Protein-ligand binding site and affinity prediction method and system

Abstract

The invention discloses a protein-ligand binding site and affinity prediction method and system, and belongs to the technical field of bioinformatics and computer deep learning. The method comprises the following steps: acquiring and processing a PDB format file of protein and an SDF format file of a ligand to obtain a batch-processed graph data tensor; constructing a composite encoder, and inputting the image data tensor after batch processing into the composite encoder to obtain node features and edge features in molecules and between molecules; constructing a mixed block attention module, and processing the node features and the edge features in the molecules and between the molecules based on the mixed block attention module to obtain attention output; and obtaining a prediction result based on the attention output, and predicting the affinity value based on the prediction result. According to the method, an MoBA adapter is introduced, large-scale molecular data are effectively processed through block processing and a Top-K attention mechanism, and meanwhile the calculation efficiency is kept.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and deep learning technology, specifically relating to a method and system for predicting protein-ligand binding sites and affinity. Background Technology

[0002] In drug discovery, accurately predicting the binding sites of proteins and small molecules is crucial, directly impacting the potency and specificity of candidate drugs. Traditionally, experimental methods such as X-ray crystallography and nuclear magnetic resonance (NMR) have been relied upon to elucidate binding modes. While these methods provide highly accurate structural information, they also have significant limitations, such as lengthy processing times (often several months), high costs (expensive per sample), and difficulty in applying them to complex systems like membrane proteins. With the development of computational biology, structure-based virtual screening technology has gradually become an important tool for accelerating drug development. The core of this technology lies in constructing a prediction algorithm capable of simultaneously modeling the dynamic changes in the three-dimensional conformation of proteins and the characteristics of the chemical space of small molecules. However, existing computational methods still face many challenges in modeling long-range residue interactions, fusing multi-scale structural features, and balancing physical plausibility with computational efficiency. The conformational flexibility of the protein binding pocket, the rotational degrees of freedom of ligand molecule functional groups, and the dynamic adaptation process between the two are intertwined factors, constituting the theoretical bottleneck for high-precision prediction.

[0003] Existing molecular docking technologies are mainly divided into two categories: traditional force field methods (such as AutoDockVina and Glide) and deep learning methods (such as EquiBind and DiffDock). Traditional methods are based on empirical force fields, which have good physical plausibility but low computational efficiency and limited accuracy. Deep learning methods predict and combine conformations through neural networks, which significantly improves speed in some areas, but still suffer from defects in physical plausibility and cannot balance speed and accuracy.

[0004] To address these challenges, it is particularly urgent to develop a novel prediction framework that balances atomic-level precision and computational efficiency, effectively captures intramolecular / intermolecular interactions, and incorporates physical constraints to ensure the reasonableness of the results. Such a framework is of great significance for tackling difficult-to-drug targets and shortening drug development cycles. Summary of the Invention

[0005] This invention aims to address the shortcomings of existing technologies by proposing a method and system for predicting protein-ligand binding sites and affinity. Employing a Transformer-based deep learning architecture, it mainly includes data preprocessing, data loading and batch processing, a composite encoder, hybrid attention MoBA, a Vina scoring head, and prediction generation. This architecture design enables the model to simultaneously capture both local and global features of molecules, improving prediction accuracy.

[0006] To achieve the above objectives, the present invention provides the following solution: a method for predicting protein-ligand binding sites and affinity, comprising the following steps:

[0007] S1. Collect and process the PDB format files of proteins and the SDF format files of ligands to obtain batch-processed graph data tensors;

[0008] S2. Construct a composite encoder by inputting the batch-processed graph data tensor into the composite encoder to obtain the node features and edge features inside and between molecules;

[0009] S3. Construct a hybrid block attention module, and process the node features and edge features inside and between molecules based on the hybrid block attention module to obtain attention output;

[0010] S4. Obtain the prediction result based on the attention output, and predict the affinity value based on the prediction result.

[0011] More preferably, in S1, the processing method includes: data preprocessing, data loading, and batch processing;

[0012] The data preprocessing method includes: using professional processing tools to preprocess the PDB format file of the protein and the SDF format file of the ligand to obtain a CSV file of the sample information to be predicted, as well as the processed ligand SDF file, binding pocket PDB file and initial ligand three-dimensional conformation UFF file;

[0013] The data loading method includes: loading and parsing the CSV file containing the information of the sample to be predicted and the processed ligand SDF file one by one, combining the sample data with the pocket PDB file and the initial ligand three-dimensional conformation UFF file to obtain graph structure data containing atom type, coordinates and bond connection relationship;

[0014] The batch processing method includes: aggregating the graph structure data into a batch to obtain the batch-processed graph data tensor containing node features, edge features, and edge indices.

[0015] More preferably, the method by which the hybrid block attention module obtains attention output includes:

[0016] The query vectors, key vectors, and value vectors of the node features and edge features inside and between the input molecules are grouped into one group for global self-attention calculation and the other group for MoBA attention calculation.

[0017] The method for calculating MoBA attention includes:

[0018] The query vector, key vector, and value vector are divided into blocks according to the block-sharing strategy, and the key vectors after block-sharing are subjected to mean pooling. The mean pooled key vectors are multiplied with the query vector to obtain the gating score. Based on the gating score and the strategy of selecting the top k items, the attention block is obtained. The global attention result is obtained based on the variable-length attention mechanism and the attention block. The global self-attention and the global attention result are fused to obtain the attention output.

[0019] More preferably, the block partitioning strategy includes:

[0020] The input is processed by a feature processing layer to project node features and process edge features. The structural interaction importance score is calculated by combining the node features and the edge features.

[0021] For each sample, find the indexes of all residues belonging to the pocket region; calculate the minimum distance from each residue to the pocket; obtain the core region and shell region based on the minimum distance, and merge the pocket mask, core region, and shell region; if the length of the merged region is greater than the set maximum allowed length, retain the residues with the maximum set value based on the structural interaction importance score and the minimum distance; otherwise, update the block size to obtain the block size and maximum sequence length for each batch.

[0022] More preferably, the method for calculating the structural interaction importance score of the i-th node includes:

[0023]

[0024] In the formula, σ() represents the sigmoid activation function; w represents the weight vector of the importance projection layer; f i represents the combined features of the i-th node; b represents the bias term of the importance projection layer.

[0025] More preferably, the strategy for selecting the top k items includes:

[0026] Feature extraction is performed after encoding each node feature and edge feature; attention score is calculated based on the node feature, and node expression importance score is calculated based on the attention score; contact frequency and centrality score are obtained based on the edge feature; the node feature and edge feature are fused to obtain comprehensive feature; the comprehensive feature is mapped to a probability distribution based on the Softmax activation function, so that the scores of each node feature are on the same scale; the top k nodes with the highest scores are selected based on the normalized scores.

[0027] More preferably, the method for obtaining the predicted affinity value in S4 includes:

[0028] The predicted energy function is obtained based on the attention output, and the docking posture is generated based on the energy function, combined with the pocket file and ligand file, using the Monte Carlo algorithm. The affinity value is obtained based on the docking posture.

[0029] This invention also provides a protein-ligand binding site and affinity prediction system, comprising:

[0030] The acquisition and processing module is used to acquire and process PDB format files of proteins and SDF format files of ligands to obtain batch-processed graph data tensors.

[0031] The first construction module is used to construct a composite encoder, which inputs the batch-processed graph data tensor into the composite encoder to obtain the node features and edge features inside and between molecules;

[0032] The second construction module is used to construct a hybrid block attention module, which processes the node features and edge features inside and between molecules based on the hybrid block attention module to obtain attention output;

[0033] The prediction module is used to obtain a prediction result based on the attention output, and to predict the affinity value based on the prediction result.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] The hybrid block attention module proposed in this invention demonstrates industry-leading technological advantages in the field of protein-small molecule binding site prediction. Its core advantages are reflected in three aspects: prediction accuracy, reliability, and computational efficiency. In terms of prediction accuracy, its Top-1 success rate reaches 65.5%, and its success rate in Top-5 evaluation further improves to 78.4%, both of which are superior to existing mainstream methods. While maintaining excellent accuracy, it also achieves a significant efficiency breakthrough. Its single computation time is only 23.1 seconds, showing a clear advantage over mainstream models in both prediction accuracy and computational efficiency. This "fast and accurate" characteristic makes it particularly suitable for large-scale virtual screening applications. This technological breakthrough, which perfectly combines high accuracy and high efficiency, not only establishes its leading position in this field but also provides a powerful tool for accelerating the drug discovery process.

[0036] This invention achieves several key technological innovations in the field of molecular docking: it pioneeringly integrates the Transformer architecture and the MoBA sparse attention framework, creating a new paradigm for predicting protein-small molecule binding sites. The core lies in the design of a dedicated MoBA adapter, which is embedded in the Transformer encoder and outputs feature representations processed by sparse attention. To effectively integrate the two attention mechanisms, the model employs a periodic insertion strategy during encoding, applying MoBA processing after each Transformer encoding layer to ensure the model can collaboratively capture global dependencies and local sparse attention patterns.

[0037] In the feature fusion stage, a multi-step, refined processing flow was implemented. The features output by MoBA were first normalized to stabilize their distribution. The normalized MoBA features were then concatenated with the original Transformer features to form a more informative enhanced representation. This concatenation operation aimed to preserve the advantages of both attention mechanisms, providing a sufficient information foundation for subsequent fusion. Furthermore, a specially designed Multilayer Perceptron (MLP) fusion module was designed to learn the optimal feature fusion strategy. This module includes linear layers, a ReLU activation function, and a Dropout layer.

[0038] To enhance the model's expressive power, a residual connection strategy based on a gating mechanism is further introduced. This strategy dynamically learns weight coefficients through a gating network to adjust the contribution ratio of MoBA features and original Transformer features to the final representation. This design not only achieves adaptive adjustment of the importance of the two attention mechanisms but also effectively mitigates the potential gradient vanishing problem. Finally, the fused features are optimized through a task-specific processing module to adapt to the needs of different downstream tasks.

[0039] This framework design synergistically leverages the global modeling capabilities of Transformer and the sparse computational efficiency of MoBA, significantly improving computational performance while maintaining powerful expressiveness. A carefully constructed feature fusion strategy ensures effective collaboration between the two mechanisms, avoiding feature conflicts. Furthermore, the introduced gating mechanism and task-specific modules enhance the model's adaptability and robustness across different application scenarios. This design not only improves the model's computational efficiency but also enhances its ability to learn molecular structural features, providing a more powerful feature representation for molecular docking tasks. Attached Figure Description

[0040] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a framework diagram of the protein-ligand binding site and affinity prediction method according to an embodiment of the present invention;

[0042] Figure 2 This is a flowchart of the block-segmentation strategy according to an embodiment of the present invention;

[0043] Figure 3 The flowchart for selecting the first k strategies in this embodiment of the invention is shown. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Example 1:

[0047] like Figure 1 As shown, this embodiment provides a method for predicting protein-ligand binding sites and affinity, mainly including data preprocessing, data loading and batch processing, a composite encoder, a hybrid block attention module (MoBA), a Vina scoring head, and prediction generation. This architecture design enables the model to simultaneously capture both local and global features of molecules, improving prediction accuracy. The introduction of the MoBA adapter, through block processing and a Top-K attention mechanism, effectively handles large-scale molecular data while maintaining computational efficiency.

[0048] Physical constraints and numerical stability mechanisms were introduced during training. Physical constraints included collision detection and minimum distance constraints to ensure that the prediction results conformed to physical laws. Numerical stability mechanisms improved the model's training stability through gradient pruning, loss function smoothing, and outlier handling. These mechanisms not only improved the model's prediction accuracy but also ensured the physical plausibility of the prediction results.

[0049] A multi-task learning framework is adopted, and a dynamic weight adjustment mechanism is used to balance the losses of different tasks. During training, techniques such as learning rate scheduling, gradient accumulation, and mixed precision training are used to improve training efficiency and model performance. At the same time, the model also implements optimization strategies such as dynamic batch size and block processing to effectively handle large-scale data and improve computational efficiency.

[0050] Specifically, the prediction method includes the following steps:

[0051] S1. Collect and process the PDB format files of proteins and the SDF format files of ligands to obtain batch-processed graph data tensors.

[0052] In this embodiment, the protein's PDB format file and the ligand's SDF format file were downloaded from a protein database website. The processing methods included: data preprocessing, data loading, and batch processing.

[0053] Specifically, ligand pretreatment includes: using protonation and hydrogenation tools (such as Obabel), protonating the ligand according to a set physiological pH value and adding hydrogen atoms to generate a ligand file containing complete hydrogen information, which serves as a reference ligand for subsequent pocket identification. Then, using a 3D conformation generation tool (based on the RDKit library and UFF force field), an initial 3D conformation is generated for the ligand, and energy minimization is performed to obtain a stable initial ligand structure.

[0054] Protein preprocessing: Using a whole protein preprocessing tool (such as Reduce), hydrogen atoms are added to the protein, its protonation state is adjusted, and structural anomalies are corrected to obtain a preprocessed whole protein file. Then, a binding pocket extraction tool is used, taking the preprocessed whole protein file and a reference ligand file as input, to extract a specific radius around the reference ligand (e.g., ...). Protein atoms within a certain range form a binding pocket file for the target protein.

[0055] After the above preprocessing, a CSV file containing information about the sample to be predicted is finally generated, as well as a processed ligand SDF file, a binding pocket PDB file, and an initial ligand 3D conformation UFF file.

[0056] The data loading method includes: using a dedicated dataset class module, loading the corresponding sample data one by one based on the sample information in the CSV file and the file paths of preprocessed ligands, pockets, initial conformations, etc., and parsing it into graph structure data containing information such as atom type, coordinates, and bond connection relationships.

[0057] The batch processing method involves aggregating multiple sample graph data into a single batch. During batch processing, data truncation and padding operations are performed to ensure that all graphs within the batch have a uniform dimension. Simultaneously, attention biases are generated to indicate the interactive relationships between nodes in the graph. The final output is a batch-processed graph data tensor containing information such as node features, edge features, and edge indices.

[0058] S2. Construct a composite encoder. Input the batch-processed graph data tensor into the composite encoder to obtain the node features and edge features inside and between molecules.

[0059] The composite encoder comprises an embedding layer and a multi-head attention mechanism. The composite encoder embeds features into the batch-processed graph data tensor, mapping node and edge features to a high-dimensional embedding space, and then invokes the multi-head attention mechanism. This mechanism captures inter-atomic relationships from different "perspectives," weighting and aggregating node features to more fully consider bond interaction information. The node features processed by the multi-head self-attention mechanism are further refined through other feature update operations to include complex intramolecular and intermolecular interaction information, outputting both intramolecular and intermolecular node and edge features.

[0060] S3. Construct a hybrid block attention module. Based on the hybrid block attention module, process the node features and edge features inside and between molecules to obtain the attention output.

[0061] To fully integrate the advantages of global self-attention and local sparse attention, this embodiment introduces a Hybrid Block Attention (MoBA) module. In MoBA, an adapter module is used for in-depth processing of features within and between molecules. Specifically, the input query vector, key vector, and value vector are first divided into two parts: one part is used to calculate global self-attention, and the other part is used to calculate MoBA attention. During the MoBA attention calculation, the query vector, key vector, and value vector are first divided into blocks according to a block-sharing strategy. Then, the block-sharing key vectors are average-pooled, and the resulting key vectors are multiplied by the query vector to obtain a gating score. Based on the gating score and the top k selection strategy, the most interested block is selected. Based on this most interested block, the corresponding query vector block, key vector block, and value vector block are obtained in the index selection. The selected block efficiently calculates attention in a variable-length attention mechanism to obtain the global attention result. Finally, a numerically stable online Softmax mechanism is used to fuse the output of global self-attention and the output of MoBA attention, returning an attention output that combines the advantages of both. This design allows MoBA to dynamically balance the focus on global and local information, thereby processing different data more accurately.

[0062] In this embodiment, considering the need for a method that can dynamically adapt and accurately process different data, a dynamic partitioning strategy based on protein chemical properties and spatial relationships is designed, such as... Figure 2 As shown, the process includes: First, the input atomic features are transformed through a feature processing layer, including node feature projection and edge feature processing. The features are then combined to calculate the structural interaction importance score, and invalid positions are filtered using a sequence mask. Next, the region mask is calculated. For each sample, the indices of all residues belonging to the pocket region are first found, and then the minimum distance from each residue to the pocket is calculated. For distances less than... The residues were marked as core regions, with a distance greater than or equal to 1. and less than or equal to Residues are labeled as shell regions. Region merging refers to the union of the pocket mask, core region, and shell region, ensuring that all critical regions are covered. If the length of the merged region exceeds the set maximum allowed sequence length, residues closer to the pocket and with high structural interaction importance scores are preferentially retained. Specifically, distance and structural interaction importance scores are weighted and sorted, leaving residues with the previously set maximum length. The final block size is the larger of the merged region length and the base block size, and does not exceed the maximum sequence length. The final block strategy outputs the block size and maximum sequence length for each batch to ensure that important structural information is not lost while maintaining computational efficiency. This dynamic block strategy provides the foundation for MoBA's efficient processing, enabling it to more accurately adapt to the characteristics of different data.

[0063] The method for calculating the structural interaction importance score of the i-th node includes:

[0064]

[0065] In the formula, σ() represents the sigmoid activation function; w represents the weight vector of the importance projection layer; f i represents the combined feature of the i-th node (the concatenation of node features and edge features); b represents the bias term of the importance projection layer.

[0066] To enable MoBA to focus on key information while reducing computational complexity when processing large-scale data, this embodiment proposes a strategy for selecting the top k items, such as... Figure 3 As shown. Specific methods include:

[0067] First, each node and edge feature is encoded and then processed. This processing includes the extraction of node and edge features: node features are first used to calculate an attention score, measuring the node's relevance and information interaction ability with other nodes globally; based on this, a node importance score is further extracted, reflecting the node's criticality in the graph. Edge features are used to obtain contact frequency (reflecting the frequency of interaction between nodes) and centrality score (measuring the node's central position in the entire network). These features are integrated using a feature fusion multilayer perceptron, deeply fusing node and edge features to transform them into comprehensive features. Subsequently, the comprehensive features are normalized using the Softmax activation function, mapping them to a probability distribution to ensure that the scores of each node are on the same scale and comparable. Within dynamic blocks or the global scope, the top k nodes with the highest scores are dynamically selected based on the normalized scores; these nodes become the focus of subsequent attention mechanisms. This process not only effectively reduces computational complexity and model computational burden but also ensures that the model focuses on the most representative key nodes, concentrating resources on modeling important information, thereby improving the overall modeling effect and efficiency, making the model more targeted and accurate, and better suited to the needs of practical application scenarios.

[0068] The methods for calculating attention scores include:

[0069]

[0070] In the formula, This means that the average attention score of the i-th node is obtained by averaging the sum of the correlations. This represents the dot product of the features of the i-th node and the j-th node, used to measure the correlation between them; This represents summing over all nodes to obtain the total correlation between the i-th node and all other nodes.

[0071] The methods for calculating the importance score of node representation include:

[0072] Importance Score i =LayerNorm(((W imp ·h i +b imp )×m i ),

[0073] In the formula, h i W represents the feature encoding vector of the i-th node, with dimension d; imp The linear projective weights representing the importance scores of nodes are 1×d; b imp The bias term represents the linear projection; m iThis represents the sequence mask; it is 1 if the i-th position is valid, and 0 otherwise. LayerNorm(): Normalizes the scores of all nodes to improve numerical stability. W imp ·h i +b imp To perform a linear transformation on the features of the i-th node and obtain its original importance score; ×m i This means that only valid node positions are retained, while invalid (padding) positions are hidden.

[0074] Methods for calculating contact frequency include:

[0075] For each protein node i:

[0076]

[0077] In the formula, d i,j represents the distance from node i to ligand node j; I represents the indicator function, which is 1 if less than a threshold, and 0 otherwise; l len This represents the ligand length, which is the total number of nodes in the ligand.

[0078] The methods for calculating the center score include:

[0079]

[0080] In the formula, ∈ is a constant used to prevent division by zero errors.

[0081] S4. Obtain the prediction result based on the attention output, and predict the affinity value based on the prediction result.

[0082] The Vina scoring head inputs the adapter-processed features into the energy prediction module (VinaScoreHead). Based on mechanisms such as neural networks and Gaussian mixture models, the energy prediction module maps intermolecular interaction features to continuous energy values, outputting the predicted energy function, typically including the energy contribution of each protein-ligand atom pair and the total energy after polymerization. Output results: Generates a file containing the predicted energy function, as well as copies of the binding pocket, reference ligand, and initial ligand conformation, providing data support for subsequent docking posture generation and affinity prediction.

[0083] Based on energy function files, binding pocket files, and ligand files, random perturbations are applied to the ligands within the pocket space according to the predicted energy function, including translation, rotation, and bond angle twisting. After each perturbation, the energy state of the new conformation is evaluated using the energy function, and a Monte Carlo algorithm is used to determine whether to accept the new conformation, gradually guiding the ligands to converge to a lower energy attitude, generating a series of possible docking attitudes. All generated docking attitudes are collected, and their relevant statistical information, such as RMSD (root mean square deviation), predicted energy value, and number of twist angles, is calculated to provide a quantitative basis for subsequent analysis.

[0084] Based on docking attitudes generated by Monte Carlo sampling, docking attitudes with high affinity were selected. Potential drug molecule candidates.

[0085] Example 2:

[0086] This embodiment provides a protein-ligand binding site and affinity prediction system, comprising: an acquisition and processing module for acquiring and processing PDB format files of proteins and SDF format files of ligands to obtain batch-processed graph data tensors; a first construction module for constructing a composite encoder, inputting the batch-processed graph data tensors into the composite encoder to obtain node features and edge features within and between molecules; a second construction module for constructing a hybrid block attention module, processing the node features and edge features within and between molecules based on the hybrid block attention module to obtain attention output; and a prediction module for obtaining prediction results based on the attention output and predicting affinity values ​​based on the prediction results.

[0087] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting protein-ligand binding sites and affinity, characterized in that, Includes the following steps: S1. Collect and process the PDB format files of proteins and the SDF format files of ligands to obtain batch-processed graph data tensors; S2. Construct a composite encoder by inputting the batch-processed graph data tensor into the composite encoder to obtain the node features and edge features inside and between molecules; S3. Construct a hybrid block attention module, and process the node features and edge features inside and between molecules based on the hybrid block attention module to obtain attention output; S4. Obtain the prediction result based on the attention output, and predict the affinity value based on the prediction result.

2. The method for predicting protein-ligand binding sites and affinity according to claim 1, characterized in that, In S1, the processing methods include: data preprocessing, data loading, and batch processing; The data preprocessing method includes: using professional processing tools to preprocess the PDB format file of the protein and the SDF format file of the ligand to obtain a CSV file of the sample information to be predicted, as well as the processed ligand SDF file, binding pocket PDB file and initial ligand three-dimensional conformation UFF file; The data loading method includes: loading and parsing the CSV file containing the information of the sample to be predicted and the processed ligand SDF file one by one, combining the sample data with the pocket PDB file and the initial ligand three-dimensional conformation UFF file to obtain graph structure data containing atom type, coordinates and bond connection relationship; The batch processing method includes: aggregating the graph structure data into a batch to obtain the batch-processed graph data tensor containing node features, edge features, and edge indices.

3. The method for predicting protein-ligand binding sites and affinity according to claim 1, characterized in that, The method by which the hybrid block attention module obtains attention output includes: The query vectors, key vectors, and value vectors of the node features and edge features inside and between the input molecules are grouped into one group for global self-attention calculation and the other group for MoBA attention calculation. The method for calculating MoBA attention includes: The query vector, key vector, and value vector are divided into blocks according to the block-sharing strategy, and the key vectors after block-sharing are subjected to mean pooling. The mean pooled key vectors are multiplied with the query vector to obtain the gating score. Based on the gating score and the strategy of selecting the top k items, the attention block is obtained. The global attention result is obtained based on the variable-length attention mechanism and the attention block. The global self-attention and the global attention result are fused to obtain the attention output.

4. The method for predicting protein-ligand binding sites and affinity according to claim 3, characterized in that, The block partitioning strategy includes: The input is processed by a feature processing layer to project node features and process edge features. The structural interaction importance score is calculated by combining the node features and the edge features. For each sample, find the indexes of all residues belonging to the pocket region; calculate the minimum distance from each residue to the pocket; obtain the core region and shell region based on the minimum distance, and merge the pocket mask, core region, and shell region; if the length of the merged region is greater than the set maximum allowed length, retain the residues with the maximum set value based on the structural interaction importance score and the minimum distance; otherwise, update the block size to obtain the block size and maximum sequence length for each batch.

5. The method for predicting protein-ligand binding sites and affinity according to claim 4, characterized in that, The method for calculating the structural interaction importance score of the i-th node includes: In the formula, σ() represents the sigmoid activation function; w represents the weight vector of the importance projection layer; f i represents the combined features of the i-th node; b represents the bias term of the importance projection layer.

6. The method for predicting protein-ligand binding sites and affinity according to claim 3, characterized in that, The strategy for selecting the top k items includes: Feature extraction is performed after encoding each node feature and edge feature; attention score is calculated based on the node feature, and node expression importance score is calculated based on the attention score; contact frequency and centrality score are obtained based on the edge feature; the node feature and edge feature are fused to obtain comprehensive feature; the comprehensive feature is mapped to a probability distribution based on the Softmax activation function, so that the scores of each node feature are on the same scale; the top k nodes with the highest scores are selected based on the normalized scores.

7. The method for predicting protein-ligand binding sites and affinity according to claim 1, characterized in that, The method for S4 to obtain the predicted affinity value includes: The predicted energy function is obtained based on the attention output, and the docking posture is generated based on the energy function, combined with the pocket file and ligand file, using the Monte Carlo algorithm. The affinity value is obtained based on the docking posture.

8. A protein-ligand binding site and affinity prediction system, said system being used to implement the method according to any one of claims 1-7, characterized in that, include: The acquisition and processing module is used to acquire and process PDB format files of proteins and SDF format files of ligands to obtain batch-processed graph data tensors. The first construction module is used to construct a composite encoder, which inputs the batch-processed graph data tensor into the composite encoder to obtain the node features and edge features inside and between molecules; The second construction module is used to construct a hybrid block attention module, which processes the node features and edge features inside and between molecules based on the hybrid block attention module to obtain attention output; The prediction module is used to obtain a prediction result based on the attention output, and to predict the affinity value based on the prediction result.

Images