as the 100D latent space, and it is followed by the decoded network.
  • a data augmentation strategy is employed to enrich the training dataset with additional small molecule candidates, which has been widely recognized in the literature as a practice that can potentially lead to more stable models with improved generalizability and a smoother latent space.
  • the VAE is trained over a training library consisting of compounds of interest augmented with molecules extracted from the known libraries, such as the ZINC library of commercially available small molecules 59 and/or other commercially available compound libraries for virtual screening.
  • the VAE is trained only once at the beginning of the active learning search to produce a latent space embedding of all immunomodulator candidates that was held fixed throughout all subsequent iterations of the process.
  • the molecular screen can discover novel compounds to enhance or suppress a target of interest to elicit a biological response of interest.
  • a functional goal of interest such as suppression or activation of a target of interest
  • the goal may be optimized over a generated latent space, which includes embedding of the candidates learned by the VAE.
  • the method includes training independent GPR surrogate models employing Gaussian (a.k.a. radial basis function) kernels to learn an empirical mapping from the coordinates of each molecule within the 100D latent space embedding to each of the functional goals (FIG. 1C).
  • the GPR models can be trained over all candidates for which the experimental measurements of the level of modulation of the target exist. Then, this information can be used to predict the performance of all remaining candidates for which experiments have not yet been performed.
  • the trained GPR models enable the interpolation/extrapolation from the experimentally measured performance of a small number of candidates to predict the performance of all unmeasured candidates before actually conducting the experimental measurements.
  • the surrogate models guide a prospective traversal of the candidate space by allowing the focus the time, labor, and expense of experimentation toward the most promising molecules.
  • the performance predictions of the GPR models in each of the functional goals are also equipped with uncertainty estimates. As such, one can account for the typically higher model uncertainties when making extrapolative predictions to molecules that lie far away in the latent space (i.e., are more chemically dissimilar) from those that have already been measured.
  • the GPR models are then interfaced with a multi-objective Bayesian optimization (BO) framework to select the best compound candidates for experimental testing in the next round of active learning (FIG. ID).
  • BO Bayesian optimization
  • performance predictions can be made on all as-yet-untested molecules and scored each one according to the Expected Improvement (El) acquisition function.
  • El acquisition function can account for both the mean and uncertainty of the GPR predictions to balancing exploitation and exploration to identify candidates most likely to lead to improved performance.
  • a multi-objective Kriging believer batched sampling protocol to define a batch of 720 molecules for experimental testing.
  • each of the GPR models are collated with the molecule that had not previously been selected for testing with the largest acquisition function value.
  • This group of molecules, with duplicates removed, can then be used to retrain all GPR models under a Kriging believer approach and the GPR models polled again for their next top-ranked molecules. This process can be repeated until a desired batch of molecules, such as approximately 200, 300, 400, 500, 600, 700, 800, 900, or 1000 molecules (or any range derivable therein) are selected and sent for experimental testing.
  • parallel batched selection methods are employed that mimic a sequential selection policy such as local penalization (LP)61 and parallel knowledge gradient (q-KG).
  • One cycle of VAE embedding, GPR training, BO sequence selection, and experimental screening completes one loop of the active learning cycle (FIGs. 1A-1D).
  • one of two methods are used to monitor and determine convergence of the active learning loop.
  • a stabilizing predictions method is employed to evaluate stabilization of the specialist GPR model predictions. To do so, one can set aside a randomly selected 100,000 candidate stop set and measure the average Bhattacharyya distance DB between the GPR posterior evaluated over this stop set in successive rounds of the active learning screen.
  • the performance difference method can be employed to assess the specialist GPR predictive performance by conducting 5-fold cross-validation over the accumulated labeled samples (i.e., candidates for which experimental assay measurements are available). 67 When the absolute value of the crossvalidated mean average error (MAE) on the labeled data reaches an acceptably low level and/or plateaus over the course of successive rounds, this can indicate that the predictive power of the trained GPR is no longer changing with the accumulation of additional screening data and can be taken as an indication of model convergence.
  • MAE mean average error
  • Methods disclosed herein can include those described in Tang et al. (Data-driven discovery of innate immunomodulators via machine learning-guided high throughput screening, Chemical Science, 2023), which is hereby incorporated by reference in its entirety.
  • the active compounds include compounds discovered by methods disclosed herein.
  • the active compounds may be biologically active compounds.
  • the active compounds may have immunomodulatory activity.
  • Disclosed herein are compounds of any of the following
  • compositions including pharmaceutical compositions comprising one or more compounds described herein, are administered to a subject.
  • Different aspects may involve administering an effective amount of a composition to a subject.
  • a pharmaceutical composition may be administered to the subject to protect against or treat a condition (e.g., cancer).
  • an expression vector encoding one or more such antibodies or polypeptides or peptides may be given to a subject as a preventative treatment.
  • such compositions can be administered in combination with an additional therapeutic agent (e.g., a chemotherapeutic, an immunotherapeutic, a bio therapeutic, etc.).
  • Such compositions will generally be dissolved or dispersed in a pharmaceutically acceptable carrier or aqueous medium.
  • phrases “pharmaceutically acceptable” or “pharmacologically acceptable” refer to molecular entities and compositions that do not produce an adverse, allergic, or other untoward reaction when administered to an animal or human.
  • pharmaceutically acceptable carrier includes any and all solvents, dispersion media, coatings, anti-bacterial and anti-fungal agents, isotonic and absorption delaying agents, and the like. The use of such media and agents for pharmaceutical active substances is well known in the art. Except insofar as any conventional media or agent is incompatible with the active ingredients, its use in immunogenic and therapeutic compositions is contemplated. Supplementary active ingredients, such as other anti-infective agents and vaccines, can also be incorporated into the compositions.
  • the active compounds can be formulated for parenteral administration, e.g., formulated for injection via the intravenous, intramuscular, subcutaneous, or intraperitoneal routes.
  • parenteral administration e.g., formulated for injection via the intravenous, intramuscular, subcutaneous, or intraperitoneal routes.
  • such compositions can be prepared as either liquid solutions or suspensions; solid forms suitable for use to prepare solutions or suspensions upon the addition of a liquid prior to injection can also be prepared; and, the preparations can also be emulsified.
  • the pharmaceutical forms suitable for injectable use include sterile aqueous solutions or dispersions; formulations including, for example, aqueous propylene glycol; and sterile powders for the extemporaneous preparation of sterile injectable solutions or dispersions.
  • the form must be sterile and must be fluid to the extent that it may be easily injected. It also should be stable under the conditions of manufacture and storage and must be preserved against the contaminating action of microorganisms, such as bacteria and fungi.
  • a pharmaceutical composition can include a solvent or dispersion medium containing, for example, water, ethanol, polyol (for example, glycerol, propylene glycol, and liquid polyethylene glycol, and the like), suitable mixtures thereof, and vegetable oils.
  • a solvent or dispersion medium containing, for example, water, ethanol, polyol (for example, glycerol, propylene glycol, and liquid polyethylene glycol, and the like), suitable mixtures thereof, and vegetable oils.
  • the proper fluidity can be maintained, for example, by the use of a coating, such as lecithin, by the maintenance of the required particle size in the case of dispersion, and by the use of surfactants.
  • the prevention of the action of microorganisms can be brought about by various anti-bacterial and anti-fungal agents, for example, parabens, chlorobutanol, phenol, sorbic acid, thimerosal, and the like.
  • isotonic agents for example, sugars or sodium chloride.
  • Prolonged absorption of the injectable compositions can be brought about by the use in the compositions of agents delaying absorption, for example, aluminum mono stearate and gelatin.
  • Sterile injectable solutions are prepared by incorporating the active compounds in the required amount in the appropriate solvent with various other ingredients enumerated above, as required, followed by filtered sterilization or an equivalent procedure.
  • dispersions are prepared by incorporating the various sterilized active ingredients into a sterile vehicle which contains the basic dispersion medium and the required other ingredients from those enumerated above.
  • the preferred methods of preparation are vacuum-drying and freeze-drying techniques, which yield a powder of the active ingredient, plus any additional desired ingredient from a previously sterile-filtered solution thereof.
  • compositions will typically be via any common route. This includes, but is not limited to oral, or intravenous administration. Alternatively, administration may be by orthotopic, intradermal, subcutaneous, intramuscular, intraperitoneal, or intranasal administration. Such compositions would normally be administered as pharmaceutically acceptable compositions that include physiologically acceptable carriers, buffers or other excipients.
  • solutions Upon formulation, solutions will be administered in a manner compatible with the dosage formulation and in such amount as is therapeutically or prophylactically effective.
  • the formulations are easily administered in a variety of dosage forms, such as the type of injectable solutions described above.
  • the systems and methods for analyzing, parsing, and displaying scientific research can be implemented via computer software or hardware.
  • the therapeutic composition may comprise one or more of the compounds disclosed herein.
  • the therapeutic composition may be administered in any suitable manner known in the art.
  • compositions and methods comprising therapeutic compositions.
  • the different therapies may be administered in one composition or in more than one composition, such as 2 compositions, 3 compositions, or 4 compositions.
  • Various combinations of the agents may be employed.
  • the therapeutic agents of the disclosure may be administered by the same route of administration or by different routes of administration.
  • the cancer therapy is administered intravenously, intramuscularly, subcutaneously, topically, orally, transdermally, intraperitoneally, intraorbitally, by implantation, by inhalation, intrathecally, intraventricularly, or intranasally.
  • the antibiotic is administered intravenously, intramuscularly, subcutaneously, topically, orally, transdermally, intraperitoneally, intraorbitally, by implantation, by inhalation, intrathecally, intraventricularly, or intranasally.
  • the appropriate dosage may be determined based on the type of disease to be treated, severity and course of the disease, the clinical condition of the individual, the individual's clinical history and response to the treatment, and the discretion of the attending physician.
  • the treatments may include various “unit doses.”
  • Unit dose is defined as containing a predetermined-quantity of the therapeutic composition.
  • the quantity to be administered, and the particular route and formulation, is within the skill of determination of those in the clinical arts.
  • a unit dose need not be administered as a single injection but may comprise continuous infusion over a set period of time.
  • a unit dose comprises a single administrable dose.
  • the therapy is administered at a dose of between 1 mg/kg and 5000 mg/kg. In some aspects, the therapy is administered at a dose of at least, at most, or about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102
  • the quantity to be administered depends on the treatment effect desired.
  • An effective dose is understood to refer to an amount necessary to achieve a particular effect. In the practice in certain aspects, it is contemplated that doses in the range from 10 mg/kg to 200 mg/kg can affect the protective capability of these agents.
  • doses include doses of about 0.1, 0.5, 1, 5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90, 100, 105, 110, 115, 120, 125, 130, 135, 140, 145, 150, 155, 160, 165, 170, 175, 180, 185, 190, 195, and 200, 300, 400, 500, 1000 pg/kg, mg/kg, pg/day, or mg/day or any range derivable therein.
  • doses can be administered at multiple times during a day, and/or on multiple days, weeks, or months.
  • the effective dose of the pharmaceutical composition is one which can provide a blood level of about 1 pM to 150 pM.
  • the effective dose provides a blood level of about 4 pM to 100 pM.; or about 1 pM to 100 pM; or about 1 pM to 50 pM; or about 1 pM to 40 pM; or about 1 pM to 30 pM; or about 1 pM to 20 pM; or about 1 pM to 10 pM; or about 10 pM to 150 pM; or about 10 pM to 100 pM; or about 10 pM to 50 pM; or about 25 pM to 150 pM; or about 25 pM to 100 pM; or about 25 pM to 50 pM; or about 50 pM to 150 pM; or about 50 pM to 100 pM (or any range derivable therein).
  • the dose can provide the following blood level of the agent that results from a therapeutic agent being administered to a subject: about, at least about, or at most about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29,
  • the therapeutic agent that is administered to a subject is metabolized in the body to a metabolized therapeutic agent, in which case the blood levels may refer to the amount of that agent.
  • the blood levels discussed herein may refer to the unmetabolized therapeutic agent.
  • Precise amounts of the therapeutic composition also depend on the judgment of the practitioner and are peculiar to each individual. Factors affecting dose include physical and clinical state of the patient, the route of administration, the intended goal of treatment (alleviation of symptoms versus cure) and the potency, stability and toxicity of the particular therapeutic substance or other therapies a subject may be undergoing.
  • the administrations can be at 1, 2, 3, 4, 5, 6, 7, 8, to 5, 6, 7, 8, 9, 10, 11, or 12 week intervals, including all ranges there between.
  • pharmaceutically acceptable refer to molecular entities and compositions that do not produce an adverse, allergic, or other untoward reaction when administered to an animal or human.
  • pharmaceutically acceptable carrier includes any and all solvents, dispersion media, coatings, anti-bacterial and anti-fungal agents, isotonic and absorption delaying agents, and the like. The use of such media and agents for pharmaceutical active substances is well known in the art. Except insofar as any conventional media or agent is incompatible with the active ingredients, its use in immunogenic and therapeutic compositions is contemplated. Supplementary active ingredients, such as other anti-infective agents and vaccines, can also be incorporated into the compositions.
  • the active compounds can be formulated for parenteral administration, e.g., formulated for injection via the intravenous, intramuscular, subcutaneous, or intraperitoneal routes.
  • parenteral administration e.g., formulated for injection via the intravenous, intramuscular, subcutaneous, or intraperitoneal routes.
  • such compositions can be prepared as either liquid solutions or suspensions; solid forms suitable for use to prepare solutions or suspensions upon the addition of a liquid prior to injection can also be prepared; and, the preparations can also be emulsified.
  • the pharmaceutical forms suitable for injectable use include sterile aqueous solutions or dispersions; formulations including, for example, aqueous propylene glycol; and sterile powders for the extemporaneous preparation of sterile injectable solutions or dispersions.
  • the form must be sterile and must be fluid to the extent that it may be easily injected. It also should be stable under the conditions of manufacture and storage and must be preserved against the contaminating action of microorganisms, such as bacteria and fungi.
  • a pharmaceutical composition can include a solvent or dispersion medium containing, for example, water, ethanol, polyol (for example, glycerol, propylene glycol, and liquid polyethylene glycol, and the like), suitable mixtures thereof, and vegetable oils.
  • a solvent or dispersion medium containing, for example, water, ethanol, polyol (for example, glycerol, propylene glycol, and liquid polyethylene glycol, and the like), suitable mixtures thereof, and vegetable oils.
  • the proper fluidity can be maintained, for example, by the use of a coating, such as lecithin, by the maintenance of the required particle size in the case of dispersion, and by the use of surfactants.
  • the prevention of the action of microorganisms can be brought about by various anti-bacterial and anti-fungal agents, for example, parabens, chlorobutanol, phenol, sorbic acid, thimerosal, and the like.
  • isotonic agents for example, sugars or sodium chloride.
  • Prolonged absorption of the injectable compositions can be brought about by the use in the compositions of agents delaying absorption, for example, aluminum mono stearate and gelatin.
  • Sterile injectable solutions are prepared by incorporating the active compounds in the required amount in the appropriate solvent with various other ingredients enumerated above, as required, followed by filtered sterilization or an equivalent procedure.
  • dispersions are prepared by incorporating the various sterilized active ingredients into a sterile vehicle which contains the basic dispersion medium and the required other ingredients from those enumerated above.
  • the preferred methods of preparation are vacuum-drying and freeze-drying techniques, which yield a powder of the active ingredient, plus any additional desired ingredient from a previously sterile-filtered solution thereof.
  • compositions will typically be via any common route. This includes, but is not limited to oral, or intravenous administration. Alternatively, administration may be by orthotopic, intradermal, subcutaneous, intramuscular, intraperitoneal, or intranasal administration. Such compositions would normally be administered as pharmaceutically acceptable compositions that include physiologically acceptable carriers, buffers or other excipients.
  • solutions Upon formulation, solutions will be administered in a manner compatible with the dosage formulation and in such amount as is therapeutically or prophylactically effective.
  • the formulations are easily administered in a variety of dosage forms, such as the type of injectable solutions described above.
  • Any apparently unfulfilled valency is to be understood to be properly filled by hydrogen atom(s).
  • a compound with a substituent of -O or -N is to be understood to be -OH or -NH2, respectively.
  • the claimed invention is also intended to encompass salts of any of the compounds of the present invention.
  • salt(s) as used herein, is understood as being acidic and/or basic salts formed with inorganic and/or organic acids and bases.
  • Zwitterions are understood as being included within the term “salt(s)” as used herein, as are quaternary ammonium salts such as alkylammonium salts.
  • Nontoxic, pharmaceutically acceptable salts are preferred, although other salts may be useful, as for example in isolation or purification steps during synthesis.
  • Salts include, but are not limited to, sodium, lithium, potassium, amines, tartrates, citrates, hydrohalides, phosphates and the like.
  • a salt may be a pharmaceutically acceptable salt, for example.
  • pharmaceutically acceptable salts of compounds of the present invention are contemplated.
  • pharmaceutically acceptable salts refers to salts of compounds of this invention that are substantially non-toxic to living organisms.
  • Typical pharmaceutically acceptable salts include those salts prepared by reaction of a compound of this invention with an inorganic or organic acid, or an organic base, depending on the substituents present on the compounds of the invention.
  • Non-limiting examples of inorganic acids which may be used to prepare pharmaceutically acceptable salts include: hydrochloric acid, phosphoric acid, sulfuric acid, hydrobromic acid, hydroiodic acid, phosphorous acid and the like.
  • organic acids which may be used to prepare pharmaceutically acceptable salts include: aliphatic mono- and dicarboxylic acids, such as oxalic acid, carbonic acid, citric acid, succinic acid, phenyl- heteroatom-substituted alkanoic acids, aliphatic and aromatic sulfuric acids and the like.
  • Pharmaceutically acceptable salts prepared from inorganic or organic acids thus include hydrochloride, hydrobromide, nitrate, sulfate, pyrosulfate, bisulfate, sulfite, bisulfate, phosphate, monohydrogenphosphate, dihydrogenphosphate, metaphosphate, pyrophosphate, hydroiodide, hydrofluoride, acetate, propionate, formate, oxalate, citrate, lactate, p- toluenesulfonate, methanesulfonate, maleate, and the like.
  • Suitable pharmaceutically acceptable salts may also be formed by reacting the agents of the invention with an organic base such as methylamine, ethylamine, ethanolamine, lysine, ornithine and the like.
  • Pharmaceutically acceptable salts include the salts formed between carboxylate or sulfonate groups found on some of the compounds of this invention and inorganic cations, such as sodium, potassium, ammonium, or calcium, or such organic cations as isopropylammonium, trimethylammonium, tetramethylammonium, and imidazolium.
  • derivatives of compounds of the present invention are also contemplated.
  • “derivative” refers to a chemically modified compound that still retains the desired effects of the compound prior to the chemical modification. Such derivatives may have the addition, removal, or substitution of one or more chemical moieties on the parent molecule.
  • Non-limiting examples of the types modifications that can be made to the compounds and structures disclosed herein include the addition or removal of lower alkanes such as methyl, ethyl, propyl, or substituted lower alkanes such as hydroxymethyl or aminomethyl groups; carboxyl groups and carbonyl groups; hydroxyls; nitro, amino, amide, and azo groups; sulfate, sulfonate, sulfono, sulfhydryl, sulfonyl, sulfoxido, phosphate, phosphono, phosphoryl groups, and halide substituents.
  • lower alkanes such as methyl, ethyl, propyl, or substituted lower alkanes
  • carboxyl groups and carbonyl groups hydroxyls; nitro, amino, amide, and azo groups
  • sulfate, sulfonate, sulfono, sulfhydryl, sulfonyl s
  • Additional modifications can include an addition or a deletion of one or more atoms of the atomic framework, for example, substitution of an ethyl by a propyl; substitution of a phenyl by a larger or smaller aromatic group.
  • heteroatoms such as N, S, or O can be substituted into the structure instead of a carbon atom.
  • Compounds of the present invention may contain one or more asymmetrically- substituted carbon or nitrogen atoms, and may be isolated in optically active or racemic form. Thus, all chiral, diastereomeric, racemic form, epimeric form, and all geometric isomeric forms of a structure are intended, unless the specific stereochemistry or isomeric form is specifically indicated. Compounds may occur as racemates and racemic mixtures, single enantiomers, diastereomeric mixtures and individual diastereomers. In some aspects, a single diastereomer is obtained.
  • the chiral centers of the compounds of the present invention can have the S- or the R-configuration, as defined by the IUPAC 1974 Recommendations.
  • Compounds may be of the D- or L-form, for example. It is well known in the art how to prepare and isolate such optically active forms. For example, mixtures of stereoisomers may be separated by standard techniques including, but not limited to, resolution of racemic form, normal, reverse-phase, and chiral chromatography, preferential salt formation, recrystallization, and the like, or by chiral synthesis either from chiral starting materials or by deliberate synthesis of target chiral centers.
  • Compounds of the present invention may occur as a hydrate, a compound containing an equivalent of water in the form of an H2O molecule, or polyhydrate, a compound containing more than one equivalent of water in the form of H2O molecules.
  • atoms making up the compounds of the present invention are intended to include all isotopic forms of such atoms.
  • Isotopes include those atoms having the same atomic number but different mass numbers.
  • isotopes of hydrogen include tritium and deuterium
  • isotopes of carbon include 13 C and 14 C.
  • prodrug is intended to include any covalently bonded carriers which release the active parent drug or compounds that are metabolized in vivo to an active drug or other compounds employed in the methods of the invention in vivo when such prodrug is administered to a subject.
  • prodrugs are known to enhance numerous desirable qualities of pharmaceuticals (e.g., solubility, bioavailability, manufacturing, etc.), the compounds employed in some methods of the invention may, if desired, be delivered in prodrug form.
  • the invention contemplates prodrugs of compounds of the present invention as well as methods of delivering prodrugs.
  • Prodrugs of the compounds employed in the invention may be prepared by modifying functional groups present in the compound in such a way that the modifications are cleaved, either in routine manipulation or in vivo, to the parent compound.
  • prodrugs include, for example, compounds described herein in which a hydroxy, amino, or carboxy group is bonded to any group that, when the prodrug is administered to a subject, cleaves to form a free hydroxyl, free amino, or carboxylic acid, respectively.
  • alkyl, carbocyclic, aryl, and alkylaryl esters such as methyl, ethyl, propyl, iso-propyl, butyl, isobutyl, sec -butyl, tert-butyl, cyclopropyl, phenyl, benzyl, and phenethyl esters, and the like.
  • QSAR quantitative structure activity relationship
  • QSPR quantitative structure property relationship
  • This work reports the development of a data-driven pipeline integrating machine learning and in vitro high throughput screening (HTS) to accelerate the discovery of small molecule immunomodulators of the innate immune response.
  • the inventors curated a library of 139,998 candidate small molecules from commercial chemical screening libraries readily available for purchase. These chemical screening libraries are pre-filtered to be structurally- diverse and drug-like.
  • the inventors constructed a datadriven QSAR model integrating deep representational learning, 37,40 Gaussian process regression (GPR), 41 and Bayesian optimization 42 to guide subsequent rounds of molecular selection and experimental HTS.
  • GPR Gaussian process regression
  • the inventors identified 167 novel immunomodulators with at least 2-fold enhancement or suppression over transcription factor activity of interest, which represents a 105% increase in the total number of known immunomodulators with this level of activity, and nine novel immunomodulators with at least 10-fold activity modulation.
  • the inventors conducted additional characterization assays of 17 topperforming candidates identified in the screen to measure their cytokine release profiles in primary cells. These molecules demonstrated significant capacity to modulate the secretion of various cytokines, including upregulating TNF-a production by over 10-fold, downregulating TNF-a production by over 16-fold, and upregulating IFN-/1 production by over 6-fold.
  • One of the top candidates demonstrated a 3-fold enhancement in IFN-/1 production when delivered with cGAMP, which is a strong stimulator of interferon genes (STING) agonist.
  • a deficiency of the nonlinear QSAR models based on deep representational learning is that despite their predictive power in guiding the HTS campaign, the design rules mediating the mapping from small molecule chemical structure to immunomodulatory profile is not readily available from these relatively “black box” models. Accordingly, the inventors performed a post hoc analysis of the molecules considered in the screen using interpretable “glass box” linear regression models that are expected to possess lower predictive accuracy than the nonlinear QSAR models but can transparently identify particular chemical fragments that are principal drivers of specific immunomodulation behaviors. The inventors found that the presence of halogen moiety in immunomodulators is highly predictive of suppression of NF-KB activity regardless of the type of agonist.
  • the carbonyl and carboxyl moiety is predictive of suppression of the immune responses in NF-KB pathway activated by TLR4 agonists such as LPS and MPLA, and predictive of enhancement of the immune responses in the NF-KB pathway activated by CpG. It was also discovered that aromatic heteroatom moieties are predictive of enhancement in NF-KB activity and of suppression in IRF activity. Even though IRF pathway suppression is of limited clinical interest, this information could be used to guide the development of IRF enhancing immunomodulators by either avoiding the inclusion of aromatic heteroatom moieties or removing such chemical fragments to enhance their potency. This analysis offers design rules to modify the structure of immunomodulators for achieving practical immunomodulation goals in vaccine design or immunotherapy development, enriching candidate libraries with molecules predicted to have promising immunomodulatory behaviors.
  • NF-KB and IRF activity were chosen as the two measures of immunomodulator performance when delivered in combination with one of four PRR agonists: lipopolysaccharides (LPS), 12 Monopho sphoryl-Lipid A (MPLA),13 CpG ODN 1826 (CpG), 14 and 3’3’-cGAMP (cGAMP).
  • LPS lipopolysaccharides
  • MPLA Monopho sphoryl-Lipid A
  • CpG CpG ODN 1826
  • cGAMP 3’3’-cGAMP
  • 15 NF-KB transcription levels are correlated with inflammatory responses and should typically be minimized to reduce reactogenicity in prophylactic vaccinations. Under certain circumstances, enhanced NF-KB activity is instead required to optimize the efficacy of vaccines.
  • 7,8 IRF activity which is related to interferon production, should typically be maximized for a productive antiviral response.
  • RAW-DualTM macrophages as a reporter cell line that can quantitatively report NF-KB and IRF activity via secreted alkaline phosphatase (SEAP) and Lucia luciferase using the proprietary substrates QUANTI-BlueTM and QUANTI-LucTM for absorbance and luminescence readings.
  • SEAP secreted alkaline phosphatase
  • Lucia luciferase using the proprietary substrates QUANTI-BlueTM and QUANTI-LucTM for absorbance and luminescence readings.
  • the inventors seeded 50,000 RAW-DualTM macrophages in 384-well plates in 45 qL of complete media using a MultiDropTM Combi liquid handler, then incubated for 1 hour at 37 °C.
  • the inventors transferred immunomodulator compounds from source plates (lOmM in DMSO) to a final concentration of 10 /rM using a Janus® G3 via pintool.
  • One column and two rows at the edge of each plate were left blank and filled with water to avoid possible systematic errors due to evaporation.
  • one of the four PRR agonists was added in 5 /iL of media to achieve the desired concentration.
  • Cells were incubated with or without agonist overnight until transcription factor activity was analyzed. This activity was measured via absorbance at 620 nm or luminescence readings using a BioTek SynergyTM Neo2 Hybrid multimode microplate reader.
  • the inventors were able to measure the net enhancement or suppression in activity of the screened compounds on the cells in the absence of agonists.
  • the inventors were able to simultaneously observe the NF-KB and IRF activity within a single well.
  • the raw readings of absorbance and luminescence of each well were then divided by the average reading of the positive controls (presence of agonists and absence of immunomodulators) which are on the same plate to define the fold change associated with each modulator relative to the baseline of the corresponding agonist.
  • This plate-based normalization ensures measurement consistency by eliminating plate-to-plate and day-to-day variance.
  • Each immunomodulator was incubated in two replicated plates and the results were averaged. Experimental errors are calculated from the standard deviation of the mean calculated from the two replicates with the same immunomodulator using standard propagation of errors.
  • a schematic illustration of the experimental screening process is presented in FIG. 16.
  • An illustration of the plate layout used in the HTS experiments is presented in FIG. 17.
  • the small molecule library source plates were hosted by the University of Chicago Cellular Screening Center (Chicago, Illinois, USA) and purchased from various vendors.
  • the RAWDualTM cells, QUANTI-BlueTM, QUANTI-LucTM and the PRR agonists were purchased from InvivoGen (San Diego, California, USA).
  • modulators may be cytostatic or cytotoxic to cells
  • viability was monitored after overnight addition by monitoring confluency via IncuCyte® imaging.44, 45 Two confluency masks were generated using IncuCyte® software over all imaged wells, and modulators were determined non-viable if both sets of confluency masks were lower than 70% of those of resting cells. This methodology was validated with select library plates using a traditional Promega® CellTiter-Glo® assay.
  • Immunomodulators considered non-viable are not necessarily cytostatic or cytotoxic in other cell lines or clinical settings, so immunomodulator candidates exhibiting extraordinary behaviors were not categorically ruled out on the basis of this viability assay alone. However, unless explicitly stated otherwise, the inventors only report on immunomodulators measured to be viable under this assay.
  • BMDCs murine bone marrow derived dendritic cells
  • Monocytes were harvested from 6-week-old C57BL/6 mice and were differentiated into dendritic cells using supplemented culture medium: RPMI 1640 (Life Technologies, Waltham, Massachusetts, USA), 10% HIFBS (Sigma-Aldrich, Burlington, Massachusetts, USA), Recombinant Mouse GM-CSF (carrier-free; 20 ng/ml; BioLegend, San Diego, California, USA), 2 mM L-glutamine (Life Technologies ®), 1% antibiotic-antimycotic (Life Technologies), and 50 /rM [>- mercaptoethanol (Sigma- Aldrich®).
  • BMDCs were plated at 100,000 cells per well and incubated with modulator (10 /rM). After 1 hour, one of the three tested agonists (LPS, CpG, cGAMP) was added. Cells were incubated for 24 hours at 37°C and 5% CO2. Supernatant cytokines were measured using LEGENDplexTM Mouse Inflammation Cytokine Kit (Biolegends®) or a VeriKineTM IFN-/1 ELISA (PBL Assay Science, Piscataway, New Jersey, USA).
  • the cytokines measured by LEGENDplexTM are TNF-a, IFN-/A IFN-y, IL- 6, IL-la, IL-1 IL-10, IL-12p70, IL-17A, IL-23, IL-27, MCP-1, and GM-CSF.
  • the inventors developed a data-driven active learning framework to guide experimental screening of immunomodulator candidates.
  • An overview of the integrated screening approach is illustrated schematically in Figure 1.
  • the machine learning component of the framework comprises a deep representational learned projection of all 139,998 molecular candidates into a smooth and low-dimensional latent space suitable for regression and optimization using a variational autoencoder (VAE; Figure IB), the construction of learned QSAR models over this space using Gaussian process regression (GPR, Figure 1C), and the selection of new candidate molecules from the design space using Bayesian optimization (BO, Figure ID).
  • the selected candidate molecules are then passed to experimental HTS by automated robotics (Figure 1A).
  • Experimental measurements are then used to retrain and update the GPR models and inform subsequent rounds of BO candidate selection within a virtuous cycle of QSAR model training and model-guided experimental screening. Multiple passes through the active learning loop are conducted until desired performance metrics are reached or the inventors encounter a performance ceiling evinced by a lack of improvement over a particular number of cycles.
  • VAEs variational autoencoders
  • VAE variational auto-encoders
  • Candidates that exist as salts or with other components are represented as a single SELFIES string in which the multiple molecules are concatenated using a special linking token.
  • This featurization implicitly adopt a 2D representation of the candidate molecules in their electrically neutral state that reflects their molecular topology (i.e., atomic connectivity) but does not capture 3D information on their conformation, protonation state, or partial charge distribution. It was anticipated that this representation would prove sufficient for the active learning screen, and, as shown below, this featurization does enable us to identify a number of high-performing new molecular candidates.
  • the loss function for network training comprises two objectives that are simultaneously optimized: accurate reconstruction of the chemical representations by decoding from their latent space projections and preservation of a prior distribution - frequently a multi-dimensional Gaussian - over the latent space.
  • training of the network discovers a low-dimensional fixed-length vector representation of each molecule in the training data that preserves sufficient information to permit its accurate reconstruction within a compact distribution that promotes generalization to unseen data and from which it is easy to sample.
  • the latent space embedding furnishes a smooth, low-dimensional embedding in which chemical similarity of the training molecules is related to latent space proximity and that is well-suited to traversal and optimization by active learning.
  • the inventors constructed and trained the VAE model in PyTorch50 and its parameters were optimized under 5-fold cross-validation (CV). The inventors achieve good performance employing a 500-200-100 fully-connected feedforward network architecture as the encoder, and a stack of two gated recurrent units (GRUs) as the decoder. An additional 100- node layer following the encoder network serves as the 100D latent space, and it is followed by the decoded network.
  • GRUs gated recurrent units
  • the inventors were motivated to employ a data augmentation strategy to enrich the training dataset with additional small molecule candidates, which has been widely recognized in the literature as a practice that can potentially lead to more stable models with improved generalizability and a smoother latent space. 31 33 Accordingly, the inventors trained the VAE over an training library consisting of the 139,998 immunomodulator candidates augmented with 1,108,666 molecules extracted from the ZINC library of commercially available small molecules 34 37 and other commercially available compound libraries for virtual screening. The VAE was trained only once at the beginning of the active learning search to produce a latent space embedding of all immunomodulator candidates that was held fixed throughout all subsequent iterations of the process. Training was conducted on a NVIDIA RTX 2080 GPU card requiring 9032 epochs and 1344 GPU-hours of training over the 1,248,664 candidate augmented training library.
  • GPR Gaussian process regression
  • the primary goals of the molecular screen are to discover novel immunomodulators to enhance or suppress the NF-KB response as a means of, respectively, upregulating innate immune stimulation or downregulating inflammation in prophylactic vaccination to promote efficacy and safety, and to enhance the IRF response to upregulate production of type-I interferons and promote antiviral responses in cancer immunotherapies.
  • the inventors did not search for molecules to inhibit IRF pathway, since this is presently of limited clinical significance.
  • LPS LPS
  • MPLA MPLA
  • CpG CpG
  • cGAMP targets STING. 12 15
  • the inventors sought to discover molecules that are specialists in achieving large immunomodulatory effects when delivered with one particular agonist, and those that are generalists in doing so when delivered with any one of a particular group of agonists.
  • a molecule that enhances the NF-KB response when delivered with LPS would be regarded as a specialist enhancer of NF/cB operating through the TLR4 receptor, whereas one that suppresses the NF-KB response when delivered with any one of LPS, MPLA, or CpG would be regarded as a generalist suppressor operating through the TLR4 and TLR9 receptors.
  • generalists may not be as good at immunomodulation in concert with any single agonist, but offer better immunomodulation profiles across multiple agonists.
  • the inventors quantify the performance of a specialist for a particular agonist via the fold change in the NF-KB or IRF transcription factor activity induced by co-delivery of the immunomodulator with the agonist relative to delivery of the agonist alone.
  • the inventors quantify the performance of a generalist over a group of agonists as the average fold-change over that group.
  • the eight agonist-pathway combinations representing the 12 functional immunomodulatory goals of the immunomodulator screen.
  • the rows of the matrix comprise the agonists or groups of agonists, and the columns comprise the signaling pathway and enhancement / suppression thereof.
  • downregulation of the IRF pathway is of limited clinical interest and so is not included in the screening goals.
  • the inventors seek immunomodulatory specialists that exert large enhancement or suppressive effects when co-delivered with LPS, MPLA, CpG, or cGAMP agonists alone.
  • the inventors also seek an NF-KB generalist capable of enhancement or suppression of the NF-KB pathway when delivered in concert with any one of the LPS, MPLA, or CpG agonists, and an IRF generalist capable of enhancing the IRF pathway when delivered in concert with any one of the LPS, MPLA, or cGAMP agonists.
  • the cGAMP agonist is known to primarily affect the IRF pathway and so was not considered as either a specialist or generalist for NF-KB immunomodulation.
  • CpG is known to primarily affect the NF-KB pathway, and so was not considered as either a specialist or generalist for IRF immunomodulation.
  • the trained GPR models enable us to interpolate/extrapolate from the experimentally measured performance of a small number of candidates to predict the performance of all unmeasured candidates before actually conducting the experimental measurements.
  • the surrogate models guide a prospective traversal of the candidate space by allowing us to focus the time, labor, and expense of experimentation toward the most promising molecules.
  • the performance predictions of the GPR models in each of the 12 design goals are also equipped with uncertainty estimates. As such, the inventors can account for the typically higher model uncertainties when making extrapolative predictions to molecules that lie far away in the latent space (i.e., are more chemically dissimilar) from those that have already been measured.
  • the GPR models were then interfaced with a multi-objective Bayesian optimization (BO) framework to select the best compound candidates for experimental testing in the next round of active learning 35 (Figure ID).
  • BO Bayesian optimization
  • the inventors For each of the 12 GPR models, the inventors made performance predictions on all as-yet-untested molecules and scored each one according to the Expected Improvement (El) acquisition function. 58 The El acquisition function accounts for both the mean and uncertainty of the GPR predictions to balancing exploitation and exploration to identify candidates most likely to lead to improved performance.
  • the inventors integrated the 12 GPR models with a multi-objective Kriging believer batched sampling protocol to define a batch of 720 molecules for experimental testing.
  • the inventors collated from each of the 12 GPR models the molecule that had not previously been selected for testing with the largest acquisition function value. This group of 12 molecules, with duplicates removed, was then used to retrain all 12 GPR models under a Kriging believer approach and the GPR models polled again for their next topranked molecules. The inventors repeated this process until a batch of at least 720 molecules were selected and sent for experimental testing. Importantly, all selected molecules, regardless of their source GPR models, were subjected to experimental testing to evaluate their immunomodulatory profiles across all agonist types in both the NF-KB and IRF pathways. Hence, molecules selected by one GPR model were not only used to retrain that particular GPR model, but also used to retrain all other models.
  • One cycle of VAE embedding, GPR training, BO sequence selection, and experimental screening completes one loop of the active learning cycle ( Figure 1A-D).
  • the inventors use two methods to monitor and determine convergence of the active learning loop.
  • the inventors employ a stabilizing predictions method to evaluate stabilization of the specialist GPR model predictions.
  • the inventors set aside a randomly selected 100,000 candidate stop set and measure the average Bhattacharyya distance64 DB between the GPR posterior evaluated over this stop set in successive rounds of the active learning screen. Large average DB values indicate that the GPR posterior is still being updated over the course of additional screening rounds and thus the convergence has not been reached yet, whereas small values indicate that additional rounds are not changing the GPR predictions and can be seen as an indicator for convergence.
  • the inventors employ the performance difference method to assess the specialist GPR predictive performance by conducting 5-fold cross- validation over the accumulated labeled samples (i.e., candidates for which experimental assay measurements are available).
  • 65 When the absolute value of the cross-validated mean average error (MAE) on the labeled data reaches an acceptably low level and/or plateaus over the course of successive rounds, this indicates that the predictive power of the trained GPR is no longer changing with the accumulation of additional screening data and can be taken as an indication of model convergence.
  • Generalist GPR models are not considered in this convergence assessment, because the immunomodulatory profiles of generalists are defined as linear combinations of specialists with different agonists in a particular group.
  • the convergence of generalist GPR models is closely correlated with the convergence of specialist GPR models, and the inventors choose to only assess the convergence of specialist GPR models. Using these criteria, the inventors terminated the active learning screen after four rounds, during which the inventors experimentally assayed a total of 2,880 compounds comprising ⁇ 2% of the 139,998-candidate molecular candidate space.
  • the inventors attempted to extract human interpretable design rules relating simple chemical properties of the immunomodulator candidates to their measured performance. To do so, the inventors constructed simple “glass box” linear regression models linking the occurrence of particular structural fragments to the measured fold-change in the immunomodulatory response using least absolute shrinkage and selection operator (LASSO) regression. 66,67 This resulted in sparse linear models with relatively few non-zero linear coefficients that identify those structural fragments that are the principal discriminants of the measured immunomodulatory activity.
  • LASSO least absolute shrinkage and selection operator
  • the inventors combined the dataset consisting of 2880 molecules tested in this study with the 2674 molecules obtained in a prior study. 24 The inventors then eliminated any nonviable and redundant compounds to arrive at a dataset of 3560 distinct and viable molecular structures, along with their associated immunomodulatory profiles. For consistency with the prior analysis, candidates that exist as salts or with other components are represented as a single entity for the purposes of generating molecular descriptors. The inventors used the open-source cheminformatics software RDKit68 to featurize each of the 3560 experimentally assayed immunomodulators as a numerical vector of 85 substructure occurrences.
  • the inventors Given the 3560x70 feature matrix F , the inventors then constructed LASSO regression models to predict the activity change for each agonist-pathway combination: (A) NF-KB-LPS, (B) NF-KB-MPLA, (C) NF-zcB-CpG, (D) NF-zcB-Generalist, (E) IRF-LPS, (F) IRF-MPLA, (G) IRF-cGAMP and (H) IRF-Generalist ( Figure IE). Each LASSO regression model corresponding to an immunological objective is trained to predict the log2-fold change in immunomodulatory activity for an immunomodulator using the corresponding normalized feature vector.
  • This training involves minimizing the LI regularized loss, where the LI penalization prevents overfitting by retaining only a small number of generalizable features present in the training dataset.
  • the optimal number of features to use in the model is determined by 5-fold cross-validation on the LI regularization weight.
  • the inventors identify the regularization weight values that result in the lowest generalization error, as well as the number of non-zero coefficients and mean absolute error (MAE) for predicting immunomodulation in log2-fold change corresponding to that optimal regularization weight. By examining the coefficients with the largest magnitudes in this optimal linear model, the inventors can rank the molecular descriptors based on their immunomodulatory effect.
  • Example 4 Experimental Results of Aspects Herein
  • the inventors conducted four rounds of active learning-guided experimental screening of a library of 139,998 putative immunomodulators. In each round, the inventors trained a GPR surrogate model over the experimental screening data collected to date as a surrogate predictor of immunomodulatory activity along 12 functional goals over eight agonistpathway combinations. The inventors then conducted BO to select a total of 720 candidates predicted to strongly enhance the NF-KB response, inhibit the NF-KB response, or enhance the IRF response in the presence of one particular agonist (i.e., a specialist) or in the presence of any one of a group of agonists (i.e., a generalist). The inventors terminated the screen after four rounds, corresponding to a screening of 2880 immunomodulator candidates comprising ⁇ 2% of the molecular candidate search space.
  • Figure 2A the inventors quantify the extent of immunomodulation for each of the 12 functional goals by reporting the fold change in immune activation induced by the combination of PRR agonists and immunomodulators relative to that induced by agonists alone for the molecules considered in each round of the active learning screen. Importantly, the inventors validate that immunomodulators alone do not stimulate an immune activation in the absence of agonists (FIG. 12).
  • the goal of the active learning screen was to perform on-the-fly learning of a QSAR model to guide the optimal selection of the most promising immunomodulator candidates and achieve round-on-round improvements in the identification of top performers.
  • the inventors observe the preponderance of molecules are clustered around a fold-change of unity, meaning that they have a very limited effect on immune activation, and it is the rarer molecules in the tails of the distributions that are of primary interest and to which active learning directs the screen. Looking at the most potent immunomodulator in different function goals (i.e., the maximum or minimum of the distribution in fold change illustrated as orange and purple bands in Figure 2A), the inventors observe clear round-on-round improvements in 11/12 functional goals, indicating that the screen is resolving novel high- performing candidates. Only the LPS specialist to enhance the IRF response shows no significant improvement after the first round, perhaps indicative of the relative paucity of immunomodulators for these goals within the candidate space.
  • Round 0 represents the compounds that were experimentally tested prior to the active learning-assisted screen obtained in previous work, 24 which was also the labeled training data used to train the initial active learning models.
  • Round 0 compound libraries featured compounds known to be relevant to immune signaling pathways, while the compound libraries the inventors used in the active learning discovery search in Rounds 1-4 are more generic chemical screening libraries.
  • NF-KB enhancers MPLA specialist
  • LPS specialist NF-KB suppressors
  • CpG specialist NF-KB suppressors
  • NF-KB Generalist NF-KB Generalist
  • IRF enhancers cGAMP specialist
  • the inventors In addition to discovering those top-performing immunomodulators, the inventors also seek to expand the number of immunomodulators to the tails of the distributions to identify multiple novel high performing immunomodulators.
  • Figure 2B in addition to the topperforming immunomodulators, the inventors show the log2-fold change of the 5th and the 20th strongest enhancer and suppressor for each functional goal with respect to each round. Similarly, the inventors observe round-on-round improvements in 11/12 functional goals for the 5th and the 20th strongest immunomodulator curve, again with the LPS specialist to enhance the IRF response being the only exception. This indicates that the screen is exposing high-profile immunomodulators to enrich the tails.
  • the inventors employed the stabilizing predictions method by computing the average Bhattacharyya distance DB between GPR posteriors in successive rounds over a randomly selected stop set of 100,000 points and employed the performance difference method by computing the 5-fold cross-validated mean average error (MAE) over the accumulated labeled data collected to date as a function of screening round.
  • MAE mean average error
  • the inventors then proceed to conduct a deeper analysis of the top-performing candidates identified by the active learning screen.
  • the inventors filtered 2880 experimentally assayed candidates for cytostatic or cytotoxic behavior using the confluency mask scores, as described herein
  • the top-performing enhancers of the NF-KB response achieved fold improvements relative to agonist alone of 2.6-fold (LPS specialist), 5.5-fold (MPLA specialist), 2.8-fold (CpG specialist) and 2.9-fold (LPS, MPLA, CpG generalist).
  • the top-performing suppressors of the NF-KB response achieved fold improvements relative to agonist alone of 0.1 -fold (LPS specialist), 0.23-fold (MPLA specialist), 0.06-fold (CpG specialist), and 0.15-fold (LPS, MPLA, CpG generalist).
  • the top-performing enhancers of the IRF response achieved fold improvements relative to agonist alone of 5.9-fold (LPS specialist), 6.0-fold (MPLA specialist), 3.2-fold (cGAMP specialist), and 3.6-fold (LPS, MPLA, cGAMP generalist).
  • NF-KB enhancers MPLA specialist
  • LPS specialist NF-KB suppressors
  • CpG specialist NF-KB suppressors
  • IRF enhancers cGAMP specialist
  • the screen identified a number of previously unknown strongly enhancing or suppressing immunomodulators.
  • the inventors report in FIG. 23 the number of immunomodulators with 1.5x, 2x, 5x, and lOx enhancement or suppression identified in the present work compared to the prior screen. 24
  • the inventors observed a substantial increase in the number of known candidates, and in five instances identified candidates with activity levels that were not achieved in the prior screen.
  • the inventors identified 554, 167, 36, and 9 novel immunomodulators capable of mediating 1.5x, 2x, 5x, and lOx enhancement or suppression of at least one of the 12 objective functions, relative to the 382, 159, 23, and 0 identified in previous studies.
  • the nine immunomodulators observed to downregulate NF-KB stimulation by more than 10-fold (i.e., fold change lower than 0.1) representsan unprecedented level of inhibition. 24
  • FIG. 4A The inventors present in Figure 4A the top two molecules for each of the 12 functional objectives identified by the screen, where the inventors show the chemical structures and experimentally measured immunomodulatory profiles for each molecule.
  • PME-4119 was identified as a top-performing NF-KB suppressor generalist, as well as a top-performing MPLA NF-KB suppressor specialist and a CpG NF-KB suppressor specialist, showing that some potent generalists can also function as potent specialists.
  • PME-5246 does not strongly enhance NF- KB stimulation of any particular agonist, but it enhances NF-KB stimulation with every agonist, meaning that it is a good generalist. However, potent IRF enhancer specialists appear to be poorer generalists due to their stronger specificities.
  • the inventors also computed the Tanimoto molecular similarity between each high- profile molecule with a >2-fold enhancement/suppression to identify the most similar molecule within the 2674 molecule data from a previous screen that was used to train the initial GPR model.24
  • the Tanimoto similarity metric quantifies the proportion of chemical substructures shared in a pair of molecules as a value between 0 and 1 and was computed between 2048-bit ECFP4 molecular fingerprints using RDKit.68 The higher the Tanimoto similarity, the more substructures are shared between the molecules.
  • a histogram of the Tanimoto similarity scores between the top performers and the most similar initial training molecule demonstrates significant support at low similarity values indicating that the active learning search has moved into new regions of space and is not simply sampling in the close vicinity of the training data (FIG. 15). Furthermore, a comparison of the top-performers to their closest analog in the 2674 molecules from the prior screen constituting the labeled training data illustrates that the active learning process is not simply performing a local search in the vicinity of the previously identified top performers, but rather learning over the iterative design rounds and venturing into new regions of chemical space (FIG. 16).
  • the active learning screen furnished immunomodulation measurements for 2880 new candidate molecules. Combining these with the 2674 compounds screened in previous study, 24 the inventors possess a rich data set of labeled immunomodulatory activity for 3560 compounds after removing non- viable and duplicated compounds. The inventors then sought to interrogate these data to extract interpretable design rules for the immunomodulatory activity based on the molecular structure. It can be challenging to extract interpretable understanding of structure-function relationship learned by the GPR surrogate model. To furnish more comprehensible structure-function relations, the inventors employed LASSO regression to train an interpretable linear model regressing the log2-fold change in immunomodulatory activity conditioned upon the presence or absence of particular chemical fragments or functional groups.
  • the inventors exchange nonlinearity, complexity, and accuracy of the GPR for interpretability in the LASSO model predictions.
  • the simple structure of the model means that the sign of the learned non-zero weights indicates the direction of immune response regulation. Specifically, chemical groups with a value of Ok > 0 are associated with enhancing modulation and can be regarded as enhancer promoters. Conversely, chemical groups with a value of Ok ⁇ 0 are associated with suppressing modulation and can be regarded as suppressor promoters.
  • the inventors show the performance of LASSO regression models in FIG. 11.
  • the inventors present in non-ascending order of magnitude, the up to six non-zero regression coefficients for LASSO models fitted to each of the eight agonist-pathway combinations of interest. These weights can be interpreted as being associated with the features that have the highest predictive power for the log2-fold change in immunomodulatory activity in each of the eight agonist-pathway combinations.
  • the features reflect the number of occurrences of particular chemical substructures within each candidate molecule and so can lead to actionable design rules on how to modulate immunological behavior by enriching or depleting a molecule with particular chemical groups.
  • the chemical fragments pertaining to each regression coefficient are denoted by codes starting with “fr_” and followed by letters denoting the chemical groups they are quantifying.
  • halogen moiety “fr_halogen” appears as a negative, top-ranked fragment in all four of the NF-KB specialist and generalist LASSO models.
  • NF-KB specialist and generalist LASSO models it ranked among the top two. This indicates that the presence of halogen groups in immunomodulators is predictive of suppression of the activity of this pathway, especially immune responses activated by TLR4 agonists such as LPS and MPLA.
  • a topperforming NF-KB suppressor generalist and LPS specialist PME-4426
  • PME-4426 has a fluoride group
  • PME-3873 and PME-4392 which are both NF-KB suppressors, have chloride groups.
  • aromatic heteroatom moieties including aromatic nitrogen “fr_Ar_N”, aromatic amine “fr_Ar_NH”, and aromatic hydroxyl group “fr_Ar_OH” appear frequently in multiple LASSO models with at least one of them retained in seven out of eight LASSO models, with the MPLA NF-KB specialist being the only exception.
  • the carbonyl moiety “fr_C_O_noCOO” appears as the third most influential fragment in the LPS NF-KB specialist LASSO model.
  • the sum of carbonyl and carboxyl moiety “fr_C_O” appears as a top-ranked fragments in the MPLA NF-KB specialist and CpG NF-KB specialist models with negative and positive weights, respectively. (Chemical fragments not ranked among the top six illustrated in Figure 5 are presented in FIG. 13).
  • TLR4 agonists such as LPS and MPLA
  • PME-4873 having three carbonyl groups in the structure, can greatly enhance NF-KB response with CpG while it is a weak suppressor for NF-KB response with LPS.
  • the high-throughput active learning screen already demonstrated the capacity of these immunomodulators to enhance or inhibit NF-KB and IRF responses.
  • these transcriptional activity measurements only provide an overall representation of the immunomodulatory behavior.
  • the inventors subjected 17 of the top-performing immunomodulator candidates identified in the active learning to a low-throughput assay measuring cytokine release profiles within primary cells (Figure 4). Specifically, the inventors measured modulator’ s ability to change cytokine secretion of murine bone marrow derived dendritic cells (BMDCs) stimulated with LPS, CpG and cGAMP.
  • BMDCs murine bone marrow derived dendritic cells
  • the 17 top candidates were primarily selected from the top-performing immunomodulators in each of the 12 objectives - PME-5071, PME-4855, PME-4671, PME-4633, PME-4873, PME3873, PME-5149, PME-4425, PME-3465, PME- 5246, PME-5839, PME-3808 and PME-5084 plus four molecules that were determined to be non-viable in the confluency mask test for cytostatic or cytotoxic behavior, but exhibit an exceptional immunomodulatory profiles
  • NF-KB activation is correlated with increases in proinflammatory cytokines such as TNF-a
  • IRF activation is related to production of IFN-/1. 6 3
  • the inventors hypothesized that the enhancement or suppression of transcriptional activity induced by the immunomodulators should be associated with the increase or decrease in the production of relevant cytokines.
  • the inventors focused on the immunomodulation of the release of TNF-a and IFN-/1.
  • PME-3878 and PME- 3386 are two candidates discovered in the active learning screen as top-performing generalist inhibitors of NF-KB as well as top-performing specialist inhibitors for NF-KB when treated with LPS.
  • PME-4007 is a candidate that is a top-performing specialist enhancer of the IRF when treated with cGAMP.
  • PME-3878 and PME-3386 inhibit TNF-a, IL-6, and IFN-/1 production for nearly all agonists considered (Figure 6A).
  • Suppressing immunomodulators like these can be used as potential adjuvants for prophylactic vaccines or therapeutics that benefit from minimizing pro-inflammatory cytokines.
  • PME-4007 is a moderate to strong enhancer of the TNF-a, IFN-/A IL-la, and/or IL-17A responses in the presence of LPS, CpG, or cGAMP (Figure 6A).
  • cGAMP is a pattern recognition receptor agonist that acts through the STING pathway.15 Immunomodulators that enhance IFN- ? production through the STING pathway are of particular interest in promoting antiviral defense and anti-tumor immunity thorough T cell cross priming. 70
  • the inventors further subjected the leading IFN- ? inducing compound, PME-4007, to additional comparisons of its cytokine profile in the presence of cGAMP to MSA-2, a recently identified STING agonist.71 MSA-2 was discovered via a high throughput process involving over two million compounds and is more potent than cGAMP. As illustrated in Figure 6B, the inventors observed MSA-2 to induce an IFN- ? secretion that is significantly higher than that induced by cGAMP at the same concentration (10 /rg/mL). Furthermore, when PME-4007 was added in a low concentration (2 /rM) in combination with cGAMP, it increased IFN- ?
  • the in vitro screen identified an immunomodulator that can be combined with a commonly used, naturally occurring STING agonist to induce similar immunological profiles to a best-in- class STING agonist.
  • Co-delivery of immunomodulators with PRR agonists presents a powerful means to reduce inflammation or otherwise modulate innate immune stimulation by enhancing or suppressing innate immune signaling pathways, and offers a route to improving vaccines by reducing adverse side-effects and cancer therapies by enhancing the magnitude of the immune response.
  • Small molecules present attractive immunomodulator candidates with high synthetic accessibility and reduced immunogenic potential compared to biologies.
  • the vast size of the drug-like small molecule design space makes strategies to maximize the utility of each experimental assay extremely valuable in rationally and effectively traversing this space.
  • the inventors combined the inventors constructed a data-driven QSAR model combining deep representational learning, Gaussian process regression, and Bayesian optimization to guide high throughput experimental screening of a library of 139,998 commercially available candidate small molecules. After conducting four rounds of an active learning search that screened 2880 molecules ( ⁇ 2% of the search space) the inventors identified novel immunomodulator candidates capable of suppressing NF-KB activity by up to 15-fold, elevating NF-KB activity by up to 5-fold, and elevating IRF activity by up to 6-fold.
  • the top-performing candidates furnished a 110% improvement in NF-KB activity, 83% improvement in elevating IRF activity, and 128% improvement in suppressing NF-KB activity, and the inventors also identified 167 novel immunomodulators with at least 2-fold enhancement or suppression over transcription factor activity of interest - representing a 105% increase in the total number of known immunomodulators with this level of activity - and nine novel immunomodulators with at least 10-fold activity modulation, while this level of activity modulation is not previously observed with the immunomodulator candidates in the previous work. 24 Additional characterization of the cytokine release profiles of the top 17 candidates demonstrated their ability to substantially modulate key cytokines such as TNF-a, IL-6, and IFN-/A in combination with particular PRR agonists.
  • the inventors defined an augmentation of the ZINC small molecule library 1 4 to train the VAE network for deep representational embeddings of the immunomodulator candidates.
  • the inventors combined the 2,674 molecules employed in the prior screening work, 5 with a subset of 924,870 molecules from the ZINC libraries 1 4 comprising those molecules that containing a biphenyl scaffold, inspired by the structure of previous small molecule immunomodulator discovery of Honokiol. 6
  • the inventors also incorporated some other generic molecular libraries from various vendors.
  • the augmented ZINC library consists of 1,262,866 small molecule compounds in total.
  • a list of the molecular libraries compiled to define the initial augmented ZINC library is provided in Table SI.
  • the inventors then filtered the library under a number of criteria.
  • SILES simplified molecular-input line-entry system
  • the inventors then eliminated compounds that produced an inconsistent SMILES string upon back-translation from the SELFIES representation to ensure a strict one-to- one mapping between SMILES representations and SELFIES representations of the compounds, thus enforcing a strict one-to-one mapping between SELFIES representations and the structure of the compounds.
  • the inventors With the first two steps of filtering, the inventors removed 8734 entries and were left with 1,254,132 compounds in the library.
  • the inventors capped the maximum SELFIES string length to 137 characters and eliminated compounds containing tokens that appear in fewer than 500 compounds.
  • the inventors digitized the SELFIES representations into one-hot matrices of dimension 137- by-55, where 137 represents the maximum possible string length and 55 are the number of possible SELFIES tokens, including the null tokens. SELFIES strings less than the maximum length are padded with null tokens. The inventors then flattened the one-hot matrices into 7535-element vectors to provide a fixed length representation of each molecule to be passed to the VAE.
  • Table SI Compound libraries used to assemble the augmented ZINC library.
  • the inventors selected seven generic commercial small molecule compound libraries to define an initial pool of 139,998 candidate immunomodulators to draw candidates from and bring to high throughput screening experimentation. These libraries were a subset of the previously assembled 1 ,248,664 compound augmented ZINC library. These screening compound libraries were designed with the intention of enabling cell-based and target-based high throughput screening initiatives by making a diverse range of small molecules readily available, which make it easier for us to access the molecules for screening experiments.
  • This immunomodulator candidate library went through the same filtering steps as introduced herein, with the only exception that the inventors were not using a new set of SELFIES tokens, but the inventors used the token dictionary the inventors used for the augmented ZINC library.
  • the inventors transformed the SELFIES representations into one-hot matrices, with a dimension of 137- by- 55 then flattened the one-hot matrices into 7535-element vectors to obtain a standardized representation of each molecule.
  • a list of the molecular libraries compiled to define the initial immunomodulator candidate library is provided in Table S2.
  • Table S2 Compound libraries used to assemble the immunomodulator candidate library.
  • VAEs variational autoencoders
  • the inventors employ a VAE architecture inspired by prior work by Aspuru-Guzik and co-workers 11 12 and implemented in PyTorch. 13
  • the encoder consists of three fully-connected (FC) layers that passes into a fourth fully-connected bottleneck layer corresponding to the latent space embedding.
  • the inventors employed a stack of gated recurrent units (GRU). 14
  • the architecture of the VAE is illustrated in FIG. 7. The hyperparameters of the model architecture were optimized over the ranges reported in Table S3.
  • Table S3 Architecture and hyperparameter optimization ranges for the VAE.
  • Tftg y E loss function £ V AE comprises two components (Equation SI): (1) the reconstruction loss £ Rec measured by the cross-entropy between the input and output one-hot SELFIES vectors to enforce reconstruction fidelity (Equation S2) and (2) the Kullback-Leibler divergence (KLD) 15 LKLD of the latent vectors relative to the standard normal distribution to regularize the latent space (Equation S3), 12 16
  • N is the number of samples
  • D is the dimension of the data
  • Xij is the j-th component of the z-th input vector
  • x t j is the corresponding reconstructed output of the autoencoder
  • Network training is conducted by minimizing Equation SI using the Adam optimizer. 17
  • the inventors first train the VAE over the augmented ZINC library for 10,000 epochs.
  • the training hyperparameters comprising batch size and learning rates were optimized over the ranges reported in Table S4.
  • the inventors assess model performance by computing the exact reconstruction accuracy defined as the fraction of molecules in the training set whose SELFIES strings can be reconstructed with 100% fidelity.
  • the inventors use the decoder in service of learning this latent embedding but, in the present work, do not make use of its generative capacity to extend the model beyond the training data and produce novel synthetic molecules. As such, the inventors assess model performance on the training data rather than the usual practice of employing a hold-out test set.
  • the model achieves an exact reconstruction accuracy of 97.3% at epoch 9032 after more than 1344 GPU-hours of continuous training. This high accuracy indicates that the network has discovered a 100D latent space embedding that preserves the salient information necessary for accurate molecular reconstruction, and the inventors terminate training at this point.
  • the training curves for the model are illustrated in Figure S2.
  • Table S4 VAE training hyperparameters and optimization ranges.
  • GPR Gaussian process regression
  • the inventors seek to identify immunomodulators capable of enhancing or suppressing the immune activity of a certain pathway when the pathway is activated by a certain agonist. This leads to the optimization in maximizing or minimizing the immunomodulation values as fold changes.
  • the inventors train a surrogate model to predict the fitness of immunomodulator candidates that have not been experimentally tested.
  • the inventors include an uncertainty measure that reflects the confidence in the predictions, enabling us to select points in the space that balance both high predicted fitness (exploitation) and high uncertainty (exploration).
  • GPR Gaussian Process Regression
  • RBF radial basis function
  • the inventors were seeking immunomodulators with various capabilities - different kinds of immunomodulation (enhancement, suppression), different pathway (NF-KB, IRF), and different agonist (LPS, MPLA, CpG, cGAMP) - the inventors pursue a multiobjective optimization. As such, the inventors constructed multiple GPR models, with each GPR corresponding to enhancement/suppression with a specific agonist on a specific pathway. In addition to specialists - immunomodulators that induce large changes in enhancement or suppression when co-delivered with one particular agonist - the inventors also sought to identify generalists - immunomodulators that lead to large effects when co-delivered with any one of a group of agonists.
  • the inventors constructed GPR models for generalists using the arithmetic mean of the immunomodulation fold change values of corresponding specialist objectives.
  • the immunomodulation of generalist over LPS, MPLA and CpG agonists can be expressed as njNFkB Enhancer > 1 znjNFkB Enhancer , n/i NFkB Enhancer , n/i NFkB Enhancer ⁇ l v
  • Enhancers of the NF-KB response (4 objectives): specialist for LPS agonist, specialist for MPLA agonist, specialist for CpG agonist, and generalist over LPS, MPLA, and CpG agonists;
  • Inhibitors of the NF-KB response (4 objectives): specialist for LPS agonist, specialist for MPLA agonist, specialist for CpG agonist, and generalist over LPS, MPLA, and CpG agonists;
  • Enhancers of the IRF response (6 objectives): specialist for LPS agonist, specialist for MPLA agonist, specialist for CpG agonist, specialist for cGAMP agonist, generalist over LPS, MPLA, and CpG agonists, and generalist over LPS, MPLA, CpG and cGAMP agonists.
  • the inventors did not seek to identify inhibitors of the IRF response because of limited clinical significance of such immunomodulation.
  • the inventors did not optimize for enhancers or inhibitors of the NF-KB response for specialist for cGAMP agonist because knowledge from prior work that cGAMP does not strongly stimulate the NF-KB pathway.22 Following the screening experiment, the inventors found that the specific CpG agonist the inventors were using (CpG ODN 1826) demonstrated minimal IRF stimulation, as can be seen in FIG. 13. As a result, the inventors only report specialist for CpG agonist as enhancers and inhibitors of the NF-KB response, and the 14 initial objectives were reduced to the 12 objectives as shown in Figure IE.
  • x is the input point
  • p(x) and o(x) are the mean and standard deviation of the surrogate model prediction at x
  • f(x+) is the best function value observed so far, is a tradeoff parameter that balances exploration and exploitation
  • ⁇ I> and ⁇ j) are the standard normal Ll x) — f(x + ) — £ cumulative distribution function and probability density function
  • Z - j - is the standard normal random variable.
  • x* argmax ⁇ E ⁇ x ⁇ ; Xk £ ⁇ xi, x , . . . , x botanical ⁇ ).
  • Traditional multi-objective Bayesian optimization (MOBO) strategies typically aim to balance multiple objectives to efficiently map out Pareto optimal solutions.26, 27 In this work, this would be valuable in identifying potent generalists, but may sacrifice the discovery of potent specialists that reside at the “corners” of the Pareto frontier.
  • the inventors adopted an alternative MOBO strategy in which the inventors polled each of the 12 GPR models independently to collect the best candidates recommended by each model under the El acquisition function and eliminated duplicates.
  • the inventors followed a batched sampling Kriging believer approach28,29 wherein the inventors asked each model for its next top-ranked candidates until the inventors had collected a batch of 720 molecules for experimental testing.
  • the inventors temporarily augment the training data with the molecules selected by the GPRs annotated with the GPR activity predictions, the GPR models are retrained on this augmented data, and the models polled again for their next top- ranked prediction under the El acquisition function. This process is repeated until a batch of 720 molecules has been collated for experimental testing.
  • This batched sampling approach sacrifices efficiency from an information theoretic perspective - each model is asked to choose a series of candidates without first receiving back experimental information on the previously selected candidates - but increases temporal efficiency by testing multiple molecules in an experimental batch.
  • all selected molecules regardless of their source GPR models, were subjected to comprehensive experimental testing to evaluate their immunomodulatory profiles across all agonist types and activation of both NF-KB and IRF pathways.
  • the activity entries in the augmented training data that were completed with the GPR activity predictions in service of the Kriging believer batching are corrected with the measured activities, and these data used to retrain all GPR model.
  • molecules selected by one GPR model are not only used to retrain that particular GPR model, but also used to retrain all other models.
  • the inventors conducted a simple ablation test in which the inventors evaluated model performance upon replacing the learned VAE featurization with a simple and popular 2048-bit Morgan topological fin- gerprints (FP) ECFP4 43 computed using RD Kit, 44 and replacing the GPR regression model with a simple linear regression (LR).
  • FP Morgan topological fin- gerprints
  • ECFP4 43 computed using RD Kit, 44
  • LR simple linear regression
  • the VAE featurization possesses two significant advantages over the Morgan fingerprint featurizations.
  • the 100D VAE embedding has a relatively modest dimensionality compared to the 2048D Morgan fingerprints.
  • the relatively poor scaling of GPR models with dimensionality means that this results in a ⁇ 3x slowdown in training and deployment of the FP-GPR model relative to the VAE-GPR.
  • the VAE embeddings are invertible in the sense that molecules can be generated from VAE vectors, whereas it is not generally considered possible to straightforwardly invert Morgan fingerprints into molecular structures. 43,45
  • the inventors will also use the model to generate new synthetic molecular candidates with potentially superior performance than those contained in the current screening libraries.
  • the first term on the right side of the equation is the mean squared error between the predicted and experimental immunomodulation values
  • the second term is the LI regularization penalty term with hyperparameter «L.
  • 2 and 11 ⁇ 111 represent the L2 and LI norms, respectively.
  • the LI regularization penalty term encourages the model to have fewer nonzero coefficients or parameters, thereby promoting a sparse model that can offer a more concise and interpretable representation of the data.
  • the optimal value for the hyperparameter is chosen using cross-validation, and the resulting nonzero coefficients in 0 can be interpreted as the most critical features for immunomodulation, as shown in FIG. 9.
  • Equation S7 makes it easy to interpret the sign of the learned coefficients: large negative weights 0k ⁇ 0 indicate features that are negatively correlated with immunomodulation, while large positive weights 0k > 0 indicate features that are positively correlated with immunomodulation.
  • FIG. 11 by examining the rank-ordering of the coefficients with the largest absolute values, the inventors can identify structural fragments that are most informative for predicting immunomodulation. There were, in total, 33 chemical fragment descriptors that were retained by at least one of the LASSO models for different immunological objectives.
  • the trained LASSO model can be used to predict immunomodulation for larger molecules and/or molecules not contained within the training data.
  • Table S5 Source library of screened molecules. The table shows (1) the number of com- pounds in each chemical screening library, (2) the number of compounds experimented with HTS in each library, (3) the number of good modulators (with at least 2-fold modulation) found in each library and (4) the ratio of good modulators (the 3rd column divided by the 1st column).
  • the last three libraries namely Microsource Spectrum Collection, Prestwick Chemical Library and Selleckchem FDA- approved Drug Library, have significantly higher ratio of good modulators identified. The statistics excluded non-viable compounds.
  • Table S6 List of Agonists studied and their working concentration.
  • Vargas-Caraveo A.; Sayd, A.; Robledo-Montana, J.; Caso, J. R.; Madrigal, J. L. M.; Garcia-Bueno, B.; Leza, J. C. Toll-like receptor 4 agonist and antagonist lipopolysaccharides modify innate immune response in rat brain circumventricular organs. Journal of Neuroinflammation 2020, 17. 6.
  • Hyodo, M.; Hayakawa, Y.; Vance, R. E. STING is a direct innate immune sensor of cyclic di-GMP. Nature 2011, 478, 515-518.
  • Vargas-Caraveo A.; Sayd, A.; Robledo-Montana, J.; Caso, J. R.; Madrigal, J. L. M.; Garcia-Bueno, B.; Leza, J. C. Toll-like receptor 4 agonist and antagonist lipopolysaccharides modify innate immune response in rat brain circumventricular organs. Journal of Neuroinflammation 2020, 17, 6.

Landscapes

  • Chemical & Material Sciences (AREA)
  • Organic Chemistry (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biochemistry (AREA)
  • Immunology (AREA)
  • Biotechnology (AREA)
  • Genetics & Genomics (AREA)
  • Biomedical Technology (AREA)
  • Medicinal Chemistry (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Hematology (AREA)
  • Veterinary Medicine (AREA)
  • Urology & Nephrology (AREA)
  • Public Health (AREA)
  • Pharmacology & Pharmacy (AREA)
  • Physics & Mathematics (AREA)
  • Animal Behavior & Ethology (AREA)
  • General Physics & Mathematics (AREA)
  • Epidemiology (AREA)
  • Cell Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Medical Informatics (AREA)
  • Analytical Chemistry (AREA)
  • Food Science & Technology (AREA)
  • Microbiology (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Pathology (AREA)
  • Tropical Medicine & Parasitology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Toxicology (AREA)

Abstract

Aspects herein relate to methods and systems for identifying drug candidates. Also disclosed are candidate molecules identified using the methods and systems disclosed herein.

Description

METHODS AND COMPOSITIONS FOR PHARMACEUTICALLY RELEVANT INTERACTIONS
[0001] This application claims priority of U.S. Provisional Application Nos. 63/521,617 filed June 16, 2023 and 63/510,130 filed June 25, 2023, both of which are hereby incorporated by reference in their entirety.
BACKGROUND OF THE INVENTION
[0002] This invention was made with government support under 75N93019C00041 awarded by the National Institutes of Health. The government has certain rights in the invention.
I. Field of the Invention
[0003] This invention relates to the field of chemical biology, medicinal chemistry, and machine learning.
II. Background
[0004] The efficacy of prophylactic vaccination and immunotherapeutic s is predicated on effective stimulation of innate immune responses. In vaccination, for example, helper molecules known as adjuvants are often required to stimulate innate pathways involved in antigen presentation and processing that are critical in invoking a productive adaptive immune response. Despite the necessity of such signaling events to maximize the potency, excessive activation of signaling pathways by adjuvants can cause undesirable systemic inflammation, and limit tolerability and dosage in a clinical setting.1 3 Conversely in immunotherapy, it is often essential to have strong stimulation to improve the immunogenicity and mitigate suppression from a tumor micro-environment.4,5
[0005] Two major effectors of the innate immune response are the nuclear factor K-light- chainenhancer of activated B-cells (NF-KB) pathway and the interferon regulatory factors (IRF) pathway. The NF-KB pathway plays an essential role in inflammation as well as immune activation, while the IRF pathway produces type-I interferons that are essential for a productive antiviral response.6 3 Among these signaling pathways, pattern recognition receptors (PRRs) are cellular receptors expressed on immune cells that identify pathogen-associated molecular patterns (PAMPs) and initiate a cascading immune response. PRRs are necessary for the activation of antigen-presenting cells (APCs) that act as a link between the innate and adaptive immune responses and play a critical role in detecting and responding to pathogens.9, 10 PRR agonists are molecules that bind to PRRs, mimicking the effects of pathogenic molecules and triggering an immune response. As such, PRR agonists have recently been used as adjuvants to activate both NF-KB and IRF pathways and are the most common targets for manipulating the innate immune response.i l A breadth of PRR agonist-based adjuvants comprised of pathogenic motifs have been used as adjuvants in vaccines and immunotherapies, such as lipopolysaccharides (LPS)12 and Monophosphoryl Lipid A (MPLA)13 that target toll-like receptor (TLR) 4, synthetic oligodeoxynucleotides that contain unmethylated cytosine- phosphate-guanine dinucleotide motifs (CpG-ODN) targeting TLR9 14 cyclic guanosine monophosphate- adenosine monophosphate (cyclic GMPAMP, cGAMP) that binds and activates the stimulator of interferon genes (STING),15 polyriboisosinic:polyribocytidylic acid [poly(LC)] that are recognized by TLR3 16 an imidazoquinoline synthetic small-molecule R848 that targets TLR7/8,17 and flagellin that activates TLR5.18 Although these adjuvants can be potent activators of immune responses, a well-known limitation of current popular adjuvants is excessive and uncontrolled inflammation.1 11 19 This has motivated efforts to discover novel adjuvants with reduced inflammation profiles,20,21 but it has proved challenging to develop novel PRR agonists capable of specifically tuning the level of stimulation in inflammatory pathways without disrupting the desired stimulation along immune activation pathways.
[0006] An alternative approach to regulating the innate immune response is through immunomodulators-molecules co-delivered with PRR agonists to reduce inflammation or otherwise modulate innate immune stimulation by enhancing or suppressing innate immune signaling pathways. Moser et al.20 have demonstrated that a selective NF-KB inhibitor known as SN50 has such immunomodulation capacity. SN50 is a cell permeable peptide that consists of nuclear localization sequence of the NF-KB subunits p50 and blocks the import of p50- containing dimers into the nucleus.22,23 It was found that SN50 can reduce the levels of inflammatory cytokines TNF-a and IL-6 while enhancing antigen- specific antibody titers when delivered with the TLR9 agonist CpG.20 As compared to peptides like SN50, small molecules present attractive candidates for immunomodulators due to their better synthetic accessibility and reduced potential for immunogenicity. To this end, Moser et al.21 also demonstrated that a small molecule NF-KB inhibitor honokiol and its derivatives can be used as immunomodulators with similar functions to SN50. More recently, studies conducted a targeted experimental screen of a small molecule library of ~3000 compounds, many of which were known to influence the immune system, to discover, after removing cytotoxic compounds using the same viability filter as applied in this study, novel molecules capable of suppressing NF-KB activity by up to 9-fold, elevating NF-KB activity by up to 7-fold and elevating IRF activity by up to 7- fold. 24
SUMMARY OF THE INVENTION
[0007] Aspects of the disclosure relate to the discovery that the integration of laboratory screening data with machine learning methods can lead to an efficient drug identifying algorithm. Further aspects of the disclosure relate to compositions identified using the integrated algorithm, and methods of using the compositions for treating a disease.
[0008] Disclosed are compounds including compounds of formula PME-4855, formula PME-4426, formula PME-4119, formula PME-3974, formula PME-5149, formula PME-4637, formula PME-4800, formula PME-5839, formula PME-5084, formula PME-4974, formula PME-4873, formula PME-5246, formula PME-4695, formula PME-3465, formula PME-4633, formula PME-4392, formula PME-5071, formula PME-3878, formula PME-3386, formula PME-4671, formula PME-3873, formula PME-4425, formula PME-5920, formula PME-4007, formula PME-3808, formula S8195, formula S1454, formula S4218, formula S7833, formula S8434, formula S8272, formula HY-W009732, formula H8429, formula S7113, formula S8773, formula S7856, formula S2582, formula S8034, formula S3673, formula S1482, or formula S2217.
[0009] Also disclosed are pharmaceutical compositions comprising one or more of the compounds disclosed herein. In certain aspects, the pharmaceutical composition further comprises one or more pattern recognition receptor agonists. The pattern recognition receptor agonists may target one or more of TLR4, TLR9, and STING. The pattern recognition receptor agonists may be one or more of LPS, MPLA, CpG, and cGAMP.
[0010] Also disclosed are methods of modulating activity of immune response proteins in a cell. Also disclosed are methods of inhibiting NF-kB activity in a cell. In certain aspects, the method comprises delivering to the cell an effective amount of at least one compound disclosed herein, including one or more of PME-4855, PME-4426, PME-4119, PME-3974, and PME- 5149, or any combination thereof. In certain aspects, the effective amount results in a concentration of about, at least, or at most, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 pM. The concentration may be a concentration in the cell and/or a concentration in an environment surrounding the cell. In certain aspects, the method further comprises delivering to the cell an effective amount of LPS, MPLA, and/or CpG. In certain aspects, the method further comprises delivering to the cell a pattern recognition agonist. In some aspects, the cell is an immune cell.
[0011] Also disclosed are methods of elevating NF-kB activity in a cell. In certain aspects, the method comprises delivering to the cell an effective amount of one or more compounds disclosed herein including one or more of PME-4637, PME-4800, PME-5839, PME-5084, PME-4974, PME-4873, PME-5246, PME-4695, PME-3465, PME-4633, PME-4392, and PME-5071, or any combination thereof. In certain aspects, the effective amount results in a concentration of about, at least, or at most, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 pM. The concentration may be a concentration in the cell and/or a concentration in an environment surrounding the cell. In certain aspects, the method further comprises delivering to the cell an effective amount of LPS, MPLA, and/or CpG. In some aspects, the method further comprises delivering to the cell a pattern recognition agonist. In some aspects, the cell is an immune cell.
[0012] Also disclosed are methods of elevating IFR activity in a cell. In some aspects, the method comprises delivering to the cell an effective amount of at least one compound disclosed herein, including one or more of PME-4855, PME-4426, PME-4119, PME-3974, PME-5149, PME-4637, PME-4800, PME-5839, PME-5084, PME-4974, PME-4873, PME-5246, PME- 4695, PME-3465, PME-4633, PME-4392, and PME-5071, or any combination thereof. In certain aspects, the effective amount results in a concentration of about, at least, or at most, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 pM. The concentration may be a concentration in the cell and/or a concentration in an environment surrounding the cell. In certain aspects, the method further comprises delivering to the cell an effective amount of LPS, MPLA, and/or CpG. In some aspects, the method further comprises delivering to the cell a pattern recognition agonist. In some aspects, the cell is an immune cell.
[0013] Also disclosed are methods of modulating one or more immune-related pathways in a cell the method comprising delivering to the cell an effective amount of one or more of PME-4855, PME-4426, PME-4119, PME-3974, PME-5149, PME-4637, PME-4800, PME- 5839, PME-5084, PME-4974, PME-4873, PME-5246, PME-4695, PME-3465, PME-4633, PME-4392, and PME-5071, or any combination thereof. In certain aspects, the effective amount results in a concentration of about, at least, or at most, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 pM. The concentration may be a concentration in the cell and/or a concentration in an environment surrounding the cell. In certain aspects, the method further comprises delivering to the cell an effective amount of LPS, MPLA, and/or CpG. In some aspects, the method further comprises delivering to the cell a pattern recognition agonist. In some aspects, the cell is an immune cell.
[0014] Also disclosed are methods of modulating an immune response in a patient. The method can comprise one or more steps including administering to the patient at least one of the compounds described herein and/or at least one of the pharmaceutical compositions described herein, administering a vaccine to the patient, administering an additional therapeutic composition to the patient, and/or administering an anti-viral composition and/or an anti-cancer composition. In some aspects, pro-inflammatory cytokines act antagonistically to the additional therapeutic composition. In some aspects, the patient has, has been diagnosed with, or is suspected of having cancer, an infection, and/or an autoimmune disease.
[0015] Also disclosed are methods comprising administering to a human patient at least one of the compounds described herein, at least one of the pharmaceutical composition described herein, or any combination thereof.
[0016] Also disclosed are methods for predicting a drug-target interaction. Also disclosed are methods of training a model for predicting drug-target interactions. In some aspects, the method comprises 1, 2, 3, 4, 5, or more steps including: generating a fixed-length input feature vector for each compound in a plurality compounds of interest; compiling each fixed-length input feature vectors into a latent space; receiving high-throughput screening data for each compound of the plurality of compounds of interest; associating the fixed-length input feature vectors in the latent space with the high- throughput screening data to generate a set of association information; and training a machine learning model by iteratively minimizing error to within a predetermined threshold using the set of association information.
[0017] In some aspects, the method further comprises identifying a second plurality of compounds of interest that is refined from the plurality of compounds of interest based on the trained machine learning model. In some aspects, the method further comprises predicting drug-target interactions for one or more compounds that are not in the plurality of compounds. In some aspects, the method further comprises predicting activity against the target for one or more compounds that are not in the plurality of compounds. In some aspects, the activity is agonistic activity. In some aspects, the activity is antagonistic activity. In some aspects, the activity comprises affinity to the target.
[0018] In some aspects, the plurality of compounds comprises a library of drug candidates. In some aspects, the plurality of compounds are used in an assay to generate the high- throughput screening data. In some aspects, only a subset of the plurality of compounds are used to generate the high-throughput training data.
[0019] In some aspects, the method further comprises performing one or more iterations of the method, wherein a subsequent iteration uses the second plurality of compounds of interest as the plurality of compounds of interest. As a non-limiting example, the method may be performed once to identify a set of compounds, that set is then used in a second iteration of the method to identify another set of compounds. This process can be repeated multiple times in order to identify candidate molecules. In some aspects, the set of compounds and/or the candidate molecules are then used in a high-throughput screen, which generates data that is then used in another iteration of the method.
[0020] In certain aspects, the high-throughput screening data comprises a signal from a plurality of vessels, wherein each vessel of the plurality of vessels comprises one compound of the compounds of interest and a target. The plurality of vessel may be a well plate, such as a 384-well plate. In certain aspects, the target is a protein of interest. In certain aspects, the target is in a cell. In certain aspects, the cell comprises a reporter system capable of producing the signal. In certain aspects, the signal is fluorescence produced by the cell. In certain aspects, the high-throughput screening data does not comprise data from compounds of interest that are cytostatic or cytotoxic to the cell.
[0021] In certain aspects, the generating a fixed-length input feature vector is performed by a variational autoencoder. In certain aspects, the variational autoencoder is trained by feedforward network architecture. In certain aspects, the compiling is performed via a feed forward architecture to generate a defined-node layer defining the latent space. In certain aspects, the feed forward architecture is a 500-200-100 fully-connected feed forward architecture. In certain aspects, the defined-node layer is a 100-node layer. In certain aspects, the variational autoencoder is trained with the plurality of compounds of interest and an additional set of compounds. The additional set of compounds may be compounds from a known library, such as a commercial library including the ZINC library. The additional set may comprise at least, at most, or approximately 1,000, 2,000, 3,000, 4,000, 5,000, 6,000, 7,000, 8,000, 9,000, 10,000, 50,000, 100,000, 200,000, 300,000, 400,000, 500,000, 600,000, 700,000, 800,000, 900,000, 1,00,000 known compounds (or any range derivable therein).
[0022] In certain aspects, the associating comprises an empirical mapping, which may be an automated empirical mapping, of each coordinate of each input vector to at least one biological response measured in the high-throughput screening data. In some aspects, the empirical mapping comprises an amount of calculations that cannot be done by hand. In certain aspects, the training the machine learning model comprises training at least one Gaussian procession regression models for each measured biological response in the high-throughput screening data. In certain aspects, the method further comprises decoding the latent space into a plurality of interpretable compound vectors. In certain aspects, the decoding is performed by a variational autoencoder. In certain aspects, the decoding is performed by at least one gated recurrent units.
[0023] Also disclosed are methods for identifying a compound pharmacologically active against a target of interest, the method comprising 1, 2, 3, 4, 5, 6, or more steps including: generating a fixed-length input feature vector for each compound in a plurality compounds of interest; compiling each fixed-length input feature vectors into a latent space; receiving high-throughput screening data for each compound in the plurality of compounds of interest; associating the high-throughput screening data with the latent space to generate a set of association information; applying, into a trained machine learning model, the latent space to generate an output feature vector predicting the pharmacological activity against a target of interest for each compound in the plurality of compounds of interest; and identifying one or more compounds with pharmacological activity over a determined threshold based on the output feature vector.
[0024] Also disclosed are methods of normalizing biological data. Also disclosed are methods of normalizing biological data in a high-throughput assay, which may be performed in a multi- well plate. In some aspects, raw readings of a reporter of biological activity, such as absorbance and/or luminescence from a reporter molecule, produced in each well are divided by the average reading of the positive controls (which may be the presence of agonists and absence of immunomodulators) which are on the same plate to define the fold change associated with each modulator relative to the baseline of the corresponding agonist.
[0025] Throughout this application, the term “about” is used to indicate that a value includes the inherent variation of error for the measurement or quantitation method.
[0026] The use of the word “a” or “an” when used in conjunction with the term “comprising” may mean “one,” but it is also consistent with the meaning of “one or more,” “at least one,” and “one or more than one.” [0027] The phrase “and/or” means “and” or “or”. To illustrate, A, B, and/or C includes: A alone, B alone, C alone, a combination of A and B, a combination of A and C, a combination of B and C, or a combination of A, B, and C. In other words, “and/or” operates as an inclusive or.
[0028] The words “comprising” (and any form of comprising, such as “comprise” and “comprises”), “having” (and any form of having, such as “have” and “has”), “including” (and any form of including, such as “includes” and “include”) or “containing” (and any form of containing, such as “contains” and “contain”) are inclusive or open-ended and do not exclude additional, unrecited elements or method steps.
[0029] The compositions and methods for their use can “comprise,” “consist essentially of,” or “consist of’ any of the ingredients or steps disclosed throughout the specification. Compositions and methods “consisting essentially of’ any of the ingredients or steps disclosed limits the scope of the claim to the specified materials or steps which do not materially affect the basic and novel characteristic of the claimed invention.
[0030] It is contemplated that any aspect discussed in this specification can be implemented with respect to any method or composition of the invention, and vice versa. Furthermore, compositions of the invention can be used to achieve methods of the invention.
[0031] Other objects, features and advantages of the present invention will become apparent from the following detailed description. It should be understood, however, that the detailed description and the specific examples, while indicating specific aspects of the invention, are given by way of illustration only, since various changes and modifications within the spirit and scope of the invention will become apparent to those skilled in the art from this detailed description.
BRIEF DESCRIPTION OF THE DRAWINGS
[0032] The following drawings form part of the present specification and are included to further demonstrate certain aspects of the present invention. The invention may be better understood by reference to one or more of these drawings in combination with the detailed description of specific aspects presented herein.
[0033] FIGs. 1A-1E show data-driven active learning framework for immunomodulator discovery. (A) Immunomodulator candidates are subjected to in vitro high throughput screening (HTS) using automated liquid handling platforms. (B) Deep representational learning with a variational autoencoder (VAE) is used to learn an embedding of all 139,998 immunomodulator candidates into a smooth, low-dimensional latent space amenable to regression and optimization. (C) A supervised surrogate Gaussian process regression (GPR) model is trained to predict the immunomodulation of transcription factor levels in the NF-KB and IRF path- ways from all accumulated experimental measurements to date. (D) The trained surrogate model is interrogated using Bayesian optimization (BO) to select the most promising batch of as-yet- untested immunomodulators for the next round of experimental screening. Experimental measurements are then used to retrain and update the GPR models and inform subsequent rounds of BO candidate selection within a virtuous cycle of QSAR model training and model- guided experimental screening. (E) Optimization objectives comprising eight agonist-pathway combinations and twelve functional goals (12 green arrows). Upregulation and downregulation of the NF- B response are both important immunomodulation objectives, leading us to split these agonist-pathway combinations into two goals associated with enhancement and suppression. Only upregulation of the IRF response is a preferred objective, meaning that only enhancement of this response is targeted. We seek specialist immunomodulators capable of delivering large immunomodulatory effects when co-delivered with one particular agonist, and generalist immunomodulators capable of delivering large immunomodulatory effects when co-delivered with any one of a particular agonist group. Specifically, we seek an NF-KB generalist capable of enhancement or suppression of the NF-K pathway when delivered in concert with any one of the LPS, MPLA, or CpG agonists, and an IRF generalist capable of enhancing the IRF pathway when delivered in concert with any one of the LPS, MPLA, or cGAMP agonists.
[0034] FIGs. 2A-2C show results of the data-driven immunomodulator discovery active learn- ing screen. (A) Violin plots of the active learning screen over the 12 functional objectives over the eight agonist-pathway combinations of interest. Fold changes in the NF- KB and IRF responses are measured relative to delivery of the agonist alone. Each violin for Round 1-4 contains the 720 molecules experimentally assayed in each round along the eight agonistpathway combinations, where as the violin for Round 0 denotes the 2674 molecules experimentally assayed prior to the active learning-assisted screen. The orange and purple lines show, respectively, the top enhancer and inhibitor identified by the mean of the two independent experimental measurements, and the shading shows the range of the two measurements. We are most interested in molecules populating the tails of the distributions exhibiting strong ability to alter the immune response by enhancing the NF-KB response, inhibiting the NF-KB response, or enhancing the IRF response in the presence of one particular agonist (i.e., specialists) or in the presence of any one of a group of agonists (i.e., generalists). The NF-K generalist comprises the LPS, MPLA, and CpG agonists, whereas the IRF generalist comprises the LPS, MPLA and cGAMP agonists. (B) Fold changes in the kh strongest enhancer/inhibitor in the 12 functional goals as a function of active learning round for k = 1 (red), k = 5 (purple), and k = 20 (brown). Data are reported as the mean of two independent measurements and error bars show the range of the two measurements. (C) Convergence assessment of the active learning screen for the specialist prediction goals. The Bhattacharyya distance DB between successive GPR posteriors over a randomly selected stop set of 100,000 points plateaus near a value of zero between Rounds 3 and 4 (upper row). Error bars report the standard error in DB estimated over the stop set. The 5-fold cross validated MAE in the predicted log2-fold change in immunomodulatory activity over all labeled data collected to date (lower row) shows a decreasing trend in all predictive goals indicating that additional predictive performance of the GPR may be gained under additional screening rounds. The MAE of the terminal model is reported on the right of last two panels with error bars estimated by 5-fold cross validation.
[0035] FIGs. 3A-3D show Measuring the progress of the active learning screen over 2D projections of the 100-dimensional latent space. (A) Probability density function estimated by kernel density estimation of a projection of the 142,672 molecules, consisting of 139,998 candidate molecules and 2,674 molecules from previous study24 into a 2D t-distributed Stochastic Neighbor Embedding (t-SNE) embedding of the 100D VAE latent space. (B) A contour plot of the 2D pdf presented in panel A. (C) Identification of the newly selected molecules in each of the four rounds of the active learning screen within the 2D t-SNE embedding: 139,998 molecules defining the complete candidate space (grey), 2674 molecules screened in prior work24 used to train the initial GPR models (black), 720 molecules selected in Round 1 (blue), 720 molecules selected in Round 2 (orange), 720 molecules selected in Round 3 (green), and 720 molecules selected in Round 4 (red). (D) Measured immunomodulatory effects of all molecules for which experimental assay measurements are available projected into the 2D t-SNE embeddings.
[0036] FIGs. 4A-4B show top-performing immunomodulator candidates. (A) The two topperforming immunomodulator candidates in each of the 12 functional objectives. We present for each molecule its chemical structure along with their code names. A bar chart shows the ex- perimentally measured log2-fold change in the immunomodulatory profile along all eight agonist-pathway combinations of interest. We highlight the immunomodulatory property that makes the candidate highly ranked in terms of activity enhancement (red) or sup- pression (blue). The 17 candidates with names high-lighted in bold text were selected for additional characterization of their cytokine release profiles (note that PME-3465 and PME-5839 each appear top-ranked twice). (B) An additional eight candidates with outstanding immunomodulatory profiles that were selected for additional cytokine characterization. Although four of these molecules were determined to be non- viable in our confluency mask test for cytostatic or cytotoxic behavior (PME-3878, PME-3386, PME-5920, PME-4007), their exceptional immunomodulation capacity induced us to subject them to additional characterization.
[0037] FIGs. 5A-5H show immunomodulator design rules for each of the eight agonistpathway combinations exposed by LASSO regression. Illustration of the up to six largest magnitude non-zero regression coefficients for LASSO linear regression models to predict the log2-fold change in immunomodulatory activity as a function of the presence or absence of particular molecular fragments or functional groups. The features with positive weights are displayed in black text, while the ones with negative weights are displayed in red text. Positive weights imply that there is a positive correlation between the feature values and enhanced immunomodulation, while negative weights indicate a positive correlation between the feature values and inhibitory immunomodulation.
[0038] FIGs. 6A-6B show low-throughput measurement of cytokine release profiles within pri- mary cells of 17 top-performing candidates. (A) Immunomodulation of 17 selected top-performing candidates over the release profiles of 13 cytokines activated by LPS, CpG and cGAMP, shown as suppression and enhancement, repectively. The extent of immunomodulation is visualized as color-maps depicting the log2-fold change values, in form of heat-maps contrasting corresponding cytokines and immunomodulator indices. We are primarily interested in TNF-a and IFN-/A since NF-KB activation is correlated with increases in proin- flammatory cytokines such as TNF-a, whereas IRF activation is related to production of IFN- >. The top ranked immunomodulators show substantial capabilities in enhancing and inhibiting the production of TNF-a and IFN-/A as well as other cytokines such as IL-6, IL- 27 and IL- la. (B) Co-delivery of PME-4007 with cGAMP increases IFN-/1 secretion by more than three-fold relative to cGAMP alone. MSA-2 is a potent STING agonist that can stim- ulate IFN-/1 more strongly than cGAMP. While with the modulation induced by PME-4007 at a low concentration (2 /rM), the stimulation of cGAMP (10 /ig/mL) is enhanced to be significantly stronger than that induced by MSA-2 (10 /ig/mL). PME-4007 also enhances TNF-a secretion and helps reach a comparable level of that induced by MSA-2. Statistical analyses between agonist and modulator versus agonist alone were performed by a one-way ANOVA test (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001). [0039] FIG. 7 shows architecture of VAE for deep representational learning of molecular latent space. The input and output of the VAE model are 7535-element one-hot vector that are one-to-one with the SELFIES string representation of each molecule. The encoder stack comprises three fully-connected feedforward layers with a 500-200-100 architecture. A bottleneck layer comprising 100 neurons defines the 100D latent space embedding. The decoder comprises three stacked GRUs each containing 200 neurons.
[0040] FIGs. 8A-8C shows VAE training curves over the augmented ZINC training library. (A) VAE loss function (Equation SI). (B) KLD component of VAE loss function (Equation S3, second term). The dashed vertical lines in each panel denote the epochs at which the KLD loss coefficient a was tuned from an initial value of 10-2 to 10-3 then 10-4. (C) Exact reconstruction accuracy reporting the fraction of molecules in the training data whose SELFIES representations are exactly reconstructed by the trained network.
[0041] FIG. 9 shows correlation heatmap and cluster dendrogram of highly correlated descriptors. We selected 13 descriptors that had a Pearson correlation coefficient greater than 0.95 with at least one other descriptor. Clustering analyses were conducted to organize highly correlated descriptors into larger groups/families. In this way, we identified six de- scriptor groups with high correlation (red blocks shown in the clustered heat map). For each of the six descriptor groups, we select one representative descriptor. Thus, we retained six descriptors, namely “fr_Al_OH”, “fr_Ar_OH”, “ ‘fr_phos_acid”, “fr_Ar_NH”, “fr_COO”, “fr_nitro”.
[0042] FIG. 10 shows the performance of LASSO regression models evaluated with respect to sparsity regularization parameter //,. The plot presents the number of molecular descriptors that have non-zero learned coefficient values, which were identified by training the model using a range of //. values (displayed on the left y-axis in orange). Additionally, the plot shows the 5-fold cross-validation mean absolute error (MAE) score of the LASSO regression model at each corresponding //, value (displayed on the right y-axis in purple). The optimal //, is shown in red points where the corresponding MAE is the lowest. The title of each plot denotes the corresponding immune signaling pathway and agonist used for stimulation. The MAE of the model with the optimal //, is reported in the upper-right comer of each panel with error bars estimated by 5-fold cross validation. In all cases, the MAE of the LASSO regression models are poorer than those for the corresponding GPR (c/. FIG. 2), but the performance differential is at most only 32.3%.
[0043] FIG. 11 shows full accounting of nonzero learned coefficient weights 0k associated with substructure features ranked by their magnitudes of weights in units of log2 fold change. For each molecule, we generate a standardized feature vector of length 70, and all together we have a feature matrix represented as Fn,k. Using this feature matrix, we apply LASSO regression to predict the calculated immunomodulation acquired from high throughput screening experiments and identifying the learned nonzero coefficient weights Ok with the highest magnitude from the reduced feature set retained by the LASSO model. The features with positive weights are displayed in black text, while the ones with negative weights are displayed in red text. Positive weights imply that there is a positive correlation between the feature values and enhanced immunomodulation, while negative weights indicate a positive correlation between the feature values and inhibitory immunomodulation.
[0044] FIG. 12 shows results of active learning screen as raw data. NF-KB activity was measured as absorbance readings at 620 nm, and IRF activity was measured as luminescence readings in the units of relative light unit (RLU). The higher the absorbance or luminescence, the stronger the corresponding immune activity. The CpG agonist shows minimal stimulation of the IRF pathway relative to no agonist present and was dropped from subsequent analyses. [0045] FIG. 13 shows modulators (10/rM) alone minimally affects cytokine production 24 hours after addition. Each panel shows the secreted concentration for a specific cytokine stimulated by PBS (as negative control) and 17 selected top-performing immunomodulator candidates (without the addition of agonists), addressed for in total 13 cytokines. The cytokine production result from the addition of modulators in the absence of agonists is within the same order of magnitude as that of the PBS negative control, showing that the addition of modulators alone minimally affect the production of cytokines. Although for IL-6 production, there is one molecule PME-3465 that has much higher IL-6 secretion than PBS control, the magnitude of the amount of IL-6 being released is still minimal immunologically. The cytokine production is measured by LegendPlex.
[0046] FIG. 14 shows PME-4007 induces minimal cytotoxicity while enhancing the secretion of IL-6 and IL-27 stimulated by cGAMP. (A) MTT Assay indicates that MSA-2, cGAMP, and the combination of cGAMP and PME-4007 induce minimal cytotoxicity to the cells as the cell viability is all higher than the cutoff 70%, and close to the level of negative control resting cells. (B) PME-4007 slightly enhances the secretion of IL-6. MSA-2 strongly stimulates the secretion of IL-6 and is much stronger than the stimulation induced by cGAMP. (C) PME-4007 significantly enhances the secretion of IL-27 with cGAMP, and it is close to the level of stimulation induced by MSA-2.
[0047] FIG. 15 shows maximum pairwise Tanimoto similarities between the 2674 previously screened molecules used to train our initial GPR models and the top-performing candidates capable of >2-fold enhancement/inhibition identified by our active learning screen. Tanimoto similarities between all pairs associating one molecule in the previous screen and one molecule in the active learning screen were computed with ECFP4 molecular fingerprint using RDKit.31 The maximum Tanimoto similarities for each of the molecules identified in the active learning screen were determined by max; Tanimoto(xz, x;), where xz denoted the zlh molecule in the active learning screen and x; denoted the /lh molecule in the previous screen. The Tanimoto similarity quantifies the proportion of chem- ical substructures shared by a pair of molecules and it is a continuous number between 0 and 1.0, where 1.0 denotes complete identity. Despite the peak near 1.0 showing that the model samples molecules with similar structure to those known good immunomodulators, the large support for the histogram in the 0-0.6 region indicates that the active learning screen is exploring molecules that are substantially diverse from the initial training data.
[0048] FIG. 16 shows maximum pairwise Tanimoto similarities between the 2674 previously screened molecules used to train our initial GPR models and the top-performing candidates capable of >2-fold enhancement/inhibition identified by our active learning screen. Tanimoto similarities between all pairs associating one molecule in the previous screen and one molecule in the active learning screen were computed with ECFP4 molecular fingerprint using RDKit.31 The maximum Tanimoto similarities for each of the molecules identified in the active learning screen were determined by max; Tanimoto(xz, x;), where xz denoted the zlh molecule in the active learning screen and x; denoted the
Figure imgf000015_0001
molecule in the previous screen. The Tanimoto similarity quantifies the proportion of chem- ical substructures shared by a pair of molecules and it is a continuous number between 0 and 1.0, where 1.0 denotes complete identity. Despite the peak near 1.0 showing that the model samples molecules with similar structure to those known good immunomodulators, the large support for the histogram in the 0-0.6 region indicates that the active learning screen is exploring molecules that are substantially diverse from the initial training data.
[0049] FIG. 17 shows 384-Well plate layout. One column and two rows on edges are not used to test and filled with cell culture media to avoid error induced by evaporation of water.
[0050] FIG. 18 shows immunomodulation of 17 selected top-performing candidates over cytokine release profiles activated by LPS, CpG and cGAMP. The extent of immunomodulation is visualized as bar plots with error bars indicating standard errors associated with each experiment. The bars are also color-coded by the log2 fold change values for better clarity in showing those immunomodulation with large extent, representing potent enhancers or suppressors. [0051] FIG. 19 shows a block diagram of an example computer implemented method encompassing methods disclosed herein herein.
[0052] FIG. 20 shows a block diagram of an example drug- screening method encompassing methods disclosed herein
[0053] FIG. 21 shows analysis of the influence of featurization and regression model upon predictive performance of the surrogate model. All combinations of featurizations - VAE and Morgan fingerprints (FP) - and regression model - GPR and linear regression (LR) were compared: (1) VAE-GPR; (2) VAE-LR; (3) FP-GPR; and (4) FP-LR. Performance is compared based on the MAE in the log2-fold change in six immunomogical functional goals. The mean MAE is shown as the bars and the standard deviation of the MAE computed from 5-fold cross validation is shown as the error bars. The simplest model, FP-LR, has a poor MAE compared to the other three. The three remaining model combinations exhibit similar performance.
[0054] FIG. 22 shows top performers identified in the active learning-assisted screen and their closest analog in the 2674 molecules from a prior screen. The two top-performing immunomodulator candidates in each of the 12 functional objectives. Presented, for each molecule, is its chemical structure along with their code names. Below each best performer or second-best performer, we present the chemical structure, code names, and the Tanimoto similarity score of the closest analog to it. The Tanimoto similarity quantifies the proportion of chemical substructures shared by a pair of molecules and it is a continuous number between 0 and 1.0, where 1.0 denotes complete topological identity. Of the twelve closest analogs to the top performing molecules, 6/12 (50%) possess a Tanimoto similarity score to the top performer of less than 0.4 and 10/12 (83%) possess a score of less than 0.5, indicative of a significant level of dissimilarity in the chemical structure.
[0055] FIG. 23 shows comparison of the number of immunomodulators with 1.5x, 2x, 5x, and lOx enhancement or suppression identified in the present work compared to a prior screen. For each immunological functional goal (rows), the number of compounds meeting the foldchange threshold (columns) is presented. In each cell, the left number represents the number of candidates meeting this criterion identified from our prior work, and the right number represents the number identified in the present work. In many combinations of functional goals and cutoffs, the active learning-guided screen yielded a significant increase in the population of immunomodulator candidates, and in five instances identified candidates with activity levels that were not achieved in our prior screen. DETAILED DESCRIPTION OF THE INVENTION
[0056] The innate immune response is vital for the success of prophylactic vaccines and immunotherapies. Control of signaling in innate immune pathways can improve prophylactic vaccines by inhibiting unfavorable systemic inflammation and immunotherapies by enhancing immune stimulation. Aspects herein developed a machine learning-enabled active learning pipeline to guide in vitro experimental screening and discovery of small molecule immunomodulators that improve immune responses by altering the signaling activity of innate immune responses stimulated by traditional pattern recognition receptor agonists. In aspects herein, molecules were tested by in vitro high throughput screening (HTS) where modulation of the nuclear factor K-light-chain-enhancer of activated B-cells (NF-KB) and the interferon regulatory factors (IRF) pathways were measured. In certain aspects, these data were used to train data-driven predictive models linking molecular structure to modulation of the NF-KB and IRF responses using deep representational learning, Gaussian process regression, and Bayesian optimization. In certain aspects, by interleaving successive rounds of model training and in vitro HTS, an active learning-guided traversal of a 139,998 molecule library was performed.
[0057] In certain aspects, a small portion of a library, including less than or approximately 1, 2, or 3 percent (or any range derivable therein) of the library can be used to discover viable molecules with desired activity, including those capable of suppressing NF-KB activity by up to 15-fold, elevating NF-KB activity by up to 5-fold, and elevating IRF activity by up to 6-fold. In certain aspects, chemical design rules can be extracted to identify particular chemical fragments as principal drivers of activity, such as specific immunomodulation behaviors. Aspects herein validate the immunomodulatory effect of a subset of top candidates by measuring activity in vitro or in vivo, including measuring cytokine release profiles. For example, one molecule induced a 3-fold enhancement in IFN-/1 production when delivered with a cyclic di-nucleotide stimulator of interferon genes (STING) agonist. In certain aspects, machine learning-enabled screening approaches described herein lead to an efficient discovery pipeline that has furnished a library of novel small molecules with desired activity, including a strong capacity to enhance or suppress innate immune signaling pathways to shape and improve prophylactic vaccination and immunotherapies.
I. Data Acquisition
[0058] Input data for the methods described herein may comprise data collected from cell screening methods. In various aspects, data is collected for performing methods disclosed herein. In some aspects, high-throughput screening data is collected. The data can be collected by incubating a target population, such as a population of cells, with a library of candidate compounds. The data may be collected by any method disclosed herein. In certain aspects, the data is collected in a high-throughput manner, including using high-throughput culture plates and high-throughput analyzing apparatuses.
[0059] In some aspects, the method comprises contacting a population of cells and a library of candidate compounds. The population of cells may comprise a reporting system, such as one or more reporter genes. The reporting system may provide a signal when a target of interest, such as a drug target of interest, is affected by a candidate compound. In certain aspects, the contacting produces an interpretable signal. In certain aspects, the interpretable signal comprises luminescence produced by the population of cells. In certain aspects, the interpretable signal is detected by a plate reader or other apparatus capable of detecting the signal. The signal can be used in one or more of the methods described herein.
High-throughput screening assays
[0060] High-throughput screening assays can be performed using methods disclosed herein. In some aspects, reporter cells, such as RAW-Dual™ macrophages as a reporter cell line,t can quantitatively report target of interest activity via a reporter molecule, such as secreted alkaline phosphatase (SEAP) and/or Lucia luciferase using substrates such as QUANTI-Blue™ and QUANTI-Luc™ for absorbance and luminescence readings. In some aspects, cells can be seed into well-plates, such as 384-well plate, which are then contacted with the compounds of interest. In some aspects, the activity is measured, such as by absorbance or luminescence readings. In some aspects, the raw readings of absorbance and luminescence of each well are divided by the average reading of the positive controls (presence of agonists and absence of immunomodulators) which are on the same plate to define the fold change associated with each modulator relative to the baseline of the corresponding agonist. This plate-based normalization can ensure measurement consistency by eliminating plate-to- plate and day-to-day variance. Each immunomodulator can be incubated in two replicated plates and the results can be averaged. Experimental errors can be calculated from the standard deviation of the mean calculated from the two replicates with the same compound of interest using standard propagation of errors. A schematic illustration of the experimental screening process is presented in FIG. 16. An illustration of the plate layout used in the HTS experiments is presented in FIG. 17. [0061] In some aspects, viability is monitored after overnight addition by monitoring confluency to insure the compounds of interest are not cytotoxic or cytostatic.
[0062]
II. Computer Implemented Systems
[0063] FIG. 19 is a block diagram illustrating a computer system 100 upon which aspects of the present teachings may be implemented. In various aspects of the present teachings, computer system 100 can include a bus 102 or other communication mechanism for communicating information and a processor 104 coupled with bus 102 for processing information. In various aspects, computer system 100 can also include a memory, which can be a random-access memory (RAM) 106 or other dynamic storage device, coupled to bus 102 for determining instructions to be executed by processor 104. Memory can also be used for storing temporary variables or other intermediate information during execution of instructions to be executed by processor 104. In various aspects, computer system 100 can further include a read only memory (ROM) 108 or other static storage device coupled to bus 102 for storing static information and instructions for processor 104. A storage device 110, such as a magnetic disk or optical disk, can be provided and coupled to bus 102 for storing information and instructions.
[0064] In various aspects, computer system 100 can be coupled via bus 102 to a display 112, such as a cathode ray tube (CRT) or liquid crystal display (LCD), for displaying information to a computer user. An input device 114, including alphanumeric and other keys, can be coupled to bus 102 for communication of information and command selections to processor 104. Another type of user input device is a cursor control 116, such as a mouse, a trackball or cursor direction keys for communicating direction information and command selections to processor 104 and for controlling cursor movement on display 112. This input device 114 typically has two degrees of freedom in two axes, a first axis (i.e., x) and a second axis (i.e., y), that allows the device to specify positions in a plane. However, it should be understood that input devices 114 allowing for 3-dimensional (x, y and z) cursor movement are also contemplated herein.
[0065] Consistent with certain implementations of the present teachings, results can be provided by computer system 100 in response to processor 104 executing one or more sequences of one or more instructions contained in memory 106. Such instructions can be read into memory 106 from another computer-readable medium or computer-readable storage medium, such as storage device 110. Execution of the sequences of instructions contained in memory 106 can cause processor 104 to perform the processes described herein. Alternatively, hard-wired circuitry can be used in place of or in combination with software instructions to implement the present teachings. Thus, implementations of the present teachings are not limited to any specific combination of hardware circuitry and software.
[0066] FIG. 20 is a block diagram illustrating a workflow encompassing aspects of the present teaching, including aspects related to a drug-discovery process. In various aspects, a library of compounds of interest 200 are manipulated into a latent space 204. The manipulation may be performed using aspects described herein including using a variational autoencoder. The library of compounds of interest 200 may also be used in a high-throughput screen 202. The high-throughput screen can include testing the compounds of interest 200 against one or more targets of interest to generate data. The high-throughput screen 202 may be associated with the latent space 204 then used to train a model 206. The model 206 can be used to identify or predict a set of active compounds 208. The active compounds 208 may include compounds that were not present in the library of compounds of interest 200. The active compounds 208 may then be fed back into the library of compounds 200 and the process may be repeated until a desired number or efficacy of active compounds 208 are identified.
[0067] The term “computer-readable medium” (e.g., data store, data storage, etc.) or “computer-readable storage medium” as used herein refers to any media that participates in providing instructions to processor 104 for execution. Such a medium can take many forms, including but not limited to, non-volatile media, volatile media, and transmission media. Examples of non-volatile media can include, but are not limited to, dynamic memory, such as memory 106. Examples of transmission media can include, but are not limited to, coaxial cables, copper wire, and fiber optics, including the wires that comprise bus 102.
[0068] Common forms of computer-readable media include, for example, a floppy disk, a flexible disk, hard disk, magnetic tape, or any other magnetic medium, a CD-ROM, any other optical medium, punch cards, paper tape, any other physical medium with patterns of holes, a RAM, PROM, and EPROM, a FLASH-EPROM, another memory chip or cartridge, or any other tangible medium from which a computer can read.
[0069] In addition to computer-readable medium, instructions or data can be provided as signals on transmission media included in a communications apparatus or system to provide sequences of one or more instructions to processor 104 of computer system 100 for execution. For example, a communication apparatus may include a transceiver having signals indicative of instructions and data. The instructions and data are configured to cause one or more processors to implement the functions outlined in the disclosure herein. Representative examples of data communications transmission connections can include, but are not limited to, telephone modem connections, wide area networks (WAN), local area networks (LAN), infrared data connections, NFC connections, etc.
[0070] It should be appreciated that the methodologies described herein, flow charts, diagrams and accompanying disclosure can be implemented using computer system 100 as a standalone device or on a distributed network or shared computer processing resources such as a cloud computing network.
[0071] The methodologies described herein may be implemented by various means depending upon the application. For example, these methodologies may be implemented in hardware, firmware, software, or any combination thereof. For a hardware implementation, the processing unit may be implemented within one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, micro-controllers, microprocessors, electronic devices, other electronic units designed to perform the functions described herein, or a combination thereof.
[0072] In various aspects, the methods of the present teachings may be implemented as firmware and/or a software program and applications written in conventional programming languages such as C, C++, Python, etc. If implemented as firmware and/or software, the aspects described herein can be implemented on a non-transitory computer-readable medium in which a program is stored for causing a computer to perform the methods described above. It should be understood that the various engines described herein can be provided on a computer system, such as computer system 100, whereby processor 104 would execute the analyses and determinations provided by these engines, subject to instructions provided by any one of, or a combination of, memory components 106/108/110 and user input provided via input device 114.
[0073] In describing the various aspects, the specification may have presented a method and/or process as a particular sequence of steps. However, to the extent that the method or process does not rely on the particular order of steps set forth herein, the method or process should not be limited to the particular sequence of steps described. As one of ordinary skill in the art would appreciate, other sequences of steps may be possible. Therefore, the particular order of the steps set forth in the specification should not be construed as limitations on the claims. In addition, the claims directed to the method and/or process should not be limited to the performance of their steps in the order written, and one skilled in the art can readily appreciate that the sequences may be varied and still remain within the spirit and scope of the various aspects. Similarly, any of the various system aspects may have been presented as a group of particular components. However, these systems should not be limited to the particular set of components, now their specific configuration, communication and physical orientation with respect to each other. One skilled in the art should readily appreciate that these components can have various configurations and physical orientations (e.g., wholly separate components, units and subunits of groups of components, different communication regimes between components).
[0074] Aspects herein relate to a data-driven active learning framework to guide experimental screening of immunomodulator candidates. An overview of the integrated screening approach is illustrated schematically in FIG. 1. In some aspects, the machine learning component of the framework comprises a deep representational learned projection of molecular candidates into a smooth and low-dimensional latent space suitable for regression and optimization, including by using a variational autoencoder (VAE; FIG. IB), the construction of learned QSAR models over this space, including by using Gaussian process regression (GPR, FIG. 1C), and the selection of new candidate molecules from the design space, including by using Bayesian optimization (BO, FIG. ID). The selected candidate molecules can then be passed to experimental HTS by automated robotics (FIG. 1 A). Experimental measurements can then be used to retrain and update the GPR models and inform subsequent rounds of BO candidate selection within a virtuous cycle of QSAR model training and model-guided experimental screening. Multiple passes through the active learning loop can be conducted until desired performance metrics are reached or a performance ceiling is reached, which may be evinced by a lack of improvement over a particular number of cycles.
[0075] Certain aspects utilize a variational auto-encoders (VAE), which can be used to perform deep representational learning of discrete chemical structures to embed them into lowdimensional, continuous latent space embeddings well-suited to property regression and molecular optimization. In some aspects, the VAE comprises two consecutive deep networks: an encoder converting the discrete chemical representations into fixed-length vectors defining an embedding into a latent space, and a decoder that inverts this operation to reconstruct the discrete chemical representations from their latent space projections (FIG. IB). Chemical structures can be represented to the VAE as SELFIES strings flattened into one-hot vectors. In some aspects, candidates that exist as salts or with other components are represented as a single SELFIES string in which the multiple molecules are concatenated using a special linking token. In some aspects, one or more feature rich representations (e.g, Smooth Overlap of Atomic Position (SOAP),47 Many-body Tensor Representation (MBTR),48 Coulomb matrices,49 quantum mechanically calculated full and/or partial charges and optimized atomic coordinates) are employed to furnish the learning model with additional molecular information, although these would typically entail a higher cost in both generating the featurizations and in model complexity. In some aspects, the loss function for network training comprises two objectives that are simultaneously optimized: accurate reconstruction of the chemical representations by decoding from their latent space projections and preservation of a prior distribution - frequently a multi-dimensional Gaussian - over the latent space. In some aspects, training of the network discovers a low-dimensional fixed- length vector representation of each molecule in the training data that preserves sufficient information to permit its accurate reconstruction within a compact distribution that promotes generalization to unseen data and from which it is easy to sample. As such, the latent space embedding furnishes a smooth, low-dimensional embedding in which chemical similarity of the training molecules is related to latent space proximity and that is well-suited to traversal and optimization by active learning.
[0076] In some aspects, the VAE model is constructed and trained in a defined algorithm such as PyTorch. In some aspects the parameters of the VAE model are optimized under 5-fold cross-validation (CV). In some aspects, performance is achieved by employing a 500-200-100 fully-connected feedforward network architecture as the encoder, and a stack of two gated recurrent units (GRUs) as the decoder. In some aspects, an additional 100-node layer following the encoder network serves as the 100D latent space, and it is followed by the decoded network. In some aspects, a data augmentation strategy is employed to enrich the training dataset with additional small molecule candidates, which has been widely recognized in the literature as a practice that can potentially lead to more stable models with improved generalizability and a smoother latent space. In some aspects, the VAE is trained over a training library consisting of compounds of interest augmented with molecules extracted from the known libraries, such as the ZINC library of commercially available small molecules59 and/or other commercially available compound libraries for virtual screening. In some aspects, the VAE is trained only once at the beginning of the active learning search to produce a latent space embedding of all immunomodulator candidates that was held fixed throughout all subsequent iterations of the process.
[0077] In some aspects, the molecular screen can discover novel compounds to enhance or suppress a target of interest to elicit a biological response of interest. In some aspects, a functional goal of interest, such as suppression or activation of a target of interest, is defined. After defining the functional goal, the goal may be optimized over a generated latent space, which includes embedding of the candidates learned by the VAE. In some aspects, the method includes training independent GPR surrogate models employing Gaussian (a.k.a. radial basis function) kernels to learn an empirical mapping from the coordinates of each molecule within the 100D latent space embedding to each of the functional goals (FIG. 1C). In each round of the active learning screen, the GPR models can be trained over all candidates for which the experimental measurements of the level of modulation of the target exist. Then, this information can be used to predict the performance of all remaining candidates for which experiments have not yet been performed. This is the key step in the data-driven screening process - the trained GPR models enable the interpolation/extrapolation from the experimentally measured performance of a small number of candidates to predict the performance of all unmeasured candidates before actually conducting the experimental measurements. In this way, the surrogate models guide a prospective traversal of the candidate space by allowing the focus the time, labor, and expense of experimentation toward the most promising molecules. Importantly, the performance predictions of the GPR models in each of the functional goals are also equipped with uncertainty estimates. As such, one can account for the typically higher model uncertainties when making extrapolative predictions to molecules that lie far away in the latent space (i.e., are more chemically dissimilar) from those that have already been measured.
[0078] In some aspects, the GPR models are then interfaced with a multi-objective Bayesian optimization (BO) framework to select the best compound candidates for experimental testing in the next round of active learning (FIG. ID). For each of the GPR models, performance predictions can be made on all as-yet-untested molecules and scored each one according to the Expected Improvement (El) acquisition function. The El acquisition function can account for both the mean and uncertainty of the GPR predictions to balancing exploitation and exploration to identify candidates most likely to lead to improved performance. To select molecules across the performance goals, one can integrate the GPR models with a multi-objective Kriging believer batched sampling protocol to define a batch of 720 molecules for experimental testing. In some aspects, each of the GPR models are collated with the molecule that had not previously been selected for testing with the largest acquisition function value. This group of molecules, with duplicates removed, can then be used to retrain all GPR models under a Kriging believer approach and the GPR models polled again for their next top-ranked molecules. This process can be repeated until a desired batch of molecules, such as approximately 200, 300, 400, 500, 600, 700, 800, 900, or 1000 molecules (or any range derivable therein) are selected and sent for experimental testing. In some aspects, parallel batched selection methods are employed that mimic a sequential selection policy such as local penalization (LP)61 and parallel knowledge gradient (q-KG).
[0079] One cycle of VAE embedding, GPR training, BO sequence selection, and experimental screening completes one loop of the active learning cycle (FIGs. 1A-1D). In some aspects, one of two methods are used to monitor and determine convergence of the active learning loop. First, in some aspects, a stabilizing predictions method is employed to evaluate stabilization of the specialist GPR model predictions. To do so, one can set aside a randomly selected 100,000 candidate stop set and measure the average Bhattacharyya distance DB between the GPR posterior evaluated over this stop set in successive rounds of the active learning screen. Large average DB values indicate that the GPR posterior is still being updated over the course of additional screening rounds and thus the convergence has not been reached yet, whereas small values indicate that additional rounds are not changing the GPR predictions and can be seen as an indicator for convergence. Second, in some aspects, the performance difference method can be employed to assess the specialist GPR predictive performance by conducting 5-fold cross-validation over the accumulated labeled samples (i.e., candidates for which experimental assay measurements are available).67 When the absolute value of the crossvalidated mean average error (MAE) on the labeled data reaches an acceptably low level and/or plateaus over the course of successive rounds, this can indicate that the predictive power of the trained GPR is no longer changing with the accumulation of additional screening data and can be taken as an indication of model convergence.
[0080] Methods disclosed herein can include those described in Tang et al. (Data-driven discovery of innate immunomodulators via machine learning-guided high throughput screening, Chemical Science, 2023), which is hereby incorporated by reference in its entirety.
III. Active Compounds
[0081] Aspects herein concern active compounds. The active compounds include compounds discovered by methods disclosed herein. The active compounds may be biologically active compounds. The active compounds may have immunomodulatory activity. [0082] Disclosed herein are compounds of any of the following
Figure imgf000026_0001
PME-4637 PME-5839
Figure imgf000027_0001
PME-5246
Figure imgf000028_0001
Figure imgf000029_0001
Figure imgf000030_0001
Figure imgf000031_0001
Figure imgf000032_0001
IV. General Pharmaceutical Compositions
[0083] In some aspects, pharmaceutical compositions, including pharmaceutical compositions comprising one or more compounds described herein, are administered to a subject. Different aspects may involve administering an effective amount of a composition to a subject. In some aspects, a pharmaceutical composition may be administered to the subject to protect against or treat a condition (e.g., cancer). Alternatively, an expression vector encoding one or more such antibodies or polypeptides or peptides may be given to a subject as a preventative treatment. Additionally, such compositions can be administered in combination with an additional therapeutic agent (e.g., a chemotherapeutic, an immunotherapeutic, a bio therapeutic, etc.). Such compositions will generally be dissolved or dispersed in a pharmaceutically acceptable carrier or aqueous medium.
[0084] The phrases “pharmaceutically acceptable” or “pharmacologically acceptable” refer to molecular entities and compositions that do not produce an adverse, allergic, or other untoward reaction when administered to an animal or human. As used herein, “pharmaceutically acceptable carrier” includes any and all solvents, dispersion media, coatings, anti-bacterial and anti-fungal agents, isotonic and absorption delaying agents, and the like. The use of such media and agents for pharmaceutical active substances is well known in the art. Except insofar as any conventional media or agent is incompatible with the active ingredients, its use in immunogenic and therapeutic compositions is contemplated. Supplementary active ingredients, such as other anti-infective agents and vaccines, can also be incorporated into the compositions.
[0085] The active compounds can be formulated for parenteral administration, e.g., formulated for injection via the intravenous, intramuscular, subcutaneous, or intraperitoneal routes. Typically, such compositions can be prepared as either liquid solutions or suspensions; solid forms suitable for use to prepare solutions or suspensions upon the addition of a liquid prior to injection can also be prepared; and, the preparations can also be emulsified.
[0086] The pharmaceutical forms suitable for injectable use include sterile aqueous solutions or dispersions; formulations including, for example, aqueous propylene glycol; and sterile powders for the extemporaneous preparation of sterile injectable solutions or dispersions. In all cases the form must be sterile and must be fluid to the extent that it may be easily injected. It also should be stable under the conditions of manufacture and storage and must be preserved against the contaminating action of microorganisms, such as bacteria and fungi. [0087] A pharmaceutical composition can include a solvent or dispersion medium containing, for example, water, ethanol, polyol (for example, glycerol, propylene glycol, and liquid polyethylene glycol, and the like), suitable mixtures thereof, and vegetable oils. The proper fluidity can be maintained, for example, by the use of a coating, such as lecithin, by the maintenance of the required particle size in the case of dispersion, and by the use of surfactants. The prevention of the action of microorganisms can be brought about by various anti-bacterial and anti-fungal agents, for example, parabens, chlorobutanol, phenol, sorbic acid, thimerosal, and the like. In many cases, it will be preferable to include isotonic agents, for example, sugars or sodium chloride. Prolonged absorption of the injectable compositions can be brought about by the use in the compositions of agents delaying absorption, for example, aluminum mono stearate and gelatin.
[0088] Sterile injectable solutions are prepared by incorporating the active compounds in the required amount in the appropriate solvent with various other ingredients enumerated above, as required, followed by filtered sterilization or an equivalent procedure. Generally, dispersions are prepared by incorporating the various sterilized active ingredients into a sterile vehicle which contains the basic dispersion medium and the required other ingredients from those enumerated above. In the case of sterile powders for the preparation of sterile injectable solutions, the preferred methods of preparation are vacuum-drying and freeze-drying techniques, which yield a powder of the active ingredient, plus any additional desired ingredient from a previously sterile-filtered solution thereof.
[0089] Administration of the compositions will typically be via any common route. This includes, but is not limited to oral, or intravenous administration. Alternatively, administration may be by orthotopic, intradermal, subcutaneous, intramuscular, intraperitoneal, or intranasal administration. Such compositions would normally be administered as pharmaceutically acceptable compositions that include physiologically acceptable carriers, buffers or other excipients.
[0090] Upon formulation, solutions will be administered in a manner compatible with the dosage formulation and in such amount as is therapeutically or prophylactically effective. The formulations are easily administered in a variety of dosage forms, such as the type of injectable solutions described above.
[0091] In various aspects, the systems and methods for analyzing, parsing, and displaying scientific research can be implemented via computer software or hardware. Refer to the Appendix for further information regarding the system, devices and methods provided herein, in accordance with various aspects. V. Administration of Therapeutic Compositions
[0092] Disclosed herein are methods and systems for administering a therapeutic composition to an individual. In certain aspects, the therapeutic composition may comprise one or more of the compounds disclosed herein. The therapeutic composition may be administered in any suitable manner known in the art.
[0093] Aspects of the disclosure relate to compositions and methods comprising therapeutic compositions. The different therapies may be administered in one composition or in more than one composition, such as 2 compositions, 3 compositions, or 4 compositions. Various combinations of the agents may be employed.
[0094] The therapeutic agents of the disclosure may be administered by the same route of administration or by different routes of administration. In some aspects, the cancer therapy is administered intravenously, intramuscularly, subcutaneously, topically, orally, transdermally, intraperitoneally, intraorbitally, by implantation, by inhalation, intrathecally, intraventricularly, or intranasally. In some aspects, the antibiotic is administered intravenously, intramuscularly, subcutaneously, topically, orally, transdermally, intraperitoneally, intraorbitally, by implantation, by inhalation, intrathecally, intraventricularly, or intranasally. The appropriate dosage may be determined based on the type of disease to be treated, severity and course of the disease, the clinical condition of the individual, the individual's clinical history and response to the treatment, and the discretion of the attending physician.
[0095] The treatments may include various “unit doses.” Unit dose is defined as containing a predetermined-quantity of the therapeutic composition. The quantity to be administered, and the particular route and formulation, is within the skill of determination of those in the clinical arts. A unit dose need not be administered as a single injection but may comprise continuous infusion over a set period of time. In some aspects, a unit dose comprises a single administrable dose.
[0096] In some aspects, the therapy is administered at a dose of between 1 mg/kg and 5000 mg/kg. In some aspects, the therapy is administered at a dose of at least, at most, or about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160,
161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179,
180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198,
199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217,
218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236,
237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255,
256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274,
275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293,
294, 295, 296, 297, 298, 299, 300, 301, 302, 303, 304, 305, 306, 307, 308, 309, 310, 311, 312,
313, 314, 315, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329, 330, 331,
332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 343, 344, 345, 346, 347, 348, 349, 350,
351, 352, 353, 354, 355, 356, 357, 358, 359, 360, 361, 362, 363, 364, 365, 366, 367, 368, 369,
370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388,
389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404, 405, 406, 407,
408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 420, 421, 422, 423, 424, 425, 426,
427, 428, 429, 430, 431, 432, 433, 434, 435, 436, 437, 438, 439, 440, 441, 442, 443, 444, 445,
446, 447, 448, 449, 450, 451, 452, 453, 454, 455, 456, 457, 458, 459, 460, 461, 462, 463, 464,
465, 466, 467, 468, 469, 470, 471, 472, 473, 474, 475, 476, 477, 478, 479, 480, 481, 482, 483,
484, 485, 486, 487, 488, 489, 490, 491, 492, 493, 494, 495, 496, 497, 498, 499, 500, 501, 502,
503, 504, 505, 506, 507, 508, 509, 510, 511, 512, 513, 514, 515, 516, 517, 518, 519, 520, 521,
522, 523, 524, 525, 526, 527, 528, 529, 530, 531, 532, 533, 534, 535, 536, 537, 538, 539, 540,
541, 542, 543, 544, 545, 546, 547, 548, 549, 550, 551, 552, 553, 554, 555, 556, 557, 558, 559,
560, 561, 562, 563, 564, 565, 566, 567, 568, 569, 570, 571, 572, 600, 700, 800, 900, 1000, 1100, 1200, 1300, 1400, 1500, 1600, 1700, 1800, 1900, 2000, 2100, 2200, 2300, 2400, 2500, 2600, 2700, 2800, 2900, 3000, 3100, 3200, 3300, 3400, 3500, 3600, 3700, 3800, 3900, 4000, 4100, 4200, 4300, 4400, 4500, 4600, 4700, 4800, 4900, or 5000 mg/kg.
[0097] The quantity to be administered, both according to number of treatments and unit dose, depends on the treatment effect desired. An effective dose is understood to refer to an amount necessary to achieve a particular effect. In the practice in certain aspects, it is contemplated that doses in the range from 10 mg/kg to 200 mg/kg can affect the protective capability of these agents. Thus, it is contemplated that doses include doses of about 0.1, 0.5, 1, 5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75, 80, 85, 90, 100, 105, 110, 115, 120, 125, 130, 135, 140, 145, 150, 155, 160, 165, 170, 175, 180, 185, 190, 195, and 200, 300, 400, 500, 1000 pg/kg, mg/kg, pg/day, or mg/day or any range derivable therein. Furthermore, such doses can be administered at multiple times during a day, and/or on multiple days, weeks, or months.
[0098] In certain aspects, the effective dose of the pharmaceutical composition is one which can provide a blood level of about 1 pM to 150 pM. In another aspect, the effective dose provides a blood level of about 4 pM to 100 pM.; or about 1 pM to 100 pM; or about 1 pM to 50 pM; or about 1 pM to 40 pM; or about 1 pM to 30 pM; or about 1 pM to 20 pM; or about 1 pM to 10 pM; or about 10 pM to 150 pM; or about 10 pM to 100 pM; or about 10 pM to 50 pM; or about 25 pM to 150 pM; or about 25 pM to 100 pM; or about 25 pM to 50 pM; or about 50 pM to 150 pM; or about 50 pM to 100 pM (or any range derivable therein). In other aspects, the dose can provide the following blood level of the agent that results from a therapeutic agent being administered to a subject: about, at least about, or at most about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29,
30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54,
55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79,
80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, or 100 pM or any range derivable therein. In certain aspects, the therapeutic agent that is administered to a subject is metabolized in the body to a metabolized therapeutic agent, in which case the blood levels may refer to the amount of that agent. Alternatively, to the extent the therapeutic agent is not metabolized by a subject, the blood levels discussed herein may refer to the unmetabolized therapeutic agent.
[0099] Precise amounts of the therapeutic composition also depend on the judgment of the practitioner and are peculiar to each individual. Factors affecting dose include physical and clinical state of the patient, the route of administration, the intended goal of treatment (alleviation of symptoms versus cure) and the potency, stability and toxicity of the particular therapeutic substance or other therapies a subject may be undergoing.
[0100] It will be understood by those skilled in the art and made aware that dosage units of pg/kg or mg/kg of body weight can be converted and expressed in comparable concentration units of pg/ml or mM (blood levels). It is also understood that uptake is species and organ/tissue dependent. The applicable conversion factors and physiological assumptions to be made concerning uptake and concentration measurement are well-known and would permit those of skill in the art to convert one concentration measurement to another and make reasonable comparisons and conclusions regarding the doses, efficacies and results described herein. [0101] In certain instances, it will be desirable to have multiple administrations of the composition, e.g., 2, 3, 4, 5, 6 or more administrations. The administrations can be at 1, 2, 3, 4, 5, 6, 7, 8, to 5, 6, 7, 8, 9, 10, 11, or 12 week intervals, including all ranges there between. [0102] The phrases “pharmaceutically acceptable” or “pharmacologically acceptable” refer to molecular entities and compositions that do not produce an adverse, allergic, or other untoward reaction when administered to an animal or human. As used herein, “pharmaceutically acceptable carrier” includes any and all solvents, dispersion media, coatings, anti-bacterial and anti-fungal agents, isotonic and absorption delaying agents, and the like. The use of such media and agents for pharmaceutical active substances is well known in the art. Except insofar as any conventional media or agent is incompatible with the active ingredients, its use in immunogenic and therapeutic compositions is contemplated. Supplementary active ingredients, such as other anti-infective agents and vaccines, can also be incorporated into the compositions.
[0103] The active compounds can be formulated for parenteral administration, e.g., formulated for injection via the intravenous, intramuscular, subcutaneous, or intraperitoneal routes. Typically, such compositions can be prepared as either liquid solutions or suspensions; solid forms suitable for use to prepare solutions or suspensions upon the addition of a liquid prior to injection can also be prepared; and, the preparations can also be emulsified.
[0104] The pharmaceutical forms suitable for injectable use include sterile aqueous solutions or dispersions; formulations including, for example, aqueous propylene glycol; and sterile powders for the extemporaneous preparation of sterile injectable solutions or dispersions. In all cases the form must be sterile and must be fluid to the extent that it may be easily injected. It also should be stable under the conditions of manufacture and storage and must be preserved against the contaminating action of microorganisms, such as bacteria and fungi.
[0105] A pharmaceutical composition can include a solvent or dispersion medium containing, for example, water, ethanol, polyol (for example, glycerol, propylene glycol, and liquid polyethylene glycol, and the like), suitable mixtures thereof, and vegetable oils. The proper fluidity can be maintained, for example, by the use of a coating, such as lecithin, by the maintenance of the required particle size in the case of dispersion, and by the use of surfactants. The prevention of the action of microorganisms can be brought about by various anti-bacterial and anti-fungal agents, for example, parabens, chlorobutanol, phenol, sorbic acid, thimerosal, and the like. In many cases, it will be preferable to include isotonic agents, for example, sugars or sodium chloride. Prolonged absorption of the injectable compositions can be brought about by the use in the compositions of agents delaying absorption, for example, aluminum mono stearate and gelatin.
[0106] Sterile injectable solutions are prepared by incorporating the active compounds in the required amount in the appropriate solvent with various other ingredients enumerated above, as required, followed by filtered sterilization or an equivalent procedure. Generally, dispersions are prepared by incorporating the various sterilized active ingredients into a sterile vehicle which contains the basic dispersion medium and the required other ingredients from those enumerated above. In the case of sterile powders for the preparation of sterile injectable solutions, the preferred methods of preparation are vacuum-drying and freeze-drying techniques, which yield a powder of the active ingredient, plus any additional desired ingredient from a previously sterile-filtered solution thereof.
[0107] Administration of the compositions will typically be via any common route. This includes, but is not limited to oral, or intravenous administration. Alternatively, administration may be by orthotopic, intradermal, subcutaneous, intramuscular, intraperitoneal, or intranasal administration. Such compositions would normally be administered as pharmaceutically acceptable compositions that include physiologically acceptable carriers, buffers or other excipients.
[0108] Upon formulation, solutions will be administered in a manner compatible with the dosage formulation and in such amount as is therapeutically or prophylactically effective. The formulations are easily administered in a variety of dosage forms, such as the type of injectable solutions described above.
VI. Chemical Definitions
[0109] Any apparently unfulfilled valency is to be understood to be properly filled by hydrogen atom(s). For example, a compound with a substituent of -O or -N is to be understood to be -OH or -NH2, respectively.
[0110] Any genus, subgenus, or specific compound discussed herein is specifically contemplated as being excluded from any aspect described herein.
[0111] Compounds described herein may be prepared synthetically using conventional organic chemistry methods known to those of skill in the art and/or are commercially available (e.g., ChemBridge Co., San Diego, Calif.; MCule, Palo Alto, CA; etc.).
[0112] The claimed invention is also intended to encompass salts of any of the compounds of the present invention. The term “salt(s)” as used herein, is understood as being acidic and/or basic salts formed with inorganic and/or organic acids and bases. Zwitterions (internal or inner salts) are understood as being included within the term “salt(s)” as used herein, as are quaternary ammonium salts such as alkylammonium salts. Nontoxic, pharmaceutically acceptable salts are preferred, although other salts may be useful, as for example in isolation or purification steps during synthesis. Salts include, but are not limited to, sodium, lithium, potassium, amines, tartrates, citrates, hydrohalides, phosphates and the like. A salt may be a pharmaceutically acceptable salt, for example. Thus, pharmaceutically acceptable salts of compounds of the present invention are contemplated.
[0113] The term “pharmaceutically acceptable salts,” as used herein, refers to salts of compounds of this invention that are substantially non-toxic to living organisms. Typical pharmaceutically acceptable salts include those salts prepared by reaction of a compound of this invention with an inorganic or organic acid, or an organic base, depending on the substituents present on the compounds of the invention.
[0114] Non-limiting examples of inorganic acids which may be used to prepare pharmaceutically acceptable salts include: hydrochloric acid, phosphoric acid, sulfuric acid, hydrobromic acid, hydroiodic acid, phosphorous acid and the like. Examples of organic acids which may be used to prepare pharmaceutically acceptable salts include: aliphatic mono- and dicarboxylic acids, such as oxalic acid, carbonic acid, citric acid, succinic acid, phenyl- heteroatom-substituted alkanoic acids, aliphatic and aromatic sulfuric acids and the like. Pharmaceutically acceptable salts prepared from inorganic or organic acids thus include hydrochloride, hydrobromide, nitrate, sulfate, pyrosulfate, bisulfate, sulfite, bisulfate, phosphate, monohydrogenphosphate, dihydrogenphosphate, metaphosphate, pyrophosphate, hydroiodide, hydrofluoride, acetate, propionate, formate, oxalate, citrate, lactate, p- toluenesulfonate, methanesulfonate, maleate, and the like.
[0115] Suitable pharmaceutically acceptable salts may also be formed by reacting the agents of the invention with an organic base such as methylamine, ethylamine, ethanolamine, lysine, ornithine and the like.
[0116] Pharmaceutically acceptable salts include the salts formed between carboxylate or sulfonate groups found on some of the compounds of this invention and inorganic cations, such as sodium, potassium, ammonium, or calcium, or such organic cations as isopropylammonium, trimethylammonium, tetramethylammonium, and imidazolium.
[0117] Derivatives of compounds of the present invention are also contemplated. In certain aspects, “derivative” refers to a chemically modified compound that still retains the desired effects of the compound prior to the chemical modification. Such derivatives may have the addition, removal, or substitution of one or more chemical moieties on the parent molecule. Non-limiting examples of the types modifications that can be made to the compounds and structures disclosed herein include the addition or removal of lower alkanes such as methyl, ethyl, propyl, or substituted lower alkanes such as hydroxymethyl or aminomethyl groups; carboxyl groups and carbonyl groups; hydroxyls; nitro, amino, amide, and azo groups; sulfate, sulfonate, sulfono, sulfhydryl, sulfonyl, sulfoxido, phosphate, phosphono, phosphoryl groups, and halide substituents. Additional modifications can include an addition or a deletion of one or more atoms of the atomic framework, for example, substitution of an ethyl by a propyl; substitution of a phenyl by a larger or smaller aromatic group. Alternatively, in a cyclic or bicyclic structure, heteroatoms such as N, S, or O can be substituted into the structure instead of a carbon atom.
[0118] Compounds of the present invention may contain one or more asymmetrically- substituted carbon or nitrogen atoms, and may be isolated in optically active or racemic form. Thus, all chiral, diastereomeric, racemic form, epimeric form, and all geometric isomeric forms of a structure are intended, unless the specific stereochemistry or isomeric form is specifically indicated. Compounds may occur as racemates and racemic mixtures, single enantiomers, diastereomeric mixtures and individual diastereomers. In some aspects, a single diastereomer is obtained. The chiral centers of the compounds of the present invention can have the S- or the R-configuration, as defined by the IUPAC 1974 Recommendations. Compounds may be of the D- or L-form, for example. It is well known in the art how to prepare and isolate such optically active forms. For example, mixtures of stereoisomers may be separated by standard techniques including, but not limited to, resolution of racemic form, normal, reverse-phase, and chiral chromatography, preferential salt formation, recrystallization, and the like, or by chiral synthesis either from chiral starting materials or by deliberate synthesis of target chiral centers. Compounds of the present invention may occur as a hydrate, a compound containing an equivalent of water in the form of an H2O molecule, or polyhydrate, a compound containing more than one equivalent of water in the form of H2O molecules.
[0119] In addition, atoms making up the compounds of the present invention are intended to include all isotopic forms of such atoms. Isotopes, as used herein, include those atoms having the same atomic number but different mass numbers. By way of general example and without limitation, isotopes of hydrogen include tritium and deuterium, and isotopes of carbon include 13C and 14C.
[0120] As noted above, compounds of the present invention may exist in prodrug form. As used herein, “prodrug” is intended to include any covalently bonded carriers which release the active parent drug or compounds that are metabolized in vivo to an active drug or other compounds employed in the methods of the invention in vivo when such prodrug is administered to a subject. Since prodrugs are known to enhance numerous desirable qualities of pharmaceuticals (e.g., solubility, bioavailability, manufacturing, etc.), the compounds employed in some methods of the invention may, if desired, be delivered in prodrug form. Thus, the invention contemplates prodrugs of compounds of the present invention as well as methods of delivering prodrugs. Prodrugs of the compounds employed in the invention may be prepared by modifying functional groups present in the compound in such a way that the modifications are cleaved, either in routine manipulation or in vivo, to the parent compound.
[0121] Accordingly, prodrugs include, for example, compounds described herein in which a hydroxy, amino, or carboxy group is bonded to any group that, when the prodrug is administered to a subject, cleaves to form a free hydroxyl, free amino, or carboxylic acid, respectively. Other examples include, but are not limited to, acetate, formate, and benzoate derivatives of alcohol and amine functional groups; and alkyl, carbocyclic, aryl, and alkylaryl esters such as methyl, ethyl, propyl, iso-propyl, butyl, isobutyl, sec -butyl, tert-butyl, cyclopropyl, phenyl, benzyl, and phenethyl esters, and the like.
[0122] It should be recognized that the particular anion or cation forming a part of any salt of this invention is not critical, so long as the salt, as a whole, is pharmacologically acceptable. Additional examples of pharmaceutically acceptable salts and their methods of preparation and use are presented in Handbook of Pharmaceutical Salts: Properties, Selection and Use (2002), which is incorporated herein by reference.
Examples
[0123] The following examples are included to demonstrate preferred aspects of the invention. It should be appreciated by those of skill in the art that the techniques disclosed in the examples which follow represent techniques discovered by the inventor to function well in the practice of the invention, and thus can be considered to constitute preferred modes for its practice. However, those of skill in the art should, in light of the present disclosure, appreciate that many changes can be made in the specific aspects which are disclosed and still obtain a like or similar result without departing from the spirit and scope of the invention.
Example 1: Data-Driven Discovery of Drug Candidates
[0124] The hypothesis underpinning the present work is that immunomodulators of greater potency and diversity may be discovered by screening large libraries of small molecules. A challenge in extending experimental screening efforts in this manner is the vast size of molecular space: the number of pharmacologically active molecules obeying the Lipinski rules has been estimated to be in excess of 1060.25,26 (To help place this number in context, it is estimated that there are “only” 1022 stars in the visible universe.27) Experimental screening can only probe a small fraction of these molecules due to time, labor, and materials constraints. Human intuition and experience present valuable heuristics to guide this search, but the relative infancy of immunomodulator discovery efforts, absence of detailed mechanistic understanding, and vast size of molecular space can make these heuristics limiting and subject to human preconceptions, bias, and potential blind spots. Data-driven models trained in concert with experimental screens offer a systematic means to guide traversal of molecular design space using predictive models of immunomodulation activity learned on-the-fly from the screening data. Such models, often referred to as quantitative structure activity relationship (QSAR) or quantitative structure property relationship (QSPR) models, 2X 30 have been employed in multiple applications to molecular design and discovery, including self-assembling TT- conjugated peptides,31 cardiolipin- selective small molecules,32 battery electrolytes,33 energy storage materials,34 and nanoporous materials.35 QSAR models have a venerable history dating back to the development of cheminformatics and combinatorial chemistry in the 1980’ s.36 More recently, the field of molecular design has seen significant advancements associated with the integration of powerful generative deep learning tools. These tools have played a pivotal role in enhancing molecular design workflows, where the synergy between QSAR modeling and generative deep learning has been particularly beneficial.37-39
[0125] This work reports the development of a data-driven pipeline integrating machine learning and in vitro high throughput screening (HTS) to accelerate the discovery of small molecule immunomodulators of the innate immune response. The inventors curated a library of 139,998 candidate small molecules from commercial chemical screening libraries readily available for purchase. These chemical screening libraries are pre-filtered to be structurally- diverse and drug-like. Commencing with limited experimental measurements, the inventors constructed a datadriven QSAR model integrating deep representational learning,37,40 Gaussian process regression (GPR),41 and Bayesian optimization42 to guide subsequent rounds of molecular selection and experimental HTS. After screening only 2,880 compounds comprising ~2% of the molecular search space using the QSAR/HTS framework, the inventors discovered molecules with an unprecedented ability to either up or downregulate the activity of NF-KB or IRF when activated by known agonists. The most successful candidates from the study demonstrated remarkable improvements in modulating immune activity compared to those identified in the earlier screen.24 Considering the magnitude of the immunomodulation response relative to that induced by specific agonists, these top-performing candidates realized improvement of up to 110% in elevating NF-KB activity, 83% in elevating IRF activity, and 128% in suppressing NF-KB activity. Additionally, the inventors identified 167 novel immunomodulators with at least 2-fold enhancement or suppression over transcription factor activity of interest, which represents a 105% increase in the total number of known immunomodulators with this level of activity, and nine novel immunomodulators with at least 10-fold activity modulation.
[0126] In addition to specialists-immunomodulators capable of enhancing or suppressing an immune signaling pathway when delivered in concert with a specific agonist-the inventors also discovered a number of generalists - immunomodulators capable of modulating immune signaling pathways when co-delivered with a number of agonists. These generalists, although less effective in immunomodulation for any specific agonist, represent promising versatile agents to incorporate into diverse prophylactic vaccines and immunotherapies due to their broad spectrum compatibility.
[0127] Finally, the inventors conducted additional characterization assays of 17 topperforming candidates identified in the screen to measure their cytokine release profiles in primary cells. These molecules demonstrated significant capacity to modulate the secretion of various cytokines, including upregulating TNF-a production by over 10-fold, downregulating TNF-a production by over 16-fold, and upregulating IFN-/1 production by over 6-fold. One of the top candidates demonstrated a 3-fold enhancement in IFN-/1 production when delivered with cGAMP, which is a strong stimulator of interferon genes (STING) agonist.
[0128] A deficiency of the nonlinear QSAR models based on deep representational learning is that despite their predictive power in guiding the HTS campaign, the design rules mediating the mapping from small molecule chemical structure to immunomodulatory profile is not readily available from these relatively “black box” models. Accordingly, the inventors performed a post hoc analysis of the molecules considered in the screen using interpretable “glass box” linear regression models that are expected to possess lower predictive accuracy than the nonlinear QSAR models but can transparently identify particular chemical fragments that are principal drivers of specific immunomodulation behaviors. The inventors found that the presence of halogen moiety in immunomodulators is highly predictive of suppression of NF-KB activity regardless of the type of agonist. The carbonyl and carboxyl moiety is predictive of suppression of the immune responses in NF-KB pathway activated by TLR4 agonists such as LPS and MPLA, and predictive of enhancement of the immune responses in the NF-KB pathway activated by CpG. It was also discovered that aromatic heteroatom moieties are predictive of enhancement in NF-KB activity and of suppression in IRF activity. Even though IRF pathway suppression is of limited clinical interest, this information could be used to guide the development of IRF enhancing immunomodulators by either avoiding the inclusion of aromatic heteroatom moieties or removing such chemical fragments to enhance their potency. This analysis offers design rules to modify the structure of immunomodulators for achieving practical immunomodulation goals in vaccine design or immunotherapy development, enriching candidate libraries with molecules predicted to have promising immunomodulatory behaviors.
[0129] The data-driven QSAR computation detailed in this work presents a powerful framework to integrate with experimental HTS to construct on-the-fly models of immunomodulatory activity from the screening data and use these models to efficiently traverse molecular candidate space and extract interpretable design rules. The inventors also observe that this integrated computational and experimental screening platform detailed herein is quite generic, and can be easily repurposed to other early-stage molecular discovery pipelines to guide efficient search and optimization within large molecular spaces by modular contextspecific replacement of the molecular library and experimental assays.
Example 2: Experimental Methods
High-throughput screening assays
[0130] The inventors chose transcription factor levels of NF-KB and IRF activity as the two measures of immunomodulator performance when delivered in combination with one of four PRR agonists: lipopolysaccharides (LPS),12 Monopho sphoryl-Lipid A (MPLA),13 CpG ODN 1826 (CpG),14 and 3’3’-cGAMP (cGAMP).15 NF-KB transcription levels are correlated with inflammatory responses and should typically be minimized to reduce reactogenicity in prophylactic vaccinations. Under certain circumstances, enhanced NF-KB activity is instead required to optimize the efficacy of vaccines.7,8 IRF activity, which is related to interferon production, should typically be maximized for a productive antiviral response.8,43 Following previously established protocols, the inventors employed RAW-Dual™ macrophages as a reporter cell line that can quantitatively report NF-KB and IRF activity via secreted alkaline phosphatase (SEAP) and Lucia luciferase using the proprietary substrates QUANTI-Blue™ and QUANTI-Luc™ for absorbance and luminescence readings.24 The inventors seeded 50,000 RAW-Dual™ macrophages in 384-well plates in 45 qL of complete media using a MultiDrop™ Combi liquid handler, then incubated for 1 hour at 37 °C. The inventors transferred immunomodulator compounds from source plates (lOmM in DMSO) to a final concentration of 10 /rM using a Janus® G3 via pintool. One column and two rows at the edge of each plate were left blank and filled with water to avoid possible systematic errors due to evaporation. Following 1 hour incubation at 37°C, one of the four PRR agonists was added in 5 /iL of media to achieve the desired concentration. Cells were incubated with or without agonist overnight until transcription factor activity was analyzed. This activity was measured via absorbance at 620 nm or luminescence readings using a BioTek Synergy™ Neo2 Hybrid multimode microplate reader. In the modulator-only negative control, the inventors were able to measure the net enhancement or suppression in activity of the screened compounds on the cells in the absence of agonists. By employing a supernatant based assay, the inventors were able to simultaneously observe the NF-KB and IRF activity within a single well. The raw readings of absorbance and luminescence of each well were then divided by the average reading of the positive controls (presence of agonists and absence of immunomodulators) which are on the same plate to define the fold change associated with each modulator relative to the baseline of the corresponding agonist. This plate-based normalization ensures measurement consistency by eliminating plate-to-plate and day-to-day variance. Each immunomodulator was incubated in two replicated plates and the results were averaged. Experimental errors are calculated from the standard deviation of the mean calculated from the two replicates with the same immunomodulator using standard propagation of errors. A schematic illustration of the experimental screening process is presented in FIG. 16. An illustration of the plate layout used in the HTS experiments is presented in FIG. 17.
[0131] The small molecule library source plates were hosted by the University of Chicago Cellular Screening Center (Chicago, Illinois, USA) and purchased from various vendors. The RAWDual™ cells, QUANTI-Blue™, QUANTI-Luc™ and the PRR agonists were purchased from InvivoGen (San Diego, California, USA).
Viability filter analysis
[0132] As some modulators may be cytostatic or cytotoxic to cells, viability was monitored after overnight addition by monitoring confluency via IncuCyte® imaging.44, 45 Two confluency masks were generated using IncuCyte® software over all imaged wells, and modulators were determined non-viable if both sets of confluency masks were lower than 70% of those of resting cells. This methodology was validated with select library plates using a traditional Promega® CellTiter-Glo® assay. Immunomodulators considered non-viable are not necessarily cytostatic or cytotoxic in other cell lines or clinical settings, so immunomodulator candidates exhibiting extraordinary behaviors were not categorically ruled out on the basis of this viability assay alone. However, unless explicitly stated otherwise, the inventors only report on immunomodulators measured to be viable under this assay.
Measurement of cytokine release profiles in bone marrow derived dendritic cells (BMDCs}
[0133] A subset of the top performing candidate molecules identified in the high- throughput active learning screen were subjected to low-throughput measurement of cytokine release profiles in murine bone marrow derived dendritic cells (BMDCs). Monocytes were harvested from 6-week-old C57BL/6 mice and were differentiated into dendritic cells using supplemented culture medium: RPMI 1640 (Life Technologies, Waltham, Massachusetts, USA), 10% HIFBS (Sigma-Aldrich, Burlington, Massachusetts, USA), Recombinant Mouse GM-CSF (carrier-free; 20 ng/ml; BioLegend, San Diego, California, USA), 2 mM L-glutamine (Life Technologies ®), 1% antibiotic-antimycotic (Life Technologies), and 50 /rM [>- mercaptoethanol (Sigma- Aldrich®). After 6 days of culture, BMDCs were plated at 100,000 cells per well and incubated with modulator (10 /rM). After 1 hour, one of the three tested agonists (LPS, CpG, cGAMP) was added. Cells were incubated for 24 hours at 37°C and 5% CO2. Supernatant cytokines were measured using LEGENDplex™ Mouse Inflammation Cytokine Kit (Biolegends®) or a VeriKine™ IFN-/1 ELISA (PBL Assay Science, Piscataway, New Jersey, USA). The cytokines measured by LEGENDplex™ are TNF-a, IFN-/A IFN-y, IL- 6, IL-la, IL-1 IL-10, IL-12p70, IL-17A, IL-23, IL-27, MCP-1, and GM-CSF.
Example 3: Computational Methods
[0134] The inventors developed a data-driven active learning framework to guide experimental screening of immunomodulator candidates. An overview of the integrated screening approach is illustrated schematically in Figure 1. The machine learning component of the framework comprises a deep representational learned projection of all 139,998 molecular candidates into a smooth and low-dimensional latent space suitable for regression and optimization using a variational autoencoder (VAE; Figure IB), the construction of learned QSAR models over this space using Gaussian process regression (GPR, Figure 1C), and the selection of new candidate molecules from the design space using Bayesian optimization (BO, Figure ID). The selected candidate molecules are then passed to experimental HTS by automated robotics (Figure 1A). Experimental measurements are then used to retrain and update the GPR models and inform subsequent rounds of BO candidate selection within a virtuous cycle of QSAR model training and model-guided experimental screening. Multiple passes through the active learning loop are conducted until desired performance metrics are reached or the inventors encounter a performance ceiling evinced by a lack of improvement over a particular number of cycles.
Deep representational learning using variational autoencoders (VAEs)
[0135] Previous work has shown that variational auto-encoders (VAE) can be used to perform deep representational learning of discrete chemical structures to embed them into lowdimensional, continuous latent space embeddings well-suited to property regression and molecular optimization.37,40’46 The VAE comprises two consecutive deep networks: an encoder converting the discrete chemical representations into fixed-length vectors defining an embedding into a latent space, and a decoder that inverts this operation to reconstruct the discrete chemical representations from their latent space projections40 (Figure IB). Chemical structures are represented to the VAE as SELFIES strings46 flattened into one-hot vectors. Candidates that exist as salts or with other components are represented as a single SELFIES string in which the multiple molecules are concatenated using a special linking token. This featurization implicitly adopt a 2D representation of the candidate molecules in their electrically neutral state that reflects their molecular topology (i.e., atomic connectivity) but does not capture 3D information on their conformation, protonation state, or partial charge distribution. It was anticipated that this representation would prove sufficient for the active learning screen, and, as shown below, this featurization does enable us to identify a number of high-performing new molecular candidates. It is noted that a number of more feature rich representations (e.g, Smooth Overlap of Atomic Position (SOAP),47 Many-body Tensor Representation (MBTR),48 Coulomb matrices,49 quantum mechanically calculated full and/or partial charges and optimized atomic coordinates) could be employed to furnish the learning model with additional molecular information, although these would typically entail a higher cost in both generating the featurizations and in model complexity. The loss function for network training comprises two objectives that are simultaneously optimized: accurate reconstruction of the chemical representations by decoding from their latent space projections and preservation of a prior distribution - frequently a multi-dimensional Gaussian - over the latent space. Conceptually, training of the network discovers a low-dimensional fixed-length vector representation of each molecule in the training data that preserves sufficient information to permit its accurate reconstruction within a compact distribution that promotes generalization to unseen data and from which it is easy to sample. As such, the latent space embedding furnishes a smooth, low-dimensional embedding in which chemical similarity of the training molecules is related to latent space proximity and that is well-suited to traversal and optimization by active learning.
[0136] The inventors constructed and trained the VAE model in PyTorch50 and its parameters were optimized under 5-fold cross-validation (CV). The inventors achieve good performance employing a 500-200-100 fully-connected feedforward network architecture as the encoder, and a stack of two gated recurrent units (GRUs) as the decoder. An additional 100- node layer following the encoder network serves as the 100D latent space, and it is followed by the decoded network. Although the study primarily focuses on the analysis of 139,998 immunomodulator candidates within the screening dataset, the relatively modest size of this library may present challenges in achieving robust model performance. To guard against this possibility, the inventors were motivated to employ a data augmentation strategy to enrich the training dataset with additional small molecule candidates, which has been widely recognized in the literature as a practice that can potentially lead to more stable models with improved generalizability and a smoother latent space.31 33 Accordingly, the inventors trained the VAE over an training library consisting of the 139,998 immunomodulator candidates augmented with 1,108,666 molecules extracted from the ZINC library of commercially available small molecules34 37 and other commercially available compound libraries for virtual screening. The VAE was trained only once at the beginning of the active learning search to produce a latent space embedding of all immunomodulator candidates that was held fixed throughout all subsequent iterations of the process. Training was conducted on a NVIDIA RTX 2080 GPU card requiring 9032 epochs and 1344 GPU-hours of training over the 1,248,664 candidate augmented training library.
Gaussian process regression (GPR) surrogate model training
[0137] The primary goals of the molecular screen are to discover novel immunomodulators to enhance or suppress the NF-KB response as a means of, respectively, upregulating innate immune stimulation or downregulating inflammation in prophylactic vaccination to promote efficacy and safety, and to enhance the IRF response to upregulate production of type-I interferons and promote antiviral responses in cancer immunotherapies. This motivated us to discover molecules capable of enhancing the NF-KB response, suppressing the NF-KB response, and enhancing the IRF response. The inventors did not search for molecules to inhibit IRF pathway, since this is presently of limited clinical significance.
[0138] The inventors consider four agonists in this work: LPS, MPLA, CpG, and cGAMP. LPS and MPLA both target TLR4, CpG targets TLR9, and cGAMP targets STING.12 15 The inventors sought to discover molecules that are specialists in achieving large immunomodulatory effects when delivered with one particular agonist, and those that are generalists in doing so when delivered with any one of a particular group of agonists. For example, a molecule that enhances the NF-KB response when delivered with LPS would be regarded as a specialist enhancer of NF/cB operating through the TLR4 receptor, whereas one that suppresses the NF-KB response when delivered with any one of LPS, MPLA, or CpG would be regarded as a generalist suppressor operating through the TLR4 and TLR9 receptors. It is expected that generalists may not be as good at immunomodulation in concert with any single agonist, but offer better immunomodulation profiles across multiple agonists. The inventors quantify the performance of a specialist for a particular agonist via the fold change in the NF-KB or IRF transcription factor activity induced by co-delivery of the immunomodulator with the agonist relative to delivery of the agonist alone. The inventors quantify the performance of a generalist over a group of agonists as the average fold-change over that group.
[0139] As illustrated in Figure IE the eight agonist-pathway combinations representing the 12 functional immunomodulatory goals of the immunomodulator screen. The rows of the matrix comprise the agonists or groups of agonists, and the columns comprise the signaling pathway and enhancement / suppression thereof. Considering the columns of the table, downregulation of the IRF pathway is of limited clinical interest and so is not included in the screening goals. Considering now the rows, the inventors seek immunomodulatory specialists that exert large enhancement or suppressive effects when co-delivered with LPS, MPLA, CpG, or cGAMP agonists alone. The inventors also seek an NF-KB generalist capable of enhancement or suppression of the NF-KB pathway when delivered in concert with any one of the LPS, MPLA, or CpG agonists, and an IRF generalist capable of enhancing the IRF pathway when delivered in concert with any one of the LPS, MPLA, or cGAMP agonists. It is noted that the cGAMP agonist is known to primarily affect the IRF pathway and so was not considered as either a specialist or generalist for NF-KB immunomodulation. Similarly, CpG is known to primarily affect the NF-KB pathway, and so was not considered as either a specialist or generalist for IRF immunomodulation. [0140] Having defined the 12 functional goals it was the next task to optimize these functionsover the smooth, low-dimensional latent space embedding of the 139,998 immunomodulator candidates learned by the VAE.37 To do so, the inventors trained 12 independent GPR surrogate models employing Gaussian (a.k.a. radial basis function) kernels to learn an empirical mapping from the coordinates of each molecule within the 100D latent space embedding to each of the 12 objective functions (Figure 1C). In each round of the active learning screen, the GPR models were trained over all immunomodulator candidates for which the inventors had experimental measurements of the level of modulation of the NF-KB and IRF responses, and were then used to predict the performance of all remaining candidates for which experiments had not yet been performed. This is the key step in the data-driven screening process - the trained GPR models enable us to interpolate/extrapolate from the experimentally measured performance of a small number of candidates to predict the performance of all unmeasured candidates before actually conducting the experimental measurements. In this way, the surrogate models guide a prospective traversal of the candidate space by allowing us to focus the time, labor, and expense of experimentation toward the most promising molecules. Importantly, the performance predictions of the GPR models in each of the 12 design goals are also equipped with uncertainty estimates. As such, the inventors can account for the typically higher model uncertainties when making extrapolative predictions to molecules that lie far away in the latent space (i.e., are more chemically dissimilar) from those that have already been measured. In the first round of the active learning screen, initial GPR models were trained over measurements for 2674 molecules within the candidate space for which the inventors previously conducted experimental screening in the prior work.24 In subsequent passes through the active learning loop the inventors retrained the GPR models over these original measurements plus all new measurements conducted in subsequent passes through the loop. Full details of the GPR kernel, training, and hyperparameter tuning are provided in the herein. This simple ablation study indicates that the nonlinear GPR surrogate model substantially outperforms simple linear regression in predictive accuracy, but that a simple fingerprint-based featurization can perform on par, albeit with a ~ 20-fold larger dimensionality, with the 100D VAE embedding.
Multi-objective Bayesian optimization (BO) candidate selection
[0141] The GPR models were then interfaced with a multi-objective Bayesian optimization (BO) framework to select the best compound candidates for experimental testing in the next round of active learning35 (Figure ID). For each of the 12 GPR models, the inventors made performance predictions on all as-yet-untested molecules and scored each one according to the Expected Improvement (El) acquisition function.58 The El acquisition function accounts for both the mean and uncertainty of the GPR predictions to balancing exploitation and exploration to identify candidates most likely to lead to improved performance. To select molecules across the 12 performance goals, the inventors integrated the 12 GPR models with a multi-objective Kriging believer batched sampling protocol to define a batch of 720 molecules for experimental testing.59,60 Specifically, the inventors collated from each of the 12 GPR models the molecule that had not previously been selected for testing with the largest acquisition function value. This group of 12 molecules, with duplicates removed, was then used to retrain all 12 GPR models under a Kriging believer approach and the GPR models polled again for their next topranked molecules. The inventors repeated this process until a batch of at least 720 molecules were selected and sent for experimental testing. Importantly, all selected molecules, regardless of their source GPR models, were subjected to experimental testing to evaluate their immunomodulatory profiles across all agonist types in both the NF-KB and IRF pathways. Hence, molecules selected by one GPR model were not only used to retrain that particular GPR model, but also used to retrain all other models. It is noted that while the Kriging believer is a well-established batched sampling procedure with a strong theoretical basis, the sequential nature by which candidates are selected means that it can scale poorly with dimensionality and batch size.61 In the present work, approximately 24-48 hours is required to select the 720- molecule batch using a large memory high performance compute cluster. This time scale is approximately an order of magnitude smaller than the 1-2 weeks required to conduct each experimental screening round. As such, the batched selection procedure does not throttle the overall screening process. For time- sensitive applications, or scenarios where faster candidate selection is imperative, it can be desirable to employ parallel batched selection methods that mimic a sequential selection policy such as local penalization (LP)61 and parallel knowledge gradient (q-KG).62 Full details of the BO acquisition function, integrated Kriging believer batched sampling and stopping criteria analysis are provided herein. Convergence assessment
[0142] One cycle of VAE embedding, GPR training, BO sequence selection, and experimental screening completes one loop of the active learning cycle (Figure 1A-D). The inventors use two methods to monitor and determine convergence of the active learning loop. First, the inventors employ a stabilizing predictions method to evaluate stabilization of the specialist GPR model predictions.63 To do so, the inventors set aside a randomly selected 100,000 candidate stop set and measure the average Bhattacharyya distance64 DB between the GPR posterior evaluated over this stop set in successive rounds of the active learning screen. Large average DB values indicate that the GPR posterior is still being updated over the course of additional screening rounds and thus the convergence has not been reached yet, whereas small values indicate that additional rounds are not changing the GPR predictions and can be seen as an indicator for convergence. Second, the inventors employ the performance difference method to assess the specialist GPR predictive performance by conducting 5-fold cross- validation over the accumulated labeled samples (i.e., candidates for which experimental assay measurements are available).65 When the absolute value of the cross-validated mean average error (MAE) on the labeled data reaches an acceptably low level and/or plateaus over the course of successive rounds, this indicates that the predictive power of the trained GPR is no longer changing with the accumulation of additional screening data and can be taken as an indication of model convergence. Generalist GPR models are not considered in this convergence assessment, because the immunomodulatory profiles of generalists are defined as linear combinations of specialists with different agonists in a particular group. Hence, the convergence of generalist GPR models is closely correlated with the convergence of specialist GPR models, and the inventors choose to only assess the convergence of specialist GPR models. Using these criteria, the inventors terminated the active learning screen after four rounds, during which the inventors experimentally assayed a total of 2,880 compounds comprising ~2% of the 139,998-candidate molecular candidate space.
Inference of chemical design rules
[0143] At the conclusion of the active learning screen, the inventors attempted to extract human interpretable design rules relating simple chemical properties of the immunomodulator candidates to their measured performance. To do so, the inventors constructed simple “glass box” linear regression models linking the occurrence of particular structural fragments to the measured fold-change in the immunomodulatory response using least absolute shrinkage and selection operator (LASSO) regression.66,67 This resulted in sparse linear models with relatively few non-zero linear coefficients that identify those structural fragments that are the principal discriminants of the measured immunomodulatory activity.
[0144] To train these models, the inventors combined the dataset consisting of 2880 molecules tested in this study with the 2674 molecules obtained in a prior study.24 The inventors then eliminated any nonviable and redundant compounds to arrive at a dataset of 3560 distinct and viable molecular structures, along with their associated immunomodulatory profiles. For consistency with the prior analysis, candidates that exist as salts or with other components are represented as a single entity for the purposes of generating molecular descriptors. The inventors used the open-source cheminformatics software RDKit68 to featurize each of the 3560 experimentally assayed immunomodulators as a numerical vector of 85 substructure occurrences. The inventors then eliminated eight irrelevant features that showed no variation across all the immunomodulators and seven redundant features that had a linear correlation >0.95 with any other features. 69 This left us with k = 70 features for each immunomodulator that were compiled into the feature matrix F G R(A,=356°)X(*=70) Each feature row comprising the feature values for each immunomodulator was normalized to unit length.
[0145] Given the 3560x70 feature matrix F , the inventors then constructed LASSO regression models to predict the activity change for each agonist-pathway combination: (A) NF-KB-LPS, (B) NF-KB-MPLA, (C) NF-zcB-CpG, (D) NF-zcB-Generalist, (E) IRF-LPS, (F) IRF-MPLA, (G) IRF-cGAMP and (H) IRF-Generalist (Figure IE). Each LASSO regression model corresponding to an immunological objective is trained to predict the log2-fold change in immunomodulatory activity for an immunomodulator using the corresponding normalized feature vector. This training involves minimizing the LI regularized loss, where the LI penalization prevents overfitting by retaining only a small number of generalizable features present in the training dataset. The optimal number of features to use in the model is determined by 5-fold cross-validation on the LI regularization weight. The inventors identify the regularization weight values that result in the lowest generalization error, as well as the number of non-zero coefficients and mean absolute error (MAE) for predicting immunomodulation in log2-fold change corresponding to that optimal regularization weight. By examining the coefficients with the largest magnitudes in this optimal linear model, the inventors can rank the molecular descriptors based on their immunomodulatory effect. Example 4: Experimental Results of Aspects Herein
Active learning identifies novel candidate immunomodulators
[0146] The inventors conducted four rounds of active learning-guided experimental screening of a library of 139,998 putative immunomodulators. In each round, the inventors trained a GPR surrogate model over the experimental screening data collected to date as a surrogate predictor of immunomodulatory activity along 12 functional goals over eight agonistpathway combinations. The inventors then conducted BO to select a total of 720 candidates predicted to strongly enhance the NF-KB response, inhibit the NF-KB response, or enhance the IRF response in the presence of one particular agonist (i.e., a specialist) or in the presence of any one of a group of agonists (i.e., a generalist). The inventors terminated the screen after four rounds, corresponding to a screening of 2880 immunomodulator candidates comprising ~2% of the molecular candidate search space.
[0147] In Figure 2A the inventors quantify the extent of immunomodulation for each of the 12 functional goals by reporting the fold change in immune activation induced by the combination of PRR agonists and immunomodulators relative to that induced by agonists alone for the molecules considered in each round of the active learning screen. Importantly, the inventors validate that immunomodulators alone do not stimulate an immune activation in the absence of agonists (FIG. 12). The goal of the active learning screen was to perform on-the-fly learning of a QSAR model to guide the optimal selection of the most promising immunomodulator candidates and achieve round-on-round improvements in the identification of top performers. The inventors observe the preponderance of molecules are clustered around a fold-change of unity, meaning that they have a very limited effect on immune activation, and it is the rarer molecules in the tails of the distributions that are of primary interest and to which active learning directs the screen. Looking at the most potent immunomodulator in different function goals (i.e., the maximum or minimum of the distribution in fold change illustrated as orange and purple bands in Figure 2A), the inventors observe clear round-on-round improvements in 11/12 functional goals, indicating that the screen is resolving novel high- performing candidates. Only the LPS specialist to enhance the IRF response shows no significant improvement after the first round, perhaps indicative of the relative paucity of immunomodulators for these goals within the candidate space. Two functional objectives - MPLA specialists to enhance the NF-KB response and MPLA specialists to enhance the IRF response - show a continuing upward trend after four rounds of screening, but all other functional goals appear to have reached a plateau. The distribution denoted as Round 0 represents the compounds that were experimentally tested prior to the active learning-assisted screen obtained in previous work,24 which was also the labeled training data used to train the initial active learning models. Round 0 compound libraries featured compounds known to be relevant to immune signaling pathways, while the compound libraries the inventors used in the active learning discovery search in Rounds 1-4 are more generic chemical screening libraries. Nevertheless, the inventors discovered molecules with improved immunomodulatory capacity as compared to those in the Round 0 library in five out of 12 functional goals, namely NF-KB enhancers (MPLA specialist), NF-KB suppressors (LPS specialist), NF-KB suppressors (CpG specialist), NF-KB Generalist, and IRF enhancers (cGAMP specialist).
[0148] In addition to discovering those top-performing immunomodulators, the inventors also seek to expand the number of immunomodulators to the tails of the distributions to identify multiple novel high performing immunomodulators. In Figure 2B, in addition to the topperforming immunomodulators, the inventors show the log2-fold change of the 5th and the 20th strongest enhancer and suppressor for each functional goal with respect to each round. Similarly, the inventors observe round-on-round improvements in 11/12 functional goals for the 5th and the 20th strongest immunomodulator curve, again with the LPS specialist to enhance the IRF response being the only exception. This indicates that the screen is exposing high-profile immunomodulators to enrich the tails.
[0149] To quantify convergence of the GPR surrogate models, the inventors employed the stabilizing predictions method by computing the average Bhattacharyya distance DB between GPR posteriors in successive rounds over a randomly selected stop set of 100,000 points and employed the performance difference method by computing the 5-fold cross-validated mean average error (MAE) over the accumulated labeled data collected to date as a function of screening round. 1 63 63 As illustrated in Figure 2C, all specialist GPR models exhibited convergence in DB by Round 4. Figure 2C indicates that the models have not yet fully converged with respect to the MAE, indicating that the predictive capacity could be further improved by additional rounds of screening. However, given the plateau trends in the active learning screen (Figure 2B), the convergence of the average Bhattacharyya distance, and the expensive nature of an additional round of experimental screening, the inventors elected to terminate the search after four rounds under the rationale that the marginal returns of additional screening rounds are likely to be small and that a large number of high-performing candidates in all 12 functional objectives have been discovered within the first four rounds. Nevertheless, it is possible that more performant molecules could be identified by additional rounds of screening. [0150] The inventors present in Figure 3 projections of the candidate molecules into a 2D t-distributed Stochastic Neighbor Embedding (t-SNE) compression of the 100D VAE latent space. The inventors observe that the distribution of sampled points in the learned latent space embedding is consistent with the GPR/BO driving broad exploration of the molecular candidate space and has not become overly focused or stuck in any one particular region over the course of the active learning screen (Figure 3A-C). Immunomodulators with potent enhancement or suppression performance are distributed broadly within the 2D t-SNE embeddings (Figure 3D), although it appears that there may be some localization of high-performance suppressors in the top-left corner of the space.
Analysis of the top performing candidates identified in the active learning screen
[0151] The inventors then proceed to conduct a deeper analysis of the top-performing candidates identified by the active learning screen. First, the inventors filtered 2880 experimentally assayed candidates for cytostatic or cytotoxic behavior using the confluency mask scores, as described herein
[0152] This resulted in the removal of 303/2880 (10.5%) compounds from the screen, leaving a majority 2577/2880 (89.5%) as viable candidates. All following analyses will be conducted over viable candidates only, unless otherwise stated. The viability filter used to determine the viability of immunomodulator candidates for this work as described herein is more stringent than the previous work, 24 and the inventors will only compare the viable candidates in both works under these criteria.
[0153] The top-performing enhancers of the NF-KB response achieved fold improvements relative to agonist alone of 2.6-fold (LPS specialist), 5.5-fold (MPLA specialist), 2.8-fold (CpG specialist) and 2.9-fold (LPS, MPLA, CpG generalist). The top-performing suppressors of the NF-KB response achieved fold improvements relative to agonist alone of 0.1 -fold (LPS specialist), 0.23-fold (MPLA specialist), 0.06-fold (CpG specialist), and 0.15-fold (LPS, MPLA, CpG generalist). The top-performing enhancers of the IRF response achieved fold improvements relative to agonist alone of 5.9-fold (LPS specialist), 6.0-fold (MPLA specialist), 3.2-fold (cGAMP specialist), and 3.6-fold (LPS, MPLA, cGAMP generalist). Compared to a previous screen,24 the inventors discovered specialist and generalist molecules with unprecedented immunomodulatory capacity: NF-KB enhancers (MPLA specialist) of 5.5-fold as compared to previously 2.6-fold, NF-KB suppressors (LPS specialist) of 0.1-fold as compared to previously 0.23-fold, NF-KB suppressors (CpG specialist) of 0.06-fold as compared to previously 0.15-fold, NF-KB suppressors (LPS, MPLA, CpG generalist) of 0.15- fold as compared to previously 0.17-fold, and IRF enhancers (cGAMP specialist) of 3.2-fold as compared to previously 1.74-fold. This shows that the top-performing candidates can have better suppression or enhancement over specific immune responses in comparison to the bestperforming viable molecules identified in previous screening efforts over a more modest and (putatively) more immunologically-relevant 2674-compound library.24
[0154] In addition, the screen identified a number of previously unknown strongly enhancing or suppressing immunomodulators. The inventors report in FIG. 23 the number of immunomodulators with 1.5x, 2x, 5x, and lOx enhancement or suppression identified in the present work compared to the prior screen.24 In most functional goals, the inventors observed a substantial increase in the number of known candidates, and in five instances identified candidates with activity levels that were not achieved in the prior screen. Overall, the inventors identified 554, 167, 36, and 9 novel immunomodulators capable of mediating 1.5x, 2x, 5x, and lOx enhancement or suppression of at least one of the 12 objective functions, relative to the 382, 159, 23, and 0 identified in previous studies. In particular, the nine immunomodulators observed to downregulate NF-KB stimulation by more than 10-fold (i.e., fold change lower than 0.1) representsan unprecedented level of inhibition.24
[0155] The inventors present in Figure 4A the top two molecules for each of the 12 functional objectives identified by the screen, where the inventors show the chemical structures and experimentally measured immunomodulatory profiles for each molecule. PME-4119 was identified as a top-performing NF-KB suppressor generalist, as well as a top-performing MPLA NF-KB suppressor specialist and a CpG NF-KB suppressor specialist, showing that some potent generalists can also function as potent specialists. PME-5246 does not strongly enhance NF- KB stimulation of any particular agonist, but it enhances NF-KB stimulation with every agonist, meaning that it is a good generalist. However, potent IRF enhancer specialists appear to be poorer generalists due to their stronger specificities.
[0156] The inventors next traced the source of each top-performing molecule with a >2- fold enhancement/inhibition to identify that a significant number of these molecules come from the Microsource Spectrum Collection, Prestwick Chemical Library and Selleckchem FDA- approved Drug Library (Table S5). The common feature these three libraries share is that they all have a large portion of molecules that have been approved by some regulation institute, such as U.S. Food and Drug Administration (FDA), to be used as drugs or therapeutics. This shows that there is abundant repurposing potential for approved drugs to be used as immunomodulators . [0157] The inventors also computed the Tanimoto molecular similarity between each high- profile molecule with a >2-fold enhancement/suppression to identify the most similar molecule within the 2674 molecule data from a previous screen that was used to train the initial GPR model.24 The Tanimoto similarity metric quantifies the proportion of chemical substructures shared in a pair of molecules as a value between 0 and 1 and was computed between 2048-bit ECFP4 molecular fingerprints using RDKit.68 The higher the Tanimoto similarity, the more substructures are shared between the molecules. A histogram of the Tanimoto similarity scores between the top performers and the most similar initial training molecule demonstrates significant support at low similarity values indicating that the active learning search has moved into new regions of space and is not simply sampling in the close vicinity of the training data (FIG. 15). Furthermore, a comparison of the top-performers to their closest analog in the 2674 molecules from the prior screen constituting the labeled training data illustrates that the active learning process is not simply performing a local search in the vicinity of the previously identified top performers, but rather learning over the iterative design rounds and venturing into new regions of chemical space (FIG. 16).
[0158] Taken together, these results demonstrate the value of the active learning screen over large candidate libraries in efficiently identifying large numbers of novel small molecule candidates with high immunomodulatory activity.
Data-driven inference of chemical design rules
[0159] The active learning screen furnished immunomodulation measurements for 2880 new candidate molecules. Combining these with the 2674 compounds screened in previous study, 24 the inventors possess a rich data set of labeled immunomodulatory activity for 3560 compounds after removing non- viable and duplicated compounds. The inventors then sought to interrogate these data to extract interpretable design rules for the immunomodulatory activity based on the molecular structure. It can be challenging to extract interpretable understanding of structure-function relationship learned by the GPR surrogate model. To furnish more comprehensible structure-function relations, the inventors employed LASSO regression to train an interpretable linear model regressing the log2-fold change in immunomodulatory activity conditioned upon the presence or absence of particular chemical fragments or functional groups. In doing so, the inventors exchange nonlinearity, complexity, and accuracy of the GPR for interpretability in the LASSO model predictions. The LASSO regularization term promotes sparse regression models in which many of the learned regression coefficients shrink to precisely zero (0k = 0), and the remaining non-zero coefficients can be interpreted as pertaining to those chemical features that are the strongest determinants of observed immunomodulatory behaviors. The simple structure of the model means that the sign of the learned non-zero weights indicates the direction of immune response regulation. Specifically, chemical groups with a value of Ok > 0 are associated with enhancing modulation and can be regarded as enhancer promoters. Conversely, chemical groups with a value of Ok < 0 are associated with suppressing modulation and can be regarded as suppressor promoters. The inventors show the performance of LASSO regression models in FIG. 11.
[0160] In FIG. 5 the inventors present in non-ascending order of magnitude, the up to six non-zero regression coefficients for LASSO models fitted to each of the eight agonist-pathway combinations of interest. These weights can be interpreted as being associated with the features that have the highest predictive power for the log2-fold change in immunomodulatory activity in each of the eight agonist-pathway combinations. Importantly, the features reflect the number of occurrences of particular chemical substructures within each candidate molecule and so can lead to actionable design rules on how to modulate immunological behavior by enriching or depleting a molecule with particular chemical groups. The chemical fragments pertaining to each regression coefficient are denoted by codes starting with “fr_” and followed by letters denoting the chemical groups they are quantifying.
[0161] The analysis reveals a number of interesting design rules. First, the halogen moiety “fr_halogen” appears as a negative, top-ranked fragment in all four of the NF-KB specialist and generalist LASSO models. In particular, in the LPS and MPLA specialist and NF/cB generalist categories it ranked among the top two. This indicates that the presence of halogen groups in immunomodulators is predictive of suppression of the activity of this pathway, especially immune responses activated by TLR4 agonists such as LPS and MPLA. This finding is aligned with several top performing candidates, some of which are shown in Figure 4A: a topperforming NF-KB suppressor generalist and LPS specialist, PME-4426, has a fluoride group and PME-3873 and PME-4392, which are both NF-KB suppressors, have chloride groups. Second, aromatic heteroatom moieties, including aromatic nitrogen “fr_Ar_N”, aromatic amine “fr_Ar_NH”, and aromatic hydroxyl group “fr_Ar_OH” appear frequently in multiple LASSO models with at least one of them retained in seven out of eight LASSO models, with the MPLA NF-KB specialist being the only exception. Interestingly, the weights of these aromatic heteroatom moieties are positive for all NF-KB LASSO models and negative for all IRF LASSO models. This result indicates that the presence of aromatic nitrogen/amine/hydroxyl group is predictive of enhancement of the immune activation in NFKB pathway, while it is also predictive of suppression of the activity of IRF pathway. Since the suppression of IRF pathway is of limited clinical significance, the information suggests elimination of these aromatic heteroatom moieties in IRF enhancer immunomodulators to optimize their potency. For example, PME-5246 and PME-4800, two NF-KB suppressors, both possess aromatic hydroxyl groups, and PME-4974, which possesses an aromatic amine group, enhance NF-KB responses while suppressing IRF responses in general. Third, the carbonyl moiety “fr_C_O_noCOO” appears as the third most influential fragment in the LPS NF-KB specialist LASSO model. The sum of carbonyl and carboxyl moiety “fr_C_O” appears as a top-ranked fragments in the MPLA NF-KB specialist and CpG NF-KB specialist models with negative and positive weights, respectively. (Chemical fragments not ranked among the top six illustrated in Figure 5 are presented in FIG. 13). This indicates that the presence of carbonyl and carboxyl moieties appears to be predictive of suppression of the immune activity in NF- KB pathway stimulated by TLR4 agonists such as LPS and MPLA, while it is predictive of enhancement of immune responses for CpG NF-KB specialists. As a representative example, PME-4873, having three carbonyl groups in the structure, can greatly enhance NF-KB response with CpG while it is a weak suppressor for NF-KB response with LPS.
[0162] Overall, this analysis provides insights into understanding the characteristics of a molecule that promote specific immunomodulatory behaviors. These design rules are of value in advancing understanding of the possible modes of action of these molecules, suggesting how one might modify the structure of a particular immunomodulator to boost its performance in a particular immunomodulation goal, and in guiding how one might augment future candidate libraries to enrich them in molecules predicted to have promising immunomodulatory behaviors.
Validating top-performing immunomodulators with cytokine release profile measurements
[0163] The high-throughput active learning screen already demonstrated the capacity of these immunomodulators to enhance or inhibit NF-KB and IRF responses. However, these transcriptional activity measurements only provide an overall representation of the immunomodulatory behavior. In order to obtain a more detailed understanding of the immune signaling behaviors, the inventors subjected 17 of the top-performing immunomodulator candidates identified in the active learning to a low-throughput assay measuring cytokine release profiles within primary cells (Figure 4). Specifically, the inventors measured modulator’ s ability to change cytokine secretion of murine bone marrow derived dendritic cells (BMDCs) stimulated with LPS, CpG and cGAMP. For this validation assay, MPLA was not tested as both LPS and MPLA are TLR4 agonists and one TLR4 agonist was determined to be sufficient for preliminary validation. The 17 top candidates were primarily selected from the top-performing immunomodulators in each of the 12 objectives - PME-5071, PME-4855, PME-4671, PME-4633, PME-4873, PME3873, PME-5149, PME-4425, PME-3465, PME- 5246, PME-5839, PME-3808 and PME-5084 plus four molecules that were determined to be non-viable in the confluency mask test for cytostatic or cytotoxic behavior, but exhibit an exceptional immunomodulatory profiles
[0164] PME-3878, PME-3386, PME-5920 and PME-4007 (Figure 4B). Among these four additional candidates, two demonstrated potent inhibition - PME-3878 and PME-3386 - and two are potent enhancers - PME-5920 and PME-4007. The main results of this secondary screen are presented in Figure 6. Full data, including negative controls demonstrating that immunomodulators do not modulate cytokine secretion profiles in the absence of agonist, are presented in FIGs. 13, 14, and 18.
[0165] Immunologically, NF-KB activation is correlated with increases in proinflammatory cytokines such as TNF-a, whereas IRF activation is related to production of IFN-/1.6 3 Based on the results of the previous screen,24 the inventors hypothesized that the enhancement or suppression of transcriptional activity induced by the immunomodulators should be associated with the increase or decrease in the production of relevant cytokines. Thus, the inventors focused on the immunomodulation of the release of TNF-a and IFN-/1.
[0166] The inventors observed from the secondary screening results that six immunomodulators - PME3878, PME-3386, PME-4671, PME-4425, PME-3465, and PME- 4007 - out of the 17 top-ranked candidates demonstrated significant capacity to modulate the secretion of TNF-a and IFN-/1 when co-delivered with LPS, CpG, or cGAMP (Figure 6A). Taken together, this cytokine validation assay demonstrated that the immunomodulators identified by the active learning-guided pipeline upregulate TNF-a production by over 10-fold, downregulate TNF-a production by over 16-fold, and upregulate IFN-/1 production by over 6- fold.
[0167] Three immunomodulators stand out as of particular interest. PME-3878 and PME- 3386 are two candidates discovered in the active learning screen as top-performing generalist inhibitors of NF-KB as well as top-performing specialist inhibitors for NF-KB when treated with LPS. PME-4007 is a candidate that is a top-performing specialist enhancer of the IRF when treated with cGAMP. PME-3878 and PME-3386 inhibit TNF-a, IL-6, and IFN-/1 production for nearly all agonists considered (Figure 6A). Suppressing immunomodulators like these can be used as potential adjuvants for prophylactic vaccines or therapeutics that benefit from minimizing pro-inflammatory cytokines. In contrast, PME-4007 is a moderate to strong enhancer of the TNF-a, IFN-/A IL-la, and/or IL-17A responses in the presence of LPS, CpG, or cGAMP (Figure 6A). cGAMP is a pattern recognition receptor agonist that acts through the STING pathway.15 Immunomodulators that enhance IFN- ? production through the STING pathway are of particular interest in promoting antiviral defense and anti-tumor immunity thorough T cell cross priming. 70
[0168] The inventors further subjected the leading IFN- ? inducing compound, PME-4007, to additional comparisons of its cytokine profile in the presence of cGAMP to MSA-2, a recently identified STING agonist.71 MSA-2 was discovered via a high throughput process involving over two million compounds and is more potent than cGAMP. As illustrated in Figure 6B, the inventors observed MSA-2 to induce an IFN- ? secretion that is significantly higher than that induced by cGAMP at the same concentration (10 /rg/mL). Furthermore, when PME-4007 was added in a low concentration (2 /rM) in combination with cGAMP, it increased IFN- ? levels by approximately 3-fold relative to cGAMP treated cells and IFN-// was statistically significantly elevated relative MSA-2 treated cells. PME-4007 also enhanced TNF- a secretion stimulated by cGAMP to a level that is comparable to that stimulated by MSA-2. The in vitro screen identified an immunomodulator that can be combined with a commonly used, naturally occurring STING agonist to induce similar immunological profiles to a best-in- class STING agonist.
Example 5: Discussion and Conclusions From Aspects Herein
[0169] Co-delivery of immunomodulators with PRR agonists presents a powerful means to reduce inflammation or otherwise modulate innate immune stimulation by enhancing or suppressing innate immune signaling pathways, and offers a route to improving vaccines by reducing adverse side-effects and cancer therapies by enhancing the magnitude of the immune response. Small molecules present attractive immunomodulator candidates with high synthetic accessibility and reduced immunogenic potential compared to biologies. The vast size of the drug-like small molecule design space makes strategies to maximize the utility of each experimental assay extremely valuable in rationally and effectively traversing this space. In this work, the inventors combined the inventors constructed a data-driven QSAR model combining deep representational learning, Gaussian process regression, and Bayesian optimization to guide high throughput experimental screening of a library of 139,998 commercially available candidate small molecules. After conducting four rounds of an active learning search that screened 2880 molecules (~2% of the search space) the inventors identified novel immunomodulator candidates capable of suppressing NF-KB activity by up to 15-fold, elevating NF-KB activity by up to 5-fold, and elevating IRF activity by up to 6-fold. The top-performing candidates furnished a 110% improvement in NF-KB activity, 83% improvement in elevating IRF activity, and 128% improvement in suppressing NF-KB activity, and the inventors also identified 167 novel immunomodulators with at least 2-fold enhancement or suppression over transcription factor activity of interest - representing a 105% increase in the total number of known immunomodulators with this level of activity - and nine novel immunomodulators with at least 10-fold activity modulation, while this level of activity modulation is not previously observed with the immunomodulator candidates in the previous work. 24 Additional characterization of the cytokine release profiles of the top 17 candidates demonstrated their ability to substantially modulate key cytokines such as TNF-a, IL-6, and IFN-/A in combination with particular PRR agonists. One particular candidate, PME-4007, was observed to produce a 3-fold increase in IFN-/1 production when co-delivered with the STING agonist cGAMP, which is comparable with the recently developed STING agonist MSA-2 identified via a large two-million-compound high throughput screen. 71 Finally, the inventors performed a post hoc analysis of the high-throughput screening data using interpretable linear models to extract interpretable design rules linking the presence or absence of particular functional groups to immunomodulatory performance. Of particular interest, the inventors found halogen moieties to be correlated with suppression of NF-KB activity, and carbonyl and carboxyl moieties with suppression of NF-KB activity in pathways activated by TLR4 agonists such as LPS and MPLA, but suppression of the IRF pathway.
[0170] More detailed characterization of the top performing small molecule candidates identified in this screen will be performed including in vivo testing and kinetic measurements to unveil their mechanism of action. A screen of larger candidate libraries will also be performed, including those enriched in molecules adhering to the design rules extracted from the analysis. Finally, the inventors also plan to improve the screen to incorporate additional constraints on physical properties of the immunomodulators such as water solubility and synthetic accessibility, to better facilitate their incorporation into vaccine formulations and delivery as vaccines and therapeutics. Example 6: Supplementary Computational Methods
Definition of ~1M compound augmented ZINC library
[0171] The inventors defined an augmentation of the ZINC small molecule library 1 4 to train the VAE network for deep representational embeddings of the immunomodulator candidates. The inventors combined the 2,674 molecules employed in the prior screening work,5 with a subset of 924,870 molecules from the ZINC libraries1 4 comprising those molecules that containing a biphenyl scaffold, inspired by the structure of previous small molecule immunomodulator discovery of Honokiol.6 The inventors also incorporated some other generic molecular libraries from various vendors. As a result, the augmented ZINC library consists of 1,262,866 small molecule compounds in total. A list of the molecular libraries compiled to define the initial augmented ZINC library is provided in Table SI.
[0172] The inventors then filtered the library under a number of criteria. First, the inventors represented each compound as simplified molecular-input line-entry system (SMILES) strings. 7 10 The jnven(Ors then canonicalized these SMILES representations and removed duplicates. Second, the inventors translated each SMILES string into self-referencing embedded strings (SELFIES) as a more robust representation of molecules that are guaranteed to define valid chemical structures. 11 The inventors then eliminated compounds that produced an inconsistent SMILES string upon back-translation from the SELFIES representation to ensure a strict one-to- one mapping between SMILES representations and SELFIES representations of the compounds, thus enforcing a strict one-to-one mapping between SELFIES representations and the structure of the compounds. With the first two steps of filtering, the inventors removed 8734 entries and were left with 1,254,132 compounds in the library. Third, the inventors capped the maximum SELFIES string length to 137 characters and eliminated compounds containing tokens that appear in fewer than 500 compounds. Since the SELFIES strings will ultimately be represented to the VAE input layer as fixed-length vectors, this step constrained the dimension of the representation space so as to limit the size of the network and its associated training costs. These constraints resulted in the ehmination of only 5468 compounds, corresponding to fewer than 0.5% of candidates in the immunomodulator library. After performing these filtering operations, the inventors were left with 1,248,664 compounds in the augmented ZINC library.
[0173] Finally, the inventors digitized the SELFIES representations into one-hot matrices of dimension 137- by-55, where 137 represents the maximum possible string length and 55 are the number of possible SELFIES tokens, including the null tokens. SELFIES strings less than the maximum length are padded with null tokens. The inventors then flattened the one-hot matrices into 7535-element vectors to provide a fixed length representation of each molecule to be passed to the VAE.
Table SI: Compound libraries used to assemble the augmented ZINC library.
Figure imgf000066_0001
Definition of ~140k compound immunomodulator candidate library
[0174] The inventors selected seven generic commercial small molecule compound libraries to define an initial pool of 139,998 candidate immunomodulators to draw candidates from and bring to high throughput screening experimentation. These libraries were a subset of the previously assembled 1 ,248,664 compound augmented ZINC library. These screening compound libraries were designed with the intention of enabling cell-based and target-based high throughput screening initiatives by making a diverse range of small molecules readily available, which make it easier for us to access the molecules for screening experiments. This immunomodulator candidate library went through the same filtering steps as introduced herein, with the only exception that the inventors were not using a new set of SELFIES tokens, but the inventors used the token dictionary the inventors used for the augmented ZINC library. This is to ensure the mapping between the network parameters of VAE and the one-hot encodings of compounds. The inventors transformed the SELFIES representations into one-hot matrices, with a dimension of 137- by- 55 then flattened the one-hot matrices into 7535-element vectors to obtain a standardized representation of each molecule. A list of the molecular libraries compiled to define the initial immunomodulator candidate library is provided in Table S2.
Table S2: Compound libraries used to assemble the immunomodulator candidate library.
Figure imgf000067_0001
Deep representational learning using variational autoencoders (VAEs)
Model architecture
[0175] The inventors employ a VAE architecture inspired by prior work by Aspuru-Guzik and co-workers11 12 and implemented in PyTorch.13 The encoder consists of three fully-connected (FC) layers that passes into a fourth fully-connected bottleneck layer corresponding to the latent space embedding. For the decoder, the inventors employed a stack of gated recurrent units (GRU).14 The architecture of the VAE is illustrated in FIG. 7. The hyperparameters of the model architecture were optimized over the ranges reported in Table S3.
Table S3: Architecture and hyperparameter optimization ranges for the VAE.
Figure imgf000067_0002
Model training over augmented ZINC library
[0176] Tftg y E loss function £VAE comprises two components (Equation SI): (1) the reconstruction loss £Rec measured by the cross-entropy between the input and output one-hot SELFIES vectors to enforce reconstruction fidelity (Equation S2) and (2) the Kullback-Leibler divergence (KLD)15 LKLD of the latent vectors relative to the standard normal distribution to regularize the latent space (Equation S3),12 16
^VAE = ^Rec + ^KLD (S I)
[0177] where,
Figure imgf000068_0001
[0178] where N is the number of samples, D is the dimension of the data, Xij is the j-th component of the z-th input vector, and xtj is the corresponding reconstructed output of the autoencoder, and,
Figure imgf000068_0002
log(a/) - nJ - of , (S3)
[0179] where .j and of are the mean and variance of the j-th element of the latent vector z, respectively, and J is the dimensionality of z.
[0180] Network training is conducted by minimizing Equation SI using the Adam optimizer.17 The inventors first train the VAE over the augmented ZINC library for 10,000 epochs. To guard against posterior collapse, 18 the inventors employ an initial KLD loss coefficient of a = 10“2 that the inventors schedule to reduce to 10“3 at epoch 2000, and to 10“4 at epoch 5500. The training hyperparameters comprising batch size and learning rates were optimized over the ranges reported in Table S4. The inventors assess model performance by computing the exact reconstruction accuracy defined as the fraction of molecules in the training set whose SELFIES strings can be reconstructed with 100% fidelity. Since the intended application of the trained VAE is to define a latent space embedding of the candidate molecules to support an active learning search, the inventors use the decoder in service of learning this latent embedding but, in the present work, do not make use of its generative capacity to extend the model beyond the training data and produce novel synthetic molecules. As such, the inventors assess model performance on the training data rather than the usual practice of employing a hold-out test set. The model achieves an exact reconstruction accuracy of 97.3% at epoch 9032 after more than 1344 GPU-hours of continuous training. This high accuracy indicates that the network has discovered a 100D latent space embedding that preserves the salient information necessary for accurate molecular reconstruction, and the inventors terminate training at this point. The training curves for the model are illustrated in Figure S2.
Table S4: VAE training hyperparameters and optimization ranges.
Figure imgf000068_0003
Figure imgf000069_0003
Gaussian process regression (GPR) surrogate models
[0181] In aspects herein, the inventors seek to identify immunomodulators capable of enhancing or suppressing the immune activity of a certain pathway when the pathway is activated by a certain agonist. This leads to the optimization in maximizing or minimizing the immunomodulation values as fold changes. To guide the search in the chemical design space towards profitable regions, the inventors train a surrogate model to predict the fitness of immunomodulator candidates that have not been experimentally tested. The inventors include an uncertainty measure that reflects the confidence in the predictions, enabling us to select points in the space that balance both high predicted fitness (exploitation) and high uncertainty (exploration). For this supervised regression task, the inventors use Gaussian Process Regression (GPR) as the surrogate model, as it is a non-parametric Bayesian regression technique that provides built-in uncertainty estimates. 19 The inventors have chosen to use the radial basis function (RBF) kernel (also known as the squared exponential kernel) as the covariance function within the GPR implementation in scikit-leam,20
Figure imgf000069_0001
[0182] where xi and xj are input data points, ct 2 is the signal variance, fl is the length scale of the kernel, and I ■ I represents the Euclidean distance between two points. To account for uncertainties associated with each experimental measurement ct n, the inventors adopt Tikhonov regularization21 such that an uncertainty vector G = [si, 02, . . . , on]T is added to the diagonal of the kernel matrix K during fitting. Although in each experiment, the standard deviation for each sample being tested was different, the inventors took the average of the experimental errors of all samples and fed into the GPR models as the
Figure imgf000069_0002
uncertainty vector G = [01, 02, . . . , GH |T as the source of the errors should be the same for all samples in an experiment.
[0183] As the inventors were seeking immunomodulators with various capabilities - different kinds of immunomodulation (enhancement, suppression), different pathway (NF-KB, IRF), and different agonist (LPS, MPLA, CpG, cGAMP) - the inventors pursue a multiobjective optimization. As such, the inventors constructed multiple GPR models, with each GPR corresponding to enhancement/suppression with a specific agonist on a specific pathway. In addition to specialists - immunomodulators that induce large changes in enhancement or suppression when co-delivered with one particular agonist - the inventors also sought to identify generalists - immunomodulators that lead to large effects when co-delivered with any one of a group of agonists. The inventors constructed GPR models for generalists using the arithmetic mean of the immunomodulation fold change values of corresponding specialist objectives. For example, as one of the objectives for enhancers of the NF-KB response, the immunomodulation of generalist over LPS, MPLA and CpG agonists can be expressed as njNFkB Enhancer > 1 znjNFkB Enhancer , n/i NFkB Enhancer , n/i NFkB Enhancer^ lv|generalist (LPS, MPLA, CpG) “ 3 UV,LPS 1V,MPLA 1V,CPG A
(S5)
[0184] where
Figure imgf000070_0001
are immunomodulation fold change values for modulator specialist with agonist LPS, MPLA and CpG, respectively. The inventors initially have 14 objectives, each representing a GPR model:
(1) Enhancers of the NF-KB response (4 objectives): specialist for LPS agonist, specialist for MPLA agonist, specialist for CpG agonist, and generalist over LPS, MPLA, and CpG agonists;
(2) Inhibitors of the NF-KB response (4 objectives): specialist for LPS agonist, specialist for MPLA agonist, specialist for CpG agonist, and generalist over LPS, MPLA, and CpG agonists;
(3) Enhancers of the IRF response (6 objectives): specialist for LPS agonist, specialist for MPLA agonist, specialist for CpG agonist, specialist for cGAMP agonist, generalist over LPS, MPLA, and CpG agonists, and generalist over LPS, MPLA, CpG and cGAMP agonists. The inventors did not seek to identify inhibitors of the IRF response because of limited clinical significance of such immunomodulation. The inventors did not optimize for enhancers or inhibitors of the NF-KB response for specialist for cGAMP agonist because knowledge from prior work that cGAMP does not strongly stimulate the NF-KB pathway.22 Following the screening experiment, the inventors found that the specific CpG agonist the inventors were using (CpG ODN 1826) demonstrated minimal IRF stimulation, as can be seen in FIG. 13. As a result, the inventors only report specialist for CpG agonist as enhancers and inhibitors of the NF-KB response, and the 14 initial objectives were reduced to the 12 objectives as shown in Figure IE.
SI.5 Multi-objective Bayesian optimization (BO)
[0185] After building GPR surrogate models to predict various forms of immunomodulation for every compound, the objective is to explore the design space by iteratively querying new compounds to experimentally test. This involves striking a balance between selecting the most promising candidates with high expected performance (exploitation) and potentially valuable candidates with high uncertainty associated with their prediction (exploration). To achieve this balance, the inventors utilize Bayesian optimization (BO), which is a powerful black-box optimization technique that utilizes the posterior mean and uncertainties derived from GPR predictions to identify the most promising candidates for testing.23 BO can optimize costly black-box functions by optimizing the surrogate objectives that are defined by GPR. An acquisition function is needed to determine the next best candidate for evaluation. To meet the goal of discovering more effective immunomodulators, the inventors have chosen the Expected Improvement (El) acquisition function acquisition function.24 This function guides the selection of candidates that are predicted to result in maximum expected gains in fitness. The El acquisition function can be expressed as,
Figure imgf000071_0001
[0186] where x is the input point, p(x) and o(x) are the mean and standard deviation of the surrogate model prediction at x, f(x+) is the best function value observed so far, is a tradeoff parameter that balances exploration and exploitation, <I> and <j) are the standard normal Ll x) — f(x+ ) — £ cumulative distribution function and probability density function, and Z = - j - is the standard normal random variable. The inventors select the next best candidate point x* with the highest value of the acquisition function from the set of available candidates Xk that have not yet been sampled, i.e., x* = argmax^E^x^; Xk £{xi, x , . . . , x„}). In this work, the inventors employ = 0.01 as a good general purpose choice commonly employed in the literature.24,25 [0187] Traditional multi-objective Bayesian optimization (MOBO) strategies typically aim to balance multiple objectives to efficiently map out Pareto optimal solutions.26, 27 In this work, this would be valuable in identifying potent generalists, but may sacrifice the discovery of potent specialists that reside at the “corners” of the Pareto frontier. Furthermore, some of the objectives are mutually incompatible, such as seeking to find enhancers and inhibitors for the same pathway with the same agonist. Accordingly, the inventors adopted an alternative MOBO strategy in which the inventors polled each of the 12 GPR models independently to collect the best candidates recommended by each model under the El acquisition function and eliminated duplicates. To make efficient use of the experimental assay, the inventors followed a batched sampling Kriging believer approach28,29 wherein the inventors asked each model for its next top-ranked candidates until the inventors had collected a batch of 720 molecules for experimental testing. Under this protocol, the inventors temporarily augment the training data with the molecules selected by the GPRs annotated with the GPR activity predictions, the GPR models are retrained on this augmented data, and the models polled again for their next top- ranked prediction under the El acquisition function. This process is repeated until a batch of 720 molecules has been collated for experimental testing. This batched sampling approach sacrifices efficiency from an information theoretic perspective - each model is asked to choose a series of candidates without first receiving back experimental information on the previously selected candidates - but increases temporal efficiency by testing multiple molecules in an experimental batch. Importantly, all selected molecules, regardless of their source GPR models, were subjected to comprehensive experimental testing to evaluate their immunomodulatory profiles across all agonist types and activation of both NF-KB and IRF pathways. As such, once the experiments are complete, the activity entries in the augmented training data that were completed with the GPR activity predictions in service of the Kriging believer batching are corrected with the measured activities, and these data used to retrain all GPR model. In this manner, molecules selected by one GPR model are not only used to retrain that particular GPR model, but also used to retrain all other models.
VAE-GPR surrogate model ablation test
[0188] The choice of constructing a low-dimensional embedding using VAE and training a GPR surrogate model was motivated, in part, by previous experience in the success of the VAE+GPR+BO paradigm in a number of recent applications in previous studies30 33 and others.36 39 The GPR paradigm is also a natural fit for BO due to its intrinsic uncertainty estimates. There are, however, a number of other surrogate modeling approaches that do not require learning low-dimensional embeddings, but rather can operate directly on molecular I'caturizations,40 42 and that may employ simpler regression models such as linear regression or support vector regression. To test the value of the learned low-dimensional embedding and nonlinear and nonparametric nature of the GPR in elevating the predictive performance, the inventors conducted a simple ablation test in which the inventors evaluated model performance upon replacing the learned VAE featurization with a simple and popular 2048-bit Morgan topological fin- gerprints (FP) ECFP443 computed using RD Kit, 44 and replacing the GPR regression model with a simple linear regression (LR). Considering these two ablations either independently or in combination results in four possible models: (1) VAE-GPR; (2) VAE-LR; (3) FP-GPR; and (4) FP-LR. We use the MAE calculated from a 5-fold cross-validation using each QSAR model over six immunological functional goals to compare model performance. [0189] The performance comparison is presented in FIG. 9. It is immediately apparent that the performance of the FP-LR model is significantly poorer than the other three. Between the remaining three, the VAE-LR tends to perform marginally less well than the VAE-GPR and FP-GPR. These results indicate that the simple featurization combined with the simple regression strategy (FP-LR) is outperformed by both the learned featurization employing the simple regression strategy (VAE-LR) and the simple featurization employing the sophisticated regression strategy (FP-GPR). Interestingly, the performance of the FP-GPR is on par with the VAE-GPR, suggesting that the sophisticated GPR regression model performs equally well under a simple (FP) or learned (VAE) featurization. The VAE featurization, however, possesses two significant advantages over the Morgan fingerprint featurizations. First, the 100D VAE embedding has a relatively modest dimensionality compared to the 2048D Morgan fingerprints. The relatively poor scaling of GPR models with dimensionality, means that this results in a ~3x slowdown in training and deployment of the FP-GPR model relative to the VAE-GPR. Second, the VAE embeddings are invertible in the sense that molecules can be generated from VAE vectors, whereas it is not generally considered possible to straightforwardly invert Morgan fingerprints into molecular structures.43,45 The inventors will also use the model to generate new synthetic molecular candidates with potentially superior performance than those contained in the current screening libraries.
S1.6 Inference of chemical design rules using LASSO regression
[0190] A comprehensive list of descriptors that RDKit is able to compute is available at https:// www.rdkit.org/docs/GettingStartedInPython.html#list-of-available-descriptors, which includes a total of 2082D descriptors. The inventors elected not to include 3D descriptors since 3D molecular structures were not readily available for most of the compounds in the study. The inventors selected those 85 descriptors denoting the occurrence of chemical groups for better interpretability. During feature selection, eight irrelevant descriptors were removed because they are invariant across all 3560 viable compounds in the study. To eliminate redundant descriptors, the inventors looked at the 13 descriptors with Pearson correlation coefficients higher than 0.95 with any other descriptors. The inventors then used clustering analysis to group highly correlated descriptors together into larger families or categories (FIG. 9). In each group of highly correlated descriptors, the inventors selected one descriptor as representative of the group. In this way, seven additional descriptors were marked redundant and removed. In sum, the inventors retained 85-8-7 = 70 effective descriptors that were kept for further analysis. [0191] Having defined the featurization for each immunomodulator, the inventors construct a straightforward and interpretable model to predict the log2-fold change in immunomodulatory activity log2M as a linear function of the k = 70 standardized features,
Figure imgf000074_0001
[0192] where log2 Mexp is the arithmetic mean of the log fold change in
Figure imgf000074_0002
immunomodulatory activity over all N = 3560 experimental measurements and 0 G R7° is a vector regression coefficients assigning weights to the different features. The coefficients 0 are learned by minimizing the LASSO regression loss between the predicted and experimentally measured log2-fold change in immunomodulatory activities,
Figure imgf000074_0003
[0193] The first term on the right side of the equation is the mean squared error between the predicted and experimental immunomodulation values, and the second term is the LI regularization penalty term with hyperparameter «L. The 11 ■ || 2 and 11 ■ 111 represent the L2 and LI norms, respectively. The LI regularization penalty term encourages the model to have fewer nonzero coefficients or parameters, thereby promoting a sparse model that can offer a more concise and interpretable representation of the data. The optimal value for the hyperparameter is chosen using cross-validation, and the resulting nonzero coefficients in 0 can be interpreted as the most critical features for immunomodulation, as shown in FIG. 9. The linear nature of Equation S7 makes it easy to interpret the sign of the learned coefficients: large negative weights 0k < 0 indicate features that are negatively correlated with immunomodulation, while large positive weights 0k > 0 indicate features that are positively correlated with immunomodulation. As shown in FIG. 11, by examining the rank-ordering of the coefficients with the largest absolute values, the inventors can identify structural fragments that are most informative for predicting immunomodulation. There were, in total, 33 chemical fragment descriptors that were retained by at least one of the LASSO models for different immunological objectives. In addition, the trained LASSO model can be used to predict immunomodulation for larger molecules and/or molecules not contained within the training data.
[0194] Table S5: Source library of screened molecules. The table shows (1) the number of com- pounds in each chemical screening library, (2) the number of compounds experimented with HTS in each library, (3) the number of good modulators (with at least 2-fold modulation) found in each library and (4) the ratio of good modulators (the 3rd column divided by the 1st column). The last three libraries, namely Microsource Spectrum Collection, Prestwick Chemical Library and Selleckchem FDA- approved Drug Library, have significantly higher ratio of good modulators identified. The statistics excluded non-viable compounds.
Figure imgf000075_0001
Table S6: List of Agonists studied and their working concentration.
Figure imgf000075_0002
* * *
[0195] All of the methods disclosed and claimed herein can be made and executed without undue experimentation in light of the present disclosure. While the compositions and methods of this invention have been described in terms of preferred aspects, it will be apparent to those of skill in the art that variations may be applied to the methods and in the steps or in the sequence of steps of the method described herein without departing from the concept, spirit and scope of the invention. More specifically, it will be apparent that certain agents which are both chemically and physiologically related may be substituted for the agents described herein while the same or similar results would be achieved. All such similar substitutes and modifications apparent to those skilled in the art are deemed to be within the spirit, scope and concept of the invention as defined by the appended claims.
REFERENCES
The following references, to the extent that they provide exemplary procedural or other details supplementary to those set forth herein, are specifically incorporated herein by reference. References for Background and Examples 1-5:
(1) Harandi, A. M.; Davies, G.; Olesen, O. F. Vaccine adjuvants: scientific challenges and strategic initiatives. Expert Review of Vaccines 2009, 8, 293- 298.
(2) Wu, T. Y.-H.; Singh, M.; Miller, A. T.; De Gregorio, E.; Doro, F.; D’Oro, U.; Skibinski, D. A. G.; Mbow, M. L.; Bufali, S.; Herman, A. E.; Cortez, A.; Li, Y.; Nayak, B. P.;
Tritto, E.; Filippi, C. M.; Otten, G. R.; Brito, L. A.; Monaci, E.; Li, C.; Aprea, S.; Valentini, S.; Calabro, S.; Laera, D.; Brunelli, B.; Caproni, E.; Malyala, P.; Panchai, R. G.; Warren, T. K.; Bavari, S.; O’Hagan, D. T.; Cooke, M. P.; Valiante, N. M. Rational design of small molecules as vaccine adjuvants.
Science Translational Medicine 2014, 6, 263ral60.
(3) Pulendran, B.; S. Arunachalam, P.; O’Hagan, D. T. Emerging concepts in the science of vaccine adjuvants. Nature Reviews Drug Discovery 2021, 20, 454-475.
(4) Corrales, L.; Glickman, L. H.; McWhirter, S. M.; Kanne, D. B.; Sivick, K. E.; Kati- bah, G. E.; Woo, S.-R.; Lemmens, E.; Banda, T.; Leong, J. J.; Metchette, K.; Dubensky, T. W.; Gajewski, T. F. Direct Activation of STING in the Tumor Microenvironment Leads to Potent and Systemic Tumor Regression and Immunity. Cell Reports 2015, 11, 1018-1030.
(5) Circelli, L.; Tornesello, M.; Buonaguro, F. M.; Buonaguro, L. Use of adjuvants for immunotherapy. Human Vaccines & Immunotherapeutics 2017, 13, 1774-1777.
(6) Gilmore, T. D. Introduction to NF-KB: players, pathways, perspectives. Oncogene 2006, 25, 6680-6684.
(7) Liu, T.; Zhang, L.; Joo, D.; Sun, S.-C. NF-KB signaling in inflammation. Signal Transduction and Targeted Therapy 2017, 2, 1-9.
(8) Platanitis, E.; Decker, T. Regulatory networks involving STATs, IRFs, and NFKB in inflammation. Frontiers in Immunology 2018, 9, 2542. (9) Olive, C. Pattern recognition receptors: sentinels in innate immunity and targets of new vaccine adjuvants. Expert Review of Vaccines 2012, 77, 237-
256.
(10) Vasou, A.; Sultanoglu, N.; Goodbourn, S.; Randall, R. E.; Kostrikis, L. G. Targeting Pattern Recognition Receptors (PRR) for Vaccine Adjuvantation: From Synthetic PRR Agonists to the Potential of Defective Interfering Particles of Viruses. Viruses 2017, 9 .
(11) Ong, G. H.; Lian, B. S. X.; Kawasaki, T.; Kawai, T. Exploration of Pattern Recognition Receptor Agonists as Candidate Adjuvants. Frontiers in Cellular and Infection Microbiology 2021, 77 .
(12) Vargas-Caraveo, A.; Sayd, A.; Robledo-Montana, J.; Caso, J. R.; Madrigal, J. L. M.; Garcia-Bueno, B.; Leza, J. C. Toll-like receptor 4 agonist and antagonist lipopolysaccharides modify innate immune response in rat brain circumventricular organs. Journal of Neuroinflammation 2020, 17. 6.
(13) Mata-Haro, V.; Cekic, C.; Martin, M.; Chilton, P. M.; Casella, C. R.; Mitchell, T. C. The vaccine adjuvant monophosphoryl lipid A as a TRIF- biased agonist of TLR4. Science 2007, 316, 1628-1632.
(14) Ballas, Z. K.; Krieg, A. M.; Warren, T.; Rasmussen, W.; Davis, H. L.; Waldschmidt, M.; Weiner, G. J. Divergent therapeutic and immunologic effects of oligodeoxynucleotides with distinct CpG motifs. Journal of Immunology (Baltimore, Md.: 1950) 2001, 167, 4878-4886.
(15) Zhang, X.; Shi, H.; Wu, J.; Zhang, X.; Sun, L.; Chen, C.; Chen, Z. J. Cyclic GMPAMP containing mixed phosphodiester linkages is an endogenous high-affinity ligand for STING. Molecular Cell 2013, 57, 226-235.
(16) Okahira, S.; Nishikawa, F.; Nishikawa, S.; Akazawa, T.; Seya, T.; Matsumoto, M. Interferon-beta induction through toll-like receptor 3 depends on double- stranded RNA structure. DNA and Cell Biology 2005, 24, 614-623. (17) Kieffer, M. E.; Patel, A. M.; Hollingsworth, S. A.; Seganish, W. M. Small molecule agonists of toll-like receptors 7 and 8: a patent review 2014 - 2020. Expert Opinion on Therapeutic Patents 2020, 30, 825-845.
(18) Vijay-Kumar, M.; Carvalho, F. A.; Aitken, J. D.; Fifadara, N. H.; Gewirtz, A. T. TER5 or NERC4 is necessary and sufficient for promotion of humoral immunity by flagellin. European Journal of Immunology 2010, 40, 3528- 3534.
(19) O’Hagan, D. T.; De Gregorio, E. The path to a successful vaccine adjuvant- ‘the long and winding road’. Drug Discovery Today 2009, 14, 541-551.
(20) Moser, B. A.; Steinhardt, R. C.; Escalante-Buendia, Y.; Boltz, D. A.; Barker, K. M.; Cassaidy, B. J.; Rosenberger, M.; Yoo, S.; McGonnigal, B. G.; Esser- Kahn, A. P. Increased vaccine tolerability and protection via NF-KB modulation. Science Advances 2020, 6, eaaz8700.
(21) Moser, B. A.; Escalante-Buendia, Y.; Steinhardt, R. C.; Rosenberger, M. G.; Cassaidy, B. J.; Naorem, N.; Chon, A. C.; Nguyen, M. H.; Tran, N. T.; Esser- Kahn, A. P. Small Molecule NF-KB Inhibitors as Immune Potentiators for Enhancement of Vaccine Adjuvants. Frontiers in Immunology 2020, 11, 511513.
(22) Zienkiewicz, J.; Armitage, A.; Hawiger, J. Targeting nuclear import shuttles, importins/karyopherins alpha by a peptide mimicking the NF/cB I/p50 nuclear localization sequence. Journal of the American Heart Association 2013, 2, e000386.
(23) Lin, Y. Z.; Yao, S. Y.; Veach, R. A.; Torgerson, T. R.; Hawiger, J. Inhibition of nuclear translocation of transcription factor NF-kappa B by a synthetic peptide containing a cell membrane-permeable motif and nuclear localization sequence. The Journal of Biological Chemistry 1995, 270, 14255- 14258.
(24) Kim, J. Y.; Rosenberger, M. G.; Chen, S.; IP, C. K.; Bahmani, A.; Chen, Q.; Shen, J.; Tang, Y.; Wang, A.; Kenna, E.; Son, M.; Tay, S.; Ferguson, A. L.; Esser- Kahn, A. P. Discovery of New States of Immunomodulation for Vaccine Adjuvants via High Throughput Screening: Expanding Innate Responses to PRRs. ACS Central Science 2023, 9, 427-439.
(25) Kirkpatrick, P.; Ellis, C. Chemical space. Nature 2004, 432, 823-823.
(26) Bohacek, R. S.; McMartin, C.; Guida, W. C. The art and practice of structure-based drug design: A molecular modeling perspective. Medicinal Research Reviews 1996, 16, 3-50.
(27) Manojlovic, L. M. Photometry-based estimation of the total number of stars in the Universe. Applied Optics 2015, 54, 6589-6591.
(28) Kwon, S.; Bae, H.; Jo, J.; Yoon, S. Comprehensive ensemble in QSAR prediction for drug discovery. BMC Bioinformatics 2019, 20, 521.
(29) Neves, B. J.; Braga, R. C.; Melo-Filho, C. C.; Moreira-Filho, J. T.; Muratov, E. N.; Andrade, C. H. QSAR-based virtual screening: Advances and applications in drug discovery. Frontiers in Pharmacology 2018, 9, 1275.
(30) Cherkasov, A.; Muratov, E. N.; Fourches, D.; Varnek, A.; Baskin, I. I.; Cronin, M.; Dearden, J.; Gramatica, P.; Martin, Y. C.; Todeschini, R.; Consonni, V.; Kuz’min, V. E.; Cramer, R.; Benigni, R.; Yang, C.; Rathman, J.; Terfloth, L.; Gasteiger, J.; Richard, A.; Tropsha, A. QSAR Modeling: Where Have You Been? Where Are You Going To? Journal of Medicinal Chemistry 2014, 57, 4977-5010.
(31) Shmilovich, K.; Mansbach, R. A.; Sidky, H.; Dunne, O. E.; Panda, S. S.; Tovar, J. D.; Ferguson, A. L. Discovery of Self-Assembling 7r-Conjugated Peptides by Active Learning-Directed Coarse-Grained Molecular Simulation. The Journal of Physical Chemistry B 2020, 124, 3873-3891.
(32) Mohr, B.; Shmilovich, K.; Kleinwachter, I. S.; Schneider, D.; Ferguson, A.
L.; Bereau, T. Data-driven discovery of cardiolipin-selective small molecules by computational active learning. Chemical Science 2022, 13, 4498-4511.
(33) Dave, A.; Mitchell, J.; Kandasamy, K.; Wang, H.; Burke, S.; Paria, B.; Poczos, B.; Whitacre, J.; Viswanathan, V. Autonomous discovery of battery electrolytes with robotic experimentation and machine learning. Cell Reports Physical Science 2020, 1, 100264.
(34) Agarwal, G.; Doan, H. A.; Robertson, L. A.; Zhang, L.; Assary, R. S. Discovery of Energy Storage Molecular Materials Using Quantum Chemistry-Guided Multiobjective Bayesian Optimization. Chemistry of Materials 2021, 33, 8133-8144.
(35) Deshwal, A.; Simon, C. M.; Doppa, J. R. Bayesian optimization of nanoporous materials. Molecular Systems Design & Engineering 2021, 6, 1066-1086.
(36) Xu, J.; Hagler, A. Chemoinformatics and Drug Discovery. Molecules 2002, 7, 566-600.
(37) Gomez-Bombarelli, R.; Wei, J. N.; Duvenaud, D.; Hernandez-Lobato, J. M.; SanchezLengeling, B.; Sheberla, D.; Aguilera-Iparraguirre, J.; Hirzel, T. D.; Adams, R. P.; Aspuru-Guzik, A. Automatic Chemical Design Using a Data- Driven Continuous Representation of Molecules. ACS Central Science 2018, 4, 268-276.
(38) Sanchez-Lengeling, B.; Aspuru-Guzik, A. Inverse molecular design using machine learning: Generative models for matter engineering. Science 2018, 361, 360-365.
(39) Ferguson, A. L.; Ranganathan, R. 100th anniversary of macromolecular science viewpoint: data-driven protein design. ACS Macro Letters 2021, 10, 327-340.
(40) Kingma, D. P.; Welling, M. Auto-Encoding Variational Bayes. 2022.
(41) Rasmussen, C. E.; Williams, C. K. I. In Gaussian Processes for Machine Learning', Bach, F., Ed.; MIT Press: Cambridge, MA, USA, 2005. (42) Mockus, J. On Bayesian Methods for Seeking the Extremum', Springer Berlin
Heidelberg: Berlin, Heidelberg, 1975; pp 400-404.
(43) Honda, K.; Taniguchi, T. IRFs: master regulators of signalling by Toll-like receptors and cytosolic pattern-recognition receptors. Nature Reviews Immunology 2006, 6, 644—658.
(44) Single, A.; Beetham, H.; Telford, B. J.; Guilford, P.; Chen, A. A Comparison of RealTime and Endpoint Cell Viability Assays for Improved Synthetic Lethal Drug Validation. SLAS Discovery 2015, 20, 1286-1293.
(45) Haenel, F.; Garbow, N. Cell counting and confluency analysis as quality controls in cell-based assays. 2014.
(46) Krenn, M.; Hase, F.; Nigam, A.; Friederich, P.; Aspuru-Guzik, A. Selfreferencing embedded strings (SELFIES): A 100% robust molecular string representation. Machine Learning: Science and Technology 2020, 1, 045024.
(47) Capelli, R.; Gardin, A.; Empereur-mot, C.; Doni, G.; Pavan, G. M. A Data- Driven Dimensionality Reduction Approach to Compare and Classify Lipid Force Fields. The Journal of Physical Chemistry B 2021, 725, 7785-7796.
(48) Huo, H.; Rupp, M. Unified representation of molecules and crystals for machine learning. Machine Learning: Science and Technology 2022, 3, 045017.
(49) Himanen, L.; Jager, M. O.; Morooka, E. V.; Canova, F. F.; Ranawat, Y. S.; Gao, D. Z.; Rinke, P.; Foster, A. S. DScribe: Library of descriptors for machine learning in materials science. Computer Physics Communications 2020, 247, 106949.
(50) Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; Desmaison, A.; Kopf, A.; Yang, E.; DeVito, Z.; Raison, M.; Tejani, A.; Chilamkurthy, S.; Steiner, B.; Fang, L.; Bai, J.; Chintala, S. PyTorch: An imperative style, high- performance deep learning library, Curran Associates, Inc., 2019; pp 8024- 8035.
(51) Richards, R. J.; Groener, A. M. Conditional ?-VAE for De Novo Molecular Generation. 2022; https://arxiv.org/abs/2205.01592.
(52) Lim, J.; Ryu, S.; Kim, J. W.; Kim, W. Y. Molecular generative model based on conditional variational autoencoder for de novo molecular design. Journal of Cheminformatics 2018, 10 .
(53) Tevosyan, A.; Khondkaryan, L.; Khachatrian, H.; Tadevosyan, G.; Apresyan, L.; Babayan, N.; Stopper, H.; Navoyan, Z. Improving VAE based molecular representations for compound property prediction. Journal of Cheminformatics 2022, 14 .
(54) Irwin, J. J.; Shoichet, B. K. ZINC-a free database of commercially available compounds for virtual screening. Journal of Chemical Information and Modeling 2005, 45, 177-182.
(55) Irwin, J. J.; Sterling, T.; Mysinger, M. M.; Bolstad, E. S.; Coleman, R. G. ZINC: A Free Tool to Discover Chemistry for Biology. Journal of Chemical Information and Modeling 2012, 52, 1757-1768.
(56) Sterling, T.; Irwin, J. J. ZINC 15 - Ligand Discovery for Everyone. Journal of Chemical Information and Modeling 2015, 55, 2324-2337.
(57) Irwin, J. J.; Tang, K. G.; Young, J.; Dandarchuluun, C.; Wong, B. R.; Khurelbaatar, M.; Moroz, Y. S.; Mayfield, J.; Sayle, R. A. ZINC20-A Free Ultralarge-Scale Chemical Database for Ligand Discovery. Journal of Chemical Information and Modeling 2020, 60, 6065-6073.
(58) Jones, D. R.; Schonlau, M.; Welch, W. J. Efficient Global Optimization of Expensive Black-Box Functions. Journal of Global Optimization 1998, 13, 455-492.
(59) Wang, B.; Singh, H. K.; Ray, T. Bridging kriging believer and expected improvement using bump hunting for expensive black-box optimization. Lille France, 2021; pp 211- 212.
(60) Ginsbourger, D.; Le Riche, R.; Carraro, L. A multi-points criterion for deterministic parallel global optimization based on gaussian processes; 2008.
(61) Gonzalez, J.; Dai, Z.; Hennig, P.; Lawrence, N. D. Batch Bayesian Optimization via Local Penalization. 2015; https://arxiv.org/abs/1505.08052.
(62) Wu, J.; Frazier, P. I. The Parallel Knowledge Gradient Method for Batch Bayesian Optimization. 2016; https://arxiv.org/abs/1606.04414.
(63) Bloodgood, M.; Vijay-Shanker, K. A Method for Stopping Active Learning Based on Stabilizing Predictions and the Need for User- Adjustable Stopping. 2014.
(64) Bhattacharyya, A. On a Measure of Divergence between Two Multinomial Populations. Sankhy~a: The Indian Journal of Statistics (1933-1960) 1946, 7, 401-406.
(65) Beatty, G.; Kochis, E.; Bloodgood, M. The Use of Unlabeled Data Versus Labeled Data for Stopping Active Learning for Text Classification. 2019; pp 287-294.
(66) Santosa, F.; Symes, W. W. Linear Inversion of Band- Limited Reflection Seismograms. SIAM Journal on Scientific and Statistical Computing 1986, 7, 1307-1330.
(67) Tibshirani, R. Regression Shrinkage and Selection via the Lasso. Journal of the Royal Statistical Society. Series B (Methodological) 1996, 58, 267-288.
(68) Landrum, G.; Tosco, P.; Kelley, B.; Ric,; sriniker,; gedeck,; Vianello, R.; NadineSchneider,; Kawashima, E.; Dalke, A.; N, D.; Cole, B.; Swain, M.;
Turk, S.; Cosgrove, D.; AlexanderSavelyev,; Vaucher, A.; Wdjcikowski, M.; Jones, G.; Probst, D.; Scalfani, V. F.; guillaume godin,; Pahl, A.;
Berenger, F.; JLVarjo,; stretsl23,; JP,; DoliathGavid,; Sforna, G.; Jensen, J.
H. rdkit/rdkit: 2021_09_2 (Q3 2021) Release. 2021; https://doi.org/10.5281/zenodo.5589557. (69) Guyon, I.; Elisseeff, A. An introduction to variable and feature selection.
Journal of Machine Learning Research 2003, 3, 1157-1182.
(70) Burdette, D. L.; Monroe, K. M.; Sotelo-Troha, K.; Iwig, J. S.; Eckert, B.;
Hyodo, M.; Hayakawa, Y.; Vance, R. E. STING is a direct innate immune sensor of cyclic di-GMP. Nature 2011, 478, 515-518.
(71) Pan, B.-S.; Perera, S. A.; Piesvaux, J. A.; Presland, J. P.; Schroeder, G. K.; Cumming, J. N.; Trotter, B. W.; Altman, M. D.; Buevich, A. V.; Cash, B.; Cemerski, S.;
Chang, W.; Chen, Y.; Dandliker, P. J.; Feng, G.; Haidle, A.; Henderson, T.; Jewell, J.;
Kariv, I.; Knemeyer, I.; Kopinja, J.; Lacey, B. M.; Laskey, J.; Lesburg, C. A.; Liang, R.;
Long, B. J.; Lu, M.; Ma, Y.; Minnihan, E. C.; O’Donnell, G.; Otte, R.; Price, L.; Rakhilina, L.; Sauvagnat, B.; Sharma, S.; Tyagarajan, S.; Woo, H.; Wyss, D. F.; Xu, S.; Bennett, D. J.; Addona, G. H. An orally available non-nucleotide STING agonist with antitumor activity. Science 2020, 369, eaba6098.
References for Example 6
(1) Irwin, J. J.; Shoichet, B. K. ZINC-a free database of commercially available compounds for virtual screening. Journal of Chemical Information and. Modeling 2005, 45, \T1- 182.
(2) Irwin, J. J.; Sterling, T.; Mysinger, M. M.; Bolstad, E. S.; Coleman, R. G. ZINC: A Free Tool to Discover Chemistry for Biology. Journal of Chemical Information and Modeling 2012, 52, 1757-1768.
(3) Sterling, T.; Irwin, J. J. ZINC 15 - Ligand Discovery for Everyone. Journal of Chemical Information and Modeling 2015, 55, 2324-2337.
(4) Irwin, J. J.; Tang, K. G.; Young, J.; Dandarchuluun, C.; Wong, B. R.; Khurelbaatar, M.; Moroz, Y. S.; Mayfield, J.; Sayle, R. A. ZINC20-A Free Ultralarge-Scale Chemical Database for Ligand Discovery. Journal of Chemical Information and Modeling 2020, 60, 6065-6073.
(5) Kim, J. Y.; Rosenberger, M. G.; Chen, S.; IP, C. K.; Bahmani, A.; Chen, Q.; Shen, J.; Tang, Y.; Wang, A.; Kenna, E.; Son, M.; Tay, S.; Ferguson, A. L.; Esser-Kahn, A. P. Dis- covery of New States of Immunomodulation for Vaccine Adjuvants via High Through- put Screening: Expanding Innate Responses to PRRs. ACS Central Science 2023, 9, 427-439.
(6) Moser, B. A.; Escalante-Buendia, Y.; Steinhardt, R. C.; Rosenberger, M. G.; Cas- saidy, B. J.; Naorem, N.; Chon, A. C.; Nguyen, M. H.; Tran, N. T.; Esser-Kahn, A. P. Small Molecule NF-KB Inhibitors as Immune Potentiators for Enhancement of Vaccine Adjuvants. Frontiers in Immunology 2020, 11, 511513.
(7) Weininger, D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. Journal of Chemical Information and. Computer Sciences 1988, 28, 31-36.
(8) Weininger, D.; Weininger, A.; Weininger, J. L. SMILES. 2. Algorithm for generation of unique SMILES notation. Journal of Chemical Information and Computer Sciences 1989, 29, 97-101.
(9) Weininger, D. SMILES. 3. DEPICT. Graphical depiction of chemical structures. Journal of Chemical Information and Computer Sciences 1990, 30, 237-243.
(10) O’Boyle, N. M. Towards a Universal SMILES representation - A standard method to generate canonical SMILES based on the InChi. Journal of Cheminformatics 2012, 4, 22.
(11) Krenn, M.; Hase, F.; Nigam, A.; Friederich, P.; Aspuru-Guzik, A. Self-referencing embedded strings (SELFIES): A 100% robust molecular string representation. Machine Learning: Science and Technology 2020, 1, 045024.
(12) Gdmez-Bombarelli, R.; Wei, J. N.; Duvenaud, D.; Hernandez-Lobato, J. M.; Sanchez- Lengeling, B.; Sheberla, D.; Aguilera-Iparraguirre, J.; Hirzel, T. D.; Adams, R. P.; Aspuru-Guzik, A. Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules. ACS Central Science 2018, 4, 268-276.
(13) Paszke, A. et al. PyTorch: An imperative style, high-performance deep learning library; Curran Associates, Inc., 2019; pp 8024-8035.
(14) Cho, K.; van Merrienboer, B.; Gulcehre, C.; Bahdanau, D.; Bougares, F.; Schwenk, H.; Bengio, Y. Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation. 2014.
(15) Kullback, S.; Leibler, R. A. On Information and Sufficiency. The Annals of Mathematical Statistics 1951, 22, 79-86.
(16) Kingma, D. P.; Welling, M. Auto-Encoding Variational Bayes. 2022.
(17) Kingma, D. P.; Ba, J. Adam: A Method for Stochastic Optimization. 2017. (18) Lucas, J.; Tucker, G.; Grosse, R.; Norouzi, M. Understanding Posterior Collapse in Generative Latent Variable Models. 2019,
(19) Rasmussen, C. E.; Williams, C. K. I. In Gaussian Processes for Machine Learning', Bach, F., Ed.; MIT Press: Cambridge, MA, USA, 2005.
(20) Pedregosa, F. et al. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 2011, 12, 2825-2830.
(21) Mohammadi, H.; Riche, R. L.; Durrande, N.; Touboul, E.; Bay, X. An analytic comparison of regularization methods for Gaussian Processes. 2017.
(22) Zhang, X.; Shi, H.; Wu, J.; Zhang, X.; Sun, L.; Chen, C.; Chen, Z. J. Cyclic GMP- AMP containing mixed phosphodiester linkages is an endogenous high-affinity ligand for STING. Molecular Cell 2013, 57, 226-235.
(23) Mockus, J. On bayesian methods for seeking the extremum', Springer Berlin Heidelberg: Berlin, Heidelberg, 1975; pp 400-404.
(24) Brochu, E.; Cora, V. M.; de Freitas, N. A Tutorial on Bayesian Optimization of Expensive Cost Functions, with Application to Active User Modeling and Hierarchical Reinforcement Learning. 2010.
(25) Snoek, J.; Larochelle, H.; Adams, R. P. Practical Bayesian Optimization of Machine Learning Algorithms. 2012.
(26) Van Veldhuizen, D. A.; Lamont, G. B. Evolutionary computation and convergence to a pareto front. Proceedings of the Genetic Programming Conference. 1998; pp 221-228.
(27) Irshad, F.; Karsch, S.; Dbpp, A. Expected hypervolume improvement for simultaneous multi-objective and multi-fidelity optimization. 2021.
(28) Wang, B.; Singh, H. K.; Ray, T. Bridging kriging believer and expected improvement using bump hunting for expensive black-box optimization. Lille France, 2021; pp 211- 212.
(29) Ginsbourger, D.; Le Riche, R.; Carraro, L. A multi-points criterion for deterministic parallel global optimization based on gaussian processes', 2008.
(30) Pan, B.-S. et al. An orally available non-nucleotide STING agonist with antitumor activity. Science 2020, 369, eaba6098.
(31) Landrum, G. et al. rdkit/rdkit: 2021_09_2 (Q3 2021) Release. 2021; https ://doi. org/10.5281/zenodo.5589557.
(32) Vargas-Caraveo, A.; Sayd, A.; Robledo-Montana, J.; Caso, J. R.; Madrigal, J. L. M.; Garcia-Bueno, B.; Leza, J. C. Toll-like receptor 4 agonist and antagonist lipopolysaccharides modify innate immune response in rat brain circumventricular organs. Journal of Neuroinflammation 2020, 17, 6.
(33) Mata-Haro, V.; Cekic, C.; Martin, M.; Chilton, P. M.; Casella, C. R.; Mitchell, T. C. The vaccine adjuvant monophosphoryl lipid A as a TRIF-biased agonist of TLR4. Science 2007, 316, 1628-1632.
(34) Ballas, Z. K.; Krieg, A. M.; Warren, T.; Rasmussen, W.; Davis, H. L.; Waldschmidt, M.; Weiner, G. J. Divergent therapeutic and immunologic effects of oligodeoxynucleotides with distinct CpG motifs. Journal of Immunology (Baltimore, Md.: 1950) 2001, 167, 4878-4886.

Claims

WHAT IS CLAIMED IS:
1. A compound of any of the formulas of Figure 4.
2. A compound of formula PME-4855.
3. A compound of formula PME-4426.
4. A compound of formula PME-4119.
5. A compound of formula PME-3974.
6. A compound of formula PME-5149.
7. A compound of formula PME-4637.
8. A compound of formula PME-4800.
9. A compound of formula PME-5839.
10. A compound of formula PME-5084.
11. A compound of formula PME-4974.
12. A compound of formula PME-4873.
13. A compound of formula PME-5246.
14. A compound of formula PME-4695.
15. A compound of formula PME-3465.
16. A compound of formula PME-4633.
17. A compound of formula PME-4392.
18. A compound of formula PME-5071.
19. A compound of formula PME-3878.
20. A compound of formula PME-3386.
21. A compound of formula PME-4671.
22. A compound of formula PME-3873.
23. A compound of formula PME-4425.
24. A compound of formula PME-5920.
25. A compound of formula PME-4007.
26. A compound of formula PME-3808.
27. A pharmaceutical composition comprising one or more of the compounds of any one of claims 1-26.
28. The pharmaceutical composition of claim 27, wherein the pharmaceutical composition further comprises one or more pattern recognition receptor agonists.
29. The pharmaceutical composition of claim 28, wherein the pattern recognition receptor agonists targets one or more of TLR4, TLR9, and STING.
30. The pharmaceutical composition of claim 28 or 29, wherein the pattern recognition receptor agonists is one or more of LPS, MPLA, CpG, and cGAMP.
31. A method of inhibiting NF-kB activity in a cell, the method comprising delivering to the cell an effective amount of one or more of PME-4855, PME-4426, PME-4119, PME-3974, and PME-5149, or any combination thereof.
32. The method of claim 31, wherein the effective amount results in a concentration of about, at least, or at most, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 pM.
33. The method of claim 32, wherein the concentration is a concentration in the cell and/or a concentration in an environment surrounding the cell.
34. The method of any one of claims 31 to 33, further comprising delivering to the cell an effective amount of LPS, MPLA, and/or CpG.
35. The method of any one of claims 31 to 34, further comprising delivering to the cell a pattern recognition agonist.
36. The method of any one of claims 31 to 35, wherein the cell is an immune cell.
37. A method of elevating NF-kB activity in a cell the method comprising delivering to the cell an effective amount of one or more of PME-4637, PME-4800, PME-5839, PME-5084, PME- 4974, PME-4873, PME-5246, PME-4695, PME-3465, PME-4633, PME-4392, and PME- 5071, or any combination thereof.
38. The method of claim 37, wherein the effective amount results in a concentration of about, at least, or at most, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 pM.
39. The method of claim 38, wherein the concentration is a concentration in the cell and/or a concentration in an environment surrounding the cell.
40. The method of any one of claims 37 to 39, further comprising delivering to the cell an effective amount of LPS, MPLA, and/or CpG.
41. The method of any one of claims 37 to 40, further comprising delivering to the cell a pattern recognition agonist.
42. The method of any one of claims 37 to 41, wherein the cell is an immune cell.
43. A method of elevating IFR activity in a cell the method comprising delivering to the cell an effective amount of one or more of PME-4855, PME-4426, PME-4119, PME-3974, PME- 5149, PME-4637, PME-4800, PME-5839, PME-5084, PME-4974, PME-4873, PME-5246, PME-4695, PME-3465, PME-4633, PME-4392, and PME-5071, or any combination thereof.
44. The method of claim 43, wherein the effective amount results in a concentration of about, at least, or at most, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 pM.
45. The method of claim 44, wherein the concentration is a concentration in the cell and/or a concentration in an environment surrounding the cell.
46. The method of any one of claims 43 to 45, further comprising delivering to the cell an effective amount of LPS, MPLA, and/or CpG.
47. The method of any one of claims 43 to 46, further comprising delivering to the cell a pattern recognition agonist.
48. The method of any one of claims 43 to 47, wherein the cell is an immune cell.
49. A method of modulating one or more immune-related pathways in a cell the method comprising delivering to the cell an effective amount of one or more of PME-4855, PME-4426, PME-4119, PME-3974, PME-5149, PME-4637, PME-4800, PME-5839, PME-5084, PME- 4974, PME-4873, PME-5246, PME-4695, PME-3465, PME-4633, PME-4392, and PME- 5071, or any combination thereof.
50. The method of claim 49, wherein the effective amount results in a concentration of approximately, at least, or at most, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 pM.
51. The method of claim 50, wherein the concentration is a concentration in the cell and/or a concentration in an environment surrounding the cell.
52. The method of any one of claims 49 to 51, further comprising delivering to the cell an effective amount of LPS, MPLA, and/or CpG.
53. The method of any one of claims 49 to 52, further comprising delivering to the cell a pattern recognition agonist.
54. The method of any one of claims 49 to 53, wherein the cell is an immune cell.
55. A method of modulating an immune response in a patient, the method comprising administering to the patient at least one of the compounds of any one of claims 1 to 26 and/or the pharmaceutical composition of any one of claims 27 to 30.
56. The method of claim 55, further comprising administering a vaccine to the patient.
57. The method of claim 55 or 56, further comprising administering an additional therapeutic composition to the patient
58. The method of claim 57, wherein pro-inflammatory cytokines act antagonistically to the additional therapeutic composition.
59. The method of claim 55, further comprising administering an anti-viral composition and/or an anti-cancer composition.
60. The method of any one of claims 55 to 59, wherein the patient has, has been diagnosed with, or is suspected of having cancer, an infection, and/or an autoimmune disease.
61. A method comprising administering to a human patient at least one of the compounds of any one of claims 1 to 26, at least one of the pharmaceutical composition of any one of claims 27 to 30, or any combination thereof.
62. A method for training a model to predict an interaction with a compound and a target, the method comprising: generating a fixed-length input feature vector for each compound in a plurality compounds of interest; compiling each fixed-length input feature vectors into a latent space; receiving high-throughput screening data for each compound of the plurality of compounds of interest; associating the fixed-length input feature vectors in the latent space with the high- throughput screening data to generate a set of association information; and training a machine learning model by iteratively minimizing error to within a predetermined threshold using the set of association information.
63. The method of claim 62, further comprising identifying a second plurality of compounds of interest that is refined from the plurality of compounds of interest based on the trained machine learning model.
64. The method of claim 62 or 63, further comprising performing one or more iterations of the method, wherein a subsequent iteration uses the second plurality of compounds of interest as the plurality of compounds of interest.
65. The method of any one of claims 62 to 64, further comprising predicting drug-target interactions for one or more compounds that are not in the plurality of compounds using the trained model.
66. The method of any one of claims 62 to 64, further comprising predicting activity against the target for one or more compounds that are not in the plurality of compounds using the trained model.
67. The method of any one of claims 62 to 64, wherein the high-throughput screening data comprises a signal from a plurality of vessels, wherein each vessel of the plurality of vessels comprises one compound of the compounds of interest and a target.
68. The method of claim 67, wherein the target is a protein of interest.
69. The method of claim 68, wherein the target is in a cell.
70. The method of claim 69, wherein the cell comprises a reporter system capable of producing the signal.
71. The method of any one of claims 67 to 70, wherein the signal is fluorescence produced by the cell.
72. The method of any one of claims 69 to 71, wherein the high-throughput screening data does not comprise data from compounds of interest that are cytostatic or cytotoxic to the cell.
73. The method of any one of claims 62 to 72, wherein the generating a fixed-length input feature vector is performed by a variational autoencoder.
74. The method of claim 73, wherein the variational autoencoder is trained by feedforward network architecture.
75. The method of any one of claims 62 to 74, wherein the compiling is performed via a feed forward architecture to generate a defined-node layer defining the latent space.
76. The method of claim 75, wherein the feed forward architecture is a 500-200-100 fully- connected feed forward architecture.
77. The method of claim 75 or 76, wherein the defined-node layer is a 100-node layer.
78. The method of any one of claims 73 to 77, wherein the variational autoencoder is trained with the plurality of compounds of interest and an additional set of compounds.
79. The method of any one of claims 62 to 78, wherein the associating comprises an empirical mapping of each coordinate of each input vector to at least one biological response measured in the high-throughput screening data.
80. The method of any one of claims 62 to 78, wherein the training the machine learning model comprises training at least one Gaussian procession regression models for each measured biological response in the high-throughput screening data.
81. The method of any one of claims 62 to 80, further comprising decoding the latent space into a plurality of interpretable compound vectors.
82. The method of claim 81, wherein the decoding is performed by a variational autoencoder.
83. The method of claim 81 or 82, wherein the decoding is performed by at least one gated recurrent units.
84. A method for identifying a compound pharmacologically active against a target of interest, the method comprising: generating a fixed-length input feature vector for each compound in a plurality compounds of interest; compiling each fixed-length input feature vectors into a latent space; receiving high-throughput screening data for each compound in the plurality of compounds of interest; associating the high-throughput screening data with the latent space to generate a set of association information; applying, into a trained machine learning model, the latent space to generate an output feature vector predicting the pharmacological activity against a target of interest for each compound in the plurality of compounds of interest; and identifying one or more compounds with pharmacological activity over a determined threshold based on the output feature vector.
85. The method of claim 84, wherein the high-throughput screening data comprises a signal from a plurality of vessels, wherein each vessel of the plurality of vessels comprises one compound of the compounds of interest and the target.
86. The method of claim 85, wherein the target is a protein of interest.
87. The method of claim 86, wherein the target is in a cell.
88. The method of claim 87, wherein the cell comprises a reporter system capable of producing the signal.
89. The method of any one of claims 85 to 88, wherein the signal is fluorescence produced by the cell.
90. The method of any one of claims 87 to 89, wherein the high-throughput screening data does not comprise data from compounds of interest that are cytostatic or cytotoxic to the cell.
91. The method of any one of claims 84 to 90, wherein the generating a fixed-length input feature vector is performed by a variational autoencoder.
92. The method of claim 91, wherein the variational autoencoder is trained by feedforward network architecture.
93. The method of any one of claims 84 to 92, wherein the compiling is performed via a feed forward architecture to generate a defined-node layer defining the latent space.
94. The method of claim 93, wherein the feed forward architecture is a 500-200-100 fully- connected feed forward architecture.
95. The method of claim 93 or 94, wherein the defined-node layer is a 100-node layer.
96. The method of any one of claims 91 to 95, wherein the variational autoencoder is trained with the plurality of compounds of interest and an additional set of compounds.
97. The method of any one of claims 84 to 96, wherein the associating comprises an empirical mapping of each coordinate of each input vector to at least one biological response measured in the high-throughput screening data.
98. The method of any one of claims 84 to 96, further comprising decoding the latent space.
99. The method of claim 98 wherein the decoding is performed by a variational autoencoder.
100. The method of claim 98 or 99, wherein the decoding is performed by at least one gated recurrent units.
101. The method of any one of claims 72 to 100, wherein the identifying comprises identifying a compound that is not in the plurality of compounds.
PCT/US2024/034369 2023-06-16 2024-06-17 Methods and compositions for pharmaceutically relevant interactions Ceased WO2024259429A2 (en)

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
US202363521617P 2023-06-16 2023-06-16
US63/521,617 2023-06-16
US202363510130P 2023-06-25 2023-06-25
US63/510,130 2023-06-25

Publications (2)

Publication Number Publication Date
WO2024259429A2 true WO2024259429A2 (en) 2024-12-19
WO2024259429A3 WO2024259429A3 (en) 2025-04-24

Family

ID=93852688

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2024/034369 Ceased WO2024259429A2 (en) 2023-06-16 2024-06-17 Methods and compositions for pharmaceutically relevant interactions

Country Status (1)

Country Link
WO (1) WO2024259429A2 (en)

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108354938A (en) * 2012-07-20 2018-08-03 硕腾服务有限责任公司 JANUS kinases（JAK）The dosage regimen of inhibitor
US20170354593A1 (en) * 2014-11-03 2017-12-14 Zoetis Services Llc Palatable chewable veterinary composition
NL2022471B1 (en) * 2019-01-29 2020-08-18 Vationpharma B V Solid state forms of oclacitinib
WO2023086806A1 (en) * 2021-11-09 2023-05-19 The University Of Chicago Compositions and methods for inducing immune tolerance

Also Published As

Publication number Publication date
WO2024259429A3 (en) 2025-04-24

Similar Documents

Publication Publication Date Title
Kwofie et al. In silico screening of isocitrate lyase for novel anti-buruli ulcer natural products originating from Africa
Tang et al. Data-driven discovery of innate immunomodulators via machine learning-guided high throughput screening
Mansbach et al. Machine learning algorithm identifies an antibiotic vocabulary for permeating Gram-negative bacteria
Ismail et al. Targeting multiple conformations of SARS-CoV2 Papain-Like Protease for drug repositioning: An in-silico study
Vieira et al. In silico identification of possible inhibitors for protein kinase B (PknB) of mycobacterium tuberculosis
Brueckner et al. MDFit: automated molecular simulations workflow enables high throughput assessment of ligands-protein dynamics
Mitra Drug discovery in tuberculosis: a molecular approach
Acharya et al. Antibiotic charge profile determines the extent of L3 dynamics in OmpF: an expedited passage for molecules with a positive charge
WO2024259429A2 (en) Methods and compositions for pharmaceutically relevant interactions
Tarín-Pelló et al. Novel antimicrobials from computational modelling and drug repositioning: potential in Silico strategies to increase therapeutic arsenal against antimicrobial resistance
Williams et al. Fast prediction of binding affinities of SARS-CoV-2 spike protein and its mutants with antibodies through intermolecular interaction modeling-based machine learning
Edache et al. Computational modeling and analysis of the theoretical structure of thiazolino 2-pyridone amide inhibitors for Yersinia pseudo-tuberculosis and Chlamydia trachomatis Infectivity
Fino et al. Computer-aided design and synthesis of a new class of PEX14 inhibitors: substituted 2, 3, 4, 5-tetrahydrobenzo [F][1, 4] oxazepines as potential new trypanocidal agents
Liu et al. Localized cardiac metabolic trajectories and post-infectious metabolic sequelae in experimental Chagas disease
Nematiasgarabad et al. In silico elucidation of protein-protein interaction network in fish pathogen Flavobacterium Columnare
Llanos et al. A combined ligand and target-based virtual screening strategy to repurpose drugs as putrescine uptake inhibitors with trypanocidal activity
Yao et al. Safe-in-Man broad spectrum antiviral agents
Kumar et al. Computational Prediction of Novel Broad-Spectrum Drug Targets Against Vibrio Cholerae by Integrated Genomics and Proteomics Approach.
Skelton et al. Drug repurposing prediction for COVID-19 using probabilistic networks and crowdsourced curation
Gonçalves et al. Integrative transcriptome analysis of SARS-CoV-2 human-infected cells combined with deep learning algorithms identifies two potential cellular targets for the treatment of coronavirus disease
Oztas Drug Repurposing for COVID-19 Using Molecular Docking Tools
CN118571312B (en) LCK inhibitor virtual screening method, device, equipment and storage medium based on deep learning
Zhao Structural Dynamics of L1 and L2 β-lactamase
Stachowicz-Kuśnierz et al. Nucleoside analog reverse-transcriptase inhibitors in membrane environment: molecular dynamics simulations
Haryini et al. Computational Screening and Molecular Dynamics of Natural Compounds Targeting TEM‐1, CTXM‐15, and CMY‐2 β‐Lactamases Mediated Ceftriaxone Resistance in Salmonella Typhi

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 24824355

Country of ref document: EP

Kind code of ref document: A2

NENP Non-entry into the national phase

Ref country code: DE