Protein-ligand binding affinity prediction method based on graph convolutional network

Abstract

The invention provides a protein-ligand binding affinity prediction method based on a graph convolutional network. The method comprises the following steps: firstly, constructing a protein map based on a three-dimensional structure of a protein pocket, and carrying out feature coding on sequence residues by using ESM3; meanwhile, a molecular map containing geometric information is constructed for the ligand, and chemical characteristics are extracted in combination with molecular fingerprints. And then cross-graph interaction of the protein graph and the ligand graph is realized by adopting a multi-layer message passing graph neural network, and an attention mechanism is introduced to capture key interaction sites and spatial dependency relationships. And generating a unified representation vector of a compound level through pooling, and outputting binding affinity through a prediction network. According to the method, the multi-modal association of the structure and the sequence can be automatically learned, and the accuracy and efficiency of virtual screening and lead compound optimization are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology. Background Technology

[0002] Graph Convolutional Networks (GCNs) are a class of neural network models capable of learning representations on graph-structured data. A typical GCN model consists of multiple layers of graph convolutional layers, normalization layers, activation layers, and pooling layers.

[0003] Currently, protein-ligand binding affinity prediction is mainly based on three technical routes: the first is the traditional structural bioinformatics method represented by molecular docking, which calculates the binding pattern of ligands in protein pockets and obtains a score by pre-setting an energy function; the second is the deep learning method based on three-dimensional mesh convolutional neural networks (3D CNN), which extracts spatial features by voxelizing the protein-ligand complex structure and performing convolution, normalization, activation and pooling operations in sequence; the third is the modeling method based on graph neural networks (GNN), which constructs proteins and ligands as graph structures, aggregates neighborhood information through graph convolution operations, and achieves feature update by combining normalization layers and activation layers.

[0004] Because proteins typically contain a large number of atoms and their ligands have complex molecular structures, the input graphs are often quite large. In existing technologies, graph convolution operations, normalization calculations, and nonlinear activation calculations all have high computational complexity. Feature aggregation and normalization steps, in particular, exhibit linear or even superlinear growth with the number of nodes, becoming major performance bottlenecks in affinity prediction. Furthermore, traditional methods often rely on single structural or sequence features, failing to fully integrate protein sequence, structure, and ligand chemical features, resulting in insufficient model generalization ability. Therefore, reducing the computational burden of graph convolution models, improving multi-source feature fusion capabilities, and enhancing the accuracy of protein-ligand affinity prediction are among the important technical challenges in this field. Summary of the Invention

[0005] This invention aims to address the problems of insufficient characterization of protein pocket structure, sequence information, and ligand chemical features, limited feature fusion capabilities, and difficulty in obtaining accurate prediction results in existing protein-ligand binding affinity prediction methods. It provides a graph convolutional network method that can effectively integrate multi-source features and improve affinity prediction accuracy.

[0006] The present invention proposes a protein-ligand binding affinity prediction method based on graph convolutional networks, which includes the following steps: Data selection step: Select protein-ligand complex data from the PDBbind database to ensure that the data covers different protein types and ligand chemical spaces, providing a rich and high-quality sample base for model training. At the same time, redundant or low-quality entries are removed through data preprocessing to improve the reliability and generalization ability of the prediction.

[0007] Steps for extracting protein features: Construct an atomic-level representation of the protein, using atom type, valence electron information, bond connections, and three-dimensional coordinates as node and edge features. At the same time, use a protein sequence language model to generate residue features, capturing amino acid physicochemical properties, sequence context dependencies, and structure-related information, thereby achieving effective integration of the structure and sequence information of the protein graph.

[0008] The steps for extracting ligand features are as follows: construct an atomic graph of the ligand, extract chemical features and geometric information, and generate a molecular fingerprint representation to comprehensively describe the structure and chemical properties of the ligand, providing multi-dimensional input features for subsequent graph convolutional networks.

[0009] The feature vector fusion step involves using a cross-graph messaging mechanism to enable feature interaction between the protein map and the ligand map, and combining this with an attention mechanism to highlight key binding sites and spatial dependencies. Finally, the features are fused into a joint feature vector at the complex level to fully characterize protein-ligand interaction information.

[0010] Affinity prediction steps: Input the joint feature vector into a multi-layer graph convolutional network model, extract high-dimensional representation through nonlinear activation and normalization layers, and output the binding affinity prediction results in continuous numerical form. The model can be optimized using evaluation metrics to achieve high-precision and high-generalization protein-ligand affinity prediction.

[0011] This invention utilizes a protein-ligand binding affinity prediction method based on graph convolutional networks. This method enables high-precision, efficient, and highly generalized prediction of binding affinity in virtual screening and lead compound design. Specific technical effects are as follows: The data selection step selects high-quality protein-ligand complex data from the PDBbind database and performs standardization and deduplication. This ensures the diversity and reliability of the model training data while reducing interference from low-quality or abnormal samples, improving the stability and generalization ability of the prediction model, and providing a solid data foundation for subsequent feature extraction and prediction.

[0012] The technical advantages of the protein feature extraction step: By constructing an atomic-level protein representation and combining it with a sequence language model to generate residue features, this invention can simultaneously capture the three-dimensional structural information, amino acid physicochemical properties, and sequence context dependencies of proteins, achieving a comprehensive characterization of protein pocket regions. This integration of multi-source information enhances the model's ability to identify key binding sites, providing ample feature support for accurately predicting protein-ligand interactions.

[0013] Technical benefits of the ligand feature extraction step: By constructing ligand atomic maps, extracting chemical features and molecular fingerprints, this invention can comprehensively describe the structure, electronic and geometric information of ligands, enabling the model to accurately capture the chemical properties of ligands and their conformational features in space, thereby enhancing the accuracy of ligand characterization and providing reliable input features for interaction.

[0014] The technical advantages of the feature vector fusion step: By utilizing a multi-layer graph convolutional network to achieve message passing between the protein and ligand maps, and combining an attention mechanism to highlight key binding sites and spatial dependencies, this invention can generate joint feature vectors at the complex level. This feature fusion method not only preserves the key information of proteins and ligands, but also enhances the ability to represent complex nonlinear interactions, thereby significantly improving the accuracy of affinity prediction.

[0015] Technical advantages of the affinity prediction step: By inputting the joint feature vector into the MLP layer and combining nonlinear activation and normalization operations, this invention can output continuous numerical binding affinity prediction results, which can be quantitatively evaluated using RMSE, MAE, SD, and Pearson correlation coefficient R. This method eliminates the need for manually setting energy functions or empirical rules, and features high accuracy, high generalization ability, and high computational efficiency, significantly improving virtual screening efficiency and lead compound optimization.

[0016] Overall technical benefits: By combining the above steps, the method of this invention can automatically and comprehensively characterize the key features of protein-ligand complexes, capture complex nonlinear interactions, and achieve high-precision, efficient, and highly generalizable binding affinity prediction. Compared with traditional methods, this invention avoids manual feature design and complex energy function calculations, improving the efficiency and reliability of virtual screening and drug design. Attached Figure Description

[0017] Figure 1 This is a flowchart of a protein-ligand binding affinity prediction method based on graph convolutional networks according to the present invention. Figure 2 This is a diagram of the graph attention layer module inspired by polar coordinates used in this invention; Figure 3 This is a diagram of the node-to-edge information interaction module used in this invention; Figure 4 This is a diagram of the edge-to-node information interaction module used in this invention. Detailed Implementation

[0018] like Figure 1 As shown in step S1, the present invention first inputs data information, including the three-dimensional conformation of the protein, the amino acid sequence, the three-dimensional structure of the ligand, and the corresponding experimental affinity value. The system can preprocess the structural information of the protein and ligand, including coordinate normalization, residue screening, pocketing, and charge calculation, to construct standardized input data suitable for subsequent feature generation and graph modeling.

[0019] After obtaining the protein sequence, the ESM3 model in step S2 is used to extract deep semantic embeddings at the residue level. The ESM3 model can perform multi-layer deep reasoning on the protein sequence and output a vector representation of the protein, which can capture structural semantics, evolutionary information, and long-range dependencies. Subsequently, step S3 constructs a protein pocket diagram based on the protein's three-dimensional structure, where the pocket regions can be determined based on ligand neighborhoods. The nodes of the protein pocket diagram correspond to the atoms in the pocket, and the node features are composed of the atomic physicochemical properties and their three-dimensional coordinates; the edge construction can be based on atomic geometric distance and interaction type, and the edge features can further include distance, angle, and polar coordinate feature spatial relationship encoding.

[0020] Simultaneously, step S4 constructs a ligand graph based on the ligand's three-dimensional structure, where nodes correspond to ligand atoms and edges correspond to chemical bonds. Node features include atom type, charge, hybridization property, aromaticity, and their three-dimensional coordinates; edge features include bond type and bond length. In addition to the structure graph, step S5 also generates a molecular fingerprint based on the overall ligand structure to provide prior knowledge of global chemical properties for subsequent fusion.

[0021] After constructing the protein and ligand maps, in step S6, both will be input into a polar coordinate-inspired graph attention layer, the internal structure of which is as follows: Figure 2 As shown, the attention module can be composed of multiple alternating stacked Node→Edge and Edge→Node modules. In the Node→Edge module, the system can calculate edge attention and update the edge representation based on the polar coordinate geometric relationship between node features and edges; in the Edge→Node module, edge features converge to nodes to update node representations. This bidirectional alternating structure can simultaneously capture local chemical information and directional three-dimensional spatial relationships.

[0022] The protein and ligand maps updated by the graph attention layer are input into the graph pooling module in step S7, where pooling operations are performed to obtain the corresponding low-dimensional global representations. Subsequently, these graph-level features, along with the protein sequence features obtained in step S2 and the ligand molecular fingerprint features generated in step S5, are input into the feature fusion module in step S8. In this module, multi-source information is integrated to ultimately form complex features for subsequent prediction.

[0023] After the complex features are generated, they are input into a multilayer perceptron in step S9, where several layers of nonlinear transformations are used to map the binding affinity. Finally, step S10 outputs the predicted affinity value. The overall process of this invention realizes the automatic construction and prediction of high-level features of complexes from raw structural data, sequence data, and graph structural data, and can effectively capture the sequence semantics, structural geometry, chemical bond connections, and physical complementarity between proteins and ligands, thereby improving the accuracy of binding affinity prediction.

[0024] Figure 3 This paper illustrates the two-stage edge feature update structure of a→e aggregation (Node-to-Edge) and e→e aggregation (Edge-to-Node) in this invention. The overall structure includes: a node feature input area, an a→e aggregation unit, an angle domain partitioning module (Domain Divider), multiple e→e aggregation units, and the final edge feature fusion and update part.

[0025] exist Figure 3 On the left, node feature a i a j And optional edge geometry information d ij The input is fed into the a→e aggregation unit. This a→e aggregation unit calculates the initial edge features based on the attributes of the two endpoint nodes and the geometric distance, and its output is multiple edge vectors. The a→e aggregation, as a Node→Edge update module, is used to generate the initial semantic and geometric encoding of the edges.

[0026] Subsequently, these edge features were input to Figure 3 The central element is the "Domain Divider." The Domain Divider divides neighboring edges in 3D space according to their angular domains, acting as an intermediary to assign each neighbor to a specific angular domain. The presence of the Domain Divider allows edges to undergo independent depth updates in different regions, improving the model's directionality, discriminative power, and expressiveness.

[0027] To the right of the Domain Divider, the system contains multiple e→e aggregation units. These e→e aggregation modules independently receive edge features from different domains and perform aggregation calculations within their respective domains. After processing by multiple e→e aggregation units, the updated edge features from each subdomain are converged into the final edge update output area. This stage merges the cross-domain edge encodings to obtain the final edge features, which are then used by the subsequent Edge→Node update module.

[0028] Figure 3 It embodies a two-layer message passing and geometric domain partitioning mechanism from node to edge and then from edge to edge, which can simultaneously capture local node information, paired geometric direction information, and structural differences within the boundary domain, thereby enhancing the sensitivity of the graph model to the three-dimensional structure of protein-ligand.

[0029] Figure 4 This paper demonstrates the node feature update process of the e→a aggregation (Edge-to-Node) in this invention. This process receives geometric features, semantic features, and node attributes from multiple associated edges to update the feature representation of the target node, and is the Node update part of the graph attention module.

[0030] exist Figure 4 On the left, multiple edge feature input units are in e ij e kj The form , ... provides messages from different neighboring edges. These edge characteristics can be derived from... Figure 3 The final edge representation after a→e and e→e are aggregated. Meanwhile, Figure 4 Also input the node's own features a j And input the geometric code d corresponding to each edge. ij d kj .

[0031] The above inputs will be fed together into the e→a aggregation unit ( Figure 4 (Medium gray area). This aggregation unit can converge the features of each neighboring edge through attention weighting and combine them with the node's own features to form node update candidates. The aggregation unit can implement directional weighting of edge information, so that edges with different directions and distances contribute differently in the update, thus reflecting the real protein-ligand spatial interactions. After the aggregation operation is completed, the e→a unit outputs the new node feature a. j .

[0032] Figure 4 This demonstrates a message aggregation mechanism from edge to node, which is related to... Figure 3 The Node→Edge structures are coupled together to form a complete message passing loop, enabling the model to be updated efficiently using node semantics, edge geometric direction, and spatial interaction strength information simultaneously. This is suitable for protein-ligand 3D structure modeling.

Claims

1. A method for predicting protein-ligand binding affinity based on graph convolutional networks, characterized in that, Includes the following steps: Data selection steps; Select data from the PDBbind database; Steps for extracting protein features: Constructing a protein atomic diagram representation and generating sequence residue features; Steps for extracting ligand features: Constructing ligand atomic maps and generating chemical features of ligands; Feature vector fusion step: Fusing protein feature vectors with ligand feature vectors; Affinity prediction steps: Input the joint features into the prediction model and output the protein-ligand binding affinity.

2. The protein-ligand binding affinity prediction method based on graph convolutional networks according to claim 1, characterized in that, The data selection steps include: obtaining samples containing the three-dimensional structure of protein-ligand complexes and experimentally determined affinity data from the PDBbind database; selecting high-quality data and dividing them into training and validation sets according to a ratio; and selecting the highest-quality portion as the test set.

3. The protein-ligand binding affinity prediction method based on graph convolutional networks according to claim 1, characterized in that, The steps for extracting protein features include: firstly, constructing an atomic-level representation of the protein based on the three-dimensional structure of the protein pocket region, using atom type, charge, hydrophobicity, hydrogen bond donor and acceptor properties, valence electron information, bond connections, and three-dimensional coordinates as the basic features of nodes and edges, and calculating local geometric descriptors for each atomic node, including interatomic distances, bond angles, and neighborhood topology; simultaneously, using a protein sequence language model to perform representation learning on the residues in the pocket region, extracting sequence features that reflect the physicochemical properties of amino acids, sequence context dependence, and potential structural associations.

4. The protein-ligand binding affinity prediction method based on graph convolutional networks according to claim 1, characterized in that, The steps for extracting ligand features include: constructing a ligand atomic map based on the atom types, charges, aromaticity, and bond types in the ligand molecule; calculating the geometric features of the ligand's three-dimensional conformation; and generating a high-dimensional chemical feature vector describing the functional groups, bond patterns, and local chemical environment using a molecular fingerprinting method.

5. The protein-ligand binding affinity prediction method based on graph convolutional networks according to claim 1, characterized in that, The feature vector fusion step includes: combining protein feature vectors and ligand feature vectors through vector concatenation, cross-modal attention mechanisms, or multilayer perceptron fusion strategies; assigning learnable weights to features from different sources during the fusion process to capture local interactions, global associations, and coordination relationships between proteins and ligands; and finally obtaining a joint representation vector at the complex level.

6. The protein-ligand binding affinity prediction method based on graph convolutional networks according to claim 1, characterized in that, The affinity prediction step includes: inputting the fused joint feature vector into a regression prediction model composed of a multi-layer fully connected network, extracting a high-dimensional representation through a nonlinear activation function and a normalization layer, and generating a protein-ligand binding affinity prediction result in continuous numerical form at the output layer; during model training, RMSE, MAE, SD and Pearson correlation coefficient R are used for evaluation to ensure the accuracy and stability of the prediction results.

Images