P-glycoprotein inhibitor and substrate prediction method based on multi-modal contrast learning

Abstract

The invention discloses a P-glycoprotein inhibitor and substrate prediction method based on multi-modal contrast learning. The method comprises the following steps: constructing a data set of a P-glycoprotein inhibitor and a substrate; constructing a P-glycoprotein inhibitor and substrate prediction model by using a deep learning method; inputting the comparison loss and the classification loss into the constructed P-glycoprotein inhibitor and substrate prediction model, combining the comparison loss and the classification loss to obtain a total loss function, and training and optimizing parameters of the P-glycoprotein inhibitor and substrate prediction model through the total loss function; and inputting data of a compound to be predicted into the optimized P-glycoprotein inhibitor and substrate prediction model, outputting a final classification result, and judging whether the compound is the P-glycoprotein inhibitor and the substrate or not. According to the method, the prediction precision and generalization performance of the model are improved, the interpretability of the model is realized, and a reliable scientific basis is provided for molecular design in drug development.

Description

Technical Field

[0001] The present invention relates to the field of deep learning prediction of drug properties, and in particular to a method for predicting P-glycoprotein inhibitors and substrates based on multimodal comparative learning. Background Art

[0002] P-glycoprotein is a transmembrane protein located on the cell membrane. It is widely expressed in various tissues such as the liver, kidney, small intestine, and blood-brain barrier, and is involved in the absorption, distribution, metabolism, and excretion of drugs. It is responsible for expelling many harmful substances from the cell to the extracellular space, but it also pushes many drugs out of the cell, which can significantly limit or destroy the activity of many drugs. Given the key role of P-gp in drug efflux and multidrug resistance, it is of great significance to study the prediction of P-gp inhibitors and substrates. The activity evaluation of P-gp inhibitors or substrates can be obtained through in vivo or in vitro experiments, but these methods have disadvantages such as long cycles, high costs, and unstable experimental conditions. Therefore, there is an urgent need to develop computational methods to assist in the experimental screening of P-gp inhibitors and substrates.

[0003] In recent years, with the development of artificial intelligence and cheminformatics, machine learning and deep learning methods have been increasingly applied to the prediction of P-gp inhibitors and substrates. However, current research faces several significant challenges. The primary challenge is the data bottleneck. Due to the difficulty in unifying experimental environment standards and the presence of high noise, the quality of existing data sets related to P-gp inhibitors and substrates is uneven and the scale is relatively limited. This situation directly limits the accuracy and generalization ability of the prediction model.

[0004] In addition, in terms of molecular characterization, most existing models rely on a single type of molecular characterization, which makes it difficult to fully and deeply capture the complex and changeable structural information and dynamic interaction processes within the molecule, thus limiting the model's comprehensive understanding and accurate prediction of molecular properties. The lack of interpretability of the model is also a problem that cannot be ignored. Although technologies such as deep learning have repeatedly achieved great results in prediction accuracy, their "black box" characteristics make the biochemical mechanisms behind the prediction results elusive. This not only hinders the in-depth understanding of the drug's mechanism of action, but also limits the conversion efficiency of research results in practical applications such as drug development. Therefore, in order to promote the further development of P-gp inhibitor and substrate prediction research, it is urgent to build high-quality, large-scale data sets, explore multimodal molecular characterization methods to enhance the accuracy of prediction models, and strengthen model interpretability research to identify key functional groups closely related to P-gp interactions, so as to provide more solid and reliable theoretical and technical support for drug development. Summary of the invention

[0005] To solve the above problems, the present invention provides a P-glycoprotein inhibitor and substrate prediction method based on multimodal contrastive learning, constructs a high-quality P-glycoprotein inhibitor and substrate dataset, improves the model prediction accuracy and generalization performance, and realizes the interpretability of the model, providing a reliable scientific basis for molecular design in drug development.

[0006] A method for predicting P-glycoprotein inhibitors and substrates based on multimodal contrastive learning, comprising the following steps:

[0007] 1) Construct a dataset of P-glycoprotein inhibitors and substrates;

[0008] 2) Use deep learning methods to construct prediction models for P-glycoprotein inhibitors and substrates;

[0009] 3) inputting the data set obtained in step 1) into the P-glycoprotein inhibitor and substrate prediction model constructed in step 2), introducing a graph contrast learning strategy for the molecular graph features to obtain a molecular graph after data augmentation, calculating a contrast loss between the molecular graph and the molecular graph after data augmentation, calculating a classification loss based on the fusion features output by the P-glycoprotein inhibitor and substrate predictor, combining the contrast loss and the classification loss to obtain a total loss function, training and optimizing the parameters of the P-glycoprotein inhibitor and substrate prediction model through the total loss function, and obtaining an optimized P-glycoprotein inhibitor and substrate prediction model;

[0010] 4) Inputting the data of the compound to be predicted into the optimized P-glycoprotein inhibitor and substrate prediction model, outputting the final classification result, and judging whether the compound is a P-glycoprotein inhibitor and substrate.

[0011] In step 1), a data set of P-glycoprotein inhibitors and substrates is constructed, specifically including:

[0012] Experimental data and compound SMILES sequences related to molecular compounds and P-glycoprotein inhibitors and substrates were collected from relevant databases, integrated and preprocessed to obtain data sets of P-glycoprotein inhibitors and substrates.

[0013] In step 2), the P-glycoprotein inhibitor and substrate prediction model specifically includes:

[0014] A molecular SMILES sequence feature extraction module, used to receive data sets of P-glycoprotein inhibitors and substrates and extract molecular SMILES sequence features;

[0015] A molecular fingerprint feature extraction module, used for receiving data sets of P-glycoprotein inhibitors and substrates and generating molecular fingerprint features;

[0016] A molecular graph feature extraction module, for receiving a data set of P-glycoprotein inhibitors and substrates and generating molecular graph features;

[0017] A feature fusion module constructed based on the attention mechanism is used to receive the molecular SMILES sequence features output by the molecular SMILES sequence feature extraction module, the molecular fingerprint features output by the molecular fingerprint feature extraction module, and the molecular graph features output by the molecular graph feature extraction module, and fuse the molecular SMILES sequence features, the molecular fingerprint features and the molecular graph features to output a fusion feature;

[0018] P-glycoprotein inhibitor and substrate predictor, including a fully connected layer, for receiving fused features and predicting output classification results.

[0019] In step 2), the molecular SMILES sequence features are extracted, and the specific steps include:

[0020] 2-1-1) Divide the SMILES sequence in the dataset of P-glycoprotein inhibitors and substrates into individual atoms or symbols, then look up the corresponding atom's embedding vector from the pre-trained dictionary, and if it is not in the dictionary, generate a random embedding vector; aggregate the atom's embedding vector and the random embedding vector to obtain a preliminary representation of the SMILES sequence;

[0021] 2-1-2) Preprocess the preliminary representation of the SMILES sequence through a bidirectional long short-term memory network Bi-LSTM to capture the SMILES context information representation;

[0022] 2-1-3) The attention weight of each atom in the SMILES context information representation is calculated through the attention mechanism, and the hidden state of each atom is weighted and summed according to the attention weight to obtain the feature vector of the molecular SMILES sequence;

[0023] 2-1-4) Pass the feature vector of the molecule SMILES sequence into the fully connected layer to obtain the SMILES sequence feature S of molecule i i .

[0024] In step 2), the molecular fingerprint feature is generated, which specifically includes:

[0025] Calculate the circular topological fingerprint ECFP4 fingerprint and PubChem fingerprint of the molecules in the dataset of P-glycoprotein inhibitors and substrates respectively, connect the two fingerprints to obtain a mixed fingerprint, and pass the mixed fingerprint to the fully connected layer to obtain the molecular fingerprint feature F of molecule i i .

[0026] In step 2), the molecular graph features are generated, specifically including:

[0027] 2-3-1) Extract node and edge information based on molecular SMILES in the dataset of P-glycoprotein inhibitors and substrates, and use the node and edge information to construct the adjacency matrix of the graph;

[0028] 2-3-2) Calculate the physicochemical properties of each atom in the data set of P-glycoprotein inhibitors and substrates as the initial feature vectors of the points in the molecular graph;

[0029] 2-3-3) Using the graph isomorphic network GIN, the initial feature vector obtained in step 2-3-2) is updated by mixing the node's own features and the features of neighboring nodes in the adjacency matrix of the aggregate graph. The global graph representation of the entire molecule is obtained by performing a global average pooling operation on the features of all nodes;

[0030] 2-3-4) Input the global graph representation into the fully connected layer to obtain the molecular graph feature G of molecule i i .

[0031] In step 3), the total loss function specifically includes:

[0032] The molecular representation M of the fused features is obtained through the fully connected layer i Processing is performed to obtain the final molecular representation y for the classification task i , and then calculate the binary cross entropy loss as the classification loss The calculation formula is:

[0033]

[0034] Where N is the total number of molecules, y i is the actual label of the ith molecule, is the probability that the model predicts that the i-th molecule is a positive class;

[0035] A graph contrast learning strategy is introduced for the molecular graph features to obtain the molecular graph after data augmentation. The NT-XEnt loss between the molecular graph and the molecular graph after data augmentation is calculated as the contrast loss.

[0036] The final total loss function It is defined as the weighted sum of classification loss and contrastive learning loss:

[0037]

[0038] Among them, β is an adjustable weight parameter.

[0039] Specifically, the technical solution of the present invention is as follows:

[0040] 1) Construct a dataset of P-glycoprotein (P-gp) inhibitors and substrates, including collecting experimental data related to small molecule compounds and P-glycoprotein inhibitors and substrates, compound simplified molecular linear input specification format SMILES sequences and other information from relevant databases, and perform preprocessing after integration to obtain the final dataset;

[0041] 2) Use deep learning methods to build a prediction model for P-glycoprotein inhibitors and substrates. The model architecture includes the following components: molecular SMILES sequence feature extraction module, molecular fingerprint feature extraction module, molecular graph feature extraction module, feature fusion module based on attention mechanism, and predictor based on fully connected layer;

[0042] 3) During the training process, a graph contrast learning strategy is introduced for the molecular graph features to obtain the molecular graph after data augmentation, and the contrast loss between the molecular graph and the augmented graph is calculated;

[0043] 4) using the data set obtained in step 1), and by calculating the classification loss of the predictor in step 2) and the contrast loss in step 3), training and optimizing the parameters of the P-glycoprotein inhibitor and substrate prediction model to obtain an optimized prediction model;

[0044] 5) Inputting the data of the compound to be predicted into the optimized P-glycoprotein inhibitor and substrate prediction model, outputting the final classification result, and determining whether the compound is a P-glycoprotein inhibitor and substrate;

[0045] 6) Perform interpretability analysis on the prediction model to identify key functional groups in the compounds that are closely related to P-gp interaction.

[0046] In step 1), in order to construct a P-gp inhibitor and substrate data set, the following steps are specifically included:

[0047] 1-1) Text mining and manual verification methods were used to collect experimental data related to P-gp from databases and literature, including ChEMBL, OCHEM, PubChem, DrugBank, etc.;

[0048] 1-2) Carefully check the experimental data and label the compounds as positive or negative according to the specified threshold;

[0049] 1-3) Remove duplicate and inconsistent data from the dataset;

[0050] 1-4) All compound structures are pre-processed by standardization, including removing inorganic substances and complexes, converting salts into corresponding acids or bases, and retaining only one data for stereoisomers;

[0051] 1-5) Collect additional compound data from relevant studies and databases as an external test set, apply the above data preprocessing steps to the external test set, and delete the compounds and corresponding bioactivity data that are repeated in the external test set and the data set.

[0052] In the step 2), according to the SMILES of the input data, the molecular SMILES sequence features are extracted by a molecular SMILES sequence feature extraction module, which specifically includes the following steps:

[0053] 2-1-1) Divide the given SMILES sequence into individual atoms or symbols, and then look up the corresponding atom's embedding vector from the pre-trained dictionary. If it is not in the dictionary, generate a random embedding vector; the pre-trained dictionary contains the embedding vectors of common atoms;

[0054] 2-1-2) Aggregate the embedding vectors to get the embedding matrix Where m is the number of atoms and d is the dimension of the embedding vector;

[0055] 2-1-3) Through the above steps, we can get the feature vector X of the SMILES sequence = {x1, x2, ..., x n},in is the feature vector of the tth atom. The SMILES context information is captured by the bidirectional long short-term memory network Bi-LSTM to obtain the hidden state h of atom t t :

[0056]

[0057] in and Respectively represent the forward and backward LSTM units, h t-1 is the hidden state of atom t-1;

[0058] 2-1-4) The attention mechanism is used to further improve the model's ability to capture important sequence information. The following formula is used to calculate h t The attention weights are:

[0059] e t =tanh(W1h t +b1) (2)

[0060]

[0061] where e t is the attention score of atom t, which is calculated by the tanh activation function feedforward neural network, W1 is the weight matrix, b1 is the bias vector; the attention score is normalized by the Softmax function to obtain the attention weight α of atom t t ;

[0062] 2-1-5) According to the attention weight, the hidden state of each atom is weighted and summed to obtain the feature vector of the molecular SMILES sequence The feature vector h s Pass it into the fully connected layer to obtain the SMILES sequence feature S of molecule i i .

[0063] In the step 2), according to the SMILES of the input data, a molecular fingerprint feature is generated by a molecular fingerprint feature extraction module, which specifically includes the following steps:

[0064] Calculate the circular topological fingerprint ECFP4 fingerprint and the substructure-based PubChem fingerprint of the molecule respectively, connect the two fingerprints to get the hybrid fingerprint, and pass the hybrid fingerprint into the two fully connected layers to get the molecular fingerprint feature F of molecule i i .

[0065] In the step 2), according to the SMILES of the input data, a molecular graph feature is generated by a molecular graph feature extraction module, which specifically includes the following steps:

[0066] 2-3-1) Extract node and edge information based on molecular SMILES, where each atom in the molecule is mapped as a node in the molecular graph, and the chemical bonds between atoms are mapped as edges in the molecular graph. Use the node and edge information to construct the adjacency matrix of the graph to represent the connection relationship between nodes;

[0067] 2-3-2) Calculate the physical and chemical properties of each atom as the initial feature vector of the point in the molecular graph; the physical and chemical properties specifically include atomic type, charge number, chiral carbon situation, number of connected hydrogen atoms, hybridization type, whether it is an aromatic atom, etc.;

[0068] 2-3-3) Using the graph isomorphic network GIN, the node representation is updated by aggregating the features of neighboring nodes and mixing them with its own features. The node feature update rule of the k-th layer GIN is as follows:

[0069]

[0070] Perform a global average pooling operation on all node features to obtain the global graph representation h of the entire molecule g :

[0071]

[0072] in represents the feature vector of node v at layer k, Represents the set of neighbor nodes of node v, MLP (k)represents the k-th layer of the multilayer perceptron (MLP), ∈ is a learnable or fixed scalar, K represents the number of layers of GIN, V represents the set of nodes, and global_mean_pool represents the global average pooling operation.

[0073] 2-3-4) Input the global graph representation into the fully connected layer to obtain the molecular graph representation G of molecule i i .

[0074] In step 2), the attention mechanism is used to process the multimodal features obtained in the above steps, which specifically includes the following steps:

[0075] 2-4-1) Characterization of S for three different molecules i (SMILES characterization), F i (molecular fingerprint characterization), G i (Molecular graph representation), the attention score is calculated through a tanh activation function feed-forward neural network:

[0076] e s =tanh(W s S i +b s ) (6)

[0077] e f =tanh(W f F i +b f ) (7)

[0078] e g =tanh(W g G i +b g ) (8)

[0079] where e s 、e f 、e g , W s , W f , W g and b s 、b f 、b g They are the attention score, weight matrix, and bias vector of SMILES representation, molecular fingerprint representation, and molecular graph representation;

[0080] 2-4-2) The attention weight of each representation is calculated by normalizing the attention score, and then the attention weight is applied to the corresponding molecule representation to obtain the representation M of molecule i through weighted fusion. i :

[0081]

[0082] The optimization model is trained by calculating the classification loss and the contrast loss obtained in steps 2-5). In step 3), the molecular graph contrast learning strategy specifically includes the following steps:

[0083] 3-1) For a given molecular graph, node embeddings are obtained from node attributes through GIN, and these embeddings are used to predict the enhancement method for the node, that is, to predict whether the node should be discarded, retained or masked. Then, the Gumbel-Softmax technique is used to sample nodes from the predicted distribution to obtain the molecular graph after data augmentation;

[0084] 3-2) Randomly select a batch of molecular graphs of size N, perform data augmentation through the above steps, and generate an equal number of corresponding augmented molecular graphs. The given molecular graph representation G i Augmented Molecular Graph Representation A positive sample pair is formed, and the representation of the remaining 2 (N-1) images / augmented images constitute a negative sample pair. NT-XEnt is used to calculate the loss of contrastive learning training to maximize the consistency between the representations of positive sample pairs and minimize the consistency between the representations of negative sample pairs. The contrastive learning loss is expressed as:

[0085]

[0086] Where τ is the temperature coefficient, which is used to adjust the scale of similarity and is set to 0.2.1 [k≠i] is an indicator function, which is 1 when k≠i and 0 otherwise. Represents G i and The cosine similarity between ;

[0087] 3-3) For the sample pairs of the input batch, the contrast loss It is calculated by averaging the contrast loss of all positive sample pairs in the batch:

[0088]

[0089] in and is the contrast loss between a pair of molecular graphs and the molecular graph after data augmentation, and N is the total number of molecules;

[0090] In step 4), calculating the model loss function specifically includes the following steps:

[0091] 4-1) Through the fully connected layer, the fused molecular representation M i Further processing is performed to obtain the final molecular representation y for the classification task i . Then the binary cross entropy loss (BCE loss) is calculated as the classification loss The calculation formula is:

[0092]

[0093] Where N is the total number of molecules, g i is the actual label of the ith molecule, is the probability that the model predicts that the i-th molecule is a positive class;

[0094] 4-2) Final total loss function It is defined as the weighted sum of classification loss and contrastive learning loss:

[0095]

[0096] Among them, β is an adjustable weight parameter.

[0097] In step 6), the interpretability analysis specifically includes the following steps:

[0098] 6-1) For the molecular SMILES sequence representation, the attention weight visualization technology is used to analyze the attention weights assigned to each SMILES character by the model during the prediction process to identify the key factors affecting the prediction;

[0099] 6-2) For molecular fingerprint characterization, the integral gradient method is used to identify important molecular fingerprint sites by analyzing the gradient on the path between the model input and the reference input. The higher the integral gradient, the greater the role played by the site;

[0100] 6-3) Based on the molecular graph features, the Gradient Weighted Class Activation Mapping (Grad-CAM) algorithm is applied to use the gradient information of the convolutional layer to highlight the areas in the graph that contribute most to the model prediction. By visualizing the key atoms and chemical bonds that the model focuses on, the association between these areas and drug activity can be further explored, providing guidance for drug design and optimization.

[0101] Compared with the prior art, the present invention has the following advantages:

[0102] (1) The multimodal representation learning strategy designed by the present invention integrates the molecular SMILES sequence features, molecular fingerprint features and molecular graph features, which can capture multiple dimensional information of the molecular structure and improve the accuracy and comprehensiveness of the model in identifying P-gp inhibitors and substrates. In contrast, existing single-modality or simple combination technical solutions are often difficult to fully reflect the complex nature of molecules.

[0103] (2) The present invention introduces an attention mechanism for multimodal feature fusion, which enables the model to automatically learn and strengthen the importance of key features while suppressing non-critical information, thereby improving the efficiency and accuracy of multimodal feature extraction and fusion.

[0104] (3) The present invention utilizes an automatic graph contrast learning strategy to generate a molecular augmented graph, which is used to maximize the similarity between representations of the same molecule and minimize the similarity between representations of different molecules. This not only enriches the diversity of training data, but also enhances the generalization ability of the model.

[0105] (4) The high-quality data set constructed by the present invention and the standardized data processing process adopted ensure the consistency and reliability of the data, providing a solid foundation for model training. This rigorous data preparation method is the key to improving model performance.

[0106] (5) The present invention conducts interpretable analysis on the three features and accurately identifies the key functional groups closely related to P-gp interaction, providing scientific and reliable information for researchers to optimize lead compounds and effectively improving the efficiency of drug research and development. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] Figure 1 Flowchart of the prediction scheme for P-glycoprotein inhibitors and substrates based on multimodal comparative learning.

[0108] Figure 2 Prediction model structure for P-glycoprotein inhibitors and substrates based on multimodal contrastive learning.

[0109] Figure 3 To evaluate the performance of the model in identifying novel substructures, A and D are the visualization results of principal component analysis after K-means clustering of the P-gp inhibitor and substrate datasets; B and E are the ROC-AUC comparisons of the five clusters measured on the P-gp inhibitor and substrate datasets; C and F are the Accuracy comparisons of the five clusters measured on the P-gp inhibitor and substrate datasets.

[0110] Figure 4 These are the 10 most important molecular fingerprints for the prediction of P-gp inhibitors and substrates in the interpretability analysis, A is for the prediction of P-gp inhibitors, and B is for the prediction of P-gp substrates.

[0111] Figure 5 The interpretability case study results of the model for identifying key functional groups of inhibitors, non-inhibitors, substrates, and non-substrate molecules based on (A) SMILES sequences and (B) molecular graph structures. The size and depth of the colored area represent the contribution to the prediction results. The darker the color, the greater the contribution. The circular dotted area represents the functional group with positive contribution, and the circular solid area represents the functional group with negative contribution. DETAILED DESCRIPTION

[0112] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention, but it does not limit the scope of the present invention. Any equivalent substitution, improvement or adjustment of the various materials, methods or structures mentioned in the present invention by researchers in this field, as long as it does not essentially deviate from the core ideas and technical features of the present invention, shall fall within the scope of protection of the present invention.

[0113] like Figure 1 As shown, the P-glycoprotein inhibitor and substrate prediction scheme based on multimodal contrastive learning specifically includes the following steps:

[0114] Step 1: Use text mining and manual verification methods to collect chemical or biological experimental data related to P-gp from databases and literature, integrate the collected data, and use standardized processes for data preprocessing to construct a high-quality P-gp inhibitor and substrate dataset.

[0115] The specific requirements for data collection and preprocessing in step 1 are as follows:

[0116] (1) Data screening. For experimental data, keep the in vitro cell experiment related data, including the inhibition rate (inhibition%) at a given concentration in the P-gp inhibitor assay and the half-maximal inhibitory concentration (IC) of the measured antagonist. 50 ), and the efflux ratio (ER) of the compound in the P-gp substrate assay.

[0117] (2) Data annotation. The data were annotated according to the specified threshold of the biological activity record, and the dataset was divided into two categories. Specifically, in the P-gp inhibitor dataset, if inhibition%>25% or IC 50 <10μM, the compound is marked as a P-gp inhibitor; if inhibition%<10% or IC 50 If the ER is ≥10 μM, the compound is marked as a non-P-gp inhibitor. In the P-gp substrate data set, if the ER is ≥2, the compound is marked as a P-gp substrate; if the ER is <2, it is marked as a non-P-gp substrate.

[0118] (3) Data preprocessing. To ensure the quality and uniqueness of the data set, the data was preprocessed through the following steps: ① Standardize the SMILES sequences of all molecules; ② Remove inorganic substances and complexes; ③ Convert salts into corresponding acids or bases; ④ For stereoisomers, only one data is retained; ⑤ Filter duplicate molecular structures. Use the RDkit tool for data preprocessing.

[0119] (4) Constructing an external test set. In order to evaluate the generalization ability of the model, an external test set needs to be constructed. In the present invention, compound data from related studies and databases are additionally collected as an external test set, the above data preprocessing steps are applied to the external test set, and the compounds and corresponding biological activity data that are repeated in the external test set and the data set are deleted.

[0120] In step 1, through comprehensive data collection and standardized data preprocessing, the P-gp inhibitor data set and the P-gp substrate data set were finally obtained. The P-gp inhibitor data set contains 5943 molecules, of which 1385 molecules (23.3%) are classified as non-inhibitors and 4558 molecules (76.7%) are classified as inhibitors; the P-gp substrate data set contains 4018 molecules, of which 1563 molecules (38.9%) are classified as non-substrates and 2455 molecules (61.1%) are classified as substrates. The P-gp inhibitor external test set contains 140 molecules, and the P-gp substrate external test set contains 185 molecules. The scale of the data set constructed by the present invention has been significantly improved compared with previous studies, and an external test set is provided, which can more comprehensively evaluate the biological activity of P-gp-related compounds.

[0121] Step 2: Construct a P-glycoprotein inhibitor and substrate prediction model based on multimodal contrastive learning. The model architecture includes a molecular SMILES sequence feature extraction module, a molecular fingerprint feature extraction module, a molecular graph feature extraction module, a feature fusion module based on the attention mechanism, and a predictor based on a fully connected layer. The molecular graph feature extraction module adopts a graph contrastive learning strategy to enhance representation learning.

[0122] like Figure 2 As shown in the figure, the specific process of the molecular SMILES sequence feature extraction module is as follows:

[0123] (1) Divide the given SMILES sequence into individual atoms or symbols, and then look up the corresponding atom’s embedding vector from the pre-trained dictionary. If it is not in the dictionary, generate a random embedding vector; the pre-trained dictionary contains the embedding vectors of common atoms;

[0124] (2) Aggregate the embedding vectors to obtain the embedding matrix Where m is the number of atoms and d is the dimension of the embedding vector;

[0125] (3) Through the above steps, we can obtain the feature vector X of the SMILES sequence = {x1, x2, …, x n},in is the feature vector of the tth atom. The SMILES context information is captured by the bidirectional long short-term memory network Bi-LSTM to obtain the hidden state h of atom t t :

[0126]

[0127] in and Respectively represent the forward and backward LSTM units, h t-1 is the hidden state of atom t-1;

[0128] (4) The attention mechanism is used to further improve the model's ability to capture important sequence information. The following formula is used to calculate h t The attention weights are:

[0129] e t =tanh(W1h t +b1) (2)

[0130]

[0131] where e t is the attention score of atom t, which is calculated by the tanh activation function feedforward neural network, W1 is the weight matrix, b1 is the bias vector; the attention score is normalized by the Softmax function to obtain the attention weight α of atom t t ;

[0132] (5) The hidden state of each atom is weighted and summed according to the attention weight to obtain the feature vector of the molecular SMILES sequence The feature vector h s Pass it into the fully connected layer to obtain the SMILES sequence feature S of molecule i i .

[0133] like Figure 2 As shown in the figure, the molecular fingerprint feature extraction module first calculates the ECFP4 fingerprint and PubChem fingerprint of the molecule, connects the two fingerprints to obtain a mixed fingerprint, and then passes the mixed fingerprint into the two fully connected layers to obtain the molecular fingerprint feature F of molecule i. i .

[0134] like Figure 2 As shown, the molecular graph feature extraction module specifically includes:

[0135] (1) Extract node and edge information based on molecular SMILES, where each atom in the molecule is mapped as a node in the molecular graph, and the chemical bonds between atoms are mapped as edges in the molecular graph. The adjacency matrix of the graph is constructed using the node and edge information to represent the connection relationship between nodes.

[0136] (2) Calculate the physical and chemical properties of each atom as the initial feature vector of the point in the molecular graph; the physical and chemical properties include atomic type, charge number, chiral carbon situation, number of connected hydrogen atoms, hybridization type, whether it is an aromatic atom, etc.;

[0137] (3) Using the graph isomorphic network GIN, the node representation is updated by aggregating the features of neighboring nodes and mixing them with its own features. The node feature update rule of the k-th layer GIN is as follows:

[0138]

[0139] Perform a global average pooling operation on all node features to obtain the global graph representation h of the entire molecule g :

[0140]

[0141] in represents the feature vector of node v at layer k, Represents the set of neighbor nodes of node v, MLP (k) represents the k-th layer of the multilayer perceptron (MLP), ∈ is a learnable or fixed scalar, K represents the number of layers of GIN, V represents the set of nodes, and global_mean_pool represents the global average pooling operation.

[0142] (4) Input the global graph representation into the fully connected layer to obtain the molecular graph representation G of molecule i i .

[0143] like Figure 2 As shown in the figure, the multimodal feature fusion module specifically includes:

[0144] (1) Characterization of S for three different molecules i (SMILES characterization), F i (molecular fingerprint characterization), G i (Molecular graph representation), the attention score is calculated through a tanh activation function feed-forward neural network:

[0145] e s =tanh(W s S i +b s ) (6)

[0146] e f =tanh(W f F i +b f ) (7)

[0147] e g =tanh(W g Gi +b g ) (8)

[0148] where e s 、e f 、e g , W s , W f , W g and b s , b f , b g They are the attention score, weight matrix, and bias vector of SMILES representation, molecular fingerprint representation, and molecular graph representation;

[0149] (2) The attention weight of each representation is calculated by normalizing the attention score, and then the attention weight is applied to the corresponding molecule representation to obtain the representation M of molecule i through weighted fusion. i :

[0150]

[0151] Step 3: During the training process, a graph contrast learning strategy is introduced for the molecular graph features to obtain the molecular graph after data augmentation, and the contrast loss between the molecular graph and the augmented graph is calculated.

[0152] like Figure 2 As shown in Figure 2, the molecular graph comparison learning strategy specifically includes:

[0153] (1) For a given molecular graph, node embeddings are obtained from node attributes through GIN, and these embeddings are used to predict the enhancement method for the node, that is, to predict whether the node should be discarded, retained, or masked. Then, the Gumbel-Softmax technique is used to sample nodes from the predicted distribution to obtain the molecular graph after data augmentation;

[0154] (2) Randomly select a batch of molecular graphs of size N, perform data augmentation through the above steps, and generate an equal number of corresponding augmented molecular graphs. The given molecular graph representation G i Augmented Molecular Graph Representation A positive sample pair is formed, and the representation of the remaining 2 (N-1) images / augmented images constitute a negative sample pair. NT-XEnt is used to calculate the loss of contrastive learning training to maximize the consistency between the representations of positive sample pairs and minimize the consistency between the representations of negative sample pairs. The contrastive learning loss is expressed as:

[0155]

[0156] Where τ is the temperature coefficient, which is used to adjust the scale of similarity and is set to 0.2.1 [k≠i] is an indicator function, which is 1 when k≠i and 0 otherwise. Represents G i and The cosine similarity between ;

[0157] (3) For the sample pairs in the input batch, the total contrast loss It is calculated by averaging the contrast loss of all positive sample pairs in the batch:

[0158]

[0159] in and is the contrast loss between a pair of molecular graphs and the molecular graph after data augmentation, and N is the total number of molecules;

[0160] Step 4: Using the data set obtained in step 1, and by calculating the classification loss of the predictor in step 2 and the contrast loss in step 3, the parameters of the P-glycoprotein inhibitor and substrate prediction model are trained and optimized to obtain the optimized prediction model.

[0161] like Figure 2 As shown, the calculation of the total loss function of the model specifically includes:

[0162] (1) Through the fully connected layer, the fused molecular representation M i Further processing is performed to obtain the final molecular representation y for the classification task i . Then the binary cross entropy loss (BCE loss) is calculated as the classification loss The calculation formula is:

[0163]

[0164] Where N is the total number of molecules, g i is the actual label of the ith molecule, is the probability that the model predicts that the i-th molecule is a positive class;

[0165] (2) Final total loss function It is defined as the weighted sum of classification loss and contrastive learning loss:

[0166]

[0167] Among them, β is an adjustable weight parameter.

[0168] The data set is randomly divided and skeletonized according to 8:1:1 to obtain training set, validation set and test set. The divided data set is input into the prediction model, and the Adam method is used to perform back propagation optimization to update the model parameters of the target domain model to minimize the loss function. The optimal network parameters are determined based on the optimal results on the validation set, and the test set data is tested.

[0169] ROC-AUC was selected as the evaluation index for hyperparameter optimization. The optimization of hyperparameters was completed by using the Optuna tool, which is based on the Bayesian optimization algorithm. According to the existing hyperparameter combinations and results, the posterior probability part of the existing results was calculated through Gaussian process regression to obtain the expected mean and variance of the hyperparameters at each possible value, and comprehensively judge the most suitable combination of hyperparameters. 30 iterations of optimization were performed on the validation set to accurately adjust the hyperparameters of the model, as shown in Table 1:

[0170] Table 1 Model hyperparameter optimization results

[0171]

[0172]

[0173] Step 5: Input the data of the compound to be predicted into the optimized P-glycoprotein inhibitor and substrate prediction model, output the final classification result, and determine whether the compound is a P-glycoprotein inhibitor and substrate.

[0174] The saved optimal model was used for the test set test, and the test evaluation indicators used in the present invention were ROC-AUC, accuracy, F1 score, and Matthews correlation coefficient (MCC). Compared with the baseline model, the prediction accuracy of the P-gp inhibitor and substrate of the present invention is better than that of the baseline model, as shown in Table 2 and Table 3:

[0175] Table 2 Performance comparison of the models under random partitioning on the P-gp inhibitor and substrate test sets

[0176]

[0177]

[0178] Table 3 Performance comparison of the models under skeleton partitioning in the P-gp inhibitor and substrate test sets

[0179]

[0180] MC-PGP is the method of the present invention.

[0181] In order to fully verify the generalization ability of the method in the present invention, the model was tested on an external test set, and the results were compared with the baseline model, as shown in Tables 4 and 5. MC-PGP achieved the best performance in all evaluation indicators, significantly surpassing other comparison methods. It is worth noting that MC-PGP showed a greater performance improvement on the external test set compared to the performance on the internal test set. In particular, the MC-PGP model showed a significant advantage in the MCC indicator, which emphasized its superiority in handling imbalanced data sets. In contrast, other baseline methods, especially those based on traditional machine learning, usually showed performance degradation on external test sets.

[0182] Table 4 Performance comparison of models on the P-gp inhibitor external test set

[0183]

[0184] Table 5 Performance comparison of models on the P-gp substrate test set

[0185]

[0186] To evaluate the performance of the model in identifying new analytical structures, ECFP fingerprinting was performed on the molecules in the dataset and the Tanimoto coefficient was calculated to determine the molecular similarity. K-means clustering was used to divide the molecules into five clusters based on their structural properties, one of which was designated as the test set in each iteration and the other clusters as the training set. PCA visualization was performed, such as Figure 3 As shown in A and D, we can see that the dataset is divided into five different compound clusters. The model performance of different clusters as test sets is evaluated, as shown in Figure 3 As shown in Figures B, C, E, and F, the results show that the ROC-AUC and Accuracy indicators of the MC-PGP model are significantly better than those of other models. This confirms the effectiveness of the model in the present invention in identifying new molecular structures and its adaptability to different structural changes.

[0187] Step 6: Perform interpretability analysis on the prediction model to identify key functional groups in the compound that are closely related to P-gp interaction. The interpretability analysis specifically includes the following steps:

[0188] For molecular fingerprint characterization, the integrated gradient method was applied to obtain significant and contributing molecular fingerprints.To obtain more comprehensive statistics related to the chemical groups that affect the interaction of compounds with P-gp, we performed interpretability analysis on the corresponding test sets obtained by randomly splitting the two datasets. Figure 4 The ten most important molecular fingerprints in the prediction of P-gp inhibitors and substrates determined by the integrated gradient method are listed, and the corresponding substructures and importance scores are given.

[0189] like Figure 4 As shown in A, among the top ten molecular fingerprints predicted for P-gp inhibitors, most of the substructures indicated by key fingerprints have aromatic ring structures with high lipophilicity. Studies have shown that lipophilicity is closely related to the ability of P-gp inhibitors to pass through the phospholipid bilayer. In addition, most of the substructures indicated by key fingerprints have basic tertiary nitrogen atoms and hydrogen bond donor or acceptor properties, which help inhibitors effectively bind to key amino acid residues and P-gp through hydrogen bonds and other forces.

[0190] As shown in Figure 4B, in the prediction of P-gp substrates, the key substructures indicated by molecular fingerprints include hydrogen bond donors, basic groups with cations, and polar groups. Studies have shown that these structures help form hydrogen bonds with the substrate binding region of P-gp, thereby increasing its possibility of being a substrate. In addition, Figure 4 Most of the substructures in B are flexible. Compounds with these flexible segments can easily adjust their conformations and are more likely to form favorable interactions with P-gp.

[0191] Furthermore, we selected four representative inhibitors, non-inhibitors, substrates, and non-substrate molecules, and used attention weight visualization and Grad-CAM methods to perform interpretability analysis on SMILES and molecular graph features, respectively. Figure 5 As shown. The visualization results of SMILES and molecular graph features show significant consistency, and the two representation forms can effectively complement information and enhance the interpretability of the model. We found that tertiary nitrogen atoms have positive contributions to P-gp inhibitor predictions, nitro groups and seven-membered heterocycles have negative contributions in non-inhibitor predictions, and amino groups have positive contributions in P-gp substrate predictions, and peptide bonds have negative contributions in non-substrate predictions. These results are consistent with the previous results of molecular fingerprint interpretability analysis. In addition, we can speculate through the results of interpretability analysis, for example, the positive contribution of tert-butyl groups to non-substrate predictions may be due to its steric hindrance effect, which may affect the binding of compounds to P-gp.

[0192] Through in-depth interpretability analysis, we not only confirmed that key molecular features are essential for the prediction of P-gp inhibitors and substrates, but also revealed the mechanistic effects of these features in different prediction tasks. These findings provide valuable insights for further optimizing drug design and prediction models.

Claims

1. A method for predicting P-glycoprotein inhibitors and substrates based on multimodal contrastive learning, characterized in that: The following steps are involved: 1) Construct a dataset of P-glycoprotein inhibitors and substrates; 2) Use deep learning methods to construct prediction models for P-glycoprotein inhibitors and substrates; 3) inputting the data set obtained in step 1) into the P-glycoprotein inhibitor and substrate prediction model constructed in step 2), introducing a graph contrast learning strategy for the molecular graph features to obtain a molecular graph after data augmentation, calculating a contrast loss between the molecular graph and the molecular graph after data augmentation, calculating a classification loss based on the fusion features output by the P-glycoprotein inhibitor and substrate predictor, combining the contrast loss and the classification loss to obtain a total loss function, training and optimizing the parameters of the P-glycoprotein inhibitor and substrate prediction model through the total loss function, and obtaining an optimized P-glycoprotein inhibitor and substrate prediction model; 4) Inputting the data of the compound to be predicted into the optimized P-glycoprotein inhibitor and substrate prediction model, outputting the final classification result, and judging whether the compound is a P-glycoprotein inhibitor and substrate.

2. The method for predicting P-glycoprotein inhibitors and substrates based on multimodal contrastive learning according to claim 1, characterized in that: In step 1), a data set of P-glycoprotein inhibitors and substrates is constructed, specifically including: Experimental data and compound SMILES sequences related to molecular compounds and P-glycoprotein inhibitors and substrates were collected from relevant databases, integrated and preprocessed to obtain data sets of P-glycoprotein inhibitors and substrates.

3. The method for predicting P-glycoprotein inhibitors and substrates based on multimodal contrastive learning according to claim 1, characterized in that: In step 2), the P-glycoprotein inhibitor and substrate prediction model specifically includes: A molecular SMILES sequence feature extraction module, used to receive data sets of P-glycoprotein inhibitors and substrates and extract molecular SMILES sequence features; A molecular fingerprint feature extraction module, used for receiving data sets of P-glycoprotein inhibitors and substrates and generating molecular fingerprint features; A molecular graph feature extraction module, for receiving a data set of P-glycoprotein inhibitors and substrates and generating molecular graph features; A feature fusion module constructed based on the attention mechanism is used to receive the molecular SMILES sequence features output by the molecular SMILES sequence feature extraction module, the molecular fingerprint features output by the molecular fingerprint feature extraction module, and the molecular graph features output by the molecular graph feature extraction module, and fuse the molecular SMILES sequence features, the molecular fingerprint features and the molecular graph features to output a fusion feature; P-glycoprotein inhibitor and substrate predictor, including a fully connected layer, for receiving fused features and predicting output classification results.

4. The method for predicting P-glycoprotein inhibitors and substrates based on multimodal contrastive learning according to claim 3, characterized in that: In step 2), the molecular SMILES sequence features are extracted, and the specific steps include: 2-1-1) Divide the SMILES sequence in the dataset of P-glycoprotein inhibitors and substrates into individual atoms or symbols, then look up the corresponding atom's embedding vector from the pre-trained dictionary, and if it is not in the dictionary, generate a random embedding vector; aggregate the atom's embedding vector and the random embedding vector to obtain a preliminary representation of the SMILES sequence; 2-1-2) Preprocess the preliminary representation of the SMILES sequence through a bidirectional long short-term memory network Bi-LSTM to capture the SMILES context information representation; 2-1-3) The attention weight of each atom in the SMILES context information representation is calculated through the attention mechanism, and the hidden state of each atom is weighted and summed according to the attention weight to obtain the feature vector of the molecular SMILES sequence; 2-1-4) The feature vector of the molecule SMILES sequence is passed into the fully connected layer to obtain the SMILES sequence features of the molecule.

5. The method for predicting P-glycoprotein inhibitors and substrates based on multimodal contrastive learning according to claim 3, characterized in that: In step 2), the molecular fingerprint feature is generated, which specifically includes: The circular topological fingerprints ECFP4 fingerprints and PubChem fingerprints of the molecules in the datasets of P-glycoprotein inhibitors and substrates were calculated respectively, the two fingerprints were connected to obtain a hybrid fingerprint, and the hybrid fingerprint was passed to the fully connected layer to obtain the molecular fingerprint characteristics of the molecule.

6. The method for predicting P-glycoprotein inhibitors and substrates based on multimodal contrastive learning according to claim 3, characterized in that: In step 2), the molecular graph features are generated, specifically including: 2-3-1) Extract node and edge information based on molecular SMILES in the dataset of P-glycoprotein inhibitors and substrates, and use the node and edge information to construct the adjacency matrix of the graph; 2-3-2) Calculate the physicochemical properties of each atom in the data set of P-glycoprotein inhibitors and substrates as the initial feature vectors of the points in the molecular graph; 2-3-3) Using the graph isomorphic network GIN, the initial feature vector obtained in step 2-3-2) is updated by mixing the node's own features and the features of neighboring nodes in the adjacency matrix of the aggregate graph. The global graph representation of the entire molecule is obtained by performing a global average pooling operation on the features of all nodes; 2-3-4) Input the global graph representation into the fully connected layer to obtain the molecular graph feature G of molecule i i .

7. The method for predicting P-glycoprotein inhibitors and substrates based on multimodal contrastive learning according to claim 3, characterized in that: In step 3), the total loss function specifically includes: The molecular representation M of the fusion feature is processed through the fully connected layer to obtain the final molecular representation y for the classification task i , and then calculate the binary cross entropy loss as the classification loss The calculation formula is: Where N is the total number of molecules, y i is the actual label of the ith molecule, is the probability that the model predicts that the i-th molecule is a positive class; A graph contrast learning strategy is introduced for the molecular graph features to obtain the molecular graph after data augmentation. The NT-XEnt loss between the molecular graph and the molecular graph after data augmentation is calculated as the contrast learning loss. The final total loss function It is defined as the weighted sum of classification loss and contrastive learning loss: Among them, β is an adjustable weight parameter.

Images