WO2024250185A1 - 基于矢量量化的图神经网络的药物性质预测方法和装置 - Google Patents

基于矢量量化的图神经网络的药物性质预测方法和装置 Download PDF

Info

Publication number
WO2024250185A1
WO2024250185A1 PCT/CN2023/098736 CN2023098736W WO2024250185A1 WO 2024250185 A1 WO2024250185 A1 WO 2024250185A1 CN 2023098736 W CN2023098736 W CN 2023098736W WO 2024250185 A1 WO2024250185 A1 WO 2024250185A1
Authority
WO
WIPO (PCT)
Prior art keywords
graph
drug molecule
drug
original
enhanced
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/CN2023/098736
Other languages
English (en)
French (fr)
Inventor
郑鹏飞
吴春琦
陈红阳
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.)
Zhejiang Lab
Original Assignee
Zhejiang Lab
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 Zhejiang Lab filed Critical Zhejiang Lab
Priority to PCT/CN2023/098736 priority Critical patent/WO2024250185A1/zh
Publication of WO2024250185A1 publication Critical patent/WO2024250185A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
    • G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
    • G16C20/50—Molecular design, e.g. of drugs
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
    • G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
    • G16C20/90—Programming languages; Computing architectures; Database systems; Data warehousing
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H70/00—ICT specially adapted for the handling or processing of medical references
    • G16H70/40—ICT specially adapted for the handling or processing of medical references relating to drugs, e.g. their side effects or intended usage

Definitions

  • the present invention mainly relates to the field of drug property prediction, and specifically relates to a drug property prediction method and device based on vector quantization graph neural network.
  • the present invention provides a method and device for predicting drug properties based on a graph neural network with vector quantization.
  • a method for predicting drug properties based on a graph neural network based on vector quantization comprising:
  • the original graph of drug molecules is encoded into latent variable features through a graph encoder; a codebook is constructed; the Euclidean distance between the latent variable features corresponding to each node in the original graph of drug molecules and each vector in the codebook is calculated, and the nearest neighbor vector in the codebook is used to replace the latent variable features.
  • Quantify features to obtain vectorized latent variable features construct a graph decoder, decode the vectorized latent variable features through the graph decoder, and reconstruct the drug molecule enhancement graph;
  • the graph features obtained after graph feature encoding of the drug molecule graph are input into the trained graph comparison network for prediction to obtain the drug property prediction results.
  • a drug property prediction device based on a vector quantization graph neural network
  • the device comprising: a storage module, a processing module, and a prediction module;
  • the storage module includes:
  • Drug molecule original graph database used to store drug molecule original graphs
  • a drug molecule enhancement graph database used to store drug molecule enhancement graphs
  • the processing modules include:
  • a drug molecule original graph acquisition unit used for acquiring a drug molecule original graph from a drug molecule original graph database or through an external interface
  • the drug molecule enhanced graph acquisition unit is used to encode the drug molecule original graph into latent variable features through a graph encoder; construct a codebook; calculate the Euclidean distance between the latent variable features corresponding to each node in the drug molecule original graph and each vector in the codebook, and replace the latent variable features with the nearest neighbor vector in the codebook to obtain vectorized latent variable features; decode the vectorized latent variable features through a graph decoder to reconstruct the drug molecule enhanced graph;
  • a graph instance pair construction unit which constructs a drug molecule original graph-enhanced graph instance pair based on the drug molecule original graph and the drug molecule enhanced graph;
  • a drug molecule graph comparison network unit used to convert drug molecule original graph-enhanced graph instance pairs into drug molecule graph instance pair features
  • a collaborative training unit calculates a first loss function between an original drug molecule graph and an enhanced drug molecule graph, calculates a second loss function corresponding to a drug molecule graph comparison network through a drug molecule graph instance feature, and collaboratively trains the drug molecule enhancement graph and the drug molecule graph comparison network based on the first loss function and the second loss function;
  • the prediction module is used to input the graph features obtained after graph feature encoding of the drug molecule graph into the trained graph comparison network for prediction to obtain the drug property prediction results.
  • an electronic device comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned drug property prediction method based on vector quantization graph neural network.
  • a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the above-mentioned drug property prediction method based on vector quantization graph neural network is implemented.
  • the present invention provides a method and device for predicting drug properties based on a graph neural network with vector quantization.
  • the original drug molecule graph is data enhanced by a graph encoder and a codebook to obtain a drug molecule enhanced graph, thereby avoiding the problem of poor generalization ability caused by the artificially designed graph enhancement method, and introducing a codebook in the process of data enhancement to vectorize the latent variable features of the original drug molecule graph, thereby avoiding the problem of posterior collapse.
  • the graph encoder in the drug molecule graph comparison network and the graph encoder that encodes the original drug molecule graph into latent variable features in the process of constructing the drug molecule graph enhancement network use the same graph encoder, so that the structural information and semantic information in the drug molecule graph can be more fully extracted, and the prediction accuracy of drug properties can be improved.
  • the first loss function between the original drug molecule graph and the enhanced drug molecule graph is calculated, and the second loss function corresponding to the drug molecule graph comparison network is calculated by the drug molecule graph example for the feature, and the sum of the first loss function and the second loss function is used as the overall loss function to perform collaborative training on the drug molecule enhanced graph and the drug molecule graph comparison network, which significantly improves the feature extraction effect of the drug molecule graph data and can better predict the properties of drug molecules.
  • FIG1 is a flow chart of a method for predicting drug properties based on a graph neural network using vector quantization provided by an embodiment of the present invention
  • FIG2 is a schematic diagram of a drug molecule enhancement diagram provided by an embodiment of the present invention.
  • FIG3 is a schematic diagram of a drug molecule graph comparison network provided by an embodiment of the present invention.
  • FIG4 is a schematic diagram of a drug property prediction device based on a vector quantization graph neural network provided by an embodiment of the present invention.
  • FIG. 5 is a schematic diagram of an electronic device provided by an embodiment of the present invention.
  • an embodiment of the present invention provides a drug property prediction method based on a graph neural network with vector quantization.
  • Step S1 obtaining the original image of the drug molecule.
  • the method also includes: storing the original drug molecule graph in a drug molecule original graph database.
  • Step S2 construct a graph encoder based on the graph isomorphism network.
  • the update of the graph isomorphic network is achieved by updating the node features and the edge features, and the expression is as follows:
  • the node feature update method is as follows:
  • Represents the node features of node i at the kth layer in the graph isomorphic network ( represents the initial feature of node i), MLP r represents a multi-layer perceptron, ⁇ represents a discount coefficient, represents the node features of node i in the k-1th layer of the graph isomorphic network, N(i) represents the set of neighbor nodes of node i, Represents the edge feature of the k-1th layer of the edge between node i and node j in the graph isomorphic network ( represents the initial feature of the edge between node i and node j, which can be a null value), K represents the number of graph isomorphism network layers, and in this example, K 3 is set.
  • the edge feature is updated as follows:
  • Sigmoid represents the activation function
  • ⁇ represents the Concat operation
  • Step S3 encode the original graph of drug molecules into latent variable features through a graph encoder; construct a codebook; calculate the Euclidean distance between the latent variable sub-features corresponding to each node in the latent variable features and each vector in the codebook, and replace the latent variable sub-features with the nearest neighbor vectors in the codebook to obtain vectorized latent variable features; construct a graph decoder, decode the vectorized latent variable features through the graph decoder, and reconstruct the enhanced graph of drug molecules.
  • step S3 specifically includes the following sub-steps:
  • V represents the node set (nodes are atoms)
  • E represents the edge set (edges are chemical bonds)
  • H represents the initial features (initial features are atomic features)
  • the original drug molecule graph data is aggregated and compressed to reduce information redundancy, so that the original drug molecule graph is mapped into a low-dimensional space.
  • the expression is as follows:
  • Quantize represents the quantization function
  • ⁇ 2 represents the Euclidean distance
  • Step S304 construct a graph decoder using the fully connected layer and the graph reconstruction layer.
  • Step S305 vectorize latent variable features through a graph decoder Decode based on the vectorized latent variable features The information in predicts the probability of whether there are corresponding edges between each node in the drug molecule enhancement graph G′ and reconstructs the drug molecule enhancement graph G′.
  • the expression is as follows:
  • MLP( ⁇ ) represents a multilayer perceptron
  • z i represents the latent variable sub-feature corresponding to the i-th node in the latent variable feature
  • z j represents the latent variable sub-feature corresponding to the j-th node in the latent variable feature
  • threshold represents the graph reconstruction threshold
  • ⁇ (t) represents the unit step function
  • t p(e i,j )-threshold.
  • Step S4 constructing a drug molecule original graph-enhanced graph instance pair based on the drug molecule original graph and the drug molecule enhanced graph;
  • Step S5 constructing a drug molecule graph comparison network, inputting the drug molecule original graph-enhanced graph instance pair into the drug molecule graph comparison network, and obtaining drug molecule graph instance pair features.
  • the drug molecule graph comparison network includes a graph encoder and an information mapping layer connected thereto, wherein the graph encoder is the graph encoder constructed in step S2 .
  • the drug molecule original graph-enhanced graph instance pair is input into the graph encoder for extraction to obtain the drug molecule original graph-enhanced graph instance pair.
  • Hidden variable features of strong graph-instance pairs Input the latent variable features into the information mapping layer to obtain the drug molecule graph instance pair features
  • N represents the number of nodes in the drug molecule original graph corresponding to the drug molecule original graph-enhanced graph instance pair
  • M represents the number of nodes in the drug molecule enhanced graph corresponding to the drug molecule original graph-enhanced graph instance pair
  • i ⁇ G i represents all nodes i in the drug molecule original graph G i
  • j ⁇ G′ j represents all nodes j in the drug molecule enhanced graph G′ j .
  • the drug molecule graph comparison network and the process of constructing the drug molecule graph enhancement network use the same graph encoder, which can more fully extract the structural and semantic information in the drug molecule graph and improve the prediction accuracy of drug properties.
  • Step S6 construct a first loss function between the original drug molecule graph G and the enhanced drug molecule graph G′, calculate the second loss function corresponding to the drug molecule graph comparison network through the drug molecule graph instance feature, and collaboratively train the drug molecule enhanced graph G′ and the drug molecule graph comparison network based on the first loss function and the second loss function.
  • step S6 includes the following sub-steps:
  • Step S601 constructing a first loss function L(G, G′) between the original graph G of drug molecules and the enhanced graph G′ of drug molecules, which is expressed as follows:
  • sg( ⁇ ) means stopping back propagation
  • ⁇ is a hyperparameter.
  • It is the graph reconstruction loss, which acts on the graph encoder and the graph decoder; Represents the vector quantization loss, which is used to update the codebook C; is the transfer loss used to constrain the output of the graph encoder.
  • Step S602 feature mapping using drug molecule graph examples Calculate the second loss function corresponding to the drug molecule graph comparison network
  • the drug molecule graph instance pair feature obtained from step S5 The similarity within the instance pair is calculated based on this. For the original drug molecule graph G i , the similarity between the positive instance G′ i is maximized, while the similarity between the negative instance and the original drug molecule graph is reduced. The loss function is constructed based on the similarity within the drug molecule graph instance pair.
  • Step S603 collaboratively training the drug molecule enhancement graph G′ and the drug molecule graph comparison network based on the first loss function and the second loss function.
  • the loss function of collaborative training is expressed as follows:
  • Step S7 inputting the graph features obtained after graph feature encoding of the drug molecule graph into the trained graph comparison network for prediction to obtain the drug property prediction result.
  • g represents the information mapping function
  • f represents the graph encoder
  • G represents the drug molecule graph
  • H G represents the drug molecule graph feature
  • Softmax( ⁇ ) represents the activation function
  • W and b represent the trained graph comparison network parameters
  • Pred represents the drug property prediction result.
  • the drug property prediction results include drug biological property prediction results (ADMET, toxicity, side effects, etc.), drug physicochemical property prediction results (water solubility, polar surface area, bioavailability, octanol solubility, metabolic stability, etc.) and drug quantum property prediction results (atomic coordinates, energy and partial charge), etc.
  • ADMET drug biological property prediction results
  • drug physicochemical property prediction results water solubility, polar surface area, bioavailability, octanol solubility, metabolic stability, etc.
  • drug quantum property prediction results atomic coordinates, energy and partial charge
  • an embodiment of the present invention further provides a drug property prediction device based on a vector quantized graph neural network, the device comprising: a storage module, a processing module, and a prediction module;
  • the storage module includes:
  • Drug molecule original graph database used to store drug molecule original graphs
  • a drug molecule enhancement graph database used to store drug molecule enhancement graphs
  • the processing modules include:
  • a drug molecule original graph acquisition unit used for acquiring a drug molecule original graph from a drug molecule original graph database or through an external interface
  • the drug molecule enhanced graph acquisition unit is used to encode the drug molecule original graph into latent variable features through a graph encoder; construct a codebook; calculate the Euclidean distance between the latent variable features corresponding to each node in the drug molecule original graph and each vector in the codebook, and replace the latent variable features with the nearest neighbor vector in the codebook to obtain vectorized latent variable features; decode the vectorized latent variable features through a graph decoder to reconstruct the drug molecule enhanced graph;
  • a graph instance pair construction unit which constructs a drug molecule original graph-enhanced graph instance pair based on the drug molecule original graph and the drug molecule enhanced graph;
  • a drug molecule graph comparison network unit used to convert drug molecule original graph-enhanced graph instance pairs into drug molecule graph instance pair features
  • a collaborative training unit calculates a first loss function between an original drug molecule graph and an enhanced drug molecule graph, calculates a second loss function corresponding to a drug molecule graph comparison network through a drug molecule graph instance feature, and collaboratively trains the drug molecule enhancement graph and the drug molecule graph comparison network based on the first loss function and the second loss function;
  • the prediction module is used to input the graph features obtained after graph feature encoding of the drug molecule graph into the trained graph comparison network for prediction to obtain the drug property prediction results.
  • the device is provided with an input control interface at the input end of the processing module, connected to the input interface for controlling the input; and an output control interface is provided at the output end of the prediction module, connected to the external interface for controlling the output.
  • the relevant parts can refer to the partial description of the method embodiment.
  • the system embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present application. A person of ordinary skill in the art can understand and implement it without creative work.
  • the present application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the drug property prediction method based on vector quantization graph neural network as described above.
  • any device with data processing capability in which the device in the embodiment is located is usually based on the actual function of the device with data processing capability. Other hardware may also be included, which will not be described in detail.
  • the present application also provides a computer-readable storage medium on which computer instructions are stored, and when the instructions are executed by the processor, the drug property prediction method based on the vector quantization graph neural network as described above is implemented.
  • the computer-readable storage medium can be an internal storage unit of any device with data processing capabilities described in any of the aforementioned embodiments, such as a hard disk or a memory.
  • the computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), an SD card, a flash card (Flash Card), etc. equipped on the device.
  • the computer-readable storage medium can also include both an internal storage unit and an external storage device of any device with data processing capabilities.
  • the computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store data that has been output or is to be output.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Chemical & Material Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Crystallography & Structural Chemistry (AREA)
  • Software Systems (AREA)
  • Medicinal Chemistry (AREA)
  • Pharmacology & Pharmacy (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Toxicology (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Medical Informatics (AREA)
  • Artificial Intelligence (AREA)
  • Primary Health Care (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Epidemiology (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Public Health (AREA)
  • Databases & Information Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明公开了一种基于矢量量化的图神经网络的药物性质预测方法，包括：获取药物分子原始图；构建图编码器，将药物分子原始图编码为隐变量特征；构建码本；计算各节点对应的隐变量特征与码本中各个向量的欧式距离，利用最近邻向量替换隐变量特征，得到矢量化的隐变量特征；构建图解码器，对矢量化的隐变量特征进行解码，得到药物分子增强图；构建药物分子原始图-增强图实例对，将其对输入至药物分子图对比网络中，得到药物分子图实例对特征；构建损失函数，对药物分子增强图和药物分子图对比网络进行协同训练；将药物分子图进行图特征编码后得到的图特征输入至训练好的图对比网络进行预测，得到药物性质预测结果。

Description

基于矢量量化的图神经网络的药物性质预测方法和装置 技术领域
本发明主要涉及药物性质预测领域，具体涉及一种基于矢量量化的图神经网络的药物性质预测方法和装置。
背景技术
得益于图神经网络的快速发展及药物分子天然的图结构(以原子为节点，以化学键为边)，以图数据为驱动的药物分子性质预测是加速药物研发的必然趋势。通过药物分子性质预测，可以对候选药物进行筛选，极大的减少药物研发的试错成本，缩短整个药物研发周期。图神经网络通过信息聚合机制和信息传递机制，对药物分子图中的每个节点进行特征更新，并最终聚合成整个药物分子图特征，从而进行药物分子性质预测。然而，当下的图神经网络进行药物分子性质预测时，存在以下的局限性：
(1)药物分子的真实性质数据量少，通常需要湿实验或由丰富药理学经验的专家标注获得，成本太高，这使得传统的有监督图神经网络效果不佳。
(2)现有的无监督图神经网络(如图对比网络)，往往需要引入人为设计的训练策略，这使得无监督图神经网络的泛化能力差，对新的药物分子性质的预测准确率较低。
上述局限性使得图神经网络在药物分子性质预测上效果不佳，是以图数据为驱动的药物分子性质预测的巨大挑战。
因此，亟需提出一种药物性质预测方法来克服药物真实性质数据量少的问题，同时获取更准确的药物分子图表征，提高药物分子性质的预测准确率。
发明内容
针对现有技术不足，本发明提供了一种基于矢量量化的图神经网络的药物性质预测方法和装置。
根据本发明实施例的第一方面，提供了一种基于矢量量化的图神经网络的药物性质预测方法，所述方法包括：
获取药物分子原始图；
构建图编码器；
通过图编码器将药物分子原始图编码为隐变量特征；构建码本；计算药物分子原始图中各节点对应的隐变量特征与码本中各个向量的欧式距离，利用码本中的最近邻向量替换隐变 量特征，得到矢量化的隐变量特征；构建图解码器，通过图解码器对矢量化的隐变量特征进行解码，重构得到药物分子增强图；
基于药物分子原始图和药物分子增强图构建药物分子原始图-增强图实例对；
构建药物分子图对比网络，将药物分子原始图-增强图实例对输入至药物分子图对比网络中，得到药物分子图实例对特征；
计算药物分子原始图和药物分子增强图之间的第一损失函数，通过药物分子图实例对特征计算药物分子图对比网络对应的第二损失函数，基于第一损失函数和第二损失函数对药物分子增强图和药物分子图对比网络进行协同训练；
将药物分子图进行图特征编码后得到的图特征输入至训练好的图对比网络进行预测，得到药物性质预测结果。
根据本发明实施例的第二方面，提供了一种基于矢量量化的图神经网络的药物性质预测装置，所述装置包括：存储模块、处理模块、预测模块；
存储模块包括：
药物分子原始图数据库，用于存储药物分子原始图；
药物分子增强图数据库，用于存储药物分子增强图；
处理模块包括：
药物分子原始图获取单元，用于从药物分子原始图数据库中或通过外部接口获取药物分子原始图；
药物分子增强图获取单元，用于通过图编码器将药物分子原始图编码为隐变量特征；构建码本；计算药物分子原始图中各节点对应的隐变量特征与码本中各个向量的欧式距离，利用码本中的最近邻向量替换隐变量特征，得到矢量化的隐变量特征；通过图解码器对矢量化的隐变量特征进行解码，重构得到药物分子增强图；
图实例对构建单元，基于药物分子原始图和药物分子增强图构建药物分子原始图-增强图实例对；
药物分子图对比网络单元，用于将药物分子原始图-增强图实例对转换为药物分子图实例对特征；
协同训练单元，计算药物分子原始图和药物分子增强图之间的第一损失函数，通过药物分子图实例对特征计算药物分子图对比网络对应的第二损失函数，基于第一损失函数和第二损失函数对药物分子增强图和药物分子图对比网络进行协同训练；
预测模块，用于将药物分子图进行图特征编码后得到的图特征输入至训练好的图对比网络进行预测，得到药物性质预测结果。
根据本发明实施例的第三方面，提供了一种电子设备，包括存储器和处理器，所述存储器与所述处理器耦接；其中，所述存储器用于存储程序数据，所述处理器用于执行所述程序数据以实现上述的基于矢量量化的图神经网络的药物性质预测方法。
根据本发明实施例的第四方面，提供了一种计算机可读存储介质，其上存储有计算机程序，所述程序被处理器执行时实现上述的基于矢量量化的图神经网络的药物性质预测方法。
与现有技术相比，本发明的有益效果是：本发明提供了一种基于矢量量化的图神经网络的药物性质预测方法和装置。通过图编码器和码本对药物分子原始图进行数据增强，得到药物分子增强图，避免了人为设计的图增强方式所造成的泛化能力差的问题，并在数据增强的过程中引入了码本来将药物分子原始图的隐变量特征矢量化，避免了后验坍塌问题。其中，药物分子图对比网络中的图编码器与构建药物分子图增强网络过程中将药物分子原始图编码为隐变量特征的图编码器采用了同一个图编码器，这样可以更充分的提取药物分子图中的结构信息和语义信息，提高药物性质的预测准确率。计算药物分子原始图和药物分子增强图之间的第一损失函数，通过药物分子图实例对特征计算药物分子图对比网络对应的第二损失函数，将第一损失函数与第二损失函数之和作为整体损失函数对药物分子增强图和药物分子图对比网络进行协同训练，显著提高了药物分子图数据的特征提取效果，可以更好地对药物分子进行性质预测。
附图说明
为了更清楚地说明本发明实施例中的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1为本发明实施例提供的基于矢量量化的图神经网络的药物性质预测方法的流程图；
图2为本发明实施例提供的药物分子增强图的示意图；
图3为本发明实施例提供的药物分子图对比网络的示意图；
图4为本发明实施例提供的基于矢量量化的图神经网络的药物性质预测装置的示意图；
图5为本发明实施例提供的一种电子设备的示意图。
具体实施例
下面根据附图详细说明本发明。但是应该理解，此处所描述的具体实施例仅用于解释本发明，并不用于限制本发明的范围。此外，在以下说明中，省略了对公知结构和技术的描述，以避免不必要地混淆本发明的概念。
如图1所示，本发明实施例提供了一种基于矢量量化的图神经网络的药物性质预测方法，
步骤S1，获取药物分子原始图。
进一步地，还包括：将药物分子原始图存储至药物分子原始图数据库中。
步骤S2，基于图同构网络构建图编码器。
进一步地，图同构网络的更新通过节点特征更新和边特征更新实现，表达式如下：
其中，节点特征更新方式如下：
其中，表示节点i在图同构网络中第k层的节点特征(表示节点i的初始特征)，MLPr表示多层感知机，∈表示折扣系数，表示节点i在图同构网络中第k-1层的节点特征，N(i)表示节点i的邻居节点集合，表示节点i和节点j之间的边在图同构网络中第k-1层的边特征(表示节点i和节点j之间的边的初始特征，可为空值)，K表示图同构网络层数，在本实例中，设置K＝3。
边特征更新方式如下：
其中，Sigmoid表示激活函数，"∥"表示Concat操作，和表示图同构网络参数。
步骤S3，通过图编码器将药物分子原始图编码为隐变量特征；构建码本；计算隐变量特征中各节点对应的隐变量子特征与码本中各个向量的欧式距离，利用码本中的最近邻向量替换隐变量子特征，得到矢量化的隐变量特征；构建图解码器，通过图解码器对矢量化的隐变量特征进行解码，重构得到药物分子增强图。
具体地，如图2所示，所述步骤S3具体包括以下子步骤：
步骤S301，通过图编码器将药物分子原始图G＝{V,E,H}编码为隐变量特征Z＝{z1,z2,…,zi,…,zN},Z∈RN×d。
式中，V表示节点集合(节点为原子),E表示边集合(边为化学键),H表示初始特征(初始特征为原子特征),zi表示各节点的隐变量特征，i＝1,2,…N，N表示节点数量，d表示隐变量的特征维度，在本实例中，设置d＝256。
需要说明的是，在对药物分子原始图进行编码的过程中，通过对药物分子原始图数据进行信息聚合、压缩，从而减少信息冗余，使得药物分子原始图映射到低维空间中。
步骤S302，构建码本，记为C＝{c1,c2,…,ck}。
步骤S303，将隐变量特征Z＝{z1,z2,…,zi,…,zN}中各节点对应的隐变量子特征zi与码本＝＝{c1,c2,…,ck}中各个向量的欧式距离，利用码本C中的最近邻向量替换隐变量子特征，得 到矢量化的隐变量特征表达式如下：

其中，Quantize表示量化函数，∥·∥2表示欧氏距离。
步骤S304，由全连接层和图重构层构建图解码器。
步骤S305，通过图解码器对矢量化的隐变量特征进行解码，依据矢量化的隐变量特征中的信息预测药物分子增强图G′中各个节点之间是否存在相应的边的概率重构得到药物分子增强图G′。表达式如下：
全连接层：
p(ei,j)＝MLP(zi)T·MLP(zj)
图重构层：
G′i,j＝ε(p(ei,j)-threshold)
其中，MLP(·)表示多层感知机，zi表示隐变量特征中第i个节点对应的隐变量子特征，zj表示隐变量特征中第j个节点对应的隐变量子特征，threshold表示图重构阈值，ε(t)表示单位阶跃函数，t＝p(ei,j)-threshold。当ε(t)＝1时，节点i和节点j之间存在边；当ε(t)＝0时，节点i和节点j之间不存在边。
步骤S4，基于药物分子原始图和药物分子增强图构建药物分子原始图-增强图实例对；
获取药物分子原始图集合及其对应的药物分子增强图集合
对于药物分子原始图和药物分子增强图当i＝j时，构建成正原始图-增强图实例对；而当i≠j时，则构建成负原始图-增强图实例对。
对每个药物分子图Gi都有相应的原始图-增强图实例对，其相应的集合表示为 当且仅当i＝j时，(Gi,G′j)为图正例对。
步骤S5，构建药物分子图对比网络，将药物分子原始图-增强图实例对输入至药物分子图对比网络中，得到药物分子图实例对特征。
如图3所示，所述药物分子图对比网络包括图编码器和与其连接的信息映射层，其中，图编码器即步骤S2构建的图编码器。
将药物分子原始图-增强图实例对输入至图编码器中进行提取，得到药物分子原始图-增 强图实例对的隐变量特征将隐变量特征输入至信息映射层中得到药物分子图实例对特征
表达式如下：


其中，N表示药物分子原始图-增强图实例对中对应的药物分子原始图中的节点数量，M表示药物分子原始图-增强图实例对中对应的药物分子增强图中的节点数量，i∈Gi表示药物分子原始图Gi中的所有节点i,j∈G′j表示药物分子增强图G′j中的所有节点j。
需要说明的是，药物分子图对比网络与构建药物分子图增强网络过程采用是同一个图编码器，可以更充分的提取药物分子图中的结构信息和语义信息，提高药物性质的预测准确率。
步骤S6，构建药物分子原始图G和药物分子增强图G′之间的第一损失函数，通过药物分子图实例对特征计算药物分子图对比网络对应的第二损失函数，基于第一损失函数和第二损失函数对药物分子增强图G′和药物分子图对比网络进行协同训练。
具体地，所述步骤S6包括以下子步骤：
步骤S601，构建药物分子原始图G和药物分子增强图G′之间的第一损失函数L(G,G′)，表达式如下：
式中，sg(·)表示停止反向传播，β表示超参数。具体来说，为图重构损失，作用在图编码器和图解码器上；表示矢量量化损失，用于更新码本C；为传递损失，用来约束图编码器的输出。
步骤S602，通过药物分子图实例对特征计算药物分子图对比网络对应的第二损失函数
具体地，从步骤S5中得到药物分子图实例对特征并以此计算实例对内的相似性程度。对于药物分子原始图Gi，最大化与正实例G′i之间的相似性程度，同时减少与其他负实例之间相似性，依据药物分子图实例对内的相似性程度构建损失函数
其中，药物分子图实例对之间的相似性通过余弦相似度来衡量，损失函数的计算方式如下：
式中，表示药物分子图正实例对特征之间的余弦相似度，表示药物分子图负实例对特征之间的余弦相似度，τ表示温度控制参数。具体地，可以表示为：
步骤S603，基于第一损失函数和第二损失函数对药物分子增强图G′和药物分子图对比网络进行协同训练。
协同训练的损失函数的表达式如下：
步骤S7，将药物分子图进行图特征编码后得到的图特征输入至训练好的图对比网络进行预测，得到药物性质预测结果。
表达式如下：
HG＝g(f(G))
Pred＝Softmax(HG*W+b)
式中，g表示信息映射函数，f表示图编码器，G表示药物分子图，HG表示药物分子图特征，Softmax(·)表示激活函数，W和b表示训练好的图对比网络参数，Pred表示药物性质预测结果。
进一步地，所述药物性质预测结果包括药物生物性质预测结果(ADMET、毒性、副作用等)、药物物理化学性质预测结果(水溶性、极性表面积、生物利用度、辛醇溶解度、代谢稳定性等)和药物量子性质预测结果(原子坐标、能量和部分电荷)等。
如图4所示，本发明实施例还提供了一种基于矢量量化的图神经网络的药物性质预测装置，所述装置包括：存储模块、处理模块、预测模块；
存储模块包括：
药物分子原始图数据库，用于存储药物分子原始图；
药物分子增强图数据库，用于存储药物分子增强图；
处理模块包括：
药物分子原始图获取单元，用于从药物分子原始图数据库中或通过外部接口获取药物分子原始图；
药物分子增强图获取单元，用于通过图编码器将药物分子原始图编码为隐变量特征；构建码本；计算药物分子原始图中各节点对应的隐变量特征与码本中各个向量的欧式距离，利用码本中的最近邻向量替换隐变量特征，得到矢量化的隐变量特征；通过图解码器对矢量化的隐变量特征进行解码，重构得到药物分子增强图；
图实例对构建单元，基于药物分子原始图和药物分子增强图构建药物分子原始图-增强图实例对；
药物分子图对比网络单元，用于将药物分子原始图-增强图实例对转换为药物分子图实例对特征；
协同训练单元，计算药物分子原始图和药物分子增强图之间的第一损失函数，通过药物分子图实例对特征计算药物分子图对比网络对应的第二损失函数，基于第一损失函数和第二损失函数对药物分子增强图和药物分子图对比网络进行协同训练；
预测模块，用于将药物分子图进行图特征编码后得到的图特征输入至训练好的图对比网络进行预测，得到药物性质预测结果。
特别地，该装置在处理模块的输入端设置有输入控制接口，与输入接口相连，用来控制输入；在预测模块的输出端设置有输出控制接口，与外部接口相连，用来控制输出。
关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。
对于装置实施例而言，由于其基本对应于方法实施例，所以相关之处参见方法实施例的部分说明即可。以上所描述的系统实施例仅仅是示意性的，其中所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部模块来实现本申请方案的目的。本领域普通技术人员在不付出创造性劳动的情况下，即可以理解并实施。
相应的，本申请还提供一种电子设备，包括：一个或多个处理器；存储器，用于存储一个或多个程序；当所述一个或多个程序被所述一个或多个处理器执行，使得所述一个或多个处理器实现如上述的基于矢量量化的图神经网络的药物性质预测方法。如图5所示，为本发明实施例提供的基于矢量量化的图神经网络的药物性质预测方法所在任意具备数据处理能力的设备的一种硬件结构图，除了图5所示的处理器、内存以及网络接口之外，实施例中装置所在的任意具备数据处理能力的设备通常根据该任意具备数据处理能力的设备的实际功能， 还可以包括其他硬件，对此不再赘述。
相应的，本申请还提供一种计算机可读存储介质，其上存储有计算机指令，该指令被处理器执行时实现如上述的基于矢量量化的图神经网络的药物性质预测方法。所述计算机可读存储介质可以是前述任一实施例所述的任意具备数据处理能力的设备的内部存储单元，例如硬盘或内存。所述计算机可读存储介质也可以是外部存储设备，例如所述设备上配备的插接式硬盘、智能存储卡(Smart Media Card，SMC)、SD卡、闪存卡(Flash Card)等。进一步的，所述计算机可读存储介还可以既包括任意具备数据处理能力的设备的内部存储单元也包括外部存储设备。所述计算机可读存储介质用于存储所述计算机程序以及所述任意具备数据处理能力的设备所需的其他程序和数据，还可以用于暂时地存储已经输出或者将要输出的数据。
以上实施例仅用于说明本发明的设计思想和特点，其目的在于使本领域内的技术人员能够了解本发明的内容并据以实施，本发明的保护范围不限于上述实施例。所以，凡依据本发明所揭示的原理、设计思路所作的等同变化或修饰，均在本发明的保护范围之内。

Claims (10)

  1. 一种基于矢量量化的图神经网络的药物性质预测方法，其特征在于，所述方法包括：
    获取药物分子原始图；
    构建图编码器；
    通过图编码器将药物分子原始图编码为隐变量特征；构建码本；计算药物分子原始图中各节点对应的隐变量特征与码本中各个向量的欧式距离，利用码本中的最近邻向量替换隐变量特征，得到矢量化的隐变量特征；构建图解码器，通过图解码器对矢量化的隐变量特征进行解码，重构得到药物分子增强图；
    基于药物分子原始图和药物分子增强图构建药物分子原始图-增强图实例对；
    构建药物分子图对比网络，所述药物分子图对比网络包括图编码器和与其连接的信息映射层，将药物分子原始图-增强图实例对输入至图编码器中转化为对应的隐变量特征，再经信息映射层处理得到药物分子图实例对特征；
    计算药物分子原始图和药物分子增强图之间的第一损失函数，通过药物分子图实例对特征计算药物分子图对比网络对应的第二损失函数，基于第一损失函数和第二损失函数对药物分子增强图和药物分子图对比网络进行协同训练；
    将药物分子图进行图特征编码后得到的图特征输入至训练好的图对比网络进行预测，得到药物性质预测结果。
  2. 根据权利要求1所述的基于矢量量化的图神经网络的药物性质预测方法，其特征在于，构建码本；计算药物分子原始图中各节点对应的隐变量特征与码本中各个向量的欧式距离，利用码本中的最近邻向量替换隐变量特征，得到矢量化的隐变量特征包括：
    记码本为C＝{c1,c2,…,ck}；
    将隐变量特征Z＝{z1,z2,…,zi,…,zN}中各节点对应的隐变量子特征zi与码本C＝{c1,c2,…,ck}中各个向量的欧式距离，利用码本C中的最近邻向量替换隐变量子特征，得到矢量化的隐变量特征表达式如下：

    其中，Quantize表示量化函数，||·||2表示欧氏距离。
  3. 根据权利要求1所述的基于矢量量化的图神经网络的药物性质预测方法，其特征在于，构建图解码器，通过图解码器对矢量化的隐变量特征进行解码，重构得到药物分子增强图包 括：
    由全连接层和图重构层构建图解码器；
    通过图解码器对矢量化的隐变量特征进行解码，依据矢量化的隐变量特征中的信息预测药物分子增强图G′中各个节点之间是否存在相应的边的概率重构得到药物分子增强图G′；表达式如下：
    p(ei,j)＝MLP(zi)T·MLP(zj)
    G′i,j＝ε(p(ei,j)-threshold)
    其中，MLP(·)表示多层感知机，zi表示隐变量特征中第i个节点对应的隐变量子特征，zj表示隐变量特征中第j个节点对应的隐变量子特征，threshold表示图重构阈值，ε(t)表示单位阶跃函数，t＝p(ei,j)-threshold；当ε(t)＝1时，节点i和节点j之间存在边；当ε(t)＝0时，节点i和节点j之间不存在边。
  4. 根据权利要求1所述的基于矢量量化的图神经网络的药物性质预测方法，其特征在于，基于药物分子原始图和药物分子增强图构建药物分子原始图-增强图实例对包括：
    获取药物分子原始图集合及其对应的药物分子增强图集合
    对于药物分子原始图和药物分子增强图当i＝j时，构建成药物分子正原始图-增强图实例对；而当i≠j时，则构建成药物分子负原始图-增强图实例对。
  5. 根据权利要求1所述的基于矢量量化的图神经网络的药物性质预测方法，其特征在于，构建药物分子图对比网络，将药物分子原始图-增强图实例对输入至药物分子图对比网络中，得到药物分子图实例对特征包括：
    将药物分子原始图-增强图实例对输入至图编码器中进行提取，得到药物分子原始图-增强图实例对的隐变量特征将隐变量特征输入至信息映射层中得到药物分子图实例对特征
    表达式如下：


    其中，N表示药物分子原始图-增强图实例对中对应的药物分子原始图中的节点数量，M表示药物分子原始图-增强图实例对中对应的药物分子增强图中的节点数量，i∈Gi表示药物分子原始图Gi中的所有节点i,j∈G′j表示药物分子增强图G′j中的所有节点j。
  6. 根据权利要求1所述的基于矢量量化的图神经网络的药物性质预测方法，其特征在于，构建药物分子原始图和药物分子增强图之间的第一损失函数，通过药物分子图实例对特征计算药物分子图对比网络对应的第二损失函数，基于第一损失函数和第二损失函数对药物分子增强图和药物分子图对比网络进行协同训练包括：
    计算药物分子原始图G和药物分子增强图G′之间的第一损失函数L(G,G′)，表达式如下：
    式中，sg(·)表示停止反向传播，β表示超参数，为图重构损失，表示矢量量化损失，为传递损失；
    通过药物分子图实例对特征之间的相似性计算药物分子图对比网络对应的第二损失函数表达式如下：
    式中，N表示药物分子图的节点数量，表示药物分子图正实例对特征之间的余弦相似度，表示药物分子图负实例对特征之间的余弦相似度，τ表示温度控制参数；
    协同训练损失函数为第一损失函数和第二损失函数之和，基于协同训练损失函数对药物分子增强图和药物分子图对比网络进行协同训练。
  7. 根据权利要求1所述的基于矢量量化的图神经网络的药物性质预测方法，其特征在于，药物性质预测结果包括药物生物性质预测结果、药物物理化学性质预测结果、药物量子性质预测结果。
  8. 一种基于矢量量化的图神经网络的药物性质预测装置，其特征在于，所述装置包括：存储模块、处理模块、预测模块；
    存储模块包括：
    药物分子原始图数据库，用于存储药物分子原始图；
    药物分子增强图数据库，用于存储药物分子增强图；
    处理模块包括：
    药物分子原始图获取单元，用于从药物分子原始图数据库中或通过外部接口获取药物分 子原始图；
    药物分子增强图获取单元，用于通过图编码器将药物分子原始图编码为隐变量特征；构建码本；计算药物分子原始图中各节点对应的隐变量特征与码本中各个向量的欧式距离，利用码本中的最近邻向量替换隐变量特征，得到矢量化的隐变量特征；通过图解码器对矢量化的隐变量特征进行解码，重构得到药物分子增强图；
    图实例对构建单元，基于药物分子原始图和药物分子增强图构建药物分子原始图-增强图实例对；
    药物分子图对比网络单元，用于将药物分子原始图-增强图实例对转换为药物分子图实例对特征；
    协同训练单元，计算药物分子原始图和药物分子增强图之间的第一损失函数，通过药物分子图实例对特征计算药物分子图对比网络对应的第二损失函数，基于第一损失函数和第二损失函数对药物分子增强图和药物分子图对比网络进行协同训练；
    预测模块，用于将药物分子图进行图特征编码后得到的图特征输入至训练好的图对比网络进行预测，得到药物性质预测结果。
  9. 一种电子设备，包括存储器和处理器，其特征在于，所述存储器与所述处理器耦接；其中，所述存储器用于存储程序数据，所述处理器用于执行所述程序数据以实现上述权利要求1-7任一项所述的基于矢量量化的图神经网络的药物性质预测方法。
  10. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，所述程序被处理器执行时实现如权利要求1-7中任一所述的基于矢量量化的图神经网络的药物性质预测方法。
PCT/CN2023/098736 2023-06-07 2023-06-07 基于矢量量化的图神经网络的药物性质预测方法和装置 Ceased WO2024250185A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2023/098736 WO2024250185A1 (zh) 2023-06-07 2023-06-07 基于矢量量化的图神经网络的药物性质预测方法和装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2023/098736 WO2024250185A1 (zh) 2023-06-07 2023-06-07 基于矢量量化的图神经网络的药物性质预测方法和装置

Publications (1)

Publication Number Publication Date
WO2024250185A1 true WO2024250185A1 (zh) 2024-12-12

Family

ID=93794726

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/098736 Ceased WO2024250185A1 (zh) 2023-06-07 2023-06-07 基于矢量量化的图神经网络的药物性质预测方法和装置

Country Status (1)

Country Link
WO (1) WO2024250185A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN120725010A (zh) * 2025-08-21 2025-09-30 中国科学技术大学 一种基于稀疏自编码器的词元构造方法及系统

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110970099A (zh) * 2019-12-10 2020-04-07 北京大学 一种基于正则化变分自动编码器的药物分子生成方法
WO2022040245A1 (en) * 2020-08-18 2022-02-24 Nnaisense, Sa Automatic design of molecules having specific desirable characteristics
CN115274008A (zh) * 2022-08-08 2022-11-01 苏州创腾软件有限公司 基于图神经网络的分子性质预测方法和系统
CN115762657A (zh) * 2022-09-26 2023-03-07 浙江大学 一种基于化学元素知识图谱与官能团提示的分子性质预测方法
WO2023096712A1 (en) * 2021-11-29 2023-06-01 X Development Llc Depolymerization optimization platform

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110970099A (zh) * 2019-12-10 2020-04-07 北京大学 一种基于正则化变分自动编码器的药物分子生成方法
WO2022040245A1 (en) * 2020-08-18 2022-02-24 Nnaisense, Sa Automatic design of molecules having specific desirable characteristics
WO2023096712A1 (en) * 2021-11-29 2023-06-01 X Development Llc Depolymerization optimization platform
CN115274008A (zh) * 2022-08-08 2022-11-01 苏州创腾软件有限公司 基于图神经网络的分子性质预测方法和系统
CN115762657A (zh) * 2022-09-26 2023-03-07 浙江大学 一种基于化学元素知识图谱与官能团提示的分子性质预测方法

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN120725010A (zh) * 2025-08-21 2025-09-30 中国科学技术大学 一种基于稀疏自编码器的词元构造方法及系统

Similar Documents

Publication Publication Date Title
Bose et al. Latent variable modelling with hyperbolic normalizing flows
Ju et al. Unsupervised graph-level representation learning with hierarchical contrasts
CN114610935B (zh) 一种文本控制图像风格的语义图像合成的方法及系统
US11610124B2 (en) Learning compressible features
CN110674323B (zh) 基于虚拟标签回归的无监督跨模态哈希检索方法及系统
CN110347932B (zh) 一种基于深度学习的跨网络用户对齐方法
CN108108854A (zh) 城市路网链路预测方法、系统及存储介质
JP2024541461A (ja) 構造データの生成方法、装置、デバイス、およびプログラム
CN115525771B (zh) 基于上下文数据增强的少样本知识图谱表示学习方法及系统
Shi et al. Mobile edge artificial intelligence: Opportunities and challenges
CN112287166A (zh) 一种基于改进深度信念网络的电影推荐方法及系统
CN115599984B (zh) 一种检索方法
CN118708943B (zh) 特征提取模型的训练方法、数据分类方法、装置及设备
US20230394323A1 (en) Personalized federated learning via heterogeneous modular networks
CN111368176A (zh) 基于监督语义耦合一致的跨模态哈希检索方法及系统
WO2024250185A1 (zh) 基于矢量量化的图神经网络的药物性质预测方法和装置
CN115588471A (zh) 持续学习下自校正单步逆合成方法、终端、服务器及系统
CN116662554B (zh) 基于异构图卷积神经网络的传染病方面级情感分类方法
CN119167994A (zh) 一种基于多层级自编码器的异构图节点表示学习方法
CN114091445B (zh) 实体信息处理方法、装置、计算机设备、存储介质
CN112329483A (zh) 多机制合并注意力的多路径神经机器翻译方法
CN115424660A (zh) 一种使用预测模型预测多源信息关系的方法、装置
CN116705195B (zh) 基于矢量量化的图神经网络的药物性质预测方法和装置
Shi et al. Convolutional neural network knowledge graph link prediction model based on relational memory
CN119578454A (zh) 一种用于生成正负样本节点的模型构建方法、装置和设备

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: 23940075

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE