Prediction method for identifying protein hidden binding sites

Abstract

The invention discloses a deep learning prediction method fused with multi-modal features, which can accurately identify protein hidden binding sites in a ligand-free (apoo) state. The method comprises the following steps of: firstly, constructing a protein graph by taking residues as nodes and taking C alpha distance less than or equal to 14 as edges, wherein node feature sets comprise amino acid one-hot, secondary structures, atomic attributes, protein language model embedding and BLOSUM62 evolutionary information, and edge features comprise distance and angle similarity; then capturing three-dimensional geometric equivariant features by adopting an equivariant graph neural network (EGNN), and modeling a chemical topological relation by using a graph isomorphic network (GINE) with edge features; eGNN and GINE double-branch feature fusion and global dependence integration are realized through gating cross attention and gating multi-head attention; and finally, inputting the fusion features into a Kolmogorov-Arnold network (KAN) classifier, and predicting whether each residue belongs to a hidden binding site or not. The method can adapt to large-scale conformation change without coordinate alignment, AUC and F1 on a standard data set are remarkably superior to those of an existing method, high robustness and generalization are kept for multi-chain protein and complex conformation, and the method can be widely applied to drug target discovery and structure-driven drug design.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of structural bioinformatics and computational drug design, and specifically relates to a protein cryptic binding site prediction method based on multi-modal feature fusion, which can be used to identify potential cryptic binding pockets in protein structures in the absence of ligands (apo) to assist drug target discovery and structure-driven drug design. BACKGROUND

[0002] The protein cryptic binding site refers to an important functional region that is usually invisible or difficult to identify when the protein is not bound to a ligand (apo state), but is revealed by conformational changes after ligand binding. The formation of such binding sites often results from the dynamic conformational changes of the protein, and its appearance reveals the flexibility and polymorphism of the protein molecule at the structural level. Cryptic binding sites have important value in drug research and development, protein function regulation, and target discovery. It can provide new binding regions for traditional difficult-to-drug targets, expand the druggable space for drug design, and provide new ideas for allosteric regulation and selective inhibition.

[0003] However, since the cryptic binding site is often closed, collapsed or hidden in the static structure, it is difficult to be directly observed or identified in the absence of ligand crystal structure. This makes the traditional pocket recognition, molecular docking and energy scoring methods that rely on static structure have limited effect in such tasks, and it is difficult to reflect the true binding characteristics of the protein under dynamic conditions.

[0004] Existing technologies in the identification of cryptic binding sites mainly develop along three paths: 1) Structure sampling methods based on molecular dynamics simulation. This kind of method explores the conformational changes of proteins in different energy states through long-time molecular dynamics simulation or enhanced sampling techniques, so as to identify potential pockets temporarily formed in the dynamic process. Such methods can reveal the formation mechanism of cryptic pockets from the physical level, but they have high computational cost, low sampling efficiency, and are highly dependent on potential energy functions and initial structures, making it difficult to apply to large-scale protein systems. 2) Static structure prediction methods based on machine learning. This kind of method usually extracts protein sequence, geometric shape, surface accessibility, hydrophobicity and other features, and uses classification models or ensemble learning methods to predict potential binding regions. The advantage of this route is fast calculation speed and easy promotion, but it relies on manual feature engineering and is difficult to capture the complex dynamic behavior of protein conformational changes, and has limited prediction ability for cryptic sites. 3) Multi-modal feature modeling methods based on deep learning. With the development of deep learning, researchers began to use neural networks to automatically learn multi-level structure representations of proteins. Common practices include using convolutional networks to extract features from three-dimensional structure voxel representations, or using graph neural networks to model residues as nodes and learn the topology and spatial relationship of proteins through message passing mechanisms. This kind of method has stronger automatic feature extraction ability than traditional machine learning, and can capture the geometric dependence between residues to some extent. However, traditional deep models still have several limitations: • Lack of isometric constraints for spatial rotation, translation and other geometric transformations, making it sensitive to three-dimensional conformational changes; • Lack of multi-source feature fusion (such as structure information, sequence information and evolutionary information), making it difficult to fully capture the multi-factor mechanism of cryptic site formation; • Existing training data is mostly based on ligand-bound state structures, and the model's generalization ability for real ligand-free structures is insufficient; • Some methods have poor prediction stability in multi-chain proteins or complex systems containing missing regions, with high false positive rate. SUMMARY

[0005] The purpose of the present application is to overcome the problems of low accuracy in identifying cryptic binding sites of proteins and poor adaptability to large-scale conformational changes in the prior art, and to propose a multi-modal deep learning prediction method that integrates protein structure graph features and multi-source sequence features.

[0006] The technical solution of the present application comprises the following steps: 1) Constructing protein graph structure: protein residues are taken as nodes, and edges are connected when the distance between Cα atoms is less than a threshold (14 Å); node features include amino acid one-hot encoding, DSSP secondary structure features, atom attributes, protein language model embedding (PLM), and BLOSUM62 evolutionary features; edge features include inter-residue Euclidean distance and angle similarity. 2) Feature extraction: isometric graph neural network (EGNN) is used to capture the isometric features of three-dimensional geometric structure, and graph isomorphism network with edge features (GINE) is used to model the chemical and topological relationship between nodes. 3) Feature fusion: introduce a gated cross-attention mechanism to realize the bidirectional feature interaction between EGNN and GINE branches, and integrate global dependency through a gated multi-head attention mechanism. 4) Classification and prediction: input the fused features into the classifier based on Kolmogorov-Arnold network (KAN), and improve the non-linear expression ability through the learnable B-spline activation function to predict whether each amino acid residue belongs to the cryptic binding site. Advantages

[0007] 1) The present application realizes explicit modeling of the three-dimensional geometric relationship of the protein through the isometric characteristics of EGNN, which can capture conformational changes without coordinate alignment; 2) The GINE module introduces edge features to enhance the expression ability of the graph structure, which can more accurately describe the chemical relationship between residues; 3) The gated attention mechanism realizes multi-level feature fusion, effectively balancing local geometric features and global semantic dependencies; 4) The KAN classifier has stronger function approximation and interpretability, and realizes high-precision prediction with fewer parameters;

[0008] The comprehensive results show that the AUC and F1 score of the method of the present application on the standard data set are significantly better than those of the existing mainstream methods, and the method has high robustness and generalization under multi-chain proteins and complex conformations. The present application provides a new method for predicting cryptic binding sites of proteins. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 The overall flowchart of the method of the present application, including four stages of data input, graph construction, feature extraction and prediction;

[0010] Figure 2 The structure diagram of the gated cross-attention and multi-head attention fusion mechanism; DETAILED DESCRIPTION

[0011] The specific embodiments of the present application are described in detail below with reference to the accompanying drawings. It should be noted that the accompanying drawings are for illustrative purposes only and should not be construed as limiting the present application, and are used only for reference and illustration, and do not constitute a constraint on the scope of patent protection. Without departing from the core idea and technical scope of the present application, the present application can be changed and optimized in many ways.

[0012] As Figure 1 shown, the figure shows the overall workflow of the present study. The model architecture contains three main parts, graph construction, feature extraction and classification. The feature extraction adopts a double-branch neural network structure, which extracts protein features from different angles respectively. Then, the model realizes the deep fusion of complementary features extracted by the two branches through the Gated Cross-Attention layer, which adaptively adjusts the contribution ratio of different branch information in the feature interaction process by using the gating mechanism; then through the Gated Multi-Head Attention layer, the fused features are further integrated to capture global dependency relationships and multi-scale feature patterns, so as to generate node-level embedding representations with sufficient information interaction and global consistency. Figure 2 Finally, the unified feature representation is taken as the input of the Kolmogorov-Arnold Networks (KAN) classifier to predict whether each amino acid residue is a potential buried binding site.

[0013] In order to evaluate the proposed model, we use two datasets, the first dataset is used to test its prediction ability on protein buried binding sites, and the second dataset is used to test its generalization ability on protein-protein interaction sites (PPIS). The model uses the same network structure and parameters in two experiments, the only difference is the input data.

[0014] The experiment adopts a unified data division and evaluation protocol, and the evaluation indexes include weighted F1 score, Matthews correlation coefficient (MCC), area under ROC curve (AUROC), area under precision-recall curve (AUPRC) and accuracy (ACC) and the like. The comparison results of the independent test set are shown in Table 1, and the generalization test results are shown in Table 2.

[0015] Table 1. Performance comparison based on independent test set and existing methods Method Weighted F1 MCC AUROC AUPRC ACC P2RANK 0.9049 0.3226 0.7416 0.3828 0.8829 PocketMiner 0.8500 0.1855 0.7419 0.1576 0.7990 Gated-GPS 0.9323 0.2174 0.8607 0.3109 0.9476 PLM-NN 0.9273 0.3947 0.8600 0.3621 0.9321 CrypKANet 0.9420 0.4510 0.9005 0.4039 0.9383

[0016] Table 2. Performance comparison of the model on the PPIS dataset Method F1 MCC AUROC AUPRC ACC GraphPPIS 0.451 0.333 0.786 0.429 0.776 AGAT-PPIS 0.569 0.484 0.867 0.574 0.856 Gated-GPS 0.627 0.555 0.896 0.688 0.879 CrypticKAN 0.572 0.563 0.901 0.700 0.892

Claims

1. A prediction method for identifying cryptic binding sites of proteins, characterized in that, The method comprises the following steps: 1) Protein graph construction: define the amino acid residues in the protein structure as nodes, and establish an edge connection when the Euclidean distance between the Cα atoms of any two residues is less than 14 angstroms; The node features include amino acid one-hot encoding, secondary structure features, atomic property features, protein language model embedding, and evolutionary features based on BLOSUM62; the edge features include Euclidean distance features and angle similarity features between residues; 2) Feature extraction: the feature extraction adopts a double-branch neural network structure, which extracts protein features from different angles. The EGNN branch takes the three-dimensional geometric structure of the protein as the core, and performs isometric message passing combined with node features and spatial coordinates. The node coordinates are dynamically adjusted during feature updating, so as to capture the spatial conformation mode and geometric relationship of the protein. The GINE performs message passing between nodes through an MLP with learnable weights, and introduces edge attributes to enable it to more accurately capture the relationship information in the protein graph and enhance the modeling ability of the semantic and chemical properties between protein residues; 3) Feature fusion: the model realizes the deep fusion of the complementary features extracted by the two branches through a gated cross-attention layer (Gated Cross-Attention), and uses a gating mechanism to adaptively adjust the contribution proportion of the information of different branches during feature interaction; then, the fused features are further integrated through a gated multi-head attention layer (Gated Multi-Head Attention) to capture global dependency relationships and multi-scale feature patterns, thereby generating node-level embedding representations with sufficient information interaction and global consistency; 4) Classification and prediction: input the fused features into a classifier based on the Kolmogorov-Arnold network, and use a learnable continuous function activation mechanism to realize nonlinear expression, and output the probability value of whether each amino acid residue is a cryptic binding site; 5) Model optimization method: a variable learning rate is adopted, and the learning rate ranges from 0.0002 to 0.

002. A dropout rate of 0.2-0.6 is configured to prevent model overfitting. A joint loss function is used to optimize the model, which includes a cross-entropy loss and a Tversky loss to balance the class imbalance and improve the recognition accuracy of minority class samples.

2. The method of claim 1, wherein The isometric graph neural network maintains the isometry of geometric transformations such as translation, rotation, and reflection during message passing, to realize consistent representation of protein spatial structure under different coordinate systems. The gated cross-attention module adaptively adjusts the information flow proportion from different branch features by setting a learnable gating matrix, to realize dynamic fusion of local geometric information and global semantic features. The Kolmogorov-Arnold network uses a learnable one-dimensional function based on B-spline as an activation unit to enhance the nonlinear expression ability of the network and improve the model interpretability.

3. The method of predicting a protein cryptic site according to claim 1, wherein The method can be widely applied to prediction of cryptic binding sites of ligand-free proteins, mining of potential drug binding pockets, protein function annotation and structure-driven drug design. The method provides a new method for prediction of cryptic sites of proteins.

Images