CN120992425A - A method and system for identifying and classifying electrochemical signals from single vesicle particles. - Google Patents

Abstract

The invention discloses a recognition classification method and a recognition system of electrochemical signals of single vesicle particles, wherein the method comprises the steps of S1, preparing a carbon fiber ultramicro electrode, applying voltage in a patch clamp system constant potential mode, collecting a Pian-level current response generated by oxidation of single vesicle content on the surface of the electrode, S2, carrying out smooth filtering on the collected original current signal, using a multi-parameter peak recognition method based on a threshold value to realize characteristic peak recognition and multi-dimensional physical/statistical characteristic extraction, S3, carrying out pretreatment operations including missing value treatment, standardization, main component analysis dimension reduction and class equalization on the extracted peak characteristic data, and S4, constructing and training at least one machine learning model based on the pretreated characteristic data, wherein the automatic recognition is carried out on the particle size class of vesicle particles. The invention combines the advantages of single vesicle electrochemical high-sensitivity measurement and intelligent algorithm classification, and has the advantages of high recognition efficiency, good repeatability, wide applicability and the like.

Description

Identification and classification method and identification system for electrochemical signals of single vesicle particles

Technical Field

The invention relates to the technical field of electrochemical detection, in particular to a method and a recognition system for acquiring a current signal based on a patch clamp technology and automatically recognizing and quantitatively analyzing vesicle particle size distribution by combining a machine learning algorithm.

Background

The vesicle is widely applied to the fields of biomarker detection, disease diagnosis, drug delivery and the like, and the particle size distribution directly influences the biological functions and the action effects. Traditional particle size analysis methods such as Dynamic Light Scattering (DLS), transmission electron microscopy and the like can effectively characterize particle size, but the methods are generally complex to operate, high in cost and lack in real-time performance and intelligent recognition capability. On the other hand, the electrochemical technologies such as patch clamp have millisecond-scale time resolution and can detect oxidation-reduction reaction.

However, electrochemical signals often suffer from strong noise, unstructured, and large event fluctuations, which makes classification and analysis of the signals difficult by conventional methods. Therefore, how to combine signal processing, statistical feature extraction and machine learning to realize rapid and intelligent identification and detection of vesicle particle size becomes a technical problem to be solved urgently.

Therefore, the application provides a recognition and classification method and a recognition system for electrochemical signals of single vesicle particles, and a set of efficient and reusable recognition system is constructed by combining membrane particle size regulation, electrode microstructure design, current characteristic modeling and machine learning training processes, so that the measurement and recognition level of vesicles and other nano structures can be improved, and the current technical problem is solved.

Disclosure of Invention

The invention mainly aims to provide a recognition and classification method and a recognition system for electrochemical signals of single vesicle particles, which are characterized in that current signal acquisition is carried out on vesicle samples with different particle size distributions obtained by preparing ultra-micro-scale carbon fiber electrodes and combining a patch clamp system with centrifugal purification, then multidimensional characteristic parameters are obtained by utilizing a signal processing and characteristic extraction algorithm, and further, the automatic classification and recognition of the vesicle particle sizes and the quantitative calculation of the number of molecules contained in the vesicle are realized by a Principal Component Analysis (PCA), an oversampling algorithm (SMOTE) and a plurality of machine learning models, so that the technical problems in the background technology are solved.

The invention adopts the following technical scheme to solve the technical problems:

the method for identifying and classifying the vesicle particle electrochemical signals specifically comprises the following steps:

S1, preparing a microscale carbon fiber electrode, connecting the microscale carbon fiber electrode with a patch clamp system, and applying a set voltage to acquire an oxidation current signal caused by a vesicle on the surface of the electrode;

s2, carrying out smooth filtering on the original data of the collected current signals, and realizing characteristic peak identification and multidimensional physical/statistical characteristic extraction by using a multi-parameter peak identification method based on a threshold value;

S3, carrying out preprocessing operations including missing value processing, standardization and Principal Component Analysis (PCA) dimension reduction (retaining 95% variance contribution rate) and class equalization on peak characteristic data in the extracted characteristic data;

S4, constructing and training at least one machine learning model based on the preprocessed characteristic data, and automatically identifying the particle size class of the vesicle particles.

Preferably, the preparation process of the microscale carbon fiber electrode in the step S1 comprises the steps of connecting a carbon fiber wire with a copper wire at a specified micron level by using conductive silver paste, embedding the carbon fiber wire into a glass capillary tube, filling epoxy resin and ensuring that the carbon fiber wire is exposed, and shearing and polishing the front end of the microscale carbon fiber electrode to form a micropore structure with a smooth surface, wherein the micropore structure is used as a working electrode of a patch clamp system.

Preferably, in the step S1, the extracted vesicles are continuously extruded by an extruder and centrifugally purified in combination, a specified nm series of filter membranes are adopted for extrusion for multiple times in the extrusion process, the pore diameters of the vesicles are gradually reduced to 800nm, 400nm, 200nm and 100nm, at least 10 rounds of extrusion are carried out on each round of extrusion by a filter, so that sample groups with obvious particle size distribution differences are obtained, and corresponding labels are prepared.

Preferably, the current signal is acquired by using a AxonMultiClamp B patch clamp system by taking a carbon fiber microelectrode as a working electrode, and is carried out under +600mV bias, the sampling frequency is 100kHz, and the data is derived into a CSV format.

Preferably, the specific operation flow of the step S2 includes:

s21, carrying out Gaussian filter smoothing treatment on an original current signal, and selecting the first N peaks according to saliency sequencing based on multi-parameter combined identification characteristic peaks including peak height (> average +20pA), saliency (> 30pA), width (> 3 time points) and interval (> 10 time points);

the existence condition is that the peak height exceeds the base line mean value plus a specified threshold value, the significance is more than 30pA, the peak width is not less than 3 time points, and the peak-to-peak distance is not less than 10 time points;

s22, calculating physical and statistical characteristics, including charge integration (unit conversion to nC/C) and molecular number conversion, calculating peak height, peak width and current statistics (mean value, standard deviation, skewness and kurtosis), and then calculating the ratio of peak area time normalization value to baseline current;

the feature extraction comprises at least one of charge integration, peak height, peak width, average current, standard deviation, skewness, kurtosis, converted molecule number converted by charge and normalized area;

S23, generating a signal overview chart containing peak position marks and a unimodal analysis chart.

Preferably, in the step S21 of identifying the characteristic peaks, a signal interval with a fixed width is respectively intercepted from left to right of each group of characteristic peaks, and meanwhile, a short interval estimated baseline average current is extracted outside the section;

If there is no data about the characteristic peak, the opposite side estimation value is used.

Preferably, in the calculating process of the physical and statistical characteristics in the step S22:

the charge integration formula is:

Wherein, the Is the total amount of charge that is present,To be over timeThe current of the current that is varied,And (3) withA start time and an end time respectively specified for each peak segment;

And calculating the number of charge conversion molecules according to the Faraday constant and the Avgalileo constant, wherein the formula is as follows:

wherein N is the number of molecules wrapped by vesicles corresponding to a single current peak, Q is the charge, Is the Avoldlo constant, F is Faraday constant, n is electron transfer number;

Calculating a peak area time normalization value according to the residence time;

The baseline current ratio is the peak-to-peak current difference Δi to the baseline ratio.

Preferably, in the step S3, an accumulated variance contribution rate of not less than 95% is reserved in the process of principal component analysis dimension reduction, and the accumulated variance contribution rate is used for feature dimension compression and noise removal, so that the data quality is further improved.

Preferably, the machine learning model in the step S4 comprises any one or more of a decision tree, a random forest, a support vector machine, a K nearest neighbor or XGBoost, and is verified by adopting a multi-model integrated classification scheme.

Preferably, the machine learning model is constructed by adopting an optimization strategy that a training set/test set (80 percent: 20 percent) is divided by hierarchical sampling, grid searching and optimizing key super-parameters are adopted in the training process, and generalization capability is ensured by 5-fold cross validation.

Preferably, in the model training process of the step S4, a SMOTE algorithm is used to perform minority oversampling on the training data in the cross-validation process, so as to ensure balance of the training data.

Preferably, the model evaluation in the step S4 includes constructing a confusion matrix, a classification report, a multi-class ROC curve, and calculation and visual presentation of AUC values to evaluate the recognition classification performance of the machine learning model.

Preferably, a group of thresholding association formulas are constructed in the execution process of the step S4, and ratio parameters existThe method comprises the following steps:

When the ratio parameters are in different threshold intervals, respectively calling different machine learning models to classify the appointed samples;

Extracting physical and statistical features including peak height and peak width from a portion of the training set;

calculating a ratio parameter k of the peak height to the peak width, wherein the ratio parameter k is used for representing the peak morphological characteristics of the detection signal;

Based on the thresholding association formula, performing performance comparison on classification results of the machine learning models under different threshold conditions, and determining an optimal model for the ratio interval;

And outputting a prediction result of the optimal model to realize optimal discrimination of the vesicle particle size class.

A recognition system for single vesicle electrochemical signals, configured to perform any one of the above methods for recognizing and classifying vesicle electrochemical signals, specifically comprising:

The current acquisition module comprises a carbon fiber electrode, a patch clamp amplifier and a data acquisition card and is used for recording electric signals caused by vesicle oxidation reaction;

The signal processing and feature extraction module is used for performing signal smoothing, peak detection and multidimensional feature construction operations based on the electric signal data of the current acquisition module so as to acquire processed feature extraction data;

The data preprocessing module is used for executing missing value processing, feature standardization, PCA dimension reduction and class equalization processing operations based on the feature extraction data of the signal processing and feature extraction module;

And the classification and prediction module is used for loading a training model and carrying out particle size prediction classification on the data processed by the data preprocessing module by adopting the loading training model.

Preferably, the classifying and predicting module supports a graphical interface to display an ROC analysis chart, and realizes automatic parameter adjustment and optimization of multi-model grid search of data through Python language so as to improve classifying effect.

According to the technical scheme, the invention provides a vesicle electrochemical signal identification and classification method and an identification system. Compared with the prior art, the invention has the following advantages:

1. According to the invention, in the data characteristic extraction process, by combining a multi-parameter dynamic threshold method and Gaussian filtering, the noise interference of a current signal can be obviously reduced, the signal peak identification precision is improved, and the characteristics extracted from an electrochemical signal are ensured to be more accurate and reliable.

2. According to the invention, by fusing physical characteristics such as charge integration, molecular number conversion and the like and combining a machine learning model for deep learning and classification, the interpretability and the accuracy of a classification model are enhanced, and particularly in a complex signal analysis task, the classification accuracy of vesicle particle sizes can be effectively improved by combining a machine learning technology.

3. According to the invention, through combining the PCA dimension reduction and SMOTE data enhancement strategies, the recognition robustness of small sample types can be remarkably improved, and especially when the problem of data unbalance is faced, the adaptability and generalization capability of a model can be effectively improved by further combining a machine learning method, so that the method can adapt to the diversity of data in different experimental environments.

4. According to the invention, classification tasks are carried out by integrating a plurality of machine learning models, and classification accuracy and robustness of the models are remarkably improved by cross-validation, super-parameter tuning and other optimization means, so that not only is the accuracy of signal classification enhanced, but also complex electrochemical signal data can be processed, the intelligent level and the automation degree of the system are improved, and the method is suitable for large-scale vesicle particle size classification tasks.

5. The invention combines the advantages of electrochemical high-sensitivity measurement and intelligent algorithm classification, has the advantages of high recognition efficiency, good repeatability, strong applicability and the like, and can be applied to the fields of nano-carriers, membrane channel mechanisms, biomedical material analysis and the like.

It should be understood that the description in this section is not intended to identify key or critical features of the embodiments of the invention or to delineate the scope of the invention. Other features of the present invention will become apparent from the description that follows. Of course, it is not necessary for any one product to practice the invention to achieve all of the advantages set forth above at the same time.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

FIG. 1 is a schematic diagram of the overall operation flow of the present invention;

FIG. 2 is an overview of the peak distribution in the feature extraction step of the present invention;

FIG. 3 is a visual illustration of a single peak analysis of the current peaks extracted in the feature extraction step of the present invention;

FIG. 4 is a schematic view of a thermal analysis of the correlation of features in a data preprocessing step of the present invention;

FIG. 5 is a graph illustrating the cumulative variance of PCA in the data preprocessing step of the present invention;

FIG. 6 is a graph showing the comparison of the accuracy of different models in the model training and prediction steps of the present invention;

FIG. 7 is a matrix of confusion probabilities for a decision tree model in the model training and prediction step of the present invention;

FIG. 8 is a matrix of confusion probabilities for a random forest model in the model training and prediction step of the present invention;

FIG. 9 is a confusion probability matrix of the SVM model in the model training and prediction step of the present invention;

FIG. 10 is a matrix of confusion probabilities for the XGboost model in the model training and prediction step of the present invention;

FIG. 11 is a confusion probability matrix of the KNN model in the model training and prediction step of the present invention;

FIG. 12 is a graph of ROC of a decision tree model in the model training and prediction step of the present invention;

FIG. 13 is a graph of ROC of a random forest model in the model training and prediction step of the present invention;

FIG. 14 is a ROC graph of an SVM model during the model training and prediction step of the present invention;

FIG. 15 is a graph of the ROC of XGboost models in the model training and prediction step of the present invention;

FIG. 16 is a ROC graph of a KNN model in the model training and prognosis step of the invention.

Detailed Description

The following description of the embodiments of the present application will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all embodiments of the application. Embodiments of the application and features of the embodiments may be combined with each other without conflict. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.

In the examples, see fig. 1 to 16 in detail.

As shown in fig. 1, the method for identifying and classifying electrochemical signals of single vesicle particles provided by the embodiment of the invention comprises the following operation procedures in the specific implementation process:

S1, preparing a microscale carbon fiber electrode, connecting the microscale carbon fiber electrode with a patch clamp system, and applying a set voltage to acquire an oxidation current signal caused by a vesicle on the surface of the electrode.

In a specific implementation process, the microelectrode is prepared by using carbon fiber filaments with the diameter of about 4 μm and through glass capillary encapsulation, and the specific steps comprise:

(a1) Selecting commercial carbon fiber filaments, and cutting about 1cm;

(a2) One end of the copper wire is connected with the copper wire by conductive silver paste for conduction, and then the copper wire is dried in a high-temperature drying box at 80 ℃ for 20 min;

(a3) Inserting the copper wire connected with the carbon fiber wire into a capillary glass tube, filling epoxy resin into the glass tube until the copper wire part is completely covered, ensuring that a part of the carbon fiber wire is exposed out of the glass tube, and drying in a high-temperature drying box at 80 ℃ for 40 min;

(a4) Polishing the front end carbon fiber with sand paper to form a smooth electrode surface structure;

(a5) Optionally subjecting the electrode to an electrochemical activation treatment (acid treatment or alkali treatment) to increase the sensitivity;

(a6) The prepared electrode connection patch clamp system is used as a working electrode.

Furthermore, the application adopts vesicles as identified objects, wherein the vesicles are extruded and separated by an extruder (Genizer), and the specific sample acquisition process comprises the following steps:

(b1) Continuously extruding vesicle samples through an extruder (Genizer), adopting a series of filter membranes with different specifications, wherein the pore diameters of the filter membranes are respectively 100nm, 200nm, 400nm and 800nm, extruding each sample through the filter for at least 10 rounds, and then performing gradient centrifugation and purification to obtain four samples with obvious main peak particle diameter difference, wherein the four samples are used as classification basis for subsequent machine learning model training.

It should be noted that the extruded components of each stage of the vesicle sample are stored in separate samples, the particle size distribution is mainly concentrated in small, medium, large and medium particles, and the labels are made as target variable columns, such as vc_50, vc_100, vc_150, and vc_200, so as to correspond to 100nm, 200nm, 400nm, and 800nm vesicle sample groups, respectively.

In addition, in the signal acquisition process, the corresponding used electrical signals are acquired by a AxonMultiClamp700B patch clamp system, and the main setting parameters are as follows:

(c1) Using 4 μm+ -carbon fiber microelectrode;

(c2) The data acquisition card is connected with the amplifier system and matched with the amplifier system;

(c3) Dropping the vesicle sample into a centrifuge tube at room temperature;

(c4) Applying +600mV bias voltage to the system, and recording a current signal generated by oxidation reaction of the vesicle on the surface of the microelectrode;

(c5) The sampling frequency is set to 100kHz;

(c6) The picked signals are derived in text format (. Txt/. Csv) for subsequent data analysis.

S2, carrying out smooth filtering on the original data of the collected current signals, and realizing characteristic peak identification and multi-dimensional physical/statistical feature extraction by using a multi-parameter peak identification method based on a threshold value, wherein the single peak analysis of the characteristic peak identification (extracted current peak) is referred to as figure 2.

In a preferred embodiment of the present application, the peak value recognition and feature extraction process is performed on the original current signal measured by the patch clamp or the grinding plate line by adopting the following steps:

(S21) importing original current-time sequence data from the electrochemical signal acquisition equipment, wherein the data format is double-channel text (time/ms, current/pA). To reduce noise interference, the original signal is smoothed using gaussian filtering (σ=3).

The peak identification is carried out by a dynamic threshold method (S22), wherein the peak identification is carried out by using the following conditions that (1) peak height is larger than a smoothed signal mean value plus a threshold value (such as +20pA), (2) peak significance (prominence) is not smaller than 30pA, (3) peak width is not smaller than 3 time points, (4) distance between adjacent peaks is not smaller than 10 time points, and (5) when the candidate peak number is larger than a specified upper limit, the most significant top N peaks are selected according to prominence.

(S23) a section of the fixed width signal section is cut out for each peak and left and right, respectively, while a short section estimated baseline current is extracted outside the section (averaged). If there is no data left or right, the baseline stability is ensured using the opposite side estimates.

(S24) calculating the characteristic value, namely, carrying out the following physical and statistical characteristic calculation on each peak segment:

(a) Thresholding correlation formulas, with ratio parameters The method comprises the following steps:

When the ratio parameters are in different threshold intervals, respectively calling different machine learning models to classify the samples;

the conversion formula of the charge conversion molecular number is as follows:

Q is the total charge (unit: pA.s);

the conversion formula of the charge conversion molecular number is as follows:

wherein N is the number of vesicle wrapped molecules corresponding to a single current peak, Q is the charge (unit C), Is the Avwherem F is Faraday constant, n is electron transfer number,Is a time-varying current (unit: pA),AndIs the start and end time (in s) of the integration.

(B) Peak height, peak width, average current, standard deviation;

(c) Skewness (skewness), kurtosis (kurtosis);

(d) Current peak area normalization (in residence time);

(e) Peak-to-peak current difference (Δi) to baseline ratio;

(f) The number of charge-transfer molecules (calculated based on faraday constant and avogalileo constant).

(S25) drawing an independent graph for each peak, marking the peak start-stop time and the current curve, storing the graph in a form of an image, visually representing the analysis of a single peak as shown in FIG. 3, and storing characteristic data of all peaks as a structured CSV file to provide training input for a subsequent machine learning model.

(S26) drawing an overlay of the original signal and the smooth signal, marking each peak position and number, and realizing visual evaluation of the overall signal identification effect.

S3, carrying out preprocessing operations including missing value processing, standardization, main component analysis dimension reduction and class equalization on peak characteristic data in the extracted characteristic data.

In particular, in another embodiment of the application, further preprocessing and feature engineering are performed on the extracted feature value data of the vesicle current signal, so that the robustness and the discrimination capability of a subsequent machine learning model are improved. The method specifically comprises the following steps:

And (S31) detecting missing values of the acquired vesicle peak characteristic data, and filling by using an average value of a numerical value column to eliminate the influence of sample incompleteness on the model performance.

(S32) analyzing the linear correlation between the features by using a Pearson correlation coefficient matrix, and displaying the high-correlation feature pairs by using a thermodynamic diagram form as shown in fig. 4, so as to provide reference for the subsequent redundant feature processing and dimension reduction.

And S33, all input features are of a numerical type, and a Z-score standardization method STANDARDSCALER is adopted, so that the features of different dimensions are ensured to have uniform distribution, and the training of a distance sensitive model (such as SVM) is facilitated.

(S34) dividing the data set into a training set and a testing set according to the proportion of 80%/20%, and adopting a hierarchical sampling mode (STRATIFIED SPLIT) with balanced label distribution to prevent bias caused by uneven classification.

(S35) mapping the original high-dimensional feature data into a lower-dimensional space using Principal Component Analysis (PCA) with a 95% cumulative variance contribution ratio retained, denoising while compressing the feature dimensions, and as a result, referring to FIG. 5, showing the contribution of different principal components to the cumulative interpretation variance. The process automatically selects the appropriate dimension without manual setting.

And (S36) identifying high redundancy feature pairs by using the absolute correlation coefficient larger than 0.85 as a threshold value, wherein the high redundancy feature pairs can be used as optional feature reduction basis, and feature redundancy is reduced on the premise of maintaining model performance.

(S37) saving the preprocessed and dimension-reduced training data set as a structured CSV file for the model training and deployment module to call.

S4, constructing and training at least one machine learning model based on the preprocessed characteristic data, and automatically identifying the particle size class of the vesicle.

The core of the multi-model integrated training and classifying identification method for the vesicle electrochemical signal characteristic data is to construct a plurality of mainstream machine learning models, and train, optimize, evaluate and deploy by adopting a unified interface, and specifically comprises the following steps:

(S41) a data loading and preprocessing interface comprising:

(S411) data loading, loading CSV format data generated by the aforesaid feature engineering stage.

And (S412) selecting a plurality of characteristic components after Principal Component Analysis (PCA) as input characteristics, retaining useful information and removing redundancy.

(S413) extracting labels, namely extracting target classification labels, such as VC_50, VC_100, VC_150 and VC_200, labeled according to vesicle size categories.

(S414) tag coding, converting the category tag from the character string to an integer by adopting LabelEncoder, and ensuring data compatibility.

(S415) filling in missing values, namely filling in missing values in the data by using a SimpleImputer mean value strategy, and ensuring the data integrity.

And (S416) normalizing, namely performing normalization processing on all feature dimensions by adopting STANDARDSCALER, and eliminating the influence caused by different feature dimensions.

(S42) data set partitioning strategy

(S421) training set and test set division the data set is divided into training set and test set according to the ratio of 80% to 20% by using the train_test_split method.

(S422) hierarchical sampling, namely enabling stratify parameters, ensuring that the vesicle size category distribution in the training set and the testing set is consistent, and avoiding model bias.

(S423) oversampling enhancement to solve the sample imbalance problem, the training set is oversampled enhanced using the SMOTE method.

(S43) model selection and training framework to improve classification accuracy and enhance adaptability of the system, the embodiment respectively builds and trains the following five types of machine learning models. Each model is well-defined and is realized in a targeted manner in combination with a specific scene.

(S431) designing and constructing a group of association formulas based on physical and statistical characteristics to realize intelligent discrimination of vesicle particle size categories. Specifically, a thresholding formula is constructed by taking the ratio of the peak height to the peak width of the signal as a key distinguishing characteristic to carry out model shunt. And when the ratio parameter k is in different threshold intervals, respectively calling different machine learning models to classify the samples.

At this time, as shown in fig. 6, the model training and prediction process is different in machine learning accuracy.

Thus, multiple sets of binarization formulas can be constructed through combinations of different thresholds, thereby establishing dynamic correlations between physical and statistical features and machine learning models. Taking the experimental result that K takes a value of 0.25 to 2.25 as an example, as shown in fig. 6, when k=0.25, the accuracy of the support vector machine, XGboost and the random forest all reach about 0.905, and the accuracy of the K neighbor is slightly lower, which is 0.895. As the threshold increases, the K-nearest neighbor accuracy gradually decreases, while the random forest sum XGBoost remains at a steady level around 0.905. The result shows that the robustness and generalization capability of the model can be improved while the classification accuracy is ensured by constructing a reasonable threshold formula and calling different models under different conditions.

According to the rule of experimental results in the k epsilon [0.25,2.25] interval, the following shunt and priority discrimination strategies are established:

when K is less than or equal to 1.0, five types of models including K neighbor and decision tree are allowed to be used so as to consider the flexibility of a low threshold interval.

When K >1.0, K nearest neighbor is disabled, avoiding rapid drop in accuracy.

When k >1.25, the decision tree is further disabled, and only the random forest, the support vector machine and XGBoost are reserved, so that the classification accuracy is ensured to be stabilized above 0.90.

In each threshold interval where k >1.0, random forests and XGBoost are preferentially invoked, as they exhibit the highest stability and robustness in that interval.

(S432) super parameter tuning (HYPERPARAMETER TUNING)

The super-parameter tuning is an indispensable part in the model optimization process, and the super-parameters of the models are tuned through grid search (GRIDSEARCHCV) so as to improve the performance of each model. The aim of the super-parameter tuning is to find the optimal super-parameter set capable of improving the model performance by searching different parameter combinations.

(1) And the decision tree model optimizes the decision tree structure by adjusting parameters such as the maximum depth (max_depth), the minimum division sample number (min_samples_split) and the minimum leaf node sample number (min_samples_leaf) of the tree.

(2) Random forest model, number of optimized trees (n_ estimators), maximum depth of tree (max_depth), minimum number of split samples per tree (min_samples_split).

(3) Support vector machine model (SVM) by adjusting kernel type (kernel), penalty coefficient (C) and kernel width (gamma) parameters.

(4) XGBoost model optimization learning rate (learning_rate), maximum tree depth (max_depth), and number of trees (n_ estimators).

(5) The K nearest neighbor model (KNN) optimizes the number of neighbors (n_neighbors), the weight strategy (weights) and the distance measurement mode (p).

(S433) Cross-Validation (Cross-Validation)

In order to ensure that each model performs stably on different data subsets, and avoid model overfitting or performance deviation caused by the accidental data partitioning, the invention adopts K-fold cross validation (such as 5-fold cross validation). The method comprises the following specific steps:

L1. the data set is divided into K subsets, each subset being a validation set once, the remaining K-1 subsets being training sets.

And L2, training and evaluating each model, and ensuring the generalization capability of the model (3) by calculating the average value of the performances of a plurality of training and verification sets.

Through cross validation, the stability of the model can be effectively evaluated, and accidental errors caused by data division can be reduced.

Aiming at the problem of non-uniform vesicle size category distribution, SMOTE (synthetic minority class oversampling technology) is introduced in the cross validation process, new samples are synthesized by interpolation in a feature space, the representativeness of small categories is enhanced, and the generalization capability of the model is improved.

At this time, in the model training stage, the original data set is first divided into a plurality of folds for cross-validation. In each round of cross-validation, only the training fold data of the current round is subjected to oversampling processing by adopting an SMOTE algorithm to generate an equalized training sample set, and the fold data is validated not to be subjected to oversampling processing so as to maintain the objectivity of evaluation. The model is then trained using the SMOTE-processed training sample set, and model performance is assessed using the corresponding validation fold data.

(S434) model training and evaluation

In the model selection and training stage, five machine learning models are respectively constructed, and each model is trained and evaluated. The training process for each model combines hyper-parametric tuning and cross-validation to ensure optimal performance is achieved.

(1) Decision Tree model (precision Tree)

The model definition is that a decision tree is a tree-shaped structure classifier, a plurality of decision paths are constructed through splitting of characteristic conditions, and classification results are output at leaf nodes.

Implementation is to use DecisionTreeClassifier and combine grid search optimization hyper-parameters (e.g., max_depth, min_samples_split, and min_samples_leaf) to select the optimal decision tree structure. The model has stronger interpretability and calculation efficiency, and is suitable for vesicle particle size classification tasks.

(2) Random Forest model (Random Forest)

The model definition is that the random forest is an integrated learning model composed of a plurality of decision trees, and the generalization capability of the model is improved through a Bagging technology.

The implementation mode is that RandomForestClassifier is used, and the robustness of the model is enhanced through grid search of super parameters such as the number of the optimization trees, the maximum depth, the minimum split sample number and the like. The model can effectively cope with noise and data imbalance problems.

(3) Support vector machine model (Support Vector Machine, SVM)

Model definition the SVM aims to build a hyperplane that maximizes the inter-class spacing. For the non-linearity problem, the input is mapped to a high-dimensional space by means of a kernel function.

The implementation mode is that an SVC model is used, probability prediction options are started, and multi-category classification tasks are adapted. The classification effect is optimized by adjusting kernel function types (linear, rbf), penalty coefficients C and kernel width parameters gamma.

(4) K nearest neighbor model (K-Nearest Neighbors KNN)

Model definition KNN is an example-based learning method, and classification is performed according to the distance between a sample and a training set.

Implementation is to use KNeighborsClassifier and enhance the robustness of the model by optimizing the number of neighbors n _ neighbors and the weighting scheme (equal weight/distance weighting). The method is suitable for classification tasks with clear feature distribution.

(5) XGBoost model (eXtreme Gradient Boosting)

Model definition XGBoost is an integrated learning method based on a tree structure, and regularization term and efficient feature selection are combined.

Training is carried out through XGBClassifier, super parameters such as learning rate, maximum tree depth, tree number and the like are optimized, and excellent performance of the model in high-dimensional sparse data is ensured.

Furthermore, based on the five machine learning models, the vesicle particle size classification method also needs to be subjected to multidimensional assessment, and specific indexes include accuracy, confusion matrix, classification report, ROC curve, AUC value and the like, which are used for comprehensively assessing the performance of the model, wherein:

(a) The Accuracy (Accuracy) is the final output result of each model of the terminal, and the Accuracy of each model on the test set is as follows:

Decision tree model 0.8875

Random forest model 0.9062

SVM model 0.9234

XGBoost model 0.9187

KNN model 0.9142

The accuracy rate shows that the vesicle particle size classification method can effectively distinguish different particle size categories.

(B) The confusion matrix (Confusion Matrix) is a tool for evaluating the performance of classification models. It is a square matrix in which rows represent the actual categories (true labels) and columns represent the predicted categories (predicted labels). Each element represents the comparison of the predicted outcome of a certain class with the actual label. In a binary problem, the confusion matrix typically contains four values, true (TP), false (FP), true (TN) and False (FN) cases.

For multi-class classification problems, the confusion matrix is expanded into a square matrix, and comprises a plurality of classes, wherein each element represents the matching condition of the actual class and the predicted class. Elements on the diagonal represent the proportion of correct classifications, while elements on the non-diagonal represent the proportion of incorrect classifications.

The confusion matrix not only provides classification accuracy, but also helps evaluate the performance of the model on each category, especially in the case of category imbalance, by more comprehensive evaluation through other metrics such as F1 score, accuracy, and recall.

The confusion matrix of each model respectively shows a classification result, the confusion number matrix is normalized to form a confusion probability matrix, see fig. 7-11 for details, which are used for representing specific actual results of performing model training and prediction operations in a specific implementation mode of multi-model integrated classification of vesicle electrochemical signal characteristic data of the application, wherein fig. 7 is the confusion probability matrix of a decision tree model, fig. 8 is the confusion probability matrix of a random forest model, fig. 9 is the confusion probability matrix of an SVM model, fig. 10 is the confusion probability matrix of a XGboost model, fig. 11 is the confusion probability matrix of a KNN model, and classification performance of the model can be clearly seen by comparing prediction results and actual categories of different categories, so that prediction accuracy of each category under each model is provided.

(C) A classification report (Classification Report) for displaying precision, recall, f-score for each class of model, in one embodiment, for the vesicle particle size classification procedure of the present application, the classification report for each model is as follows:

Decision tree classification report:

Random forest classification report:

SVM classification report:

XGboost classification report:

KNN classification report:

In conclusion, the macro average f1-score of the decision tree model is 0.89, which shows that the model has balanced classification performance, the macro average f1-score of the random forest model is 0.91, the macro average f1-score of the SVM model is 0.92, excellent classification performance is shown, the macro average f1-score of the XGBoost model is 0.92, the excellent effect of the integrated model in vesicle particle size classification is shown, the macro average f1-score of the KNN model is 0.92, and good stability and accuracy are shown.

(D) ROC Curve and AUC index (Multi-class)

ROC curves are tools used to evaluate the performance of classification models, which show the False Positive Rate (FPR) versus true rate (TPR) of classification models at different decision thresholds. The closer the ROC curve is to the upper left corner, the better the classification of the model, meaning a high true positive rate and a low false positive rate.

False Positive Rate (FPR) the rate that is erroneously predicted as positive sample in the case of all actual negative samples, there areWherein FP is the number of false positives and TN is the number of true negatives.

True Positive Rate (TPR) the ratio of correctly predicted positive samples, also called Sensitivity (Sensitivity) or Recall (Recall), in the case of all actually positive samples, isWherein TP is true positive number, and FN is false negative number.

AUC (Area Under the Curve) the value is the area under the ROC curve, which represents a comprehensive index of model classification performance. The range of AUC values is 0 to 1, the closer the value is to 1, the stronger the classification ability of the model, and when the AUC value is 0.5, the classification ability of the model is equivalent to random guess. The macro average AUC is the overall performance of the model in the multi-class classification task by calculating the AUC for each class and averaging.

The ROC curves of the models show the AUC values of each category, and the macro-average AUC values, for comparison of fig. 12-16, for showing the specific actual results of performing model training and prediction operations in a specific embodiment of the multi-model integrated classification of vesicle electrochemical signal feature data of the present application, wherein fig. 12 is a ROC graph of a decision tree model, fig. 13 is a ROC graph of a random forest model, fig. 14 is a ROC graph of an SVM model, fig. 15 is a ROC graph of a XGboost model, and fig. 16 is a ROC graph of a KNN model, based on which it can be derived:

Decision tree model macro average AUC is 0.96.

Random forest model macro average AUC is 0.98.

SVM model macro average AUC is 0.97.

XGBoost model macro average AUC was 0.98.

KNN model macro average AUC was 0.96.

In summary, after training multiple models, the accuracy of each model and the macro average AUC value were compared laterally. According to the evaluation result of the test set, the random forest model and the XGBoost model are outstanding in performance, the macro average AUC value is 0.98, and meanwhile, the XGBoost model also obtains higher accuracy rate 0.9187 on the test set. Immediately following the SVM model, the macro average AUC was 0.97, and the test set accuracy was 0.9234, demonstrating its excellent classification performance. The macro average AUC values of the KNN model and the decision tree model were 0.96, respectively, which, although slightly lower than the previous models, showed better classification stability and accuracy.

In combination, all models are excellent in vesicle particle size classification tasks, particularly in macro average AUC and accuracy, and the method has wide applicability and superior classification capability under various machine learning models.

On the other hand, the invention also discloses a recognition system of the single vesicle electrochemical signal, which is used for executing the recognition classification method of the vesicle electrochemical signal in the embodiment, and specifically comprises the following steps:

(1) The current acquisition module comprises a carbon fiber electrode, a patch clamp amplifier and a data acquisition card and is used for recording electric signals caused by vesicle oxidation reaction;

(2) The signal processing and feature extraction module is used for performing signal smoothing, peak detection and multidimensional feature construction operations based on the electric signal data of the current acquisition module so as to acquire processed feature extraction data;

(3) The data preprocessing module is used for executing missing value processing, feature standardization, PCA dimension reduction and class equalization processing operations based on the feature extraction data of the signal processing and feature extraction module;

(4) The classification and prediction module is used for loading a training model and carrying out particle size prediction classification on the data processed by the data preprocessing module by adopting the loading training model, in addition, the classification and prediction module also supports a graphical interface to display an ROC analysis chart, and the multi-model grid search automatic parameter adjustment optimization of the data is realized through Python language so as to improve the classification effect.

In yet another embodiment of the present application, there is also provided a computer program product comprising instructions which, when run on a computer, cause the computer to perform the method of identifying and classifying an electrochemical signal of a single vesicle particle of any of the above embodiments.

It may be understood that the system provided by the embodiment of the present invention corresponds to the method provided by the embodiment of the present invention, and explanation, examples and beneficial effects of the related content may refer to corresponding parts in the above method.

The embodiment of the application also provides an electronic device, which comprises a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory are communicated with each other through the communication bus,

A memory for storing a computer program;

and the processor is used for realizing the identification and classification method of the single vesicle particle electrochemical signals when executing the program stored in the memory.

The communication bus mentioned for the electronic device may be a peripheral component interconnect standard bus or an extended industry standard architecture bus, etc. The communication bus may be classified as an address bus, a data bus, a control bus, or the like.

The communication interface is used for communication between the electronic device and other devices.

The memory may include random access memory or may include non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one memory device located remotely from the aforementioned processor.

In the above embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When loaded and executed on a computer, produces a flow or function in accordance with embodiments of the present application, in whole or in part. The computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by a wired (e.g., coaxial cable, fiber optic, digital subscriber line), or wireless (e.g., infrared, wireless, microwave, etc.). The computer readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that contains an integration of one or more available media. The usable medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk), etc.

The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, alternatives, and improvements that fall within the spirit and scope of the invention.

In addition, if a directional indication (such as up, down, left, right, front, and rear) is referred to in the embodiment of the present invention, the directional indication is merely used to explain a relative positional relationship between the components, a movement condition, and the like in a specific posture, and if the specific posture is changed, the directional indication is changed accordingly.

In addition, if there is a description of "first", "second", etc. in the embodiments of the present invention, the description of "first", "second", etc. is for descriptive purposes only and is not to be construed as indicating or implying a relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include at least one such feature. In addition, the meaning of "and/or" as it appears throughout includes three parallel schemes, for example "A and/or B", including the A scheme, or the B scheme, or the scheme where A and B are satisfied simultaneously. In addition, in the embodiment of the present invention, "a plurality of" means two or more. In addition, the technical solutions of the embodiments may be combined with each other, but it is necessary to base that the technical solutions can be realized by those skilled in the art, and when the technical solutions are contradictory or cannot be realized, the combination of the technical solutions should be considered to be absent and not within the scope of protection claimed in the present invention.

Claims (9)

1. A method for identifying and classifying electrochemical signals of single vesicle particles, comprising the steps of:

S1, preparing a carbon fiber ultramicroelectrode, applying voltage in a patch clamp system constant potential mode, and collecting a picoampere-level current response generated by oxidizing a single vesicle content on the surface of the electrode;

s2, carrying out smooth filtering on the collected original current signals, and realizing characteristic peak identification and multidimensional physical/statistical characteristic extraction by using a multi-parameter peak identification method based on a threshold value;

S3, carrying out preprocessing operations including missing value processing, standardization, main component analysis dimension reduction and class equalization on peak characteristic data in the extracted characteristic data;

s4, constructing and training at least one machine learning model based on the preprocessed characteristic data, wherein the machine learning model is used for automatically identifying the particle size class of the vesicle particles.

2. The method for identifying and classifying electrochemical signals of single-vesicle particles according to claim 1, wherein the preparation process of the micro-scale carbon fiber electrode in the step S1 comprises the steps of connecting a carbon fiber wire with a copper wire at a specified micron level by using conductive silver paste, embedding the carbon fiber wire into a glass capillary tube, filling epoxy resin and ensuring that the carbon fiber wire is exposed, and shearing and polishing the front end of the micro-scale carbon fiber electrode to form a disc structure with a smooth surface to be used as a working electrode of the single-vesicle electrochemical measurement system.

3. The method for identifying and classifying electrochemical signals of single vesicle particles as recited in claim 1, wherein the specific operational procedure of step S2 comprises:

S21, carrying out Gaussian filter smoothing treatment on an original current signal, and selecting the first N peaks according to saliency sequencing based on multi-parameter joint identification characteristic peaks including peak height, saliency, width and spacing;

S22, calculating physical and statistical characteristics, including charge integration and conversion of molecular numbers, calculating peak height, peak width and current statistics, and then calculating the ratio of a peak area time normalization value to a base line current;

S23, generating a signal overview chart containing peak position marks and a unimodal analysis chart.

4. The method for identifying and classifying electrochemical signals of single vesicle particles according to claim 3, wherein in the step of identifying the characteristic peaks of S21, a signal interval with a fixed width is intercepted for each group of characteristic peaks, and a short interval is extracted outside the interval to estimate a baseline average current;

If there is no data about the characteristic peak, the opposite side estimation value is used.

5. The method for identifying and classifying electrochemical signals of single vesicle particles as recited in claim 4, wherein in the step of S22, the physical and statistical features are calculated as follows:

the charge integration formula is:

Wherein, the Is the total amount of charge that is present,To be over timeThe current of the current that is varied,And (3) withA start time and an end time respectively specified for each peak segment;

And calculating the number of charge conversion molecules according to the Faraday constant and the Avgalileo constant, wherein the formula is as follows:

wherein N is the number of molecules wrapped by vesicles corresponding to a single current peak, Q is the charge, Is the Avoldlo constant, F is Faraday constant, n is electron transfer number;

Calculating a peak area time normalization value according to the residence time;

The baseline current ratio is the peak-to-peak current difference Δi to the baseline ratio.

6. The method for identifying and classifying electrochemical signals of single vesicle particles as recited in claim 1, wherein the machine learning model in step S4 comprises a parallel training decision tree, a random forest, and a SVM, XGBoost, KNN model;

The construction and training process of the machine learning model adopts the following optimization strategies that a training set and a testing set are divided according to 80 percent, 20 percent of hierarchical sampling, grid searching, tuning and optimizing key super-parameters are adopted, and 5-fold cross verification ensures generalization capability;

And S4, evaluating the recognition classification performance of the machine learning model by adopting a confusion matrix.

7. The method for identifying and classifying electrochemical signals of single vesicle particles as recited in claim 6, wherein a set of thresholding associated formulas are constructed during the execution of step S4, and ratio parameters are presentThe method comprises the following steps:

and when the ratio parameter k is in different threshold intervals, respectively calling different machine learning models to classify the appointed sample so as to realize the optimal discrimination of the vesicle particle size class.

8. A system for identification of single vesicle electrochemical signals, characterized in that it is adapted to perform the method for identification and classification of vesicle electrochemical signals according to any of the preceding claims 1-7, comprising in particular:

The current acquisition module comprises a current amplifier and an analog-to-digital converter, wherein the carbon fiber is used as a working electrode for recording an electric signal generated by oxidation of vesicle contents;

The signal processing and feature extraction module is used for performing signal smoothing, peak detection and multidimensional feature construction operations based on the electric signal data of the current acquisition module so as to acquire processed feature extraction data;

The data preprocessing module is used for executing missing value processing, feature standardization, PCA dimension reduction and class equalization processing operations based on the feature extraction data of the signal processing and feature extraction module;

And the classification and prediction module is used for loading a training model and carrying out particle size prediction classification on the data processed by the data preprocessing module by adopting the loading training model.

9. The recognition system of single vesicle electrochemical signals of claim 8, wherein the classification and prediction module supports a graphical interface to display ROC analysis charts and implements multi-model grid search automatic parameter tuning optimization of data via Python language to improve classification.