Drug-target binding affinity prediction model training method and prediction method based on contrastive learning

Abstract

The application discloses a drug-target binding affinity prediction model training method and a prediction method based on contrast learning, and belongs to the technical field of drug-target binding affinity prediction. In order to solve the problem that the prediction effect of the existing DTA prediction based on contrast learning needs to be improved, the application selects positive and negative samples according to the similarity between drug molecular structures, the similarity between target sequences, and the similarity between drugs and targets in an affinity graph to optimize the effect of contrast learning, extracts drug-target features of three different scales of sequences, molecular structures and affinity graphs, fully utilizes drug and target information, captures potential relationships between cross-scale feature information through a multi-scale feature contrast learning framework, maximizes mutual information between different scales, and performs feature alignment and feature fusion, and then obtains a prediction model based on contrast loss, and the prediction model is used to realize drug-target binding affinity prediction.

Description

Technical Field

[0001] This invention belongs to the field of drug-target binding affinity prediction technology, and relates to a training method and prediction method for a drug-target binding affinity prediction model. Background Technology

[0002] Existing methods for predicting drug-target (protein) binding affinity (DTA) can be broadly categorized into structure-based and non-structure-based paradigms. Structure-based methods utilize three-dimensional structural data of drug-target complexes. They typically employ advanced architectures, such as CNNs and FCNNs, to extract hierarchical features and model molecular interactions, thereby improving prediction performance. However, obtaining the atomic-level three-dimensional structures of drug molecules and target proteins remains a significant technical challenge.

[0003] Non-structure-based deep learning methods can be mainly divided into: sequence-based methods, graph-based methods, and multi-scale feature fusion methods. Sequence-based methods take drug SMILES strings and target sequences as input, and then use deep learning methods such as CNN, Transformer, and RNN to extract features from the SMILES and target sequence data, which can automatically capture local patterns, conserved structural domains, and global dependencies in the sequence. After pooling, the sequence features of the drug and target are combined and input into a fully connected layer or used for subsequent affinity prediction.

[0004] Graph-based methods treat molecular or protein structures as graphs, where nodes represent atoms or amino acid residues, and edges represent chemical bonds or spatial adjacency relationships. Typically, tools such as RDKit and Pconsc4 are used to convert drug samples and target sequences into molecular graphs. Then, deep learning methods such as GCN, GAT, or other GNN architectures are used for feature extraction. These methods can fuse the internal topological structure of molecules and the chemical properties of atoms, capturing the complex spatial relationships of molecules and proteins to obtain more comprehensive feature representations of drugs and proteins. After pooling or readout functions, the drug / target graph features are concatenated or interoperated before being input into the affinity prediction module.

[0005] Multi-scale feature fusion methods first require obtaining diverse information about the drug and target, as well as their interaction networks, such as sequence information, molecular structure information (molecular maps / protein maps), and association network information. Then, different deep networks (e.g., CNNs for sequences, GCNs for structures) are used to obtain corresponding feature embeddings. Next, the dimensionality of feature vectors from different sources is unified. Then, attention mechanisms, concatenation fusion, and weighted fusion are used to fuse the drug and target features obtained at different scales, maximizing the learning of richer feature representations of the drug and target. Finally, the fused multimodal representation is input into the affinity prediction main module (e.g., MLP) to output a predicted score.

[0006] The aforementioned methods have all achieved certain successes, but some shortcomings remain. Sequence-based methods are simple in structure, easy to operate and extend, but they struggle to reflect the three-dimensional space or topological structure between molecules, lack sufficient understanding of high-level structure-related interactions, and are limited by the sequence information itself. Graph-based methods can reflect complex molecular structures and spatial relationships, but the algorithms and implementations are more complex, computationally intensive, and the accuracy of the molecular / protein three-dimensional structure directly affects performance. Multi-scale feature fusion methods can more comprehensively express complex molecular properties, have strong generalization, and better prediction results, but they still face problems such as insufficient and inaccurate characterization of drug and target information and feature redundancy due to improper fusion strategies. Summary of the Invention

[0007] This invention aims to address the problem that the prediction performance of existing contrastive learning-based DTA prediction models needs improvement.

[0008] A method for training a drug-target binding affinity prediction model based on contrastive learning is proposed. For a network model predicting drug-target binding affinity based on sequence data of drugs and targets, as well as associated affinity data, the model is trained using contrastive learning. During the contrastive learning process, positive and negative samples are determined as follows:

[0009] Calculate the similarity between drug molecular structures to obtain similarity Calculate the similarity between target sequences to obtain similarity. The similarity between drugs is calculated based on the number of targets they share in common at the affinity graph scale. The similarity between targets is calculated based on the number of drugs commonly connected to each other in the affinity graph. ;

[0010] Will and Add them together, and Add them together, and then sort the sums in descending order. For each drug and target, take the first K samples as positive samples and the rest as negative samples.

[0011] Then, a contrastive learning loss is calculated based on positive and negative samples and used for model training.

[0012] Furthermore, the drug-target binding affinity prediction model processing procedure is as follows:

[0013] For drug SMILES string data, a drug feature extraction network is used to obtain drug sequence feature representations. For the target amino acid sequence, a target amino acid feature extraction network is used to obtain the target sequence features. ;

[0014] For drug molecular structure diagrams obtained by converting drug SMILES string data, a graph convolutional neural network is used to obtain drug structural feature representations. For the target molecular structure map obtained by converting target amino acid sequence data, a graph convolutional neural network is used to obtain the target structural feature representation. ;

[0015] For the drug-target affinity map corresponding to drug-target affinity information, a graph convolutional neural network is used to extract features, and drug affinity map features are obtained based on the extracted features. Target affinity map features .

[0016] Furthermore, the process of calculating the contrastive learning loss based on positive and negative samples includes:

[0017] Multiscale contrastive learning was used to study the multiscale features of drugs. , , Perform pairwise comparisons and calculate drug values. Contrastive learning loss:

[0018]

[0019] in, Indicates drug The positive sample set; M is the number of all drug molecules; It is a temperature parameter that can be changed; Represents the cosine similarity function; That is, medicine corresponding , That is, medicine corresponding , That is, medicine corresponding ;

[0020] Multi-scale contrastive learning is used to analyze the multi-scale features of the target. , , Perform the same processing to obtain the target. Contrastive learning loss ;

[0021] Will As the overall contrastive learning loss.

[0022] Furthermore, the drug feature extraction network and the target amino acid feature extraction network are both Bi-LSTM.

[0023] Furthermore, a graph convolutional neural network is used to obtain the drug structure feature representation. The process includes:

[0024] First, the GCN network is used to learn the feature representations of drugs and targets from the drug molecule structure:

[0025]

[0026] in, This represents the hidden state of node v at level l. , It is the neighborhood set of node v; This represents the degree of node v in the graph. ; This represents the weight matrix of the l-th layer;

[0027] After processing by the GCN network, the features of each drug molecule are obtained through the readout function. .

[0028] Furthermore, a graph convolutional neural network is used to obtain the target structural feature representation. The process of obtaining drug structure feature representation using graph convolutional neural networks The process is the same.

[0029] Furthermore, the similarity between drugs is calculated based on the number of targets commonly connected between drugs at the affinity graph scale. During the process, utilizing Calculate drugs With drugs The similarity between them For drugs The number of edges in an affinity graph. For drugs With drugs The number of targets that are connected together.

[0030] Furthermore, the similarity between targets is calculated based on the number of drugs commonly connected between targets at the affinity graph scale. The method and calculation of drug similarity The method is the same.

[0031] Furthermore, the contrastive learning loss is calculated based on positive and negative samples. The process for training the model includes:

[0032] Predicting drug target binding affinity using MLP: , , Fusion is performed to obtain the final drug embedding feature. ;right , , Fusion is performed to obtain the final target embedding features. The final features of the fused drug target are input into the MLP for prediction. To obtain the predicted loss ;

[0033] Based on predicted loss Comparison of learning loss with total Calculate total loss Utilizing total losses The trained network model is obtained for predicting drug-target binding affinity.

[0034] A drug-target binding affinity prediction method based on contrastive learning is proposed, which uses a network model for drug-target binding affinity prediction obtained by the aforementioned contrastive learning-based drug-target binding affinity prediction model training method to predict drug-target binding affinity.

[0035] Beneficial effects:

[0036] This invention innovatively proposes a contrastive learning framework for multi-scale features to capture the latent relationships between cross-scale feature information. During the contrastive learning process, the representation of positive and negative samples is optimized, and mutual information between different scales is maximized. Furthermore, feature alignment and feature fusion are effectively performed. Therefore, this invention can fully utilize multi-scale feature information and improve feature fusion performance, thereby enhancing prediction accuracy. Compared with known methods, the method of this invention achieves state-of-the-art results across all evaluation metrics on two benchmark datasets. Specifically, on the Davis dataset, compared to the best performance baseline, this invention reduces MSE by 15.1% and improves CI by 1%. This represents a 4.1% improvement; compared to the best performance baseline on the KIBA dataset, this invention reduces MSE by 5.5% and improves CI by 0.5%. It increased by 0.7%. Attached Figure Description

[0037] Figure 1 A flowchart for training and predicting drug-target binding affinity prediction models.

[0038] Figure 2 A scatter plot of the predicted performance on the Davis and KIBA datasets.

[0039] Figure 3 This is a comparison of ablation experiments performed on the Davis dataset.

[0040] Figure 4 For visualization and analysis. Detailed Implementation

[0041] On the one hand, existing contrastive learning-based DTA predictions suffer from the influence of positive and negative sample selection on the training of the prediction model, thus hindering its improvement. On the other hand, current methods still face challenges in multi-scale feature information fusion, including difficulties in feature alignment and poor fusion results, leading to inaccurate predictions. Furthermore, current methods often only utilize two scales for drug target feature information, underutilizing known drug target information and resulting in incomplete learning of drug and target features. This inadequate utilization of multi-scale feature information also contributes to inaccurate predictions.

[0042] To address the aforementioned issues, this invention selects positive and negative samples during contrastive learning based on the similarity between drug molecular structures, the similarity between target sequences, and the similarity between drugs and targets at the affinity map scale. This optimizes the contrastive learning effect. Furthermore, this invention is the first to integrate data from sequence, molecular structure, and affinity maps. Figure 3 This approach utilizes drug target features at different scales, fully leveraging known drug target information. A multi-scale feature contrastive learning framework is employed to capture the potential relationships between cross-scale features, maximizing mutual information across different scales and effectively performing feature alignment and fusion. Specific implementation details are provided below.

[0043] Specific implementation method one: Combining Figure 1 This implementation method is described below.

[0044] This implementation method is a training method and prediction method for a drug-target binding affinity prediction model based on contrastive learning, including the following steps:

[0045] S1. Collect and download sequence data of drugs and targets, as well as associated affinity data;

[0046] The invention was validated using two publicly available drug target datasets, Davis and KIBA, as benchmark datasets. The main data collected from these two datasets included:

[0047] ① Drug SMILES string data;

[0048] ② Target amino acid sequence data;

[0049] ③ Known drug target binding affinity data;

[0050] The Davis dataset includes information on 68 drugs, 442 targets, and 30,056 dissociation constants Kd values, and is presented in [the format of the dataset]. The value is used as the affinity value; the KIBA dataset includes 52,498 drug information and 467 target information. After filtering, the KIBA dataset also contains 2,111 drug information, 229 target information, and 118,254 affinity values.

[0051] S2. Process drug and target sequence data, converting them into molecular structures and affinity maps:

[0052] S21. Convert the drug SMILES string into a drug molecular structure diagram using RDKit software;

[0053] To effectively capture the chemical properties and topological structure information of drug molecules, this invention employs the open-source cheminformatics tool RDKit to perform molecular structure analysis on the drug SMILES encoding. Chemical atoms are mapped as graph nodes, and chemical bonds are defined as edges between nodes, constructing an undirected graph structure of the drug molecule. ,in and These represent the set of atomic nodes and the set of edges formed by chemical bonds, respectively.

[0054] S22. Convert the target amino acid sequence into a target molecular structure diagram using Pconsc4 software;

[0055] To obtain molecular-scale feature information of the target protein and achieve more comprehensive target characterization, we first preprocessed the target sequence, including sequence alignment and filtering. Then, we used Pconsc4 to predict the sequence information into a contact graph. In the contact graph, target residues are treated as nodes; if the Euclidean distance between two residues is less than a preset threshold of 0.5, an edge is considered to exist between them. This method effectively preserves the spatial structure information of the target protein, thus converting the target sequence into a two-dimensional molecular graph, which can be represented as follows: ,in and These represent the set of nodes and the set of edges, respectively.

[0056] S23. Construct a drug-target affinity map based on known drug-target affinity information;

[0057] We constructed a drug-target affinity map based on known drug-target affinity information. This affinity map can be defined as follows: ,in Indicates the number of drug molecules, Indicates the number of target molecules. This represents the known drug-target affinity. We normalize the affinity; a higher affinity value indicates a stronger connection between the drug and the target, meaning the drug is more likely to be effective against the target. If the affinity is 0, it means the affinity between the drug and target is unknown, meaning there is no edge between them in the affinity network graph. From this, we obtain the adjacency matrix of the affinity graph Ga. M represents the number of drug molecules, and N represents the number of target molecules.

[0058] S3. Use the multi-scale feature extraction module to extract features of the drug target from sequence, molecular structure, and affinity map respectively:

[0059] S31. Extracting drug sequence features using Bi-LSTM;

[0060] S32. Extract target sequence features using Bi-LSTM;

[0061] S33. Extract drug structural features from drug molecule structure diagrams using the GCN model;

[0062] S34. Use the GCN model to extract target structural features from the target molecular structure diagram;

[0063] S35. Use the GCN model to extract the interaction features of drug targets on the affinity graph, that is, the affinity graph features between drugs and targets.

[0064] In steps S31 and S32, the SMILES coding sequence of the drug molecule and the target amino acid sequence are used as basic representation units, respectively, and feature interaction learning is achieved through a bidirectional temporal association mechanism. Each BiLSTM unit contains two temporal channels, forward and backward, and the temporal information from both directions is dynamically fused through a gating mechanism. This architecture not only enables each marker to capture its own temporal features, but also allows it to obtain contextual semantic information through the bidirectional temporal association mechanism.

[0065] For drug SMILES sequence In terms of its positive time series characteristics The extraction process can be represented as:

[0066]

[0067]

[0068]

[0069]

[0070]

[0071]

[0072] in: , , , These represent the activation values ​​of the input gate, forget gate, output gate, and memory cell, respectively. Represents memory cells, Save Yes Status information at any given moment; This indicates a hidden state. Used with token Interact with the input to generate Moment and , This represents the output of a forward LSTM; express Input the time token; and These are learnable weights and bias parameters. , , , The weights for the input gate, forget gate, output gate, and memory cells. , , , Bias for the input gate, forget gate, output gate, and memory cells; It is a non-linear activation function; This indicates element-wise multiplication;

[0073] Obtain the positive LSTM hidden state for each drug or target. Hidden states of inverse LSTM Then, the final sequence feature representation can be obtained by concatenation:

[0074]

[0075] in, This indicates a splicing operation.

[0076] Finally, the drug sequence feature representation is obtained. and target sequence feature representation .

[0077] In steps S33 and S34, this embodiment employs a three-layer graph convolutional network architecture: First, iteratively fusing the local chemical environment information of atomic nodes through a neighborhood aggregation mechanism, i.e., computing the neighbor information of the nodes (the information aggregated by the first-layer GCN); second, achieving layer-by-layer abstraction of global topological features through multi-layer propagation, ultimately generating an atomic-level representation vector with chemical fingerprint characteristics; essentially, it first aggregates the features of the neighbors around the central node, and then aggregates the features of the neighbors' neighbors (i.e., global topological features) through multiple layers. This framework not only successfully extracts atomic connection matrix information based on chemical bonds, but also preserves the two-dimensional spatial topological features of molecular structures. The formula for learning the feature representations of drugs and targets from drug molecule structures and target molecule structures using GCN is:

[0078]

[0079] in, This represents the hidden state of node v at layer l (the characteristics of the node after processing by layer l of GCN). , It is the neighborhood set of node v; This represents the degree of node v in the graph. ; This represents the weight matrix of the l-th layer;

[0080] After the GCN convolution operation, we learned the feature representation of each atom at the molecular structure scale. After three layers of GCN processing, the final feature representation of each drug molecule can be obtained through the readout function. The readout function operation can be expressed as:

[0081]

[0082] in, It is a differentiable readout function; it is an addition operation. d Represents a set of nodes.

[0083] Drug fusion by reading function formulas All atomic information is used to represent the drug structure characteristics. Based on fusion target All amino acid residue information was obtained using the same method to represent the target structure features. .

[0084] In step S35, the formula for extracting the feature representations of the drug and target from the affinity map using GCN is as follows:

[0085]

[0086] in, It is obtained by performing Laplace normalization on the adjacency matrix A. , It is a diagonal matrix; This is the initial input to the affinity diagram; and These are trainable weight parameters;

[0087] After the above processing, the affinity diagram of the drug and target is obtained based on the characteristic representation of the drug and target. and The initial input is formed by concatenating drug target matrices vertically. That is, each of the first M rows of the initial input represents a drug, and each of the M+1 to M+N rows represents a target. The final Hn is similarly constructed by taking... A single line will give you the representation of each drug or target you want.

[0088] S4. Design a multi-scale feature contrastive learning strategy and calculate the contrastive learning loss:

[0089] S41. Compare the learned drug sequence characteristics, molecular structure characteristics, and affinity diagram characteristics in pairs;

[0090] For contrastive learning, determining positive and negative samples is a key challenge. Generally, features learned at a certain scale are used as anchors, embeddings of the same drug learned at different scales are considered positive samples, and embeddings of other drugs are considered negative samples. In this invention, we employ a novel method for defining positive samples:

[0091] For molecular structure-scale data, we used the PubChem tool to calculate the similarity between drug molecule structures. We believe that drug molecules with similar structures may have similar functions, thus obtaining the similarity between drug molecules. ;

[0092] For sequence-scale data, we use the Smith-Waterman algorithm to calculate the similarity between target sequences. ;

[0093] For affinity graph-scaled data, we use the following formula to calculate the drug... With drugs Similarity between :

[0094]

[0095] in, For drugs The number of edges in the affinity graph (i.e., the number of edges related to the drug) (Number of connected targets) For drugs With drugs The number of targets they share suggests that two drugs sharing the same target are likely similar. Therefore, we calculate the similarity among all drugs, expressed as: ;

[0096] Using the same method, the similarity between targets is calculated based on the number of edges possessed by the targets in the affinity graph (the number of drugs that are commonly connected between targets). .

[0097] Next, we will obtain structure-based drug-drug similarity. Compared with network-based drug-drug similarity Add them together, and The sums are also added together, and then the sums are sorted in descending order. For each drug and target, the first K samples are taken as positive samples, and the rest are taken as negative samples. It should be noted that the K value used for drugs and targets may be different in actual experiments. In this embodiment, after parameter experiments, the K value used for drugs and targets is set to 5.

[0098] Multi-scale contrastive learning was used to study the three scale features of the drug. , , Perform pairwise comparisons and calculate the contrast loss using InfoNCE for drugs. The specific definition of contrastive learning loss is as follows:

[0099]

[0100] in, Indicates drug The positive sample set; M is the number of all drug molecules; It is a temperature parameter that can be changed; Represents the cosine similarity function;

[0101] S42. Compare the learned target sequence features, molecular structure features, and affinity map features in pairs;

[0102] Similar to step S41, using the same calculation method, we can calculate the target in step S42. Contrastive learning loss .

[0103] S5. Obtain the final fusion feature representation of the drug and target, calculate the total loss, and use MLP to predict the drug-target binding affinity.

[0104] S51. Integrating the three scale characteristics of the drug; This embodiment integrates the three scale characteristics of the drug. , , Fusion is performed to obtain the final drug embedding feature:

[0105]

[0106] S52. Fusion of the three scale features of the target; [This refers to the fusion of the three scale features of the target.] , , The features are fused to obtain the final target embedding features:

[0107]

[0108] S53. Predicting affinity scores using MLP:

[0109] The fused final features of the drug target are then input into the MLP for prediction.

[0110]

[0111] S54. Calculate the MSE loss;

[0112] Using predicted scores and experimental measurement scores Calculate MSE loss:

[0113]

[0114] in, This is the total number of drug target pairs used for training; Indicates the affinity value measured in the experiment;

[0115] S55. Calculate the total loss;

[0116] We compared the multi-scale contrast loss between the drug and the target. , and MSE loss Add them together to get the final total loss. :

[0117]

[0118] This invention integrates multi-scale contrastive training objectives into the DTA prediction task, jointly optimizing the two objectives to improve drug and target performance during training. It utilizes the total loss... The trained model is used to predict drug-target binding affinity, which includes the network model corresponding to the S3 to S53 processes.

[0119] S6. In actual prediction, obtain the sequence data of the drug and target to be predicted, as well as the associated affinity data, and then process them using the process in step S2. Finally, use the model for predicting drug-target binding affinity to obtain the prediction result.

[0120] like Figure 2 As shown, in order to intuitively evaluate the predictive performance of the present invention, Figure 2 Figures (a) and (b) show scatter plots on the Davis and KIBA datasets, respectively. These plots illustrate the relationship between predicted and experimental measurements on the Davis and KIBA datasets. The x-axis corresponds to the experimental measurements, and the y-axis corresponds to the model's predictions. Data points closer to the red dashed line indicate higher prediction accuracy. Notably, most samples are symmetrically distributed around the center line, indicating a high degree of consistency between predicted and experimental values. Furthermore, the predicted points on the KIBA dataset are more densely clustered along the central axis, suggesting that this invention achieves more accurate predictions on that dataset.

[0121] To evaluate the performance gap between the method of this invention and other methods, comparative experiments were further conducted on the Davis and KIBA datasets. The experiments used widely used evaluation metrics to measure performance: mean squared error (MSE), consistency index (CI), and regression mean (…). Among them, MSE is used to measure the deviation between the true value and the predicted value, and is a commonly used quantitative indicator in regression tasks; CI is used to evaluate the consistency between the predicted value and the true value, and the value ranges from 0 to 1. The higher the value, the better the prediction effect of the model. This reflects the model's predictive ability and relevance; a higher value means the model is more reliable.

[0122] In the experiment, each dataset was randomly divided into a training set and a test set at a 5:1 ratio. The proposed method and all comparative methods were trained on the training set, and all methods were evaluated on the test set. Furthermore, to select the optimal hyperparameters, 5-fold cross-validation (5-CV) was performed on the training set data. The experimental results are shown in Tables 1 and 2.

[0123] Table 1. Comparative experimental results of different methods on the Davis dataset.

[0124]

[0125] Table 2. Comparative experimental results of different methods on the KIBA dataset.

[0126]

[0127] Figure 3 The figures show the comparison results of ablation experiments performed on the Davis dataset, where (a), (b), and (c) represent MSE, CI, and CI, respectively. The comparison results are as follows. To verify the rationality of the design of this invention and demonstrate the effectiveness of multi-scale feature input and multi-scale feature contrastive learning, we test the relevant hypotheses by exploring the impact of each part of the method on performance. SEQ, MOL, and NET represent the sequence, molecular structure, and affinity map input of the method, respectively, and CL represents the multi-scale feature contrastive learning module. We combined and arranged the method variants by whether or not they included these modules. Under the condition that all experimental parameters remained consistent, ablation experiments were conducted on the Davis dataset to test the method and its variants.

[0128] pass Figure 3 As can be seen, the method of this invention outperforms all other variants, proving the correctness of the inventive approach. It can also be observed that the variants MFCLDTA (w / o MOL&NET), MFCLDTA (w / o SEQ&NET), and MFCLDTA (w / o SEQ&MOL), which utilize only single-scale feature information, do not perform as well as the variants MFCLDTA (w / o NET), MFCLDTA (w / o MOL), and MFCLDTA (w / o SEQ), which utilize feature fusion at two scales. This is because multiple scales can learn more and more comprehensive drug target information, compensating for the incompleteness of single-scale information. This also demonstrates that fusing and supplementing the feature information from different scales of the drug target can improve the model's prediction accuracy. Therefore, this also explains why our model, utilizing information from three scales, achieves better prediction results than the model using two scales. Furthermore, compared to the variant MFCLDTA (w / o CL) without the contrastive learning module, our prediction metrics are also better, proving the effectiveness of the contrastive learning module.

[0129] like Figure 4 As shown, to reveal the mechanism of drug molecule and target protein binding, analyze and predict potential target binding sites, and enhance the interpretability of the model, we conducted visualization experiments and performed analysis. Specifically, we selected two known target proteins (PDB ID: 1M17 and 3NW6) from the Davis dataset and performed ligand-target interaction visualization experiments on the Molecular Manipulation Environment (MOE) software. The visualization results are shown below. Figure 4As shown, we have specially marked important atoms and structures.

[0130] Figure 4 In the diagram, 'a' represents the interaction between the ligand crystal structure and the target. The middle part of each diagram shows the two-dimensional structure of the ligand, surrounded by target residues. Different colors represent residues with different properties, and the blue area represents the important part of the ligand that is more closely attached to the target residues. Figure 4 In the diagram, b is a three-dimensional representation of ligand-target binding, where the yellow area represents the ligand and the cyan area represents the target residue. The diagram partially shows the interaction between the ligand and the target residue (dashed lines between the ligand and the residue).

[0131] Specifically, for PDB structure 1M17, the method correctly identified and predicted the binding of the amide carbonyl functional group to the ligand-target, which is essential for the binding. This functional group is a carboxamide linker between the benzene and pyrimidine rings. The provided carbonyl oxygen acts as a hydrogen bond acceptor, connecting to the amino group of the phenylalanine 699 main chain. The amide carbonyl group at the ligand terminal forms a hydrogen bond with the amino donor of the side chain of lysine 721, further emphasizing the interpretability of the model. For PDB structure 3NW6, the interpretability of this invention further reveals in detail the sites and patterns of ligand-target interactions. The imino group of the ligand main backbone targets the hydroxyl group, which acts as an acceptor, forming a hydrogen bond with the phenolic hydroxyl group of the tyrosine (Tyr1135) side chain. In addition, the π-H interaction between the π system of the aromatic six-membered ring of the ligand and the Cα-hydrogen of the Gly1125 main chain helps determine the binding site of the ligand.

[0132] Based on the visualization analysis described above, our model demonstrates excellent interpretability potential. It can not only successfully reveal the mechanism of receptor-target interaction but also analyze and predict potential binding sites. This capability may help us explore some hidden interactions and accelerate drug discovery.

[0133] In summary, this invention first converts the drug SMILES string into a drug molecular structure diagram using RDKit, and then converts the target sequence into an amino acid sequence contact diagram using Pconsc4. A drug-target affinity map is constructed using known affinity information. The sequence information is then processed through a three-layer Bi-LSTM for feature extraction, while the molecular structure diagram and affinity map are processed through a three-layer GCN for feature extraction. Next, the obtained feature information of the drug and target at three scales is input into a multi-scale feature contrast learning module to maximize the mutual information of features at different scales for feature alignment and fusion. Finally, the fused feature information of the drug and target is input into the prediction module's MLP, and the output is the predicted DTA value.

[0134] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for training a drug-target binding affinity prediction model based on contrastive learning, characterized in that, For a network model that predicts drug-target binding affinity based on drug and target sequence data and associated affinity data, a contrastive learning approach is used to train the model. During the contrastive learning process, positive and negative samples are determined as follows: Calculate the similarity between drug molecular structures to obtain similarity Calculate the similarity between target sequences to obtain similarity. The similarity between drugs is calculated based on the number of targets they share in common at the affinity graph scale. The similarity between targets is calculated based on the number of drugs commonly connected to each other in the affinity graph. ; Will and Add them together, and Add them together, and then sort the sums in descending order. For each drug and target, take the first K samples as positive samples and the rest as negative samples. Then, the contrastive learning loss is calculated based on the positive and negative samples and used for model training; The drug-target binding affinity prediction model processing procedure is as follows: For drug SMILES string data, a drug feature extraction network is used to obtain drug sequence feature representations. ; For the target amino acid sequence, a target amino acid feature extraction network is used to obtain the target sequence features. ; For drug molecular structure diagrams obtained by converting drug SMILES string data, a graph convolutional neural network is used to obtain drug structural feature representations. For the target molecular structure map obtained by converting target amino acid sequence data, a graph convolutional neural network is used to obtain the target structural feature representation. ; For the drug-target affinity map corresponding to drug-target affinity information, a graph convolutional neural network is used to extract features, and drug affinity map features are obtained based on the extracted features. Target affinity map features ; The process of calculating the contrastive learning loss based on positive and negative samples includes: Multiscale contrastive learning was used to study the multiscale features of drugs. , , Perform pairwise comparisons and calculate drug values. Contrastive learning loss: in, Indicates drug The positive sample set; M is the number of all drug molecules; It is a temperature parameter that can be changed; Represents the cosine similarity function; That is, medicine corresponding , That is, medicine corresponding , That is, medicine corresponding ; Multi-scale contrastive learning is used to analyze the multi-scale features of the target. , , Perform the same processing to obtain the target. Contrastive learning loss ;Will As the overall contrastive learning loss.

2. The method for training a drug-target binding affinity prediction model based on contrastive learning according to claim 1, characterized in that, The drug feature extraction network and the target amino acid feature extraction network are both Bi-LSTM.

3. The method for training a drug-target binding affinity prediction model based on contrastive learning according to claim 1, characterized in that, Drug structure feature representation is obtained using graph convolutional neural networks. The process includes: First, the GCN network is used to learn the feature representations of drugs and targets from the drug molecule structure: in, This represents the hidden state of node v at level l. , It is the neighborhood set of node v; This represents the degree of node v in the graph. ; This represents the weight matrix of the l-th layer; It is a non-linear activation function; After processing by the GCN network, the features of each drug molecule are obtained through the readout function. .

4. The method for training a drug-target binding affinity prediction model based on contrastive learning according to claim 1, characterized in that, Target structural feature representation is obtained using graph convolutional neural networks. The process of obtaining drug structure feature representation using graph convolutional neural networks The process is the same.

5. A method for training a drug-target binding affinity prediction model based on contrastive learning according to any one of claims 1 to 4, characterized in that, The similarity between drugs is calculated based on the number of targets they are connected to in common at the affinity diagram scale. During the process, utilizing Calculate drugs With drugs The similarity between them For drugs The number of edges in an affinity graph. For drugs With drugs The number of targets that are connected together.

6. The method for training a drug-target binding affinity prediction model based on contrastive learning according to claim 5, characterized in that, The similarity between targets is calculated based on the number of drugs commonly connected to each other in the affinity diagram. The method and calculation of drug similarity The method is the same.

7. The method for training a drug-target binding affinity prediction model based on contrastive learning according to claim 6, characterized in that, The process of calculating the contrastive learning loss based on positive and negative samples for model training includes: Predicting drug target binding affinity using MLP: , , Fusion is performed to obtain the final drug embedding feature. ;right , , Fusion is performed to obtain the final target embedding features. The final features of the fused drug target are input into the MLP for prediction. To obtain the predicted loss ; For the predicted score; Based on predicted loss Comparison of learning loss with total Calculate total loss Utilizing total losses The trained network model is obtained for predicting drug-target binding affinity.

8. A method for predicting drug-target binding affinity based on contrastive learning, characterized in that, The network model for predicting drug-target binding affinity is obtained by using the drug-target binding affinity prediction model training method based on contrastive learning as described in claim 7.

Images