A protein interaction prediction method based on a heterogeneous graph neural network

Abstract

This invention discloses a protein interaction prediction method based on heterogeneous graph neural networks, belonging to the field of protein interaction prediction technology. This invention represents proteins as heterogeneous graphs containing three types of relationships and introduces a relationship modeling mechanism using static weights and learnable scaling parameters. A bimodal contrastive learning framework for sequence pathways and structural pathways is constructed to improve the robustness and generalization of node representations. Furthermore, a relationship-level multi-head adaptive attention mechanism module is designed to adaptively adjust the contribution of different relationships under perceptual conditions. This is a unified framework that can finely express residue-level structural information, maintain controllable model complexity, and enhance generalization ability for novel proteins.

Description

Technical Field

[0001] This invention relates to the field of protein interaction prediction technology, and more specifically to a protein interaction prediction method based on heterogeneous graph neural networks. Background Technology

[0002] Protein-protein interactions (PPIs) constitute the molecular basis of almost all key biological processes within living organisms. From the regulation of genetic information expression and intracellular signal transduction cascades to the precise initiation of immune responses and the maintenance of metabolic network homeostasis, all rely on a complex and intricate dynamic network woven from numerous PPIs. A profound and systematic analysis of these interactions is not only essential for understanding the core principles of life, but also provides crucial theoretical support for the mechanistic analysis of many diseases (such as cancer and neurodegenerative diseases), the discovery of innovative drug targets, and the design of personalized treatment plans.

[0003] For a long time, the identification and characterization of protein-protein interactions (PPIs) have relied primarily on traditional biological experimental techniques, such as yeast two-hybrid assays, immunoprecipitation, or mass spectrometry. While these methods are considered the gold standard, their low throughput, long processing times, high costs, and requirement for specialized experimental environments make it exceptionally difficult to map complete protein-protein interaction at scale. More importantly, these methods are typically designed for specific cellular environments or conditions, making it difficult to capture the dynamic changes and spatiotemporal specificity of PPIs, thus limiting our understanding of protein functional networks from a global perspective.

[0004] To overcome the bottlenecks of experimental techniques, computational biology methods, especially deep learning technology that has emerged with the rise of artificial intelligence, have opened up new avenues for large-scale, high-throughput PPI prediction. Existing deep learning-based prediction models, based on the information dimensions they utilize, primarily form a prediction paradigm dependent on amino acid sequences. This method uses the protein's amino acid sequence as the sole input, automatically extracting evolutionary conserved patterns, local physicochemical properties, and other features from the sequence through models such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs), and then inferring interactions. While this method has attracted considerable attention due to its wide applicability and ease of implementation, its fundamental flaw lies in completely ignoring the three-dimensional spatial structure of proteins. The mutual recognition and binding between proteins is essentially a "lock and key" process driven by three-dimensional conformation. The lack of structural information undoubtedly causes the model to lose the crucial spatial complementarity and force details that determine the specificity of interactions, resulting in an insurmountable ceiling to prediction accuracy.

[0005] Against this research backdrop, to overcome the limitations of sequence-driven methods, a protein-to-protein (PPI) prediction paradigm centered on protein structure has emerged in recent years. This paradigm aims to utilize graph neural networks to model the topological and physical interactions between residues within proteins, thereby compensating for the lack of spatial information in pure sequence models. However, these structure-based methods still face several key bottlenecks in practical applications: First, high computational and storage costs. Deep residue-level modeling of each protein incurs significant memory and time overhead, limiting the scalability of the method on large-scale datasets or long-sequence proteins. Second, limited generalization ability to new proteins or low-coverage regions. Existing methods struggle to infer new interaction cues from limited or incomplete observations, thus affecting their generalization performance in real-world scenarios. Therefore, this paper proposes a protein interaction prediction method based on heterogeneous graph neural networks. Summary of the Invention

[0006] The technical problem to be solved by this invention is: how to extract key features that can represent proteins, and based on obtaining these representative features, accurately and efficiently predict protein-protein interactions. At the same time, the prediction is automated by machines to reduce laboratory research costs, and a protein interaction prediction method based on heterogeneous graph neural networks is provided.

[0007] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:

[0008] S1: Protein Data Preprocessing

[0009] We obtained the amino acid sequences of proteins and the corresponding publicly available datasets of protein-protein interactions. We used the AlphaFold2 structure prediction model to predict the structure of the amino acid sequences to be studied, obtained the three-dimensional coordinates of each amino acid sequence, and divided the dataset into training set, validation set and test set according to different modes and set ratios.

[0010] S2: Protein Feature Extraction and Feature Fusion

[0011] Design sequence feature extractors and structural feature extractors, input the extracted sequence and structural features into the comparison learning training module, and learn the feature differences from different perspectives;

[0012] S3: Constructing a heterogeneous protein map

[0013] Each protein is represented as a heteroproton graph, which simultaneously expresses three types of relationships: linear sequence adjacency, geometric adjacency based on spatial KNN, and contact graph based on distance threshold. The static weights and learnable scaling parameters of each type of relationship are preserved to achieve weighted fusion at the relationship level.

[0014] S4: Constructing a heterogeneous graph neural network based on multi-relationship graph attention.

[0015] A heterogeneous graph neural network is constructed, which adopts a graph convolutional structure based on multi-relation modeling and introduces a relation-level multi-head self-attention mechanism to achieve dynamic fusion of three types of relations and generate protein node embeddings.

[0016] S5: Training Heterogeneous Graph Neural Networks

[0017] A heterogeneous graph neural network is trained using a training set to obtain a protein-protein interaction prediction model;

[0018] S6: Prediction of protein-protein interactions

[0019] The unknown protein is input into the protein-protein interaction prediction model to obtain the prediction results.

[0020] Furthermore, in step S1, the specific processing procedure is as follows:

[0021] S11: Obtain the publicly available dataset of the amino acid sequence and interactions of the protein, and use the AlphaFold2 structure prediction model to infer the three-dimensional conformation of the amino acid sequence, generating spatial coordinate information of all atoms, including the main chain and side chains. The AlphaFold2 structure prediction model is based on multiple sequence alignment, attention network, structural template retrieval, and geometric reasoning mechanism of the Evoformer module. The output three-dimensional structure is stored in PDB format, containing the atomic position, residue type, and local geometric parameters of each amino acid residue.

[0022] S12: Divide the amino acid sequences into training set, validation set and test set according to a set ratio. The partitioning mode includes random mode, depth-first search mode and breadth-first search mode.

[0023] Furthermore, in step S12, the three partitioning modes are defined as follows:

[0024] Random mode: Assuming that all PPI pairs are independent of each other, without considering the topology of the protein network, proteins or interaction pairs are randomly divided according to a preset ratio to form training set, validation set and test set.

[0025] Depth-first search mode: Based on the protein network topology, long paths formed during the depth search process and their corresponding long-distance node pairs are preferentially selected as the test set and validation set. The search proceeds along a path to explore to the end and then backtracks, prioritizing long paths and long-distance connections in the protein network. The training set consists of the remaining protein pairs that were not selected for the validation set and test set.

[0026] Breadth-first search mode: Based on the topology of the protein-protein interaction network, starting from a randomly selected root node, the network is traversed in a "near to far, layer by layer" manner. The first-order neighbors of the current node are explored in breadth-first order, and then the network is expanded to the next level. The nodes selected during the traversal and their internal edges are divided into test set and validation set. Correspondingly, the training set is composed of the remaining proteins and their interaction pairs after removing the nodes of the validation set and test set.

[0027] Furthermore, the specific process of step S2 is as follows:

[0028] S21: Seven residue-level features are used as the initial feature vector of the protein. The seven residue-level features are isoelectric point, polarity, acidity / basicity, hydrogen bond acceptor, hydrogen bond donor, octanol-water partition coefficient, and topological polar surface area.

[0029] S22: The sequence feature extractor is used to learn discriminative sequence representations from the primary sequence of a protein. The specific process is as follows: local sequence patterns are modeled through multiple one-dimensional depthwise separable convolutional layers; after each one-dimensional depthwise separable convolutional layer, a squeeze-incentivized attention mechanism is introduced to adaptively recalibrate the importance of different feature channels; then, through pooling and fully connected mapping, the variable-length sequence features are compressed into a fixed-dimensional global sequence embedding vector, which is used to characterize the contextual features of the protein at the sequence level, thus obtaining the final sequence features.

[0030] S23: The structural feature extractor is used to learn the spatial and topological structural representation of a protein from its three-dimensional structural information. The specific process is as follows: A structural contact graph is constructed based on the protein's three-dimensional structural data as input to the structural feature extractor, where nodes represent proteins and edges represent interconnections determined based on spatial distance thresholds. During structural feature modeling, the structural feature extractor is constructed by cascading multiple layers of graph convolutional layers based on GraphSAGE. Skip connections are introduced between these layers. Furthermore, an adaptive graph pooling module based on SAGPooling is set after four layers of graph convolutional layers based on GraphSAGE to learn node importance and perform hierarchical downsampling. Finally, global pooling aggregates the variable-scale node features into a fixed-dimensional structural embedding vector, which is used to characterize the overall three-dimensional structure and topological properties of the protein, thus obtaining the final structural features.

[0031] S24: The extracted sequence and structural features are input into the contrastive learning training module. During the training phase, the model outputs corresponding sequence embedding vectors and structural embedding vectors for the same protein sample through the sequence feature extractor and structural feature extractor, respectively. These two vectors are used as a pair of positive samples and input into the contrastive learning module. The embedding vectors of different protein samples in the same batch constitute negative samples. After normalizing the sequence embedding vectors and structural embedding vectors, the contrastive learning module calculates the cross-view similarity matrix based on cosine similarity and uses a contrastive loss function with a temperature coefficient to perform bidirectional alignment training on the two embedding vectors. By minimizing this contrastive loss, the feature representations of the same protein in the sequence view and structural view are made close to each other in the shared representation space, while the representation distance between different protein samples is increased. This outputs the sequence feature representation and structural feature representation optimized by alignment constraints, which are then concatenated to obtain the contrastive learning feature vector, which serves as the input feature for the subsequent protein interaction prediction task.

[0032] Furthermore, in step S2, a gate mechanism for heterogeneous feature alignment and fusion is set up. Specifically, a bilinear projection layer is introduced to map the initial feature vector in step S21 and the contrastive learning feature vector in step S24 to the same latent space, and then they are spliced ​​to obtain 128-dimensional protein features.

[0033] Furthermore, in step S3, the heterogeneous protein graph is constructed by modeling three types of semantically distinct relationships. Each type of relationship corresponds to an independent edge type in the graph, as detailed below:

[0034] Sequence Edges: These represent the sequential connection of amino acids in the primary structure. Bidirectional edges are established between each pair of adjacent residues to form a chain-like sequence structure.

[0035] KNN Edges: Based on the spatial proximity relationship between amino acids in the three-dimensional structure of a protein, the K nearest neighbor strategy is used to connect the K nearest amino acid residues on the three-dimensional coordinates, reflecting local structural interactions.

[0036] Threshold Distance Edges: Connects all residue pairs whose spatial distance is less than a set threshold based on a spherical distance threshold, in order to model global structural contacts.

[0037] Furthermore, in step S3, a protein relation network is constructed by integrating Sequence Edges, KNN Edges, and Threshold Distance Edges. Static base weights are preset for each edge type to reflect its prior importance, and learnable scaling parameters are introduced for dynamic calibration to adaptively adjust the actual contribution of different relations in training. The learnable scaling parameters are used to dynamically correct the static base weights. The final weights are obtained by calculating the product of the two, and the contribution of different relations in the feature aggregation process of the heterogeneous graph neural network is adaptively adjusted using the final weights.

[0038] Furthermore, in step S4, an end-to-end heterogeneous graph neural network architecture is adopted. Cross-relational attention calculation is performed on the three types of edge types through the heterogeneous graph multi-head self-attention mechanism to dynamically learn the weight allocation of different interactions. In the training process, the protein heterogeneous graph is processed by multi-layer heterogeneous graph convolution. Each layer aggregates multi-relational features through the multi-head attention mechanism. Then, the training process is stabilized by batch normalization and residual connection. Finally, the node-level representation is transformed into protein embedding through global pooling.

[0039] Furthermore, in step S5, the loss function used for model training is the cross-entropy loss function, as shown in the following formula:

[0040] ;

[0041] in, It is the total number of samples. It is the first The true label of each sample It is the first Predict the probability of each sample.

[0042] Furthermore, in step S6, protein-protein interactions include reactions, binding, post-translational modifications, activation, inhibition, catalysis, and expression. When predicting whether there is an interaction between two proteins, each type of reaction is predicted. If the predicted value of a certain type of reaction is greater than 0.5, it is considered that the protein pair has this type of reaction. Multiple reactions may exist between the same pair of proteins.

[0043] Compared with existing technologies, this invention has the following advantages: This protein interaction prediction method based on heterogeneous graph neural networks represents proteins as heterogeneous graphs containing three types of relationships and introduces a relationship modeling mechanism with static weights and learnable scaling parameters. A bimodal contrastive learning framework for sequence pathways and structural pathways is constructed to improve the robustness and generalization of node representations. Furthermore, a relationship-level multi-head adaptive attention mechanism module is designed to adaptively adjust the contribution of different relationships under perceptual conditions. It is a unified framework that can finely express residue-level structural information, maintain controllable model complexity, and enhance generalization ability for novel proteins. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the protein interaction prediction method based on heterogeneous graph neural networks in an embodiment of the present invention.

[0045] Figure 2 These are schematic diagrams of three partitioning methods in embodiments of the present invention, wherein (a) is the Random mode, (b) is the BFS mode, and (c) is the DFS mode.

[0046] Figure 3 This is a schematic diagram of the heterogeneous graph construction process in an embodiment of the present invention;

[0047] Figure 4 This is a schematic diagram of the overall structure of the protein-protein interaction prediction model in this embodiment of the invention;

[0048] Figure 5 This is a schematic diagram of the structure of the multi-head self-attention aggregation module in an embodiment of the present invention. Detailed Implementation

[0049] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0050] like Figure 1-4 As shown, this embodiment provides a technical solution: a protein interaction prediction method based on heterogeneous graph neural networks, comprising the following steps:

[0051] S1: Protein Data Preprocessing

[0052] Step S1 specifically includes the following:

[0053] S11: In this invention, firstly, publicly available protein amino acid sequence data and their corresponding protein-protein interaction datasets are acquired. Necessary preprocessing of the sequence data is performed, including format normalization, residue character unification, missing site imputation, and invalid entry filtering, to ensure the integrity and consistency of the input data. Subsequently, to overcome the spatial structure gaps caused by relying solely on sequence information, this invention further introduces the advanced structure prediction model AlphaFold2 to infer the three-dimensional conformation of the preprocessed amino acid sequences. AlphaFold2, through collaborative reasoning of multiple deep learning modules, can infer the folding morphology of proteins in physical space based on sequences. In its specific implementation, the model first uses multiple sequence alignment to capture implicit covariation patterns in the sequence evolution history, and then combines a structure template retrieval mechanism to absorb potential conformational information from a known structure database. Subsequently, sequence features and structural features are fed into the core Evoformer module of the model. Through deep attention networks, multi-branch geometric transformations, and spatial relationship reasoning based on triplet constraints, long-term dependencies between residues, local geometric arrangements, and overall topological structure are calculated to achieve accurate prediction of protein folding trends. After multiple rounds of iterative structural optimization, the model outputs a three-dimensional protein structure model containing full atomic-level coordinates, stored in PDB format. This structure file records in detail the main chain and side chain atomic coordinates, residue type identifiers, atomic numbers, bonding relationships, and local geometric information (such as twist angles, bond lengths, and bond angles) for each amino acid residue in the protein. This provides an accurate and reliable data foundation for subsequent residue-level structural feature extraction, spatial neighborhood construction, and multi-relationship protein graph generation.

[0054] S12: The dataset was divided into training, validation, and test sets in a 6:2:2 ratio. To simulate real-world protein distribution, three partitioning modes were used: random, depth-first search, and breadth-first search. Random mode randomly selects proteins for training and testing based on the ratio; depth-first search searches as deeply as possible along the starting protein node, thus partitioning the training and test sets proportionally; breadth-first search expands outwards from the given starting protein node, also partitioning the training and test sets proportionally. The preprocessed data distribution is shown in Table 1. Protein-protein interactions include reactions, binding, post-translational modifications, activation, inhibition, catalysis, and expression. BS, ES, and NS represent two proteins being visible, only one protein being visible, and neither protein being visible during training, respectively.

[0055] Table 1. Statistical data of the preprocessed dataset

[0056]

[0057] S2: Protein feature extraction and feature fusion;

[0058] Step S2 includes the following sub-steps:

[0059] S21: This invention employs a seven-dimensional residue-level physicochemical property vector as the initial feature vector for nodes in the protein graph, including isoelectric point, polarity, pH, number of hydrogen bond acceptors, number of hydrogen bond donors, octanol-water partition coefficient, and topological polar surface area. These features are abstracted from the molecular structure and physicochemical behavior of amino acids, and can fully reflect the key functional characteristics of residues in protein folding, energy distribution, and intermolecular interactions. Through this feature construction method, this invention achieves high-density encoding of the microscopic physicochemical properties of residues in the graph construction stage, enabling subsequent models to effectively capture the energy coupling relationships and topological constraints between amino acid residues, providing a solid underlying characterization foundation for protein sequence-structure fusion learning.

[0060] S22: This invention proposes a lightweight yet efficient sequence feature extraction module that extracts local sequence patterns without increasing computational complexity through depthwise separable convolution (Depth-ConV). Depthwise convolution is used to capture pattern associations in the local amino acid neighborhood, while pointwise convolution further achieves cross-channel feature fusion, thereby significantly reducing the model parameter size while maintaining expressive power.

[0061] This invention introduces an embedded SE attention mechanism (SELayer), which calculates the importance of each channel using global contextual information, enabling the model to dynamically enhance key sequence features and suppress irrelevant pattern noise. This integrated convolution-attention structure allows the sequence feature extractor to have stronger selective representation capabilities while learning local linguistic patterns, improving the accuracy and discriminativeness of sequence representation.

[0062] S23: This invention designs a structural feature extractor composed of GraphSAGE convolutional layers with four different neighborhood aggregation strategies. Through a multi-branch aggregation mechanism, it captures the topological semantics of protein structure maps at different scales from four dimensions: weighted mean, max pooling, LSTM aggregation, and attention aggregation. This heterogeneous aggregation approach allows the model to simultaneously learn local secondary structure features, long-range spatial dependencies, and geometric constraints between residues. Subsequently, this invention utilizes SAGPool to achieve graph-level hierarchical structure compression and filters key residue nodes through self-attention scores, enabling the model to focus on structurally decisive regions. To improve the stability of information flow in deep graph networks, the module further combines residual connections, batch normalization, and nonlinear activation, ensuring gradient stability and enhanced expression even in deep stacking scenarios.

[0063] S24: The extracted sequence and structural features are input into the contrastive learning training module. By learning the feature differences from different perspectives, contrastive learning can uncover the subtle relationships between sequence and structural features, thereby improving the model's generalization ability and robustness. Contrastive learning, as a self-supervised learning method, is widely used in the representation and analysis of molecular features.

[0064] Define a cosine similarity function with temperature scaling:

[0065]

[0066] The purpose of a similarity function is to measure the degree of matching, that is, to bring positive sample pairs closer together and push negative sample pairs further apart. Temperature coefficient. Control the smoothness of the distribution.

[0067] Introduction To optimize the comparison of similarity scores between positive and negative samples and improve the model's learning performance:

[0068]

[0069] in, The negative sample set, i.e., all samples not equal to Protein sample index, From the same protein For positive sample pairs, their similarity scores should be boosted by the model. These are negative sample pairs from different proteins. To block gradient propagation, terms in the denominator do not participate in gradient descent.

[0070] In this embodiment, a gating mechanism for heterogeneous feature alignment and fusion is also designed. A bilinear projection layer is introduced to map the initial feature vector and the contrastive learning feature vector to the same latent space, and then concatenate them to obtain 128-dimensional protein features.

[0071] S3: Constructing a heterogeneous protein map

[0072] Step S3 includes the following sub-steps:

[0073] S31: In the structural graph construction stage of this invention, to fully characterize the multi-level and multi-scale interaction features between protein residues, a heterogeneous protein graph containing multiple relationship types is independently constructed for each amino acid sequence. Specifically, Sequence Edges are first defined based on the linear topological relationships of amino acids in the primary sequence. By establishing bidirectional connections between each pair of adjacent residues, the structural graph fully preserves the main structural continuity of the sequence chain in a graph theory sense. These edges reflect the most basic chemical bonding relationships of the protein at the primary structure level and are an important component in preserving local sequence dependence and main chain constraints.

[0074] Furthermore, to capture local conformational features in the three-dimensional space of proteins, KNNEdges based on the K-nearest neighbor strategy are constructed. Specifically, using the three-dimensional center coordinates of residues as a reference, the K nearest spatial neighbors of each residue by Euclidean distance are retrieved, and graph connections are established between them. In addition, to further capture spatial contact features at the global scale, this invention introduces Threshold Distance Edges based on the spatial distance between residue centers. Under this strategy, any two residues whose three-dimensional distance is below a preset threshold are considered to have a potential global contact relationship, and a corresponding connection is established in the graph. These distance-threshold-based edges can effectively model long-range dependencies and global structural constraints, and have irreplaceable value in describing the overall topological structure of folded proteins. Through the joint modeling of the above three edge types, the constructed heterogeneous protein graph can simultaneously cover primary structural continuity, local geometric neighborhoods, and global spatial contact relationships, thereby comprehensively expressing the natural structural complexity of proteins at the graph structure level.

[0075] S32: After constructing a multi-relationship heterogeneous protein graph, to further improve the effectiveness of information propagation and the controllability of relationship modeling, this invention proposes a hybrid edge weight modulation mechanism combining prior knowledge and learnable parameters. Specifically, for the three types of relationships—Sequence Edges, KNN Edges, and Threshold Distance Edges—this invention first assigns static base weights to each edge type based on their biological importance in protein conformation formation and functional association, to encode the prior contribution of different relationships. For example, sequence connections, as main chain constraints, can have their base weights set to higher values; while long-range structural contacts are relatively sparse in some proteins, they can be assigned appropriate base weights to balance their participation. Based on the static base weights, this invention further introduces learnable relationship scaling parameters, dynamically calibrating the importance of different edge types through an end-to-end training process to adapt to the differences between different proteins, different structural regions, and different task objectives. These scaling parameters, as learnable multiplicative factors, together with the base weights of the edges, determine the final influence of the relationship, enabling the graph neural network to automatically adjust the weight ratio of various structural relationships in the message passing process during training, thereby achieving adaptive fusion of multi-relationship information. This static-dynamic combined edge weighting mechanism not only effectively prevents overfitting or information bias caused by a certain type of relationship dominating in the early stages of training, but also allows the model to automatically optimize the relative importance of different relationships after fully learning residue interaction patterns. Ultimately, this mechanism significantly enhances the expressive power of multi-relationship graphs, enabling heterogeneous graph neural networks to learn high-quality protein residue representations from more reasonable and balanced structural signals.

[0076] S4: Constructing a heterogeneous graph neural network based on multi-relationship graph attention.

[0077] This invention designs a relation-level multi-head self-attention aggregation module (such as...) Figure 5 As shown in the diagram, this module serves as the aggregation core of heterogeneous graph convolution, used to perform cross-relation attention allocation between different edge types. This module encodes three types of residue relations—Sequence, KNN, and Distance—in parallel through a multi-head query-key-value mapping mechanism, and introduces relation bias parameters and gating adjustment structures, enabling the model to adaptively learn the contribution levels of different structural relations.

[0078] In the heterogeneous graph encoding layer, this invention employs a multi-level heterogeneous graph convolution module (this module is a built-in module of the dgl library). Internally, each edge type has an independent graph convolution operator, ensuring that sequence information, local spatial configuration, and global structural contacts maintain independent feature flows during message passing. The features after convolution for each edge type are then uniformly aggregated through the aforementioned relational attention module, forming a heterogeneous representation capable of modeling cross-relational dependencies. This invention introduces batch normalization and linear transformation after each convolutional layer to improve training stability and enhance the non-linearity of feature representation, while layer Dropout prevents feature overfitting.

[0079] S5: Training Heterogeneous Graph Neural Networks

[0080] A heterogeneous graph neural network was trained using a training set to obtain a protein-protein interaction prediction model;

[0081] Specifically, in step S5, the specific training scheme is as follows:

[0082] S51: The initial learning rate is set to 0.001. The Adam optimizer is used to iteratively update the model parameters, and the cross-entropy loss function is used as the training objective function. In this configuration, each training round contains 200 epochs. A stable gradient update strategy ensures the effectiveness and convergence performance of the feature representation learning process. A cosine annealing learning rate scheduling strategy is used to dynamically adjust the learning rate during training, and an early stopping mechanism is enabled throughout the training process to prevent overfitting.

[0083] S52: During training, to demonstrate the relative importance of each key technique we employed, ablation experiments were conducted by removing key techniques from the final combination. The results showed that the average performance of the model decreased to varying degrees when any technique was removed. When both techniques were removed, the model performance decreased significantly.

[0084] Table 2 Ablation experiments of key technologies

[0085]

[0086] S6: Prediction of protein-protein interactions

[0087] Step S6 includes the following sub-steps:

[0088] S61: Input the unknown protein into the protein-protein interaction prediction model for prediction and obtain the prediction results. The results are shown in Table 3. As can be seen from Table 3, the performance of the three partitioning methods of the present invention is better than other methods on the three datasets.

[0089] Table 3 Performance under different methods

[0090]

[0091] S62: To investigate the classification performance of different interaction types, Table 4 provides the F1 score for each category. Experimental results show that expression categories exhibit significant recognition challenges across all datasets, while the classification of reactions, binding, and catalysis consistently demonstrates strong classification performance in all experimental settings. This observation can be attributed to the fact that their biological functions often carry distinct molecular features, making it easier for the model to capture discriminative patterns.

[0092] Table 4: F1 scores for each PPI type on the three datasets

[0093]

[0094] In summary, the protein interaction prediction method based on a relation-aware heterogeneous graph neural network described in the above embodiments first performs unified preprocessing on protein sequence and 3D structural information to construct a heterogeneous protein graph containing three edge types: linear adjacency, geometric relationships based on spatial nearest neighbors, and contact relationships based on distance thresholds. Then, sequence processing and structure processing modules are used to extract the protein sequence and structural representations, respectively, and the alignment and complementarity of the two modal feature spaces are achieved through a sequence-structure contrastive learning mechanism. Next, the contrastive-learned sequence and structural features are fused with the original residue features and input into a heterogeneous graph neural network with a learnable multi-relation weighting mechanism. When processing multiple types of relationships, the network introduces intra-graph multi-head self-attention to dynamically model the importance of different relationships and improve the recognition ability of protein node representations. Based on this, by constructing a protein-protein interaction graph and training the overall model, the interaction prediction of unknown protein pairs can be achieved. The performance of different models was verified and compared on a test set. Experimental results demonstrate that these techniques can accelerate network convergence and robustness; the model proposed in this invention has high prediction accuracy. It is expected to be deployed on edge devices for production and application in the future.

[0095] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for predicting protein interactions based on heterogeneous graph neural networks, characterized in that, Includes the following steps: S1: Protein Data Preprocessing We obtained the amino acid sequences of proteins and the corresponding publicly available datasets of protein-protein interactions. We used the AlphaFold2 structure prediction model to predict the structure of the amino acid sequences to be studied, obtained the three-dimensional coordinates of each amino acid sequence, and divided the dataset into training set, validation set and test set according to different modes and set ratios. S2: Protein Feature Extraction and Feature Fusion Design sequence feature extractors and structural feature extractors, input the extracted sequence and structural features into the comparison learning training module, and learn the feature differences from different perspectives; S3: Constructing a heterogeneous protein map Each protein is represented as a heteroproton graph, which simultaneously expresses three types of relationships: linear sequence adjacency, geometric adjacency based on spatial KNN, and contact graph based on distance threshold. The static weights and learnable scaling parameters of each type of relationship are preserved to achieve weighted fusion at the relationship level. S4: Constructing a heterogeneous graph neural network based on multi-relationship graph attention. A heterogeneous graph neural network is constructed, which adopts a graph convolutional structure based on multi-relation modeling and introduces a relation-level multi-head self-attention mechanism to achieve dynamic fusion of three types of relations and generate protein node embeddings. S5: Training Heterogeneous Graph Neural Networks A heterogeneous graph neural network is trained using a training set to obtain a protein-protein interaction prediction model; S6: Prediction of protein-protein interactions The unknown protein is input into the protein-protein interaction prediction model to obtain the prediction results.

2. The protein interaction prediction method based on heterogeneous graph neural networks according to claim 1, characterized in that, In step S1, the specific processing procedure is as follows: S11: Obtain the publicly available dataset of the amino acid sequence and interactions of the protein, and use the AlphaFold2 structure prediction model to infer the three-dimensional conformation of the amino acid sequence, generating spatial coordinate information of all atoms including the main chain and side chains. The AlphaFold2 structure prediction model is based on multiple sequence alignment, attention network, structural template retrieval and geometric reasoning mechanism of Evoformer module. The output three-dimensional structure is stored in PDB format, which includes the atomic position, residue type and local geometric parameters of each amino acid residue. S12: Divide the amino acid sequences into training set, validation set and test set according to a set ratio. The partitioning mode includes random mode, depth-first search mode and breadth-first search mode.

3. The protein interaction prediction method based on heterogeneous graph neural networks according to claim 2, characterized in that, In step S12, the three partitioning modes are defined as follows: Random mode: Assuming that all PPI pairs are independent of each other, without considering the topology of the protein network, proteins or interaction pairs are randomly divided according to a preset ratio to form training set, validation set and test set. Depth-first search mode: Based on the protein network topology, long paths formed during the depth search process and their corresponding long-distance node pairs are preferentially selected as the test set and validation set. The search proceeds along a path to explore to the end and then backtracks, prioritizing long paths and long-distance connections in the protein network. The training set consists of the remaining protein pairs that were not selected for the validation set and test set. Breadth-first search mode: Based on the topology of the protein-protein interaction network, starting from a randomly selected root node, the network is traversed in a "near to far, layer by layer" manner. The first-order neighbors of the current node are explored in breadth-first order, and then the network is expanded to the next level. The nodes selected during the traversal and their internal edges are divided into test set and validation set. Correspondingly, the training set is composed of the remaining proteins and their interaction pairs after removing the nodes of the validation set and test set.

4. The protein interaction prediction method based on heterogeneous graph neural networks according to claim 1, characterized in that, The specific process of step S2 is as follows: S21: Seven residue-level features are used as the initial feature vector of the protein. The seven residue-level features are isoelectric point, polarity, acidity / basicity, hydrogen bond acceptor, hydrogen bond donor, octanol-water partition coefficient, and topological polar surface area. S22: The sequence feature extractor is used to learn discriminative sequence representations from the primary sequence of a protein. The specific process is as follows: local sequence patterns are modeled through multiple one-dimensional depthwise separable convolutional layers; after each one-dimensional depthwise separable convolutional layer, a squeeze-incentivized attention mechanism is introduced to adaptively recalibrate the importance of different feature channels; then, through pooling and fully connected mapping, the variable-length sequence features are compressed into a fixed-dimensional global sequence embedding vector, which is used to characterize the contextual features of the protein at the sequence level, thus obtaining the final sequence features. S23: The structural feature extractor is used to learn the spatial and topological structural representation of a protein from its three-dimensional structural information. The specific process is as follows: A structural contact graph is constructed based on the protein's three-dimensional structural data as input to the structural feature extractor, where nodes represent proteins and edges represent interconnections determined based on spatial distance thresholds. During structural feature modeling, the structural feature extractor is constructed by cascading multiple layers of graph convolutional layers based on GraphSAGE. Skip connections are introduced between these layers. Furthermore, an adaptive graph pooling module based on SAGPooling is set after four layers of graph convolutional layers based on GraphSAGE to learn node importance and perform hierarchical downsampling. Finally, global pooling aggregates the variable-scale node features into a fixed-dimensional structural embedding vector, which is used to characterize the overall three-dimensional structure and topological properties of the protein, thus obtaining the final structural features. S24: Input the extracted sequence and structural features into the contrast learning training module. During the training phase, the model outputs the corresponding sequence embedding vector and structural embedding vector for the same protein sample through the sequence feature extractor and structural feature extractor, respectively. The two are used as a pair of positive samples to input into the contrast learning module. Embedding vectors of different protein samples in the same batch constitute negative samples. The contrastive learning module normalizes the sequence embedding vector and the structural embedding vector, calculates the cross-view similarity matrix based on cosine similarity, and uses a contrastive loss function with a temperature coefficient to perform bidirectional alignment training on the two embedding vectors. By minimizing this contrastive loss, the feature representations of the same protein in the sequence view and the structural view are made close to each other in the shared representation space, while widening the representation distance between different protein samples. This outputs the sequence feature representation and structural feature representation optimized by alignment constraints, which are then concatenated to obtain the contrastive learning feature vector, which serves as the input feature for the subsequent protein interaction prediction task.

5. The protein interaction prediction method based on heterogeneous graph neural networks according to claim 4, characterized in that, In step S2, a gate mechanism for heterogeneous feature alignment and fusion is set up. Specifically, a bilinear projection layer is introduced to map the initial feature vector in step S21 and the contrastive learning feature vector in step S24 to the same latent space, and then they are concatenated to obtain 128-dimensional protein features.

6. The protein interaction prediction method based on heterogeneous graph neural networks according to claim 1, characterized in that, In step S3, the heterogeneous protein graph is constructed by modeling three types of semantically distinct relationships. Each type of relationship corresponds to an independent edge type in the graph, as follows: Sequence Edges: These represent the sequential connection of amino acids in the primary structure. Bidirectional edges are established between each pair of adjacent residues to form a chain-like sequence structure. KNN Edges: Based on the spatial proximity relationship between amino acids in the three-dimensional structure of a protein, the K nearest neighbor strategy is used to connect the K nearest amino acid residues on the three-dimensional coordinates, reflecting local structural interactions. Threshold Distance Edges: Connects all residue pairs whose spatial distance is less than a set threshold based on a spherical distance threshold, in order to model global structural contacts.

7. The protein interaction prediction method based on heterogeneous graph neural networks according to claim 6, characterized in that, In step S3, a protein relation network is constructed by integrating Sequence Edges, KNN Edges, and Threshold Distance Edges. Static base weights are preset for each edge type to reflect its prior importance, and learnable scaling parameters are introduced for dynamic calibration to adaptively adjust the actual contribution of different relations in training. The learnable scaling parameters are used to dynamically correct the static base weights. The final weights are obtained by calculating the product of the two, and the contribution of different relations in the feature aggregation process of the heterogeneous graph neural network is adaptively adjusted using the final weights.

8. The protein interaction prediction method based on heterogeneous graph neural networks according to claim 7, characterized in that, In step S4, an end-to-end heterogeneous graph neural network architecture is adopted. Cross-relational attention calculation is performed on the three types of edge types through the heterogeneous graph multi-head self-attention mechanism to dynamically learn the weight allocation of different interactions. In the training process, the protein heterogeneous graph is processed by multi-layer heterogeneous graph convolution. Each layer aggregates multi-relational features through the multi-head attention mechanism. Then, the training process is stabilized by batch normalization and residual connection. Finally, the node-level representation is transformed into protein embedding through global pooling.

9. The protein interaction prediction method based on heterogeneous graph neural networks according to claim 1, characterized in that, The loss function for model training in step S5 is the cross-entropy loss function, as shown in the following formula: ； in, It is the total number of samples. It is the first The true label of each sample It is the first Predict the probability of each sample.

10. The protein interaction prediction method based on heterogeneous graph neural networks according to claim 1, characterized in that, In step S6, protein-protein interactions include reactions, binding, post-translational modifications, activation, inhibition, catalysis, and expression. When predicting whether there is an interaction between two proteins, each type of reaction is predicted. If the predicted value of a certain type of reaction is greater than 0.5, it is considered that the protein pair has this type of reaction. Multiple reactions may exist between the same pair of proteins.

Images