A method and system for predicting protein interactions by fusing multimodal geometric features

Abstract

This invention relates to the field of bioinformatics, and more particularly to a method and system for predicting protein interactions based on multimodal geometric feature fusion. The method includes: extracting features from the amino acid sequence, three-dimensional structure, and three-dimensional coordinates of a target protein to generate protein sequence features, protein structural features, and protein spatial features; performing standardization and adaptive weighted fusion on these features to generate unified multimodal protein features; updating node features in a protein graph constructed based on the multimodal protein features by combining graph position encoding and a global attention mechanism; sequentially encoding the center node, subgraph, and context of the node features to generate a comprehensive node embedding; and predicting the interaction type and probability between protein pairs based on the comprehensive node embedding. This invention effectively solves the problems of feature representation bias and insufficient complex interaction modeling capabilities caused by traditional PPI prediction methods that ignore protein geometric structure information and multi-level local associations.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method and system for predicting protein interactions by fusing multimodal geometric features. Background Technology

[0002] Accurate prediction of protein-protein interactions (PPIs) is a prerequisite for elucidating disease mechanisms and developing targeted drugs. With the development of computational biology, deep learning-based prediction models have significantly outperformed traditional experimental methods such as yeast two-hybrids in terms of efficiency, but their predictive reliability is still limited by three inherent challenges.

[0003] First, existing methods struggle to effectively integrate multimodal protein characterization. While current models can resolve linear associations of amino acids and describe static three-dimensional folding, they fail to capture dynamic conformational changes during allosteric regulation and are severely inadequate in expressing spatial interactions between distant residues. Second, mainstream graph neural networks are limited by the isomorphism discrimination framework of the first-order WL test. Their topology-independent neighbor aggregation mechanism not only ignores the sensitivity of protein geometry but also destroys the topological specificity of functional micro-regions due to excessive smoothing of local features. Furthermore, long-range dependent interactions are difficult to model due to the fragmentation of graph structures. Finally, existing techniques for protein pair feature processing mostly rely on simple vector concatenation or pooling operations, which cannot characterize the fine physicochemical complementarity patterns between residues, nor can they distinguish the decision boundaries of interaction types in complex biological scenarios.

[0004] The information disclosed in this background section is intended only to enhance the understanding of the general background of this disclosure and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0005] This invention provides a method and system for predicting protein interactions by fusing multimodal geometric features, which can effectively solve the problems in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A multimodal geometric feature fusion method for predicting protein interactions, the method comprising: The amino acid sequence, three-dimensional structure, and three-dimensional coordinates of the target protein are extracted to generate protein sequence features, protein structure features, and protein spatial features. The protein sequence features, protein structural features, and protein spatial features are standardized and adaptively weighted and fused to generate unified multimodal protein features; In the protein graph constructed based on the multimodal protein features, node features are updated by combining graph position encoding and a global attention mechanism; The updated node features are sequentially encoded by center node, subgraph, and context to generate a comprehensive node embedding. Based on the comprehensive node embedding, the interaction types and probabilities between protein pairs are predicted.

[0007] Further, protein sequence characteristics are obtained, including: The amino acid sequence of the target protein is converted into a raw sequence vector representation; Local structural features and temporal correlation features are obtained in parallel from the original sequence vector representation based on one-dimensional convolution operations and bidirectional gated loop units. The local structural features and the temporal correlation features are concatenated to generate initial sequence features; Based on the initial sequence features, a mean vector and a standard deviation vector of the latent variable space are generated. The latent variable space is a low-dimensional latent space used to carry the probability distribution characteristics of the protein sequence. Based on random sampling and gradient differentiability reparameterization operations, the mean vector and the standard deviation vector are converted into the protein sequence features; The protein sequence features are reconstructed based on the decoder, and the feature discriminability is optimized by minimizing the reconstruction error.

[0008] Furthermore, protein structural features are obtained, including: Secondary structural units of amino acid residues are obtained based on the three-dimensional structure of the target protein; Label the types of the secondary structural units and construct a protein structure diagram including the spatial distance relationships of the amino acid residues; Initialize the node features of each secondary structure unit in the protein structure diagram; The node features and the features of its neighboring nodes are aggregated according to the graph neural network; The protein structure features are generated by updating the node features of the protein structure graph through multiple rounds of iteration.

[0009] Furthermore, the spatial characteristics of the protein are obtained, including: The three-dimensional coordinate data of the target protein are converted into standardized point cloud data; The standardized point cloud data is processed by a multilayer perceptron to obtain the local spatial features of each point; Perform max pooling on the local spatial features to generate a global spatial feature representation; The global spatial feature representation is mapped to the protein spatial feature.

[0010] Furthermore, the fusion of standardization and adaptive weighting includes: The protein sequence features, protein structural features, and protein spatial features are z-score normalized to eliminate differences in feature dimensions. The standardized multimodal features are transformed into initial input features based on the fully connected layer; The initial input features are sequentially subjected to nonlinear orthogonal transformations using the Tanh activation function and the ReLU activation function; The activated initial input features are compressed into a low-dimensional latent space through another fully connected layer to generate a latent representation. In the reconstruction mapping layer, the latent representation is subjected to linear transformation and ReLU nonlinear transformation, and the transformation results are mapped back to the original feature space to generate the unified multimodal protein features.

[0011] Furthermore, node features are updated by combining graph position encoding and a global attention mechanism, including: Based on the adjacency relationship of nodes in the protein graph, the feature information of each node and its corresponding neighboring nodes is aggregated through a local graph convolutional network to generate a locally updated node feature representation. In the global attention mechanism, cross-node attention weight calculation and feature weighted aggregation are performed on the locally updated node feature representation to generate a globally enhanced node feature representation; The locally updated node feature representation and the globally enhanced node feature representation are input into a multilayer perceptron for nonlinear feature fusion. The fused node features are output as the updated node features.

[0012] Furthermore, a comprehensive node embedding is generated, including: In the protein graph, a target node is selected as the center node. By embedding the features of the target node into the star subgraph of the target node and integrating the neighbor node information of the target node, a center node encoded feature representation is generated. The star-shaped subgraph is processed based on a graph convolutional network, and mean pooling is performed to aggregate the feature information of neighboring nodes in the star-shaped subgraph to generate a subgraph encoded feature representation. The feature representation of the target node in different context subgraphs is analyzed, and a context-encoded feature representation is generated by performing mean aggregation on the embedding features of the target node in multiple context subgraphs. The central node encoding feature representation, the subgraph encoding feature representation, and the context encoding feature representation are weighted and summed to generate the comprehensive node embedding.

[0013] Furthermore, based on the comprehensive node embedding, the prediction of the interaction types and probabilities between protein pairs includes: Obtain the two node embedding features corresponding to the target protein pair from the comprehensive node embedding; Perform feature interaction operations on the acquired embedding features of the two nodes to generate interaction feature representations between protein pairs; The interaction feature representation is input into a multi-layer fully connected network and subjected to nonlinear transformation to generate probabilistic input features; Apply a probabilistic activation function to the probabilistic input features to output the interaction type between the protein pairs and the corresponding probability prediction results.

[0014] A protein interaction prediction system based on multimodal geometric feature fusion, the system comprising: The feature generation module extracts features from the amino acid sequence, three-dimensional structure, and three-dimensional coordinates of the target protein, generating protein sequence features, protein structure features, and protein spatial features. The weighted fusion module performs standardization and adaptive weighted fusion on protein sequence features, protein structural features, and protein spatial features to generate unified multimodal protein features; The node update module updates node features in a protein graph constructed based on multimodal protein features by combining graph position encoding and a global attention mechanism. The encoding and embedding module sequentially encodes the center node, subgraph, and context of the updated node features to generate a comprehensive node embedding. The interaction prediction module predicts the types and probabilities of interactions between protein pairs based on comprehensive node embedding.

[0015] Furthermore, the weighted fusion module includes: The difference elimination unit performs z-score normalization on protein sequence features, protein structure features, and protein spatial features to eliminate differences in feature dimensions; The initial transformation unit transforms the standardized multimodal features into initial input features based on the fully connected layer; The orthogonal transformation unit sequentially performs nonlinear orthogonal transformations of the Tanh activation function and the ReLU activation function on the initial input features; The feature compression unit compresses the activated initial input features into a low-dimensional latent space to generate a latent representation through another fully connected layer. The unified mapping unit performs linear and ReLU nonlinear transformations on the latent representation in the reconstruction mapping layer, and maps the transformation results back to the original feature space to generate unified multimodal protein features.

[0016] The technical solution of this invention can achieve the following technical effects: By fusing multimodal features of protein amino acid sequences, three-dimensional structures, and spatial coordinates, and generating unified protein features based on adaptive weighting, the node features are updated in the protein graph by combining graph position encoding and global attention mechanism. Then, a comprehensive node embedding is generated by three-level encoding of the central node, subgraph, and context. Finally, the interaction type and probability are predicted by protein-feature interaction operations and nonlinear transformations. This effectively solves the problems of feature representation bias and insufficient ability to model complex interactions caused by ignoring protein geometric structure information and multi-level local associations in traditional PPI prediction methods.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a protein interaction prediction method based on multimodal geometric feature fusion; Figure 2 A schematic diagram of the process for standardization and adaptive weighted fusion; Figure 3 A flowchart illustrating the process of updating node features; Figure 4 A flowchart illustrating the process of generating embedded integrated nodes; Figure 5 A flowchart illustrating the process of predicting the type and probability of protein-protein pair interactions. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] Example 1; like Figure 1 As shown, this application provides a protein interaction prediction method based on multimodal geometric feature fusion, the method comprising: S10: Extract features from the amino acid sequence, three-dimensional structure, and three-dimensional coordinates of the target protein to generate protein sequence features, protein structure features, and protein spatial features. S20: Standardize and adaptively weightedly fuse protein sequence features, protein structural features, and protein spatial features to generate unified multimodal protein features; S30: In a protein graph constructed based on multimodal protein features, node features are updated by combining graph position encoding and a global attention mechanism; S40: Perform center node encoding, subgraph encoding and context encoding on the updated node features in sequence to generate a comprehensive node embedding; S50: Predicting the types and probabilities of protein-protein pair interactions based on integrated node embedding.

[0023] Specifically, the study first focuses on the target protein whose interactions are to be predicted, obtaining its primary amino acid sequence information, corresponding three-dimensional spatial structure information, and the coordinates of each atom or residue in three-dimensional space. The amino acid sequence can be derived from public protein databases or experimental sequencing results. The three-dimensional structure is preferably obtained through experimental analysis or homology modeling and structure prediction algorithms. For the amino acid sequence information, feature extraction is preferably performed by mapping amino acids to predefined biochemical attribute vectors or context-dependent sequence representations to characterize sequence conservation, functional fragments, and upstream / downstream dependencies. For the protein's three-dimensional structure information, the study preferably focuses on secondary structure composition, folding... Structural features are extracted from stacked topological relationships and residue contact patterns to reflect the overall conformation of the protein and the organization of local structural units. For three-dimensional coordinate information, residues or key atoms are preferably used as basic units. Protein spatial features are constructed by describing their relative positions, neighborhood distribution characteristics, and spatial geometric relationships in three-dimensional space, thus explicitly introducing spatial geometric constraints. After obtaining the above sequence features, structural features, and spatial features, to overcome the differences in dimensions, distribution, and information contribution of different modal features, this embodiment preferably first performs standardization processing on all types of features at a unified scale, and then performs weighted fusion based on the adaptive importance of features in the interactive prediction task, thereby generating... A unified multimodal protein feature model capable of simultaneously representing sequence semantics, structural conformation, and spatial geometry information is developed. This fusion process avoids the problems of single-modality dominance or information redundancy through adaptive weight adjustment. Subsequently, a protein graph structure is constructed based on these multimodal protein features. Nodes in the graph preferably correspond to residues or structural units within the protein, and the connections between nodes are determined based on spatial proximity or structural contact. Graph position encoding is introduced into this protein graph to explicitly distinguish the relative positions of nodes within the overall structure. Simultaneously, a global attention mechanism is incorporated, enabling each node to comprehensively consider global information from distant nodes when updating its own features, thereby breaking through the limitations of traditional local neighborhoods. The propagation limitations are addressed. Based on this, the updated node features are sequentially encoded using center node encoding, subgraph encoding, and context encoding. Center node encoding highlights the core representation of the currently interested residues in the protein structure, subgraph encoding characterizes its local structural environment, and context encoding further integrates the overall protein background information. These three methods collaboratively generate a comprehensive node embedding with multi-scale and multi-level semantics. Finally, the comprehensive node embedding is used as input for protein-protein pair interaction prediction. Through joint analysis of the relevant node embeddings in protein pairs, the prediction result of whether protein pairs interact is output, and the probability values ​​corresponding to different interaction types are further given.

[0024] The technical solution of this invention integrates the multimodal features of protein amino acid sequence, three-dimensional structure and spatial coordinates, and generates unified protein features based on adaptive weighting. After updating node features in the protein graph by combining graph position encoding and global attention mechanism, a comprehensive node embedding is generated by a three-level encoding of center node-subgraph-context. Finally, the interaction type and probability are predicted by protein interaction operation and nonlinear transformation. This effectively solves the problems of feature representation bias and insufficient complex interaction modeling ability caused by ignoring protein geometric structure information and multi-level local associations in traditional PPI prediction methods.

[0025] Furthermore, obtaining protein sequence characteristics includes: The amino acid sequence of the target protein is converted into a raw sequence vector representation; Local structural features and temporal correlation features are obtained in parallel from the original sequence vector representation based on one-dimensional convolution operations and bidirectional gated recurrent units. Initial sequence features are generated by splicing local structural features and temporal correlation features; The mean vector and standard deviation vector of the latent variable space are generated based on the initial sequence features. The latent variable space is a low-dimensional latent space used to carry the probability distribution characteristics of the protein sequence. Based on random sampling and gradient differentiability reparameterization operations, the mean vector and standard deviation vector are converted into protein sequence features; Protein sequence features are reconstructed based on a decoder, and feature discriminability is optimized by minimizing the reconstruction error.

[0026] As a preferred embodiment of the above, the amino acid sequence of the target protein is first converted into a raw sequence vector representation according to a predefined amino acid encoding rule. Preferably, each amino acid corresponds to a vector of fixed dimensions, which can comprehensively reflect the physicochemical properties or contextual semantic information of the amino acid, thereby converting the discrete symbol sequence into a computable continuous representation. After obtaining the raw sequence vector representation, in order to simultaneously capture the local structural patterns and global temporal dependencies in the protein sequence, it is preferred to process the raw sequence vector representation in parallel. One processing path extracts the local structural features formed by adjacent amino acid fragments in the sequence through one-dimensional convolution operations to reflect local patterns such as conserved motifs and functional fragments. The other processing path performs forward and reverse temporal modeling of the sequence through a bidirectional gated recurrent unit, thereby fully learning the correlation between the sequence context and avoiding the information loss caused by relying solely on modeling in a single direction. Subsequently, the obtained local structural features and temporal correlation features are concatenated along the feature dimension. Next, initial sequence features capable of simultaneously representing local and global information are formed. Further, by mapping these initial sequence features, mean and standard deviation vectors describing the probability distribution characteristics of the sequence in the latent variable space are generated. The latent variable space is preferably a low-dimensional latent space, used to compress redundant information and characterize the inherent uncertainty and diversity of the protein sequence. Based on the mean and standard deviation vectors, random sampling combined with gradient-differentiable reparameterization is preferably used to convert the probability distribution representation into a specific protein sequence feature representation, thereby introducing randomness to enhance the generalization ability of the features while ensuring trainability. Furthermore, to ensure that the protein sequence features generated in the latent variable space can realistically and effectively reflect the original sequence information, a decoding process is preferably introduced to reconstruct the protein sequence features. The entire sequence feature learning process is optimized by minimizing the difference between the reconstructed result and the original sequence vector representation, so that the obtained protein sequence features maintain discriminativeness while possessing good information fidelity.

[0027] Furthermore, obtaining protein structural features includes: Secondary structural units of amino acid residues are obtained based on the three-dimensional structure of the target protein; Label the types of secondary structure units and construct a protein structure map including the spatial distance relationships of amino acid residues; Initialize the node features of each secondary structure unit in the protein structure diagram; The graph neural network aggregates the features of nodes and their neighboring nodes. Protein structure features are generated based on node features of the protein structure map through multiple rounds of iterative updates.

[0028] As a preferred embodiment of the above, firstly, based on the three-dimensional structural information of the target protein, the amino acid residues constituting the protein are analyzed to identify the secondary structural units to which each residue belongs. These secondary structural units include, but are not limited to, α-helices, β-sheets, and random coil structures. This process can be accomplished by analyzing the main chain conformation and hydrogen bond distribution in the three-dimensional structure. Subsequently, each identified secondary structural unit is labeled with its type, and a protein structure map is constructed using the secondary structural unit or its corresponding amino acid residue as basic nodes. The connection relationship between nodes is preferably determined based on the distance relationship between residues in three-dimensional space. For example, when the spatial distance between two residues or structural units is less than a preset threshold, a connection relationship is established between the corresponding nodes, thereby enabling the protein structure map to explicitly reflect the spatial proximity and potential interaction relationships within the protein. After constructing the protein structure map, the node features corresponding to each secondary structural unit in the map are initialized. The node features preferably comprehensively consider the structural unit. The initial node features are described by type information, the number of residues included, and their relative positions in the overall structure, ensuring that they possess basic structural semantic information. Furthermore, the protein structure graph is aggregated and updated based on a graph neural network. In each update round, each node retains its own structural features while absorbing structural information carried by its neighboring nodes, thereby gradually learning the local structural environment and broader structural dependencies. This aggregation and update process is preferably executed in a multi-round iterative manner, allowing node features to evolve from initial local descriptions to higher-order representations that simultaneously integrate multi-level structural information, ultimately generating protein structural features to characterize the overall structural properties of the target protein. For example, for an α-helix structural unit located in the core region of a protein, its node features, after multiple iterations, can not only reflect its own structural type but also integrate the spatial distribution characteristics of adjacent β-sheets or other helical structural units, thus more accurately characterizing its role in the overall protein folding.

[0029] Furthermore, obtaining the spatial characteristics of proteins includes: The three-dimensional coordinate data of the target protein is converted into standardized point cloud data; Standardized point cloud data is processed by a multilayer perceptron to obtain the local spatial features of each point; Max pooling is performed on local spatial features to generate a global spatial feature representation; The global spatial feature representation is mapped to the protein spatial feature.

[0030] As a preferred embodiment of the above, the three-dimensional coordinate data of the target protein is first obtained. The three-dimensional coordinate data preferably uses the representative atomic coordinates of each residue as the basic granularity to select point cloud points. For example, it is preferable to select the main chain key atoms of each amino acid residue or use the α-carbon atom as the representative point of the residue, thereby converting the protein structure into a set composed of several three-dimensional points. Side chain key atomic points or functional site atomic points can also be further introduced to enhance the expression of local pockets and surface morphology. However, it is preferable to always ensure that the point cloud points have a clear correspondence with the protein residues or atoms, facilitating alignment with sequence features and structural features in the subsequent fusion stage. Subsequently, to reduce the differences in coordinate system, scale, and sampling density among different proteins... To mitigate interference with the learning process, the 3D coordinate data is converted into standardized point cloud data. Preferred standardization processes include at least: decentralizing all points to align the geometric centers of the point cloud to a unified reference position; normalizing the overall scale of the point cloud to ensure that point clouds of different proteins are comparable in spatial span; removing or truncating outlier points to avoid extreme points caused by experimental noise or predicted structural errors affecting feature stability; and ideally, retaining information for each point during standardization, such as the corresponding residue number, residue type, or whether it belongs to a surface residue, so that geometric information and basic semantic information can be co-encoded during subsequent point-level feature extraction, thereby improving spatial stability. The discriminative ability of features; after point cloud standardization, a multilayer perceptron is used to process the standardized point cloud data point by point and obtain the local spatial features of each point. Point-by-point processing preferably refers to isomorphic transformation of the three-dimensional coordinates and their additional identifiers of each point, so that each point outputs a fixed-dimensional point-level vector representation. To ensure that this point-level vector not only reflects the absolute position of the point but also emphasizes local geometric relationships, a relative position description of the point is preferably introduced on the input side. That is, when constructing the input of each point, information such as the relative coordinate difference of its neighboring points, the distance distribution of neighboring points, or the density description of neighboring points is combined, so that the multilayer perceptron can learn spatial morphology closely related to protein interactions, such as local surface curvature trends and pocket boundary density changes. Features; In implementation, it is preferable to first determine several neighborhood points for each point based on the principle of spatial proximity, for example, by selecting a fixed number of neighborhood points according to a preset neighborhood radius or the nearest neighbor principle. Then, the relative geometric description of the neighborhood and the coordinates of the center point itself are input into the multilayer perceptron to obtain point-level local spatial features including local structural information. Next, max pooling is performed on the local spatial features to generate a global spatial feature representation. The preferred max pooling method is to take the maximum response in each dimension of the point set dimension, so that the model can automatically select the most significant geometric response from the global scope as the representative information of the overall spatial morphology of the protein, thereby obtaining a global spatial feature representation that is insensitive to point order and more robust to local noise.Simultaneously, max pooling is preferably used to highlight sites with significant geometric or functional spatial meaning, such as residue sites located in surface grooves, binding pocket edges, or with prominent curvature changes, thus increasing their contribution to the global representation. Finally, the global spatial feature representation is further mapped to protein spatial features. This mapping preferably involves dimensional transformation and numerical constraints on the global representation to align its output dimension with the multimodal unified feature and ensure the comparability of the spatial feature across different proteins. In a preferred embodiment, the global spatial feature representation can be transformed into a fixed-length vector and normalized for subsequent standardization and adaptive weighted fusion with sequence and structural features.

[0031] Furthermore, such as Figure 2 As shown, the standardization and adaptive weighted fusion include: z-score normalization is used to eliminate differences in feature dimensions for protein sequence features, protein structure features, and protein spatial features; The standardized multimodal features are transformed into initial input features based on the fully connected layer; The initial input features are sequentially subjected to nonlinear orthogonal transformations using the Tanh activation function and the ReLU activation function; The activated initial input features are compressed into a low-dimensional latent space through another fully connected layer to generate a latent representation. In the reconstruction mapping layer, the latent representation is subjected to linear transformation and ReLU nonlinear transformation, and the transformation results are mapped back to the original feature space to generate unified multimodal protein features.

[0032] As a preferred embodiment of the above, protein sequence features, protein structure features, and protein spatial features are first input. Given that these three types of features typically have different dimensional scales, numerical distributions, and dimensions, it is preferable to perform z-score standardization on each modality before it enters the fusion network. This involves centering and scaling each dimension using the mean and standard deviation obtained from the training data, ensuring that the modalities are comparable within the same numerical range and reducing bias accumulation in the early stages of training. In implementation, it is preferable to maintain separate standardized parameters for each of the three types of features and use the parameters fixed during the training phase during the inference phase to ensure comparability across different batches and different... Reproducibility and stability among protein samples; after standardization, the standardized multimodal features are preferably combined in a preset order and input into a fully connected layer for a unified feature projection to obtain the initial input features. This projection maps features from different sources to the same latent space scale and introduces a learnable linear combination relationship, laying the foundation for subsequent adaptive weighting; further, to improve the separability of the fused representation and suppress redundant correlations between multimodal features, the initial input features are preferably subjected to a nonlinear orthogonal transformation process formed by Tanh activation and ReLU activation in sequence, where Tanh is used to symmetrically compress the input to highlight... To differentiate between positive and negative directions and reduce the interference of extreme values ​​on fusion, ReLU is then used to enhance sparse expression and significant responses, highlighting feature components more relevant to protein interactions while automatically weakening less important components during fusion. Subsequently, another fully connected layer further compresses the activated initial input features into a low-dimensional latent space to generate latent representations. This low-dimensional latent space is preferably a fixed-length vector space that interfaces with downstream graph representation learning. Its purpose is to retain the most critical cross-modal common factors for interaction prediction while compressing modality-specific noise and redundant dimensions, thereby improving generalization ability and reducing the risk of overfitting. In the training strategy... In summary, the latent representation is preferably used as the adaptive weight carrier for fusion. That is, the dimensions of the latent representation will be adjusted to different degrees during backpropagation due to the downstream prediction loss, thereby indirectly achieving adaptive allocation of contributions to different modal features. Finally, in the reconstruction mapping layer, a linear transformation is performed on the latent representation and combined with the ReLU nonlinear transformation to map it back to the original feature space to generate unified multimodal protein features. The linear transformation is used to restore the representation dimensions consistent with the original feature space so as to align with the subsequent protein graph node features, and the ReLU nonlinear transformation is used to ensure that the reconstructed fused features have stronger discriminativeness and suppress the propagation of negative noise.

[0033] Furthermore, such as Figure 3 As shown, node features are updated by combining graph position encoding and a global attention mechanism, including: Based on the adjacency relationship of nodes in the protein graph, the feature information of each node and its corresponding neighboring nodes is aggregated through a local graph convolutional network to generate a locally updated node feature representation. In the global attention mechanism, cross-node attention weight calculation and feature weighted aggregation are performed on the locally updated node feature representation to generate a globally enhanced node feature representation; The locally updated node feature representation and the globally enhanced node feature representation are input into a multilayer perceptron for nonlinear feature fusion. The fused node features are output as the updated node features.

[0034] As a preferred embodiment of the above, a protein graph is first used as input. Nodes in the graph preferably correspond to residues or structural units, and edges are preferably established based on structural contact or spatial proximity. A graph position code is configured for each node to describe its relative position in the overall topology and spatial layout. The graph position code is preferably used in subsequent updates along with the initial node features to avoid representation confusion between nodes in different locations but with similar local neighborhoods. Subsequently, local graph convolutional network aggregation is performed based on the adjacency relationships of nodes in the protein graph. In each round of local aggregation, each node preferably receives and weights the feature information of its one-hop neighbor nodes, while retaining the residual information of its own features to reduce the risk of oversmoothing. The contribution weights of the neighbor nodes are... The re-optimization is related to edge relationships, and can be jointly determined by information such as distance segments between nodes, contact strength levels, or whether they belong to the same secondary structure region. This allows local convolution to not only acquire information about neighboring nodes, but also distinguish the importance of different adjacency relationships to structural semantics, thereby generating a locally updated node feature representation. This local representation focuses on characterizing the local structural environment, local geometry, and the cooperative patterns of neighboring residues. After the local update is completed, cross-node attention weight calculation and feature weight aggregation are performed on the locally updated node feature representation in the global attention mechanism. Preferably, a global information channel is established in a node-to-node manner, so that each node can selectively absorb information from the entire graph that is functionally related or geometrically complementary to it. To enhance the interpretability and robustness of global attention, the influence of graph position encoding is preferably introduced into the attention weight calculation. This ensures that the attention intensity of distant nodes depends not only on feature similarity but also on the modulation of relative positional relationships, thereby suppressing meaningless attention to spatially unreasonable or topologically weakly related nodes. Furthermore, a moderate attention sparsity bias is preferably set during global attention aggregation, for example, making each node more inclined to aggregate a small number of highly relevant nodes rather than averaging a large number of nodes. This reduces noise propagation and improves training stability, ultimately resulting in a globally enhanced node feature representation. This global representation primarily reflects the overall conformational constraints of the protein, long-range co-variations, and potential interactions with the interaction interface. The cross-regional association is then established. Next, the locally updated node feature representation and the globally enhanced node feature representation are input into a multilayer perceptron for nonlinear feature fusion. The preferred fusion method is to combine the two types of representations in the feature dimension and then perform a nonlinear transformation, so that the network can automatically learn the optimal ratio of local information and global information in different nodes and different proteins, thereby avoiding the dominance of a certain branch, information redundancy or gradient conflict caused by using fixed weighting or simple splicing. At the same time, it is preferred to retain a short-circuit path from the local representation or the original node representation to the fused output during the multilayer perceptron fusion, so as to ensure that the key local structural signals are not weakened in the deep fusion, and then output the fused node features as the updated node features.

[0035] Furthermore, such as Figure 4 As shown, the generation of integrated node embeddings includes: In the protein graph, a target node is selected as the central node. By embedding the features of the target node in the star subgraph of the target node and integrating the information of the target node's neighboring nodes, a central node encoded feature representation is generated. The star subgraph is processed by a graph convolutional network and mean pooling is performed to aggregate the feature information of neighboring nodes in the star subgraph to generate a subgraph encoded feature representation. The feature representation of the target node in different context subgraphs is analyzed, and the context-encoded feature representation is generated by performing mean aggregation on the embedding features of the target node in multiple context subgraphs. The central node encoding feature representation, subgraph encoding feature representation, and context encoding feature representation are weighted and summed to generate a comprehensive node embedding.

[0036] As a preferred embodiment of the above, the updated node features are first used as the basic input, and the nodes are traversed in the protein graph at the residue or structural unit granularity, or the target node is selected as the center node by sampling according to the task requirements. Preferably, the selection of the center node can cover all nodes to obtain full-graph node embedding, or only candidate interface residue nodes can be selected to reduce the amount of computation. For example, nodes with high solvent accessibility, located at the edge of surface pockets, or adjacent to functional sites are set as candidate center nodes, so that the comprehensive node embedding is more focused on the interaction-related regions. Subsequently, a star-shaped subgraph is constructed around the target node, which preferably consists of the center node and its one-hop neighbors. The node composition and edge relationships follow the spatial proximity or structural contact relationships in the protein graph. When generating the central node encoding feature representation in the star-shaped subgraph, it is preferable to use the updated features of the central node itself as the backbone information, while introducing the integration of neighbor node information to supplement the local environment of the central node. The integration of neighbor information is preferably achieved by weighted summarization of neighbor features. Furthermore, it is preferable to retain the direct connection fidelity path of the central node features when encoding the central node to avoid the dilution of the central information due to neighbor summarization, thereby obtaining a central node encoding feature representation that highlights the ontological semantics of the central node and includes the supplementation of the local neighborhood. Then, to obtain a more structured local subgraph language The process involves processing the features of a star-shaped subgraph using a graph convolutional network and performing mean pooling to generate a subgraph encoding feature representation. Preferably, the graph convolutional network performs one or more rounds of feature propagation within the star-shaped subgraph, capturing the local consistency information carried by neighboring nodes in the structural association indirectly formed through the central node. Subsequently, mean pooling is performed on the features of neighboring nodes in the star-shaped subgraph to obtain an aggregated representation that is insensitive to the number of neighbors and smoother against local noise. Preferably, this aggregated representation is combined with the convolutional features of the central node as the subgraph encoding feature representation, so that the subgraph encoding reflects both the overall features of the neighborhood and maintains alignment with the central node. In a further step, to address the issue of the same central node in a protein structure exhibiting different interactive meanings in different local structural fragments or different geometric contexts, a context-encoding feature representation generation process for the central node is introduced: preferably, for the same target node, multiple context subgraphs are constructed from the protein graph. Each context subgraph can be obtained through different strategies, such as expanding the neighborhood with different hop counts, such as expanding to two or three hops on the basis of one hop but limiting the number of nodes, truncating local regions with different spatial radii, or selecting structural sub-regions related to the central node by different structural fragment division methods, so that the same central node is repeatedly embedded in multiple context perspectives;Then, the embedding features of the target node are calculated in each context subgraph. A preferred approach is to perform lightweight graph feature propagation within each context subgraph and extract the corresponding representation of the target node. Finally, the embedding features of the target node in multiple context subgraphs are aggregated by mean to generate a context-encoded feature representation. This representation reflects the stable semantics and cross-regional dependency information of the central node within a broader structural context and reduces bias caused by inappropriate selection of a single context. Finally, the central node encoded feature representation, subgraph encoded feature representation, and context encoded feature representation are weighted and summed to generate a comprehensive node embedding. Preferably, the weighting coefficients are not fixed constants but can be learned during training. That is, different nodes may exhibit different proportions of central, local, and contextual importance in different proteins, thus making the fusion result more consistent with the interaction prediction target. For example, for residue nodes located at the edge of a clear binding pocket, the comprehensive embedding can preferably increase the weights of subgraph encoding and context encoding to emphasize local geometry and cross-segment cooperation; while for nodes located in flexible loop regions with high local noise, the weight of the central node encoding can preferably be increased to maintain the stability of its own features.

[0037] Furthermore, such as Figure 5 As shown, the prediction of protein-protein pair interaction types and their probabilities based on comprehensive node embedding includes: Extract the two node embedding features corresponding to the target protein from the comprehensive node embedding; Perform feature interaction operations on the embedded features of the two nodes to generate interaction feature representations between protein pairs; The interactive feature representation is input into a multi-layer fully connected network and subjected to nonlinear transformation to generate probabilistic input features. Apply a probabilistic activation function to the probabilistic input features to output the interaction types between protein pairs and the corresponding probabilistic prediction results.

[0038] As a preferred embodiment of the above, the comprehensive node embedding is first used as the basic input. For the target protein pair to be predicted, it is preferred to first determine the node set corresponding to each of the two proteins and the candidate node embedding acquisition method: In a preferred implementation, the comprehensive node embedding of all residue nodes in each protein graph can be regarded as the candidate set, and then two node embedding features for constructing protein pair interaction are selected according to a preset candidate strategy; the candidate strategy is preferably related to the possible interface where the interaction occurs, for example, it is preferred to select surface residues with high solvent accessibility, residues located at the edge of the pocket, residues with known functional site labels, or to sample and align according to the labeled interaction interface residues during the training phase, so as to improve the coverage of the interaction features to the real binding interface and the training efficiency; then, the two node embedding features corresponding to the target protein pair are obtained from the comprehensive node embedding, and are respectively denoted as the target node embedding of the first protein and the target node embedding of the second protein; wherein the correspondence of the target nodes can preferably be determined according to the real data labeling, candidate interface sampling or the model internal screening strategy, and it is preferred to use a multi-pair node pairing method during training to enhance robustness, that is, multiple sets of node embedding pairs can be extracted for the same protein pair and summarized in the subsequent output stage to reduce the bias caused by improper selection of a single node; then, the two obtained Each node embeds features and performs feature interaction operations to generate interactive feature representations between protein pairs. These interaction operations are preferably not limited to simple concatenation but include at least interaction forms capable of expressing complementarity and correlation, in order to more fully characterize the possible binding drivers between protein pairs. The interaction operations preferably maintain symmetry or are structurally symmetrically constrained, meaning that the semantics of the interaction features remain consistent or approximately consistent after exchanging the input order of the two proteins, thereby avoiding prediction bias caused by input order and improving generalization stability. After constructing the interactive feature representations, these representations are input into a multi-layer fully connected network for nonlinear transformation to generate probabilistic input features. The multi-layer fully connected network is preferably composed of multi-level nonlinear mappings, which makes different interaction types form a clearer separable structure in the feature space. Finally, a probabilistic activation function is applied to the probabilistic input features to output the interaction types between protein pairs and the corresponding probability prediction results: when the interaction type is multi-class, a probabilistic activation method that can output multi-class probability distribution is preferred, so that the probabilities of each type are comparable and meet the normalization requirements; when the interaction is binary classification or multi-label, a probabilistic activation method that can independently give the occurrence probability of each type is preferred, so as to express the situation that the same protein pair may have multiple interaction relationships at the same time.

[0039] Example 2; Based on the same inventive concept as the protein interaction prediction method using multimodal geometric feature fusion in the foregoing embodiments, this invention also provides a protein interaction prediction system using multimodal geometric feature fusion, the system comprising: The feature generation module extracts features from the amino acid sequence, three-dimensional structure, and three-dimensional coordinates of the target protein, generating protein sequence features, protein structure features, and protein spatial features. The weighted fusion module performs standardization and adaptive weighted fusion on protein sequence features, protein structural features, and protein spatial features to generate unified multimodal protein features; The node update module updates node features in a protein graph constructed based on multimodal protein features by combining graph position encoding and a global attention mechanism. The encoding and embedding module sequentially encodes the center node, subgraph, and context of the updated node features to generate a comprehensive node embedding. The interaction prediction module predicts the types and probabilities of interactions between protein pairs based on comprehensive node embedding.

[0040] The adjustment system described above in this invention can effectively realize the protein interaction prediction method based on the fusion of multimodal geometric features, and the technical effects it can achieve are as described in the above embodiments, and will not be repeated here.

[0041] Furthermore, the weighted fusion module includes: The difference elimination unit performs z-score normalization on protein sequence features, protein structure features, and protein spatial features to eliminate differences in feature dimensions; The initial transformation unit transforms the standardized multimodal features into initial input features based on the fully connected layer; The orthogonal transformation unit sequentially performs nonlinear orthogonal transformations of the Tanh activation function and the ReLU activation function on the initial input features; The feature compression unit compresses the activated initial input features into a low-dimensional latent space to generate a latent representation through another fully connected layer. The unified mapping unit performs linear and ReLU nonlinear transformations on the latent representation in the reconstruction mapping layer, and maps the transformation results back to the original feature space to generate unified multimodal protein features.

[0042] Similarly, the above-mentioned optimization schemes for the system can also achieve the optimization effects corresponding to the methods in Embodiment 1, which will not be repeated here.

[0043] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A protein interaction prediction method based on multimodal geometric feature fusion, characterized in that, The method includes: The amino acid sequence, three-dimensional structure, and three-dimensional coordinates of the target protein are extracted to generate protein sequence features, protein structure features, and protein spatial features. The protein sequence features, protein structural features, and protein spatial features are standardized and adaptively weighted and fused to generate unified multimodal protein features; In the protein graph constructed based on the multimodal protein features, node features are updated by combining graph position encoding and a global attention mechanism; The updated node features are sequentially encoded by center node, subgraph, and context to generate a comprehensive node embedding. Based on the comprehensive node embedding, the interaction types and probabilities between protein pairs are predicted.

2. The protein interaction prediction method based on multimodal geometric feature fusion according to claim 1, characterized in that, Obtaining protein sequence characteristics includes: The amino acid sequence of the target protein is converted into a raw sequence vector representation; Local structural features and temporal correlation features are obtained in parallel from the original sequence vector representation based on one-dimensional convolution operations and bidirectional gated loop units. The local structural features and the temporal correlation features are concatenated to generate initial sequence features; Based on the initial sequence features, a mean vector and a standard deviation vector of the latent variable space are generated. The latent variable space is a low-dimensional latent space used to carry the probability distribution characteristics of the protein sequence. Based on random sampling and gradient differentiability reparameterization operations, the mean vector and the standard deviation vector are converted into the protein sequence features; The protein sequence features are reconstructed based on the decoder, and the feature discriminability is optimized by minimizing the reconstruction error.

3. The protein interaction prediction method based on multimodal geometric feature fusion according to claim 1, characterized in that, Obtaining protein structural features, including: Secondary structural units of amino acid residues are obtained based on the three-dimensional structure of the target protein; Label the types of the secondary structural units and construct a protein structure diagram including the spatial distance relationships of the amino acid residues; Initialize the node features of each secondary structure unit in the protein structure diagram; The node features and the features of its neighboring nodes are aggregated according to the graph neural network; The protein structure features are generated by updating the node features of the protein structure graph through multiple rounds of iteration.

4. The protein interaction prediction method based on multimodal geometric feature fusion according to claim 1, characterized in that, Obtaining the spatial characteristics of proteins includes: The three-dimensional coordinate data of the target protein are converted into standardized point cloud data; The standardized point cloud data is processed by a multilayer perceptron to obtain the local spatial features of each point; Perform max pooling on the local spatial features to generate a global spatial feature representation; The global spatial feature representation is mapped to the protein spatial feature.

5. The protein interaction prediction method based on multimodal geometric feature fusion according to claim 1, characterized in that, Standardization and adaptive weighted fusion, including: The protein sequence features, protein structural features, and protein spatial features are z-score normalized to eliminate differences in feature dimensions. The standardized multimodal features are transformed into initial input features based on the fully connected layer; The initial input features are sequentially subjected to nonlinear orthogonal transformations using the Tanh activation function and the ReLU activation function; The activated initial input features are compressed into a low-dimensional latent space through another fully connected layer to generate a latent representation. In the reconstruction mapping layer, the latent representation is subjected to linear transformation and ReLU nonlinear transformation, and the transformation results are mapped back to the original feature space to generate the unified multimodal protein features.

6. The protein interaction prediction method based on multimodal geometric feature fusion according to claim 1, characterized in that, Updating node features by combining graph position encoding and a global attention mechanism includes: Based on the adjacency relationship of nodes in the protein graph, the feature information of each node and its corresponding neighboring nodes is aggregated through a local graph convolutional network to generate a locally updated node feature representation. In the global attention mechanism, cross-node attention weight calculation and feature weighted aggregation are performed on the locally updated node feature representation to generate a globally enhanced node feature representation; The locally updated node feature representation and the globally enhanced node feature representation are input into a multilayer perceptron for nonlinear feature fusion. The fused node features are output as the updated node features.

7. The protein interaction prediction method based on multimodal geometric feature fusion according to claim 1, characterized in that, Generate a comprehensive node embedding, including: In the protein graph, a target node is selected as the center node. By embedding the features of the target node into the star subgraph of the target node and integrating the neighbor node information of the target node, a center node encoded feature representation is generated. The star-shaped subgraph is processed based on a graph convolutional network, and mean pooling is performed to aggregate the feature information of neighboring nodes in the star-shaped subgraph to generate a subgraph encoded feature representation. The feature representation of the target node in different context subgraphs is analyzed, and a context-encoded feature representation is generated by performing mean aggregation on the embedding features of the target node in multiple context subgraphs. The central node encoding feature representation, the subgraph encoding feature representation, and the context encoding feature representation are weighted and summed to generate the comprehensive node embedding.

8. The protein interaction prediction method based on multimodal geometric feature fusion according to claim 7, characterized in that, Based on the comprehensive node embedding, the prediction of protein-protein pair interaction types and their probabilities includes: Obtain the two node embedding features corresponding to the target protein pair from the comprehensive node embedding; Perform feature interaction operations on the acquired embedding features of the two nodes to generate interaction feature representations between protein pairs; The interaction feature representation is input into a multi-layer fully connected network and subjected to nonlinear transformation to generate probabilistic input features; Apply a probabilistic activation function to the probabilistic input features to output the interaction type between the protein pairs and the corresponding probability prediction results.

9. A protein interaction prediction system based on multimodal geometric feature fusion, characterized in that, The system includes: The feature generation module extracts features from the amino acid sequence, three-dimensional structure, and three-dimensional coordinates of the target protein, generating protein sequence features, protein structure features, and protein spatial features. The weighted fusion module performs standardization and adaptive weighted fusion on protein sequence features, protein structural features, and protein spatial features to generate unified multimodal protein features; The node update module updates node features in a protein graph constructed based on multimodal protein features by combining graph position encoding and a global attention mechanism. The encoding and embedding module sequentially encodes the center node, subgraph, and context of the updated node features to generate a comprehensive node embedding. The interaction prediction module predicts the types and probabilities of interactions between protein pairs based on comprehensive node embedding.

10. The protein interaction prediction system based on multimodal geometric feature fusion according to claim 9, characterized in that, The weighted fusion module includes: The difference elimination unit performs z-score normalization on protein sequence features, protein structure features, and protein spatial features to eliminate differences in feature dimensions; The initial transformation unit transforms the standardized multimodal features into initial input features based on the fully connected layer; The orthogonal transformation unit sequentially performs nonlinear orthogonal transformations of the Tanh activation function and the ReLU activation function on the initial input features; The feature compression unit compresses the activated initial input features into a low-dimensional latent space to generate a latent representation through another fully connected layer. The unified mapping unit performs linear and ReLU nonlinear transformations on the latent representation in the reconstruction mapping layer, and maps the transformation results back to the original feature space to generate unified multimodal protein features.

Images