A drug-target binding affinity prediction method based on multi-feature fusion to reduce feature redundancy

Abstract

The present invention belongs to the technical field of bioinformatics, and in particular relates to a drug target binding affinity prediction method that uses multi-feature fusion to reduce feature redundancy. The PSDTA model integrates physical and chemical properties into the initial features, and explicitly introduces the structural information of amino acids to improve the generalization performance of the model. The model uses a primary neighborhood aggregation network to model the internal interactions of molecules and propagate features. In addition, the model designs two independent channels to reduce redundant node features and predict amino acids that are more critical in the binding process, so that the model can extract more accurate information. By comparing with other drug target binding affinity methods on three benchmark data sets, the model achieved optimal results in all evaluation indicators such as mean square error, mean absolute error, consistency index, adjusted coefficient of determination, and Pearson correlation coefficient.

Description

Technical Field

[0001] The present invention belongs to the technical field of bioinformatics, and particularly relates to a drug target binding affinity prediction method that reduces feature redundancy by fusing multiple features. Background Art

[0002] Accurately detecting drug-target interactions or quantifying the strength of drug-target binding (known as drug-target affinity) is of great significance. It lays the foundation for identifying target proteins at the molecular level, discovering drugs, and treating unknown diseases. However, the global pharmaceutical industry today faces significant challenges. Intense product competition, patent expirations, shorter exclusivity periods, and price constraints are forcing pharmaceutical companies to reduce costs, improve productivity, and accelerate growth. Currently, there are two main methods for identifying drug-target interactions (DTAs): biological experiments and computational methods. However, biological experiments require long timelines and significant resource investment, requiring companies to spend over $500 million and approximately 12-15 years to bring a new compound to market. Therefore, developing computational methods for accurately predicting DTAs is a critical research task in the drug discovery field.

[0003] In recent years, computer simulation methods have been recognized for their ability to utilize comprehensive chemical and proteomic libraries, overcome the shortcomings of high-throughput bioassays, and improve the efficiency of the early stages of drug discovery. Therefore, many computational methods have been explored to address the challenges of drug-target interaction affinity prediction. These prediction methods require effective features or representations of drug molecules and target proteins as input. Many studies have explored various methods to construct drug and target protein features, including: (1) using one-hot encoding generated by manually designed molecular or protein descriptors. (2) converting drug molecules into simplified molecular linear input system strings, converting amino acids in protein sequences into numerical representations, and learning drug molecule and protein representations through sequence learning models. (3) modeling drug molecules and proteins as graphs and using graph neural networks to learn graph-level representations. (4) By utilizing 3D protein structures determined experimentally or predicted by AlphaFold2, proteins are modeled as 3D graphs, so that graph neural networks can be used to learn graph-level representations. Existing protein representations are limited in capturing the fundamental structural features of proteins, particularly regarding the relative importance and contributions of different amino acids. This limitation makes it difficult for current DTA prediction models to extract important protein subgraph features, such as protein binding sites, which embody the specific mechanisms of interaction between proteins and molecular functional groups. Amino acids not directly involved in binding can introduce redundant or noisy information in certain tasks, interfering with model performance. Furthermore, the traditional large protein language models used by many current models lack explicit consideration of protein structural information.

[0004] The drug molecule R&D process is extremely lengthy, typically taking at least ten years and costing an average of $2.63 billion. Despite these years of investment, the failure rate remains as high as 92%. Traditional methods for finding drug molecules rely on a large amount of prior knowledge, while experimental results are susceptible to human interference. Existing DTA prediction models suffer from feature information redundancy, and structures predicted using AlphaFold2 may affect the accuracy of predictions and the generalizability of the model. Summary of the Invention

[0005] The present invention proposes a drug-target binding affinity prediction method that uses multi-feature fusion to reduce feature redundancy, addressing the low accuracy, generalization, and feature information redundancy issues of existing prediction models. The prediction method of the present invention uses SaProt for feature initialization, explicitly considering the structural information of amino acids, avoiding the information leakage problem caused by using AlphaFold2 predicted structure files, while also reducing redundancy and noise information during prediction through subgraph extraction.

[0006] The technical solution of the present invention is achieved as follows:

[0007] A drug-target binding affinity prediction method that uses multi-feature fusion to reduce feature redundancy is used for prediction using the PSDTA model, which includes a multi-feature fusion initialization module, a multi-level molecular interaction learning module, and a dual-channel subgraph extraction module.

[0008] The steps of prediction are as follows:

[0009] S1. After initializing the features of proteins and drugs, the multi-feature fusion initialization module uses a model that can capture protein structural features to generate amino acid representations and structure-aware information, and integrates physical and chemical properties to construct a multi-feature initialization representation of amino acids;

[0010] S2, a multi-level molecular interaction learning module, uses a principal neighborhood aggregation network to model the interactions between proteins and drug molecules, while introducing an attention mechanism to capture the interaction patterns between drugs and protein clusters;

[0011] S3, a dual-channel subgraph extraction module extracts key protein subgraphs from the perspectives of attention mechanism and information bottleneck principle to reduce redundant feature information;

[0012] S4. The two protein subgraph features obtained using the dual-channel subgraph extraction module are concatenated with the drug representation to form a unified composite representation vector, which is fed into a fully connected neural network layer as the final prediction module to regress and output the binding affinity between the drug and the target.

[0013] Through the above technical solutions, multi-feature fusion initialization can reduce the information leakage caused by the structure files predicted by AlphaFold2. Multi-level molecular interaction learning can fully integrate the features of adjacent nodes between nodes and capture the interaction patterns between drugs and proteins. Dual-channel subgraph extraction reduces redundant information and improves the generalization of the model.

[0014] Furthermore, in step S1, the multi-feature fusion initialization module includes a drug feature initialization module and a protein feature initialization module, which respectively form initial node features of drugs and proteins by fusing physical and chemical property features, atomic features, amino acid features, and amino acid structural features;

[0015] In step S2, the multi-level molecular interaction learning module includes a PNA module and an attention extraction subgraph module. The PNA module and the attention extraction subgraph module interact with the features between nodes through a graph neural network and an attention mechanism.

[0016] In step S3, the dual-channel subgraph extraction module includes an attention extraction subgraph module and an IB principle extraction subgraph module. The attention extraction subgraph module and the IB principle extraction subgraph module extract important amino acid nodes from the perspective of amino acid groups and single amino acids, respectively.

[0017] Furthermore, in step S1, the amino acid sequence is used as the input of the protein, and the SaProt model is used to construct it into a 2D graph and obtain the initialization features of the nodes. During the initialization process, the SaProt model further captures the geometric conformation information of each residue relative to its spatial neighbors by explicitly adding amino acid structure information. The three-dimensional geometric information of the protein is encoded into discrete "3D tags" through the vector quantized variational autoencoder inside foldseek, and the residue word and structure word of each residue site are combined to construct a structure-aware vocabulary.

[0018] Furthermore, in step S2, after obtaining the initialized node features, the PNA module is used to aggregate the features between drug and protein nodes. Two different aggregators and three different degree scalers are used in the aggregation process. The amplification or attenuation of the incoming message is achieved by multiplying the aggregation value with the degree scaler.

[0019]

[0020]

[0021]

[0022] Among them, α is a variable parameter, which is set to negative for attenuation, positive for amplification, and zero for no scaling. d is the degree of the current node, which represents the number of neighbor nodes. is the average logarithm of the node degree in all training sets, train represents the training set, i represents the element in the training set, Represents the degree of the i-th element. The PNA module is constructed by combining two aggregators and three degree scalers in pairs, which includes a total of six operations.

[0023]

[0024] Among them, σ represents the average value aggregator, min represents the minimum value aggregator, It is a tensor product, which is used to construct a 2D graph and initialize the features to get the representation of the drug. and protein expression , and are sets of nodes representing drugs and proteins respectively, and Represents a collection of variables for drugs and proteins, which change after being passed into the PNA module:

[0025]

[0026]

[0027] in, , E is the edge feature, A is the adjacency matrix, d is the feature length, and l represents the different levels of the PNA module; 、 Indicates the representation of the drug in the l+1th layer and the lth layer in PNA, Indicates the representation of the protein in the l+1 layer and the l layer in PNA, represents the PNA used for polymerizing drugs, and specifically refers to the PNA of the first layer, Represents PNA for polymerizing proteins, and specifically refers to the PNA of the first layer, 、 Represent the edge features of drugs and proteins at layer l respectively.

[0028] Furthermore, in step S2, a clustering algorithm is used to divide proteins and drug molecules into functional groups or structural clusters. Closely adjacent residues are classified into the same cluster, thereby forming more representative group features. Multiple clusters on the drug side are integrated into overall features. Subsequently, the attention mechanism is used to regard the drug features as queries and the various groups of the protein as keys and values. The drug molecules and protein group features can be mapped to the same space through two weight matrices. At this time, the similarity of the two features can be compared through dot product. Finally, through the cross-attention message passing mechanism, key information is transmitted between the protein and the ligand according to the group importance score, and the node features of the drug and protein are updated in this process.

[0029] Furthermore, in step S3, redundant information is reduced from the perspective of amino acid groups, key features are extracted, and different clustering algorithms are used to cluster drugs and proteins respectively. Closely contacted residues are assigned to the same cluster, thereby dividing amino acids into different functional groups. The clusters of drug molecules are integrated into an overall feature representation of the entire drug, thereby reducing computational complexity. Finally, the cross-attention message passing algorithm CAMP is used to exchange features between proteins and ligands according to the calculated protein group importance scores.

[0030] Furthermore, in step S3, redundant information is reduced and key features are extracted from the perspective of individual amino acids. In another channel, the IB principle is applied to extract subgraph modules to identify task-related subgraphs from the perspective of amino acids. In this module, the IB principle is followed to retain the features of proteins related to drug target binding as much as possible while reducing the features of proteins unrelated to drug target binding. The subgraph is obtained by optimizing the following objectives. The optimization objective function is as follows:

[0031]

[0032] Where Y represents the true value associated with P, P represents the original protein map, Represents the true value Y and subgraph The mutual information between Represents the original graph P and subgraph The mutual information between represents the IB subgraph embedding, β is the weight;

[0033] To optimize the objective function, the inverse of the regression loss between the true value and the predicted value using subgraph embedding is used as the first term The lower bound of , by minimizing the regression loss, The lower bound of maximizing is achieved. For the second term, the Donsker-Varadhan representation of the KL divergence is applied to approximate The upper bound of .

[0034] After adopting the above technical solution, the beneficial effects of the present invention are:

[0035] This paper designs a drug-target binding affinity prediction method that uses multi-feature fusion to reduce feature redundancy. Leveraging a deep neural network framework, this method can accurately predict drug-target binding affinity. This prediction method fully considers multiple levels of initial features to construct highly expressive initial node features. It utilizes a principal neighborhood aggregation network and an attention mechanism to fully capture the nonlinear nature of drug-target interactions. By leveraging the attention mechanism and the information bottleneck (IB) principle, key protein subgraphs are extracted to reduce redundant feature information. Results demonstrate that this prediction method can explore more complex chemical molecular spatial structures, further improving the accuracy of drug-target binding affinity prediction.

[0036] The PSDTA model framework consists of three key modules: a multi-feature fusion initialization module, a multi-level interactive learning module, and a dual-channel subgraph extraction module. These modules are responsible for structural representation, relationship modeling, and redundancy suppression, respectively. They feature clear logic and distinct divisions of labor. Compared to traditional single GNN or global attention models, the PSDTA model progressively models drug-target interactions from bottom to top, more systematically modeling the drug-target interaction process. Specifically, the multi-feature fusion initialization module: Traditional DTA models such as DeepDTA and GraphDTA typically use only protein sequence / drug SMILES encodings and lack structural modeling. In contrast, the PSDTA model combines graph structure, physicochemical information, and spatial representation at the input stage, resulting in a richer feature hierarchy. The PNA module: While conventional GCN or GAT models only use averaging or attention aggregation, ignoring differences in node degree distribution, the PNA module significantly enhances neighborhood modeling capabilities. The dual-channel subgraph extraction module reduces redundancy and highlights key binding features. Traditional methods often model the entire graph, which is prone to incorporating irrelevant noise. The PSDTA model compresses redundant information from two perspectives (group level + site level) to enhance the generalization and prediction accuracy of the PSDTA model.

[0037] Compared to existing DTA models, the PSDTA model incorporates structure-awareness and information-constraint mechanisms in multiple key steps, including input feature construction, graph neural network design, and feature compression and extraction. By using a dual-channel subgraph extraction module to compress irrelevant information at different scales and leveraging the PNA module to enhance effective information transfer between nodes, the model significantly improves its accuracy and robustness in drug-target affinity prediction tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 Constructing and using structure-aware vocabularies for SaProt;

[0040] Figure 2 This is the overall architecture diagram of the PSDTA model;

[0041] Figure 3 The predicted and actual values of PSDTA binding affinity;

[0042] Figure 4 The performance of PSDTA on different protein families in the PDBBind v2020 dataset;

[0043] Figure 5 Important amino acids predicted for the two channels of PSDTA. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] The embodiments of the present application disclose a method for predicting drug-target binding affinity by fusing multiple features to reduce feature redundancy.

[0046] Example, a drug-target binding affinity prediction method that reduces feature redundancy by fusion of multiple features, the specific contents are as follows:

[0047] 1. Dataset

[0048] This paper evaluated the performance of PSDTA using three publicly available benchmark datasets: PDBBind v2016, PDBBind v2020, and Davis. These datasets provide drug SMILES strings, target sequences, and binding affinities. Data preprocessing was performed on all three datasets, including the removal of redundant samples. Alphafold2 was used to obtain the target's PDB file, and the Davis dataset's affinity values were further processed. Detailed dataset information is shown in Table 1.

[0049] Table 1 Statistical details of the dataset

[0050]

[0051] 2. Overall architecture of the PSDTA model

[0052] The framework of the PSDTA model is as follows Figure 2 The PSDTA model includes a multi-feature fusion initialization module, a multi-level molecular interaction learning module, and a dual-channel subgraph extraction module. The multi-feature fusion initialization module includes a drug feature initialization module and a protein feature initialization module, the multi-level molecular interaction learning module includes a PNA module and an attention extraction subgraph module, and the dual-channel subgraph extraction module includes an attention extraction subgraph module and an IB principle extraction subgraph module.

[0053] Figure 2 A in the figure is the drug feature initialization module. After inputting a drug's SMILES string, it is first converted into a graph-level representation using the RDkit tool. Then, representative physicochemical properties are selected and converted into feature representations using one-hot encoding, along with the atomic types. The concatenated physicochemical properties are used as the features of each node. Figure 2 B in the figure is the protein feature initialization module. After inputting the amino acid sequence of the protein, it is converted into a graph-level representation through SaProt and structural information is explicitly added. Then, representative physicochemical properties are selected and the physicochemical properties and amino acid types are encoded as features through one-hot encoding. After splicing, they are used as the features of each graph node. Figure 2 C and Figure 2 The D in the figure represents the PNA module, which is used for the characteristic fusion delivery of drugs and proteins. Figure 2 The E in the figure is the attention extraction subgraph module. The attention extraction subgraph module uses the output of the PNA module as input. In the attention extraction subgraph module, a clustering algorithm is first used to cluster proteins and drugs to obtain different clustering groups. The purpose of doing this is because the time complexity of the attention mechanism is high. Using the attention mechanism on the group can effectively reduce the time complexity. The attention extraction subgraph module can output a subgraph representation and a drug representation. Figure 2 The F in the figure is the IB principle subgraph extraction module. The IB principle subgraph extraction module also uses the output of the PNA module as input, extracts protein subgraphs according to the IB principle, and finally outputs protein subgraph representation and drug representation. Figure 2 The G in the figure is the prediction module, which takes the output of the attention extraction subgraph module and the IB principle extraction subgraph module as input, concatenates them and passes them into a fully connected layer, and finally outputs the result of drug-target binding affinity.

[0054] 3. Multi-feature fusion initialization module

[0055] The multi-feature fusion initialization module includes a drug feature initialization module and a protein feature initialization module, which initializes the features of proteins and drugs to obtain feature representations with better expressiveness.

[0056] In the present invention, the molecular graph of the drug is constructed based on the SMILES string. Specifically, the RDKit tool is used to parse the molecular data to extract the two-dimensional structural information of the molecule. In the representation of the molecular graph, atoms are mapped to the nodes of the graph, and their node features are composed of a combination of multiple attributes, including atomic symbols, atomic degrees, implicit valences, number of free valence electrons, hybridization types, and aromaticity. These attributes are combined into a high-dimensional feature vector to characterize the properties of the node. The edges in the molecular graph correspond to the chemical bonds in the molecules, and the presence or absence of edges indicates whether there is an interaction between the two nodes. Based on these edges, an adjacency matrix is constructed to encode the relative position information between nodes.

[0057] In the present invention, the amino acid sequence is used as the input of the protein. It is constructed as a 2D graph through the SaProt model and the initialization features of the nodes are obtained. During the initialization process, the SaProt model further captures the geometric conformation information of each residue relative to its spatial neighbors by explicitly adding amino acid structure information. The three-dimensional geometric information of the protein is encoded into discrete "3D labels" through the vector quantization variational autoencoder (VQ-VAE) inside foldseek. These labels are the result of vector quantization and are abstract representations of the geometric conformation of each residue relative to its spatial neighbors. The residue word and structure word of each residue site are combined to construct a structure-aware (SA) vocabulary. As Figure 1 After constructing the protein's amino acid sequence into a 2D graph using the SaProt model and obtaining the initial features of the nodes, physical and chemical properties such as whether it is an aliphatic amino acid, whether it is an aromatic amino acid, relative molecular mass, acidic pKa value, basic pKb value, and isoelectric point are added to the node features.

[0058] 4. Multi-level molecular interaction learning module

[0059] The multi-level molecular interaction learning module includes a PNA module and an attention extraction subgraph module. After obtaining the initialized node features, the present invention transfers features within the protein and within the molecule, as well as between the protein and the drug. This helps to further model the spatial and structural dependencies between residues within the protein and enhance the understanding of the functional regions of the protein. At the same time, feature transfer within the drug molecule can capture the local structural patterns between different atoms and improve the accuracy of molecular representation. More importantly, establishing a feature interaction mechanism between proteins and drugs can effectively simulate the mechanism of action between them, thereby providing more refined feature support for subsequent affinity prediction.

[0060] (1) The PNA module is used to aggregate the features between drug and protein nodes respectively. Two different aggregators and three different degree scalers are used in the aggregation process. Among them, the two aggregators include the average aggregator and the minimum aggregator, and the three degree scalers include amplification, reduction, and unchanged. The amplification or attenuation of the incoming message is achieved by multiplying the aggregation value with the degree scaler. Therefore, the logarithmic scaler S(d,α) is used to better describe the neighborhood influence of a given node:

[0061]

[0062]

[0063]

[0064] Among them, α is a variable parameter, which is set to negative for attenuation, positive for amplification, and zero for no scaling. d is the degree of the current node, which represents the number of neighbor nodes. is the average logarithm of the node degree in all training sets, train represents the training set, i represents the element in the training set, Represents the degree of the i-th element. The PNA module is constructed by combining the aggregator and the three degree scalers in pairs. It includes a total of six operations, which refer to the pairwise combination of the three degree scalers and the two aggregators.

[0065]

[0066] Among them, σ represents the average value aggregator, min represents the minimum value aggregator, It is a tensor product. Through these six operations, the node can take into account the influence of the number of neighbors when aggregating neighbor features, enhance the expression and generalization ability of the PSDTA model, and adapt to graph structures with different degree distributions. By constructing a 2D graph and initializing features, the drug representation is obtained. and protein expression , and are sets of nodes representing drugs and proteins respectively, and Represents a collection of variables for drugs and proteins, which change after being passed into the PNA module:

[0067]

[0068]

[0069] The above two formulas represent the feature fusion of drugs and proteins respectively. In practice, the PNA module will have many layers, As an example, the initial representation of the drug molecule , E (edge features), A (adjacency matrix) to obtain a new drug representation , and then passed to the second layer of the PNA module to continue the above operations.

[0070] in, , E is the edge feature, A is the adjacency matrix, d is the feature length, and l represents the different levels of the PNA module; 、 Indicates the representation of the drug in the l+1th layer and the lth layer in PNA, Indicates the representation of the protein in the l+1 layer and the l layer in PNA, represents the PNA used for polymerizing drugs, and specifically refers to the PNA of the first layer, Represents PNA for polymerizing proteins, and specifically refers to the PNA of the first layer, 、 Represent the edge features of drugs and proteins at layer l respectively.

[0071] (2) A clustering algorithm is used to divide proteins and drug molecules into functional groups or structural clusters. Closely adjacent residues are grouped into the same cluster, thereby forming more representative group features. Multiple clusters on the drug side are integrated into overall features to reduce computational complexity. Subsequently, an attention mechanism is used to treat drug features as queries and each protein group as a key and value. The drug molecule and protein group features can be mapped into the same space through two weight matrices. At this time, the similarity of the two features can be compared through dot product. Feature similarity indicates the degree of chemical or structural matching between the drug molecule and the protein group. High similarity means that the group has a higher binding potential for the drug molecule. Finally, through the cross-attention message passing (CAMP) mechanism, key information is transmitted between the protein and the ligand based on the group importance score, and the node features of the drug and protein are updated in this process.

[0072] 5. Dual-channel sub-image extraction module

[0073] The dual-channel subgraph extraction module includes the attention extraction subgraph module and the IB principle extraction subgraph module. The PNA module has already integrated and transferred node features. However, when used to predict drug-target binding affinity, there are still many honorary features that are not related to binding. Therefore, the PSDTA model needs to reduce redundant features from two perspectives (amino acid group perspective and single amino acid perspective) to improve the model's generalization and accuracy.

[0074] (1) Reduce redundant information and extract key features from the perspective of amino acid groups. Different clustering algorithms are used to cluster drugs and proteins separately. Residues that are in close contact are assigned to the same cluster. This divides amino acids into different functional groups. Clusters of drug molecules are integrated into a single overall feature to represent the entire drug, thereby reducing computational complexity. Finally, features are exchanged between proteins and ligands based on the calculated protein group importance scores using the cross-attention message passing algorithm (CAMP).

[0075] (2) Reduce redundant information and extract key features from the perspective of a single amino acid. In another channel, the IB principle is applied to extract subgraph modules to identify task-related subgraphs from the perspective of amino acids. In this module, the IB principle is followed to retain the features of proteins related to drug target binding as much as possible while reducing the features of proteins unrelated to drug target binding. The subgraph is obtained by optimizing the following objectives. The optimization objective function is as follows:

[0076]

[0077] Where Y represents the true value associated with P, P represents the original protein map, Represents the true value Y and subgraph The mutual information between Represents the original graph P and subgraph The mutual information between represents the IB subgraph embedding, and β is the weight.

[0078] To optimize the objective function, the inverse of the regression loss between the true value and the predicted value using subgraph embedding is used as the first term The lower bound of , by minimizing the regression loss, The lower bound of is maximized. For the second term, the Donsker-Varadhan representation of the KL divergence is applied to approximate By optimizing this goal, we can obtain the most representative protein subgraph and effectively eliminate redundant node information.

[0079] 6. Performance comparison of PSDTA model and other prediction models

[0080] As shown in Table 2, the proposed PSDTA model framework achieved top performance across multiple dimensions compared to state-of-the-art baselines. Consequently, it surpassed other models in correctly predicting binding affinity values (lower MSE and RMSE scores) and distinguishing the rank order of binding strength between DTI pairs (higher CI and ρ scores). This increase in CI demonstrates PSDTA's superior ability to correctly assess target selectivity, which is crucial for identifying potential lead compounds and distinguishing primary from secondary interactions. Figure 3 Figures A, B, and C depict the predictions of the actual binding affinity values of the PSDTA model for the PDBBind v2016, PDBBind v2020, and Davis datasets, respectively. Significant density can be observed around the reference line (perfect model) where the prediction = true value. To verify the model's performance in different protein families, the present invention divides the proteins in the PDBBind v2020 dataset into different protein families and predicts the MSE performance of each family, as shown in Figure 5. Figure 4 As shown, the model of the present invention performs relatively consistently on all protein families, indicating that the present invention has good generalization ability.

[0081] Table 2 Performance comparison of PSDTA model and other prediction models

[0082]

[0083] 9. PSDTA model predicts amino acids important for binding

[0084] In the existing drug-target affinity (DTA) prediction methods, many methods only predict binding sites through global features, ignoring the local correlation between residues. To address this shortcoming, PSDTA introduces two channels: using attention weights and information bottleneck (IB) principles to quantify the importance of protein residues, and visualizing these two features to reflect the interpretability of the model. In order to verify the rationality of the two channels in predicting binding pockets, the present invention uses the PyMol tool to visualize the prediction results of this protein in the two channels. Figure 5 As shown, the two channels predicted important amino acids respectively. The important amino acids predicted in the two channels complement each other and are relatively consistent with the real important amino acids, indicating that the two channels have the ability to predict important amino acids.

[0085] This paper uses three of the most commonly used datasets in the field of drug-target interaction prediction, and innovates in node feature initialization, feature interaction, and subgraph extraction. Comparative experiments with multiple popular models show that the proposed model achieves optimal performance across multiple metrics. It successfully addresses the information leakage and feature redundancy issues associated with using AlphaFold2 prediction files in existing methods.

[0086] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the technical solutions of the present invention should be included in the protection scope of the present invention.

Claims

1. A drug-target binding affinity prediction method that reduces feature redundancy by fusion of multiple features, characterized in that: The prediction is performed using the PSDTA model, which includes a multi-feature fusion initialization module, a multi-level molecular interaction learning module, and a dual-channel subgraph extraction module. The steps of prediction are as follows: S1. After initializing the features of proteins and drugs, the multi-feature fusion initialization module uses a model that can capture protein structural features to generate amino acid representations and structure-aware information, and integrates physical and chemical properties to construct a multi-feature initialization representation of amino acids; S2, a multi-level molecular interaction learning module, uses a principal neighborhood aggregation network to model the interactions between proteins and drug molecules, while introducing an attention mechanism to capture the interaction patterns between drugs and protein clusters; S3, a dual-channel subgraph extraction module extracts key protein subgraphs from the perspectives of attention mechanism and information bottleneck principle to reduce redundant feature information; S4. The two protein subgraph features obtained using the dual-channel subgraph extraction module are concatenated with the drug representation to form a unified composite representation vector, which is fed into a fully connected neural network layer as the final prediction module to regress and output the binding affinity between the drug and the target.

2. The drug-target binding affinity prediction method for reducing feature redundancy by fusion of multiple features according to claim 1, characterized in that: In step S1, the multi-feature fusion initialization module includes a drug feature initialization module and a protein feature initialization module, which respectively form the initial node features of drugs and proteins by fusing physical and chemical property features, atomic features, amino acid features, and amino acid structural features; In step S2, the multi-level molecular interaction learning module includes a PNA module and an attention extraction subgraph module. The PNA module and the attention extraction subgraph module interact with the features between nodes through a graph neural network and an attention mechanism. In step S3, the dual-channel subgraph extraction module includes an attention extraction subgraph module and an IB principle extraction subgraph module. The attention extraction subgraph module and the IB principle extraction subgraph module extract important amino acid nodes from the perspective of amino acid groups and single amino acids, respectively.

3. The drug-target binding affinity prediction method for reducing feature redundancy by fusion of multiple features according to claim 2, characterized in that: In step S1, the amino acid sequence is used as the input of the protein, and the SaProt model is used to construct it into a 2D graph and obtain the initial features of the nodes. During the initialization process, the SaProt model further captures the geometric conformation information of each residue relative to its spatial neighbors by explicitly adding amino acid structure information. Through the vector quantized variational autoencoder inside foldseek, the three-dimensional geometric information of the protein is encoded into discrete "3D tags", and the residue word and structure word of each residue site are combined to construct a structure-aware vocabulary.

4. The drug-target binding affinity prediction method for reducing feature redundancy by fusion of multiple features according to claim 2, characterized in that: In step S2, after obtaining the initialized node features, the PNA module is used to aggregate the features between drug and protein nodes. Two different aggregators and three different degree scalers are used in the aggregation process. The amplification or attenuation of the incoming message is achieved by multiplying the aggregate value with the degree scaler. ； ； ； Among them, α is a variable parameter, which is set to negative for attenuation, positive for amplification, and zero for no scaling. d is the degree of the current node, which represents the number of neighbor nodes. is the average logarithm of the node degree in all training sets, train represents the training set, i represents the element in the training set, Represents the degree of the i-th element. The PNA module is constructed by combining two aggregators and three degree scalers in pairs, which includes a total of six operations. ； Among them, σ represents the average value aggregator, min represents the minimum value aggregator, It is a tensor product, which is used to construct a 2D graph and initialize the features to get the representation of the drug. and protein expression , and are sets of nodes representing drugs and proteins respectively, and Represents a collection of variables for drugs and proteins, which change after being passed into the PNA module: ； ； in, , E is the edge feature, A is the adjacency matrix, d is the feature length, and l represents the different levels of the PNA module; 、 Indicates the representation of the drug in the l+1th layer and the lth layer in PNA, Indicates the representation of the protein in the l+1 layer and the l layer in PNA, represents the PNA used for polymerizing drugs, and specifically refers to the PNA of the first layer, Represents PNA for polymerizing proteins, and specifically refers to the PNA of the first layer, 、 Represent the edge features of drugs and proteins at layer l respectively.

5. The drug-target binding affinity prediction method for reducing feature redundancy by fusion of multiple features according to claim 4, characterized in that: In step S2, a clustering algorithm is used to divide proteins and drug molecules into functional groups or structural clusters. Closely adjacent residues are classified into the same cluster, thereby forming more representative group features. Multiple clusters on the drug side are integrated into overall features. Subsequently, the attention mechanism is used to regard the drug features as queries and the various groups of the protein as keys and values. The drug molecules and protein group features can be mapped to the same space through two weight matrices. At this time, the similarity of the two features can be compared through dot product. Finally, through the cross-attention message passing mechanism, key information is transmitted between the protein and the ligand according to the group importance score, and the node features of the drug and protein are updated in this process.

6. The drug-target binding affinity prediction method for reducing feature redundancy by fusion of multiple features according to claim 2, characterized in that: In step S3, redundant information is reduced from the perspective of amino acid groups, key features are extracted, and different clustering algorithms are used to cluster drugs and proteins respectively. Closely contacting residues are assigned to the same cluster, thereby dividing amino acids into different functional groups. The clusters of drug molecules are integrated into an overall feature representation of the entire drug, thereby reducing computational complexity. Finally, the cross-attention message passing algorithm CAMP is used to exchange features between proteins and ligands based on the calculated protein group importance scores.

7. The drug-target binding affinity prediction method for reducing feature redundancy by fusion of multiple features according to claim 6, characterized in that: In step S3, redundant information is reduced and key features are extracted from the perspective of individual amino acids. In another channel, the IB principle is applied to extract subgraphs to identify task-related subgraphs from the perspective of amino acids. In this module, the IB principle is followed to retain as many protein features related to drug target binding as possible while reducing protein features unrelated to drug target binding. Subgraphs are obtained by optimizing the following objectives. The optimization objective function is as follows: ； Where Y represents the true value associated with P, P represents the original protein map, Represents the true value Y and subgraph The mutual information between Represents the original graph P and subgraph The mutual information between represents the IB subgraph embedding, β is the weight; To optimize the objective function, the inverse of the regression loss between the true value and the predicted value using subgraph embedding is used as the first term The lower bound of , by minimizing the regression loss, The lower bound of maximizing is achieved. For the second term, the Donsker-Varadhan representation of the KL divergence is applied to approximate The upper bound of .

Images