Prediction method for identifying protein-protein interaction sites

Abstract

The invention discloses a prediction method for identifying a protein-protein interaction site. According to the method, a deep learning framework is constructed based on protein chain hierarchical graph construction and double-branch graph neural network learning. The method comprises the following steps: firstly, constructing a corresponding hierarchical graph for each protein chain, including a residue-level graph and an atomic-level graph; next, the method constructs an isotropic graph neural network (EGNN) module to extract residue-level embedding with isotropic features. After a GraphSAGE module is further used for extracting atomic-scale embedding, the method integrates layered graph features by using a comparative learning strategy so as to learn residue-scale and atomic-scale consistency embedding. Finally, an improved gated multi-head attention mechanism is used for fusion embedding weighting, and residue-level interaction classification prediction is performed through a multi-layer perceptron. A comprehensive evaluation result on a plurality of data sets shows that the method is remarkably superior to an existing method. In conclusion, according to the method, various advanced technologies such as graph structure denaturation, multi-scale modeling and comparative learning are fused, and an accurate and steady solution with expansion potential is provided for protein interaction prediction. The method can assist researchers in deeply recognizing key protein acting sites and promoting analysis work of cell signal channels and metabolic networks.

Description

Technical Field

[0001] The present invention relates to the intersection of artificial intelligence and bioinformatics, and specifically to a molecular representation learning method for identifying protein-protein interaction sites. The method is suitable for multiple life science application scenarios such as drug target identification, biological network modeling, and disease mechanism research. Background Art

[0002] Proteins, core molecules that perform diverse functions within organisms, play an irreplaceable role in life processes such as signal transduction, metabolic regulation, and substance transport. Protein-protein interactions form the foundation of intracellular molecular networks and determine the dynamic regulation of complex biological processes. Accurate identification of protein-protein interaction sites is crucial for understanding disease mechanisms, designing drug targets, and regulating the function of protein complexes, particularly in the development and progression of diseases.

[0003] While traditional experimental methods such as yeast two-hybrid (Y2H) and affinity purification-mass spectrometry (AP-MS) can generate highly reliable protein interaction data, they are limited by their long experimental cycles, high costs, and limited coverage, making them inadequate for large-scale omics studies. Therefore, developing efficient and accurate computational prediction methods to identify PPI binding sites has become an important research direction in computational structural biology.

[0004] Existing methods for predicting protein-protein interaction sites can be broadly categorized into two main groups: sequence-based and structure-based. Sequence-based methods rely on amino acid sequences and evolutionary information, primarily focusing on linear sequence features. These methods use only sequence features as input and apply deep learning techniques to capture hidden residue interaction patterns within the sequence. Sequence-based methods are more versatile and are particularly effective when the protein structure is unknown. Notably, the introduction of AlphaFold provides a reliable solution for protein structure prediction. Furthermore, advances in graph neural networks (GNNs) have also promoted the efficient utilization of protein three-dimensional structural information. Conversely, the performance improvements of structure-based methods have benefited from the rich representation of protein structural features and the widespread application of graph neural networks. With the development of structure prediction tools such as AlphaFold, the richness of protein three-dimensional structures has provided a solid foundation for computational modeling. Many methods, based on graph neural networks (GNNs), model protein structures as graphs, with residues as nodes and spatial adjacencies as edges, thereby learning structural embeddings.

[0005] Although existing computational methods have made great progress, they still face some challenges. First, sequence-based methods ignore the topological structure of proteins and lack the representation of important structural information. In contrast, structure-based methods can utilize three-dimensional spatial information to provide more accurate predictions. However, existing structure-based methods are generally based on residue graph modeling, lack the representation of finer-grained atomic graphs, and ignore the information interaction between the two hierarchical features. In addition, structure-based methods generally use GCN and its variants to build models, but traditional GNN modules lose the original spatial structure of proteins while aggregating information. In contrast, equivariant graph neural networks (EGNNs) can maintain the equivariance of proteins in three-dimensional space while aggregating node information, and have been widely used in many studies.

[0006] Therefore, developing a prediction method that can not only effectively learn the hierarchical graph information of protein chains but also extract isovariant features in the graph structure has become an important direction in the research of protein interaction site prediction. Summary of the Invention

[0007] The purpose of the present invention is to solve the problem of accuracy in the prediction of protein-protein interaction sites, and to provide a method for merging hierarchical Figure 1 A deep learning architecture based on homology information is used to accurately identify potential protein-protein interaction sites.

[0008] First, this method constructs a hierarchical graph based on the protein sequence, consisting of a residue-level graph and an atom-level graph. Each residue in the residue-level graph corresponds to a set of atoms in the atom-level graph. The initial node embeddings of the residue-level graph are derived from the protein's sequence and structural features, while the atom-level node embeddings are initialized using binary encodings of the atom types. This method designs a dual-channel graph deep learning architecture to capture structural information at different levels: the residue-level graph is modeled by an EGNN module, and the atom-level graph is modeled by a GraphSAGE module. Subsequently, this method uses graph contrastive learning to align and integrate the two-level graph embeddings, thereby enhancing feature consistency across the hierarchy. The fused residue embeddings are further transferred to an improved gated multi-head attention module to update the embedding weights. Finally, a multi-layer perceptron (MLP) is used to perform the PPI prediction task. Comprehensive experimental results on two public datasets demonstrate that this method significantly outperforms existing methods.

[0009] The following is a technical solution for achieving the purpose of the present invention, comprising the following steps: 1) Benchmark dataset construction: Standard datasets for publicly available PPI prediction tasks (such as DeepPPISP and GraphPPIS) were selected and screened according to unified criteria to ensure low protein sequence homology and structural quality (e.g., resolution <3Å). Redundant or anomalous structures in the dataset were cleaned, ultimately forming a training set and an independent test set containing residue annotation labels. 2) Protein hierarchical graph construction: For each protein sequence, a corresponding hierarchical graph structure is constructed, including: • Residue-level graph: Residues are nodes, and edges are defined by the spatial distance between residues in the 3D structure (e.g., a 16Å threshold). Node features include 20-dimensional amino acid encoding, PSSM, HMM, ESM-2 embedding, and structural features such as DSSP, resAF, and pseudo-position information, for a total of 115 dimensions; • Atomic-level graph: Each atom in the protein is a node, and edges are constructed based on the 2.3Å interatomic distance. Node features are one-hot encodings of atom types (37 categories); 3) Construction of a residue-level graph feature extraction module: Using an equivariant graph neural network (EGNN) model, stacking multiple equivariant graph convolutional layers (EGCLs), this module extracts node embeddings and coordinate representations while maintaining 3D structural equivariance, thereby learning global structural representations. 4) Atomic-level graph feature extraction module construction: Using the GraphSAGE module, we implement neighbor sampling and mean aggregation strategies to learn atomic-level node embedding representations in the local structure space; 5) Molecular graph contrastive learning strategy fusion: To address the information inconsistency problem between hierarchical structures, a graph contrastive learning mechanism is introduced to fuse features from different hierarchies; 6) Residue-level embedding weighting mechanism: The fused embedding is input into an improved gated multi-head attention mechanism, which ultimately outputs a weighted residue embedding representation; 7) Model training and evaluation system: Multiple datasets are used for full validation. The final model performance is evaluated using the core evaluation metrics AUC, AUPR, and MCC to suit the recognition task under imbalanced residue categories. The model generalization ability is also verified on an independent test set.

[0010] This model is a novel graph-based representation learning method for predicting protein-protein interaction sites. It applies GNNs to learn hierarchical graph molecular representations on residue-level and atomic-level protein graphs. A contrastive learning strategy and an attention mechanism are then used to update residue embeddings. Finally, the fused embeddings are fed into an MLP to predict probability scores for protein-protein interaction sites. This architecture constructs a hierarchical graph representation based on the inherent three-dimensional hierarchical conformation of protein chains and learns residue-level embeddings through an EGNN module, providing a new approach for predicting protein-protein interaction sites. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 A flowchart of a method for predicting protein-protein interaction sites;

[0012] Figure 2 The ROC curve and PR curve of the model in multiple data sets. DETAILED DESCRIPTION

[0013] The following is a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings. It should be noted that the drawings are for illustrative purposes only and should not be construed as limiting the present invention. They are for reference and illustration purposes only and do not constitute a constraint on the scope of patent protection. The present invention can be subjected to various changes and optimizations without departing from the core concept and technical scope of the present invention.

[0014] like Figure 1 As shown in the figure, the overall workflow of this study is presented. This study consists of multiple modules: A. Generate hierarchical graphs consistent with protein sequences, including residue-level graphs and atomic-level graphs. B. Take the residue-level graph representation as input and use the EGNN module to learn the residue-level node embedding representation. C. Take the atomic-level graph representation as input and use the GraphSAGE module to learn the atomic-level node embedding representation. D. Synergistically integrate the embedding representations of the residue-level graph and the atomic-level graph through a graph contrast learning strategy. E. Use an improved gated multi-head attention mechanism to weight the residue node embedding. F. Perform graph node classification for residue-level PPIs prediction.

[0015] To objectively evaluate this method, this study selected two datasets derived from previously representative protein interaction site prediction tasks, proposed by DeepPPISP and GraphPPIS, respectively. Both datasets were obtained by screening and curating the RCSB Protein Data Bank (PDB). Furthermore, we fine-tuned the data by removing anomalous protein chains whose structures were inconsistent with the PDB database. Details of the two benchmark datasets used in the experiments are summarized in Table 1.

[0016] Table 1. Details of the benchmark datasets based on the two research tasks Task data Protein quantity Number of binding sites Number of non-binding sites Sparse ratio DeepPPISP Train352 352 11079 62102 15.14 Test70 70 2332 9459 19.78 GraphPPISP Train_335-1 334 10374 66366 13.52 Test_60 60 2075 11069 15.79 Test_315-28 287 8566 51810 14.19 Ubtest_31-6 25 711 5206 12.02

[0017] In this study, we used seven widely used machine learning evaluation indicators to verify the model performance, including accuracy (ACC), precision, recall, F1 score (F1), Markov coefficient (MCC), area under the ROC curve (AUC) and area under the PR curve (AUPR). To further evaluate the ability of this method to identify potential circRNA-disease associations, we plotted its ROC curve and PR curve on the benchmark dataset, as shown in Figure 2. Figure 2 shown.

[0018] Finally, we compared our proposed method with existing methods. The results are reported in Tables 2, 3, and 4. These results demonstrate that our proposed method significantly improves performance across all datasets, validating its ability to effectively identify protein-protein interaction sites.

[0019] Table 2. Performance comparison with existing methods on Test70 method ACC Precision Recall F1 MCC AUC AUPR <![CDATA[DeepPPISP b ]]> 0.655 0.303 0.577 0.397 0.206 0.671 0.320 <![CDATA[DELPHI a ]]> 0.667 0.320 0.604 0.418 0.236 0.690 N / A <![CDATA[HN-PPISP a ]]> 0.667 0.324 0.632 0.427 0.244 NA 0.360 <![CDATA[EnsemPPIS a ]]> 0.732 0.375 0.532 0.440 0.277 0.719 0.405 <![CDATA[EGRET b ]]> 0.715 0.358 0.516 0.438 0.270 0.719 0.405 <![CDATA[HSSPPI b ]]> 0.690 0.630 0.696 0.661 0.319 0.759 0.434 The method 0.851 0.621 0.639 0.630 0.537 0.880 0.682

[0020] Table 3. Performance comparison with existing methods on Test60 method ACC Precision Recall F1 MCC AUC AUPR <![CDATA[GraphPPIS a ]]> 0.776 0.368 0.584 0.451 0.333 0.786 0.429 <![CDATA[DELPHI a ]]> 0.697 0.276 0.568 0.372 0.225 0.699 0.319 <![CDATA[EnsemPPIS a ]]> 0.716 N / A N / A 0.401 0.266 0.726 0.352 <![CDATA[AGAT-PPIS b ]]> 0.856 0.539 0.603 0.569 0.484 0.867 0.574 <![CDATA[GACT-PPIS b ]]> 0.854 0.530 0.646 0.583 0.499 0.876 0.595 <![CDATA[HSSPPI b ]]> 0.811 0.650 0.657 0.653 0.307 0.756 0.434 The method 0.873 0.584 0.668 0.623 0.549 0.890 0.656

[0021] Table 4. Performance comparison with existing methods on Test_315-28 and Ubtest_31-6 method Test_315-28 Ubtest_31-6 MCC AUC AUPR MCC AUC AUPR <![CDATA[GraphPPIS a ]]> 0.432 N / A N / A 0.327 0.780 0.365 <![CDATA[AGAT-PPIS b ]]> 0.481 0.874 0.572 0.356 0.796 0.368 <![CDATA[HSSPPI b ]]> 0.273 0.743 0.323 0.338 0.800 0.392 The method 0.511 0.885 0.595 0.401 0.845 0.445

Claims

1. A prediction method for identifying protein-protein interaction sites, the process comprising the following steps: 1) Benchmark dataset construction: Standard datasets for publicly available PPI prediction tasks (such as DeepPPISP and GraphPPIS) were selected and screened according to unified criteria to ensure low protein sequence homology and structural quality (e.g., resolution <3Å). Redundant or anomalous structures in the dataset were cleaned, ultimately forming a training set and an independent test set containing residue annotation labels. 2) Protein hierarchical graph construction: For each protein sequence, a corresponding hierarchical graph structure is constructed, including: • Residue-level graph: Residues are nodes, and edges are defined by the spatial distance between residues in the 3D structure (e.g., a 16Å threshold). Node features include 20-dimensional amino acid encoding, PSSM, HMM, ESM-2 embedding, and structural features such as DSSP, resAF, and pseudo-position information, for a total of 115 dimensions; • Atomic-level graph: Each atom in the protein is a node, and edges are constructed based on the 2.3Å interatomic distance. Node features are one-hot encodings of atom types (37 categories); 3) Construction of a residue-level graph feature extraction module: Using an equivariant graph neural network (EGNN) model, stacking multiple equivariant graph convolutional layers (EGCLs), this module extracts node embeddings and coordinate representations while maintaining 3D structural equivariance, thereby learning global structural representations. 4) Atomic-level graph feature extraction module construction: Using the GraphSAGE module, we implement neighbor sampling and mean aggregation strategies to learn atomic-level node embedding representations in the local structure space; 5) Molecular graph contrastive learning strategy fusion: To address the information inconsistency problem between hierarchical structures, a graph contrastive learning mechanism is introduced to fuse features from different hierarchies; 6) Residue-level embedding weighting mechanism: The fused embedding is input into an improved gated multi-head attention mechanism, which ultimately outputs a weighted residue embedding representation; 7) Model training and evaluation system: Multiple datasets are used for full validation. The final model performance is evaluated using the core evaluation metrics AUC, AUPR, and MCC to suit the recognition task under imbalanced residue categories. The model generalization ability is also verified on an independent test set.

2. This method constructs a deep learning architecture that uses an equivariant graph neural network to learn hierarchical graph representations. This method builds a hierarchical graph representation based on the inherent three-dimensional hierarchical conformation of protein chains, fully extracting structural information. Secondly, unlike traditional GNNs, this method uses an EGNN as the backbone network, maintaining the spatial equivariance of protein molecules during feature learning, thereby extracting global features with equivariant-aware properties from the residue-level graph. Finally, a contrastive learning strategy is used to collaboratively optimize the consistency features of the hierarchical graph, and an enhanced gated multi-head attention mechanism is used to perform weighted updates to the residue embeddings to improve the quality of the embedding representation.

3. A prediction method for identifying circular RNA-disease association according to claim 1, characterized in that The feature encoding scheme and model construction described in step 2), step 3), step 4), step 5) and step 6) can effectively integrate multi-view similarity information and capture global structural features and local features through multi-level feature fusion and dual-branch modeling, providing a new method for protein-protein interaction prediction.

Images