Protein function prediction method and device based on multi-modal protein data

Abstract

The invention relates to the technical field of artificial intelligence, and provides a protein function prediction method and device based on multi-modal protein data, and the method comprises the steps: obtaining protein multi-source data, carrying out the feature extraction of a protein sequence in the protein multi-source data, and obtaining a protein sequence feature; constructing a heterogeneous graph based on the protein multi-source data; performing feature coding on the heterogeneous graph by adopting a graph attention mechanism to obtain protein graph features; performing multi-modal fusion on the protein sequence features and the protein map features by adopting a gating fusion mechanism to obtain fusion features; and performing multi-label classification prediction based on the fusion features to obtain a protein function annotation result. The accuracy and robustness of protein function prediction can be improved, and the problems that in the prior art, multi-source protein data cannot be effectively integrated, and the method is sensitive to data noise are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a protein function prediction method and device based on multi-modal protein data. BACKGROUND

[0002] Proteins, as the core executors of biological life activities, are widely involved in key biological processes such as metabolic regulation, signal transduction, and cell structure construction. Accurate annotation of their functions is an important basis for analyzing the nature of life phenomena and promoting disease mechanism research and innovative drug development. With the advancement of sequencing technology, the number of known proteins has grown rapidly, but the speed of functional annotation through biological experiments is far behind the speed of sequence discovery. Therefore, it is crucial to develop efficient computational prediction methods.

[0003] Existing computational methods mainly rely on protein amino acid sequences or protein-protein interaction (PPI) networks and other data. However, the existing technology has the following defects: first, sequence-based methods are heavily dependent on sequence similarity and have poor prediction results for proteins with low homology; second, PPI network-based methods are limited by the incompleteness and high noise of experimental data, affecting the reliability of the prediction. In addition, although some studies have attempted to integrate multiple data, they usually use simple feature splicing or linear fusion strategies, failing to deeply explore the complementarity between different modal information, resulting in limited representation ability and prediction accuracy of the model. It can be seen that the existing technology cannot effectively integrate multi-source protein data and is sensitive to data noise, leading to low accuracy and insufficient robustness of protein function prediction.

[0004] The preceding description is intended to provide general background and does not necessarily constitute prior art. SUMMARY

[0005] The embodiments of the present application provide a protein function prediction method and device based on multi-modal protein data, which effectively improves the accuracy and reliability of protein function prediction by integrating protein multi-source data to construct a heterogeneous graph and using a graph attention mechanism and a gating fusion mechanism.

[0006] In a first aspect, the embodiments of the present application provide a protein function prediction method based on multi-modal protein data, comprising:

[0007] Obtaining protein multi-source data and performing feature extraction on protein sequences in the protein multi-source data to obtain protein sequence features;

[0008] Heterogeneous graphs are constructed based on the protein multi-source data. The heterogeneous graphs include node features and multiple types of edges. The node features are generated based on protein domain information. The multiple types of edges include at least one of original interaction edges, original homology edges, and edges supplemented based on sequence similarity.

[0009] The heterogeneous graph is feature-encoded using a graph attention mechanism to obtain protein graph features;

[0010] A gated fusion mechanism is used to perform multimodal fusion of the protein sequence features and the protein map features to obtain fused features;

[0011] Based on the fusion features, multi-label classification prediction is performed to obtain protein function annotation results.

[0012] Furthermore, in some embodiments of this application, the step of acquiring multi-source protein data and extracting features from the protein sequences in the multi-source protein data to obtain protein sequence features includes:

[0013] Protein multi-source data are collected from public databases, including protein sequence data, protein interaction data, homology data, and functional annotation data.

[0014] The collected multi-source protein data were cleaned.

[0015] A pre-trained protein language model was used to extract features from the cleaned protein sequence data;

[0016] The output vector of the last layer of the pre-trained protein language model is extracted as the protein sequence feature, which includes at least one of the protein's evolutionary signal, structural cues, and functional information.

[0017] Furthermore, in some embodiments of this application, the step of using a pre-trained protein language model to extract features from the cleaned protein sequence data includes:

[0018] The protein sequence is preprocessed by adding special markers at the beginning and end of the sequence and converting the amino acids into corresponding numerical codes;

[0019] The preprocessed protein sequence is input into a pre-trained protein language model, and long-distance dependencies in the protein sequence are captured through a multi-head self-attention mechanism.

[0020] The pre-trained protein language model outputs a latent vector representation of the protein sequence, which includes information on the protein's evolutionary conservation and structural features.

[0021] Furthermore, in some embodiments of this application, the construction of a heterogeneity map based on the protein multi-source data includes:

[0022] Construct a set of nodes for a heterogeneous graph, using a single protein as a node in the heterogeneous graph;

[0023] The domain composition of each protein is obtained by using a domain prediction tool, the domain composition is converted into a binary feature vector, and the binary feature vector is dimensionality reduced to obtain node features.

[0024] Multiple types of edges are constructed from the aforementioned protein multi-source data to create the heterogeneous graph;

[0025] Initialize the weights of edges of various types, assigning initial weight values ​​to edges of different types.

[0026] Furthermore, in some embodiments of this application, the construction of multiple types of edges in the heterogeneous graph based on the protein multi-source data includes:

[0027] Interactions with confidence scores greater than a preset confidence threshold are selected from a protein-protein interaction database to establish primal interaction edges between proteins;

[0028] Based on homology grouping information, proteins belonging to the same homology family are connected by original homology edges; the sequence similarity between the protein sequences is calculated, and when the similarity is greater than a preset similarity threshold, supplementary edges are added between the corresponding protein nodes, including supplementary interaction edges and supplementary homology edges.

[0029] Furthermore, in some embodiments of this application, the step of using a graph attention mechanism to encode the heterogeneous graph to obtain protein graph features includes:

[0030] The node features of the heterogeneous graph are input into the input layer of the graph attention network for feature initialization.

[0031] In the hidden layer of the graph attention network, attention weights between nodes are calculated for each type of edge. These attention weights are obtained by using the node features through a learnable function and are used to characterize the strength of the association between nodes.

[0032] The features of neighboring nodes are weighted and aggregated according to the attention weights, and the feature representation of the current node is updated.

[0033] After iterative calculations by a multi-layer graph attention network, the final protein graph features are output, which are used to fuse node self-information and neighborhood association information.

[0034] Furthermore, in some embodiments of this application, the step of calculating the attention weights between nodes for each type of edge includes:

[0035] A linear transformation is performed on the features of the current node and the features of the neighboring nodes to obtain the transformed feature vector.

[0036] The initial attention score is calculated based on the transformed node features using the learnable attention function;

[0037] The initial attention scores of all neighboring nodes of the same node are normalized to obtain the attention weights.

[0038] Furthermore, in some embodiments of this application, the multimodal fusion of the protein sequence features and the protein map features using a gated fusion mechanism to obtain fused features includes:

[0039] Perform dimensional alignment on the protein sequence features and the protein graph features;

[0040] The dimension-aligned protein sequence features and protein map features are concatenated to obtain a joint feature vector.

[0041] The joint feature vector is mapped using a multilayer perceptron, and a gating factor is obtained through a nonlinear activation function. The gating factor is used to adaptively adjust the weights of the two modal features.

[0042] Based on the gating factor, the protein sequence features and protein map features are weighted and fused to obtain fused features;

[0043] The fusion features are linearly transformed, and the predicted probabilities of each functional label are obtained through an activation function, outputting the protein function annotation results.

[0044] Furthermore, in some embodiments of this application, the step of performing multi-label classification prediction based on the fusion features to obtain protein function annotation results includes:

[0045] The fused features are input into a multi-label classifier, which includes at least one fully connected layer.

[0046] The predicted probability of each feature label is calculated using an activation function;

[0047] Functional tags with predicted probabilities greater than a preset probability threshold are used as predicted functions of proteins, generating protein function annotation results that include functional tags.

[0048] Secondly, embodiments of this application provide a protein function prediction device based on multimodal protein data, comprising:

[0049] The feature extraction module is used to acquire protein multi-source data and extract features from the protein sequences in the protein multi-source data to obtain protein sequence features.

[0050] A heterogeneous graph module is used to construct a heterogeneous graph based on the protein multi-source data. The heterogeneous graph includes node features and multiple types of edges. The node features are generated based on protein domain information. The multiple types of edges include at least one of original interaction edges, original homology edges, and edges supplemented based on sequence similarity.

[0051] The feature encoding module is used to encode the heterogeneous graph using a graph attention mechanism to obtain protein graph features;

[0052] The feature fusion module is used to perform multimodal fusion of the protein sequence features and the protein map features using a gated fusion mechanism to obtain fused features;

[0053] The classification prediction module is used to perform multi-label classification prediction based on the fused features to obtain protein function annotation results.

[0054] Thirdly, embodiments of this application provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the protein function prediction method based on multimodal protein data as described in the first aspect.

[0055] Fourthly, embodiments of this application provide a storage medium storing a computer program capable of being loaded by a processor and executing the protein function prediction method based on multimodal protein data as described in the first aspect.

[0056] This application provides a method and apparatus for protein function prediction based on multimodal protein data. First, by acquiring multi-source protein data and extracting sequence features, it breaks the dependence of traditional methods on a single data source, providing basic functional information at the sequence level for the prediction task. Next, a heterogeneous graph containing domain-derived node features and various types of edges is constructed based on the multi-source data. The domain node features naturally associate with protein functional modules, while the various edge designs, including original edges and edge supplementation based on sequence similarity, effectively compensate for deficiencies such as incomplete PPI data and errors in homology information, enriching the context of protein associations. Then, a graph attention mechanism is used to encode features in the heterogeneous graph, adaptively identifying and aggregating effective information in nodes and their neighborhoods, filtering data noise interference, and generating more accurate protein graph features. Subsequently, a gating fusion mechanism dynamically adjusts the fusion weights of sequence features and graph features, rather than simply superimposing them, fully exploring the complementary value of the two modalities and forming a more comprehensive protein function representation. Finally, multi-label classification prediction is performed based on high-quality fused features, which can more accurately match multiple functional annotations corresponding to proteins, thereby improving the accuracy and robustness of protein function prediction results. As can be seen, this application achieves efficient synergistic utilization of multimodal protein data by constructing an enhanced heterogeneous graph and utilizing attention mechanism and gating fusion, thereby effectively improving the accuracy and reliability of protein function prediction and solving the problem that existing technologies are difficult to effectively integrate multi-source protein data and are sensitive to data noise. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is an application environment diagram of the protein function prediction method based on multimodal protein data provided in the embodiments of this application;

[0059] Figure 2 This is a flowchart illustrating the protein function prediction method based on multimodal protein data provided in an embodiment of this application.

[0060] Figure 3 This is a schematic diagram of the overall architecture of the protein function prediction method based on multimodal protein data provided in the embodiments of this application;

[0061] Figure 4 This is a schematic diagram of the structure of the protein function prediction device based on multimodal protein data provided in the embodiments of this application;

[0062] Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0063] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of systems and methods consistent with those detailed in the appended claims or with some aspects of this application.

[0064] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover descriptions such as non-exclusive inclusion, so that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, components, features, and elements with the same names in different embodiments of this application may have the same meaning or different meanings, the specific meaning of which must be determined by its interpretation in that specific embodiment or further in conjunction with the context of that specific embodiment.

[0065] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0066] In the following description, the use of suffixes such as "module," "part," or "unit" to denote elements is solely for the purpose of illustrative purposes and has no specific meaning in itself. Therefore, "module," "part," or "unit" may be used interchangeably.

[0067] To address the aforementioned technical problems and overcome the shortcomings of existing technologies, this application provides a method and apparatus for predicting protein function based on multimodal protein data. By integrating multi-source protein data to construct heterogeneous graphs and employing graph attention and gating fusion mechanisms, the accuracy and reliability of protein function prediction are effectively improved.

[0068] Figure 1 This is a diagram illustrating the application environment of a protein function prediction method based on multimodal protein data in one embodiment. (Refer to...) Figure 1This protein function prediction method based on multimodal protein data is applied to a protein function prediction system based on multimodal protein data. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal, specifically a mobile phone, tablet computer, laptop computer, or at least one of these. The server 120 can be a standalone server or a server cluster consisting of multiple servers. Server 120 is configured to execute the aforementioned protein function prediction method based on multimodal protein data, including: acquiring multi-source protein data and extracting features from the protein sequences in the multi-source protein data to obtain protein sequence features; constructing a heterogeneous graph based on the multi-source protein data, the heterogeneous graph including node features and various types of edges, the node features being generated based on protein domain information, and the various types of edges including at least one of original interaction edges, original homology edges, and edges supplemented based on sequence similarity; using a graph attention mechanism to encode the features of the heterogeneous graph to obtain protein graph features; using a gating fusion mechanism to perform multimodal fusion of protein sequence features and protein graph features to obtain fused features; and performing multi-label classification prediction based on the fused features to obtain protein function annotation results.

[0069] Please see Figure 2 , Figure 2 This is a flowchart illustrating a protein function prediction method based on multimodal protein data according to an embodiment of this application. Specifically, the protein function prediction method based on multimodal protein data according to an embodiment of this application may include the following steps:

[0070] S1. Obtain protein multi-source data and extract features from the protein sequences in the protein multi-source data to obtain protein sequence features;

[0071] Specifically, for step S1, multi-dimensional data covering core protein information is collected from public databases, including but not limited to protein amino acid sequence data from the UniProt database, protein-protein interaction (PPI) data from the STRING database, homology data from the eggNOG database, and functional annotation data from the GOA database. A pre-trained protein language model, such as ESM2 developed by MetaAI, specifically the esm2_t33_650M_UR50D model, is used to encode the protein sequence. The model captures long-distance sequence dependencies through a multi-head self-attention mechanism of the Transformer architecture, and the output vector of the last layer is the sequence feature, including but not limited to at least one of evolutionary signals, structural cues, and functional information. Taking the human tumor suppressor protein TP53 as an example, its 393-residue amino acid sequence is obtained from UniProt. After inputting the sequence into the ESM2 model, the 1280-dimensional vector output from the 33rd layer (final layer) is taken as the sequence feature of TP53. This vector implicitly captures signals related to domain conservation and function during pre-training.

[0072] S2. Construct a heterogeneous graph based on multi-source protein data. The heterogeneous graph includes node features and multiple types of edges. The node features are generated based on protein domain information. The multiple types of edges include at least one of the following: original interaction edges, original homology edges, and edges supplemented based on sequence similarity.

[0073] Specifically, for step S2, the domain annotations of the protein are extracted using the InterProScan tool (e.g., entries from databases such as Pfam and COILS). The "whether a certain domain is present" is converted into a binary feature vector, and then the dimensionality is reduced through a fully connected layer linear transformation to obtain the node features. The heterogeneous graph also includes the construction of various types of edges, including original interaction edges, original homologous edges, and complementary edges. For example, InterProScan detected that TP53 contains four domains, including Pfam00870 (DNA binding domain) and Pfam02277 (tetramerization domain). This was converted into a binary vector containing 10,000 InterPro entries (corresponding to a domain position of 1), and then reduced to 256-dimensional node features through a fully connected layer. The original PPI edges were constructed by taking the top 50 interacting proteins (such as MDM2) of TP53 from STRING. The original homology edges were constructed by taking its orthologous protein Trp53 (mouse) from eggNOG. There were no homology edges between TP53 and TP73, but DIAMOND comparison showed that TP53 and TP73 had significant sequence similarity (e-value less than 1e-1000), and TP53 was indirectly connected to TP73 in the protein-protein interaction network. Therefore, homology edges were added between them.

[0074] S3. The graph attention mechanism is used to encode the features of the heterogeneous graph to obtain the protein graph features;

[0075] Specifically, in step S3, the node features of the heterogeneous graph (such as the 256-dimensional structural domain features of TP53) are input into the input layer of the Graph Attention Network (GAT) to complete the initial feature mapping. For different types of edges, linear transformations are performed on the features of the current node and its neighboring nodes, and the similarity is calculated to obtain an initial score. An initial association score between nodes is then calculated using a learnable attention function, and normalized to obtain the attention weights. The features of the neighboring nodes are weighted and summed according to the attention weights to update the feature representation of the current node. After 2-3 layers of GAT iterations, a protein graph feature integrating the node's own structural domain information and the neighboring association information is output.

[0076] S4. A gated fusion mechanism is used to perform multimodal fusion of protein sequence features and protein map features to obtain fused features;

[0077] Specifically, for step S4, a linear transformation is used to map the sequence features and graph features to the same dimension, such as converting the 1280-dimensional sequence features and 256-dimensional graph features of TP53 to 512 dimensions. The aligned feature vectors of the two classes are concatenated to obtain joint features containing both self-attributes and related attributes. A multilayer perceptron (MLP) is used to perform a nonlinear mapping on the joint features, outputting a gating factor in the 0-1 interval to reflect the importance of graph features relative to sequence features. The two classes of features are weighted and summed using the gating factor to obtain the fused features.

[0078] S5. Perform multi-label classification prediction based on fusion features to obtain protein function annotation results;

[0079] Specifically, for step S5, the fused features are input into a multi-label classifier containing at least one fully connected layer. The output is mapped to the 0-1 interval using the Sigmoid activation function, with each value corresponding to the function prediction probability of a GO term. A preset probability threshold (e.g., 0.5) is set, and GO terms with probabilities greater than the threshold are used as predicted functions to generate annotation results. For example, inputting 512-dimensional fused features into the classifier outputs the probabilities of 22,916 GO terms; key probabilities: "DNA binding" (GO:0003677) 0.92, "apoptosis regulation" (GO:0042981) 0.89, "cell cycle arrest" (GO:0007050) 0.85; labels with probabilities > 0.5 are selected to generate annotation results containing the above three core functions.

[0080] This embodiment integrates multi-source protein data and utilizes graph attention and gating fusion mechanisms to fully leverage sequence and interaction information, generating high-quality sequence and graph features. Then, it captures complex relationships between proteins and suppresses noise through heterogeneous graph modeling and attention mechanisms. Finally, the gating fusion mechanism deeply integrates multimodal information to improve the accuracy and robustness of protein function prediction.

[0081] Furthermore, in some embodiments, step S1, "acquiring multi-source protein data and extracting features from the protein sequences in the multi-source protein data to obtain protein sequence features," may specifically include:

[0082] S11. Collect multi-source protein data from public databases, including protein sequence data, protein interaction data, homology data, and functional annotation data;

[0083] Specifically, for step S11, based on the multimodal information requirements for protein function prediction, four types of core data are collected from authoritative public databases. Each data type and source has a clear purpose: protein sequence data comes from the UniProt database, which integrates global protein sequences and annotation information, covering a wide range of species; protein-protein interaction (PPI) data comes from the STRING database (e.g., version 12), which integrates large-scale experimental and predicted interaction data and includes confidence scores; homology data comes from the eggNOGv5.0 database, providing cross-species orthologous and paralogous grouping annotations; and functional annotation data comes from the GeneOntology Annotation (GOA) database. The corresponding GO terminology definition file must also be downloaded from the GeneOntology official website (to clarify the functional labeling system). To adapt to real-world prediction scenarios, data collection must be screened based on the research subject (e.g., model organisms) and data reliability requirements. For example, manually compiled or experimentally supported annotation data should be prioritized to avoid low-reliability information from purely electronic inferences. Data was collected using Arabidopsis thaliana as the research subject. Amino acid sequence data of all known Arabidopsis proteins were downloaded from UniProt, including sequence length, residue composition, and other information. Interaction data between Arabidopsis proteins were obtained from STRING version 12, retaining entries with confidence scores. Orthologous grouping information of Arabidopsis proteins was extracted from eggNOGv5.0 to clarify the association with proteins of the same family. The Arabidopsis GO annotation file of January 17, 2024 was downloaded from the GOA database, and the official GO terminology definition file (including the three nomenclature spaces of molecular function (MF), biological process (BP), and cellular component (CC)) was also downloaded simultaneously.

[0084] S12. Perform data cleaning on the collected multi-source protein data;

[0085] Specifically, for step S12, noise, redundancy, and unreliable entries in the data are removed to ensure data quality to support subsequent modeling. For functional annotation data, entries with the evidence code IEA (Inferred from Electronic Annotation) are removed, and only manually compiled or experimentally supported annotations (such as entries with evidence codes EXP or IDA) are retained. For protein-protein interaction data, entries with confidence scores that meet preset standards are selected (such as retaining the top k high-confidence interactions) to reduce false positives. For homology data, the consistency of homology grouping is verified, and grouping information with ambiguous labeling or conflicting sources is removed. For sequence data, sequences containing a large number of unknown residues (such as "X" accounting for >10%) or abnormal lengths (too short <50 residues or too long >5000 residues) are removed.

[0086] S13. Use a pre-trained protein language model to extract features from the cleaned protein sequence data;

[0087] Specifically, for step S13, a pre-trained protein language model based on the Transformer architecture is used, with ESM2 (Evolutionary Scale Modeling v2) being the preferred choice. Trained on large-scale protein sequence datasets, ESM2 can capture evolutionary-scale patterns and long-distance sequence dependencies, exhibiting stronger generalization capabilities than previous models (such as ESM-1b). In practical applications, the esm2_t33_650M_UR50D model, containing 33 Transformer layers, is commonly used and can generate residue-level and protein-level features. The cleaned protein sequence is input into the model, which uses a multi-head self-attention mechanism to mine the association information of amino acid residues in the sequence (such as cross-domain interactions), outputting a latent vector representation containing biological meaning. This representation integrates deep information such as evolutionary conservation and structural features.

[0088] S14. Extract the output vector of the last layer of the pre-trained protein language model as protein sequence features. Protein sequence features include at least one of the following: protein evolutionary signals, structural cues, and functional information.

[0089] Specifically, for step S14, the output of the later layers of the pre-trained protein language model can capture more complex and comprehensive information. Therefore, the output vector of the last layer (such as the 33rd layer of esm2_t33_650M_UR50D) is selected as the protein sequence feature. This vector is a protein-level global representation, rather than a residue-level local representation. The sequence feature vector includes, but is not limited to, at least one of three types of core information: evolutionary signals, such as highly conserved amino acid sites in the sequence (reflecting evolutionary relationships between species); structural cues, such as the distribution of potential secondary structures (α-helices, β-sheets) and domain boundaries; and functional information, such as the characteristic patterns of active sites and ligand-binding regions.

[0090] This embodiment collects multi-source protein data from public databases and performs data cleaning. It then uses a pre-trained protein language model to extract features from protein sequences, which can obtain high-quality protein sequence features, including but not limited to at least one of protein evolutionary signals, structural cues, and functional information. This provides a solid data foundation for subsequent protein function prediction and improves the accuracy and reliability of the prediction model.

[0091] Furthermore, in some embodiments, step S13, "using a pre-trained protein language model to extract features from the cleaned protein sequence data," may specifically include:

[0092] S131. Preprocess the protein sequence by adding special markers at the beginning and end of the sequence and converting the amino acids into corresponding numerical codes;

[0093] Specifically, for step S131, the input format of the pre-trained protein language model (such as ESM2) is adapted to convert the unstructured amino acid character sequence into structured data that the model can compute. Simultaneously, special markers are used to define sequence boundaries, assisting the model in recognizing global information. Referring to the input specifications of natural language processing models, start markers (such as [CLS]) are added at the beginning of the protein sequence, and stop markers (such as [SEP]) are added at the end. The output of the start markers is typically used for subsequent protein-level global feature extraction. Twenty common standard amino acids (such as A, V, L, I, etc.) and special characters (such as the unknown residue "X") are mapped to unique integer codes (e.g., establishing an "amino acid-number" correspondence table, A→1, V→2, L→3…X→21), eliminating the interference of character format on model computation.

[0094] S132. Input the preprocessed protein sequence into a pre-trained protein language model and capture long-distance dependencies in the protein sequence through a multi-head self-attention mechanism;

[0095] Specifically, for step S132, a pre-trained protein language model based on the Transformer architecture (such as ESM2) is used. Its core advantage lies in overcoming the limitations of traditional models (such as RNNs) in capturing long sequence dependencies through a multi-head self-attention mechanism. This mechanism calculates the association weights of amino acids at different positions in parallel through multiple attention heads, simultaneously focusing on local sequence details (such as the physicochemical interactions of adjacent residues) and global associations (such as residue interactions across structural domains). The preprocessed digital encoded sequence is converted into a vector by the embedding layer and then input into the Transformer encoder of the model. Each attention head calculates the attention score between any two positions in the sequence (such as the i-th amino acid and the j-th amino acid). The higher the score, the stronger the functional association between the two. The results from multiple heads are concatenated and linearly transformed to ultimately achieve accurate capture of long-distance dependencies.

[0096] S133. Output the latent vector representation of the protein sequence through a pre-trained protein language model. The latent vector representation includes the evolutionary conservation and structural feature information of the protein.

[0097] Specifically, in step S133, the Transformer encoder of the pre-trained protein language model abstracts features from the input sequence layer by layer. Shallow outputs focus on basic information such as the physicochemical properties of amino acids, while deeper outputs (such as layers 20-33 of ESM2) contain more complex biological connotations. Latent vectors, represented as vectors in the intermediate or output layers of the model, contain both residue-level local features and protein-level global information. The core biological information in latent vectors includes evolutionary conservation and structural features. The model, through learning from large-scale sequence data, encodes the evolutionary stability of amino acid sites in the latent vectors. For example, the vector dimension of highly conserved residues across species (such as Arg248 in TP53) is significantly higher than that of variable residues. Latent vectors can reflect the potential structural properties of proteins; for example, the vectors corresponding to residues in α-helical regions have similar patterns, and the vectors of residues at domain boundaries show obvious inflections.

[0098] This embodiment ensures the compliance and consistency of the model input through sequence preprocessing. The pre-trained model based on the multi-head self-attention mechanism breaks through the technical bottleneck of long sequence association capture and can effectively mine residue interactions across structural domains. The output latent vector integrates core biological information such as evolutionary conservation and structural features, providing a sequence feature foundation with both depth and discriminativeness for subsequent multimodal fusion and functional prediction, and significantly improving the overall model's ability to capture key signals of protein function.

[0099] In a specific embodiment, to fully utilize the information contained in the sequence for extracting protein sequence features, this embodiment uses Evolutionary Scale Modeling v2 (ESM2) to encode the protein sequence. ESM2 is a protein language model based on the Transformer architecture developed by MetaAI, with a design philosophy similar to natural language models (such as BERT and GPT). It is trained on large-scale protein sequence datasets to capture potential patterns at the evolutionary scale. Thanks to the powerful modeling capabilities of the Transformer, ESM2 can effectively capture long-range dependencies and global contextual information in the sequence, thereby obtaining a global representation across residues while extracting local features, generating biologically meaningful embedding vectors. Compared to its predecessor, ESM-1b, ESM2 has a larger model scale and is trained on higher-quality sequence datasets, thus performing better in sequence modeling and generalization. In practical applications, ESM2 can be considered a sequence encoder that generates residue-level feature vectors for each protein. These feature vectors include, but are not limited to, at least one of evolutionary signals, structural cues, and functional information, and can be widely applied to various downstream prediction tasks.

[0100] In this embodiment, the officially released esm2_t33_650M_UR50D model is used for sequence feature extraction. This model consists of 33 Transformer layers, each generating progressively abstract feature embeddings. Later layers tend to capture more complex and comprehensive information; therefore, the output of the 33rd (final) layer is chosen as the final sequence representation for the functional prediction task. Let the input protein sequence be seq=[s1,s2,…,s…]. L ], where L represents the sequence length, s i This represents the i-th amino acid residue in the sequence. The ESM2 model processes this sequence and outputs a global protein-level embedding from layer 33. Where, vector h seq Representing the global features of a protein, this 1280-dimensional embedding vector captures at least one of the protein’s features, including but not limited to evolutionary signals, structural cues, and functional information, and serves as the sequence characterization input for subsequent multimodal fusion and functional prediction modules.

[0101] Furthermore, in some embodiments, step S2, "constructing a heterogeneous map based on multi-source protein data," may specifically include:

[0102] S21. Construct a set of nodes for a heterogeneous graph, using a single protein as a node in the heterogeneous graph;

[0103] Specifically, for step S21, the core of the heterogeneous graph is to integrate multi-source information of proteins through the topological relationships between nodes and edges, using a single protein as the basic node unit. This aligns with the biological understanding that proteins are independent basic units of functional execution, and also facilitates the direct association of protein attributes such as domains, interactions, and homology with nodes. The node set must cover all proteins in the target prediction scenario (such as all known proteins of a model organism or proteins in a specific research dataset), ensuring that no core functionally related proteins are omitted, thus providing a complete set of objects for subsequent topological feature learning.

[0104] S22. Obtain the domain composition of each protein using a domain prediction tool, convert the domain composition into a binary feature vector, and perform dimensionality reduction on the binary feature vector to obtain node features;

[0105] Specifically, for step S22, InterProScan is used as the domain prediction tool. This tool integrates multiple authoritative databases such as Pfam, COILS, and PRINTS, and can accurately identify conserved domains, functional sites, and family affiliations of proteins. Its output "InterPro entries" are the core carriers of domain features. Using the set of InterPro entries for all proteins as the dimensional space, for a single protein, if it contains a certain InterPro entry, the corresponding position of the vector is 1; otherwise, it is 0, forming a high-dimensional sparse binary vector. This vector directly reflects the differences in the domain composition of proteins. The binary vector dimension often reaches thousands to tens of thousands (e.g., if it contains 10,000 unique InterPro entries, the dimension is 10,000), resulting in computational redundancy. Dimensionality reduction is achieved through linear transformation using fully connected layers (combined with non-linear activation functions such as ReLU). While compressing the dimension, the latent semantic relationships between domains are extracted, resulting in low-dimensional dense node features.

[0106] S23. Constructing various types of edges in heterogeneous graphs based on multi-source protein data;

[0107] Specifically, for step S23, protein-protein interactions (reflecting local functional cooperation) and homology (reflecting global evolutionary conservation) are naturally complementary. Furthermore, incomplete experimental data needs to be compensated for by supplementary edges. Therefore, the edge types cover three core relationships: original interaction edges, original homology edges, and supplementary edges. Original interaction edges originate from high-confidence PPI data in the STRING database, reflecting experimentally supported dynamic cooperation between proteins. Original homology edges originate from homology groupings in the eggNOG database, reflecting cross-species evolutionary functional conservation. Edges supplemented based on sequence similarity address the issues of false negatives and misjudgments of homology in PPI data by using DIAMOND to calculate sequence similarity (e-value < 0.0001) to supplement homology or interaction edges.

[0108] S24. Initialize the weights of edges of various types, assigning initial weight values ​​to edges of different types;

[0109] Specifically, for step S24, initial weights are set based on the reliability of edge information. Original edges (original interaction edges, original homologous edges) originate from experimental / annotation data in authoritative databases, ensuring high reliability, and their initial weights are set to larger values. Supplementary edges are inferred based on sequence similarity, which carries potential noise, and their initial weights are set to smaller values. This preserves the reliability of the original data while reserving space for adaptive weight adjustment during model training. For example, the initial weight of the original edge is set to 0.8, and the initial weight of the supplementary edge is set to 0.2. Specific values ​​can be fine-tuned according to the dataset size, but the core logic of "original edge weight > supplementary edge weight" must be maintained.

[0110] This embodiment ensures object uniqueness by using a single protein as a node, and the domain feature encoding based on InterProScan endows the nodes with strong biological interpretability. Multi-type edges integrate interaction, homology, and sequence similarity information, making up for the limitations of a single data source. Edge weight initialization is based on reliability grading, reducing potential noise from supplementary information. The resulting heterogeneous graph achieves deep fusion of protein's own structural features and associated topological features, providing high-quality, low-noise structured input for feature encoding of subsequent graph attention mechanisms, and improving the accuracy of functional prediction from the network modeling level.

[0111] Furthermore, in some embodiments, step S23, "constructing multiple types of edges for a heterogeneous graph based on multi-source protein data," may specifically include:

[0112] S231. Filter the protein-protein interaction database for interaction relationships with confidence scores greater than a preset confidence threshold, and establish the original interaction edges between proteins;

[0113] Specifically, for step S231, protein-protein interaction (PPI) data is sourced from the STRING database. This database integrates multi-source PPI information, including experimental verification and computational prediction, and assigns a confidence score of 0-1 to each pair of interacting proteins. Higher scores indicate stronger reliability of the interaction. To reduce false positive interference and control the density of the graph, a strategy combining confidence ranking screening and threshold screening is adopted. The top k interaction pairs with the highest confidence scores for each protein are retained. This is essentially equivalent to using the lowest score corresponding to the "top k" as a dynamically preset confidence threshold (e.g., if the confidence score of the 50th interaction of a protein is 0.7, then the threshold is 0.7). For each selected protein pair, an undirected edge is established between their corresponding nodes, marked as the original interaction edge, directly reflecting the functional cooperation relationship verified with high confidence.

[0114] S232. Based on homology grouping information, connect proteins belonging to the same homology family through original homology edges;

[0115] Specifically, for step S232, the homology data comes from the eggNOGv5.0 database. This database, through cross-species sequence alignment and evolutionary analysis, divides proteins into multiple orthologs and paralogs. Proteins within the same group share a common evolutionary origin and similar functional potential. Primitive homology edges are established primarily based on ortholog groups, as orthologous proteins exhibit stronger functional conservation across different species (e.g., human TP53 and mouse Trp53). Paralog groups can be selectively included based on prediction needs, but the group type must be labeled. For any two protein nodes belonging to the same target homology group (e.g., ortholog groups), an undirected edge is established and labeled as a primitive homology edge, reflecting evolutionarily conserved functional associations.

[0116] S233. Calculate the sequence similarity between protein sequences. When the similarity is greater than a preset similarity threshold, add supplementary edges between the corresponding protein nodes. The supplementary edges include supplementary interaction edges and supplementary homology edges.

[0117] Specifically, for step S233, the DIAMOND tool is used to perform pairwise alignment of protein sequences. This tool combines alignment speed and accuracy, and similarity is measured by e-value (expected score). The smaller the e-value, the higher the similarity. The preset similarity threshold is set to e-value < 0.0001 (i.e., the similarity reaches a statistically significant level). To address homology relationships that may be missed in the homology database (such as misclassification due to protein variations), sequence similarity is used to supplement them: first, the top 10 interacting proteins of the target protein are selected, then the top 10 interacting proteins of each interacting protein are selected, resulting in 100 candidate proteins. The 10 candidate proteins with the highest similarity to the target protein are then selected to establish supplementary homology edges. To address the false negative issue in PPI data (such as the failure to detect true interactions in experiments), a comprehensive screening was conducted using similarity and interaction ranking: ① The top 10 proteins with the highest similarity to the target protein and their top 10 interacting proteins were selected; ② The top 10 proteins with the highest interaction with the target protein and their top 10 similar proteins were selected; these were combined to obtain 200 candidate proteins, which were then sorted by a combination of similarity and interaction ranking, and the top 10 candidate proteins were selected to establish supplementary interaction edges.

[0118] This embodiment ensures the reliability of associations by relying on the confidence screening of the STRING database for the original interaction edges, captures evolutionarily conserved functions based on the grouping information of eggNOG for the original homologous edges, and fills data gaps by inferring the sequence similarity of DIAMOND for the supplementary edges. The collaborative construction of the three types of edges enables the heterogeneous graph to retain the core associations verified by experiments, and expand the association coverage through computational inference, thereby improving the integrity, accuracy and robustness of the graph structure, and providing a high-quality topological relationship foundation for the feature encoding of the subsequent graph attention mechanism.

[0119] In a specific embodiment, to effectively integrate multimodal protein information and obtain comprehensive protein features, a heterogeneous protein graph is constructed. In the heterogeneous protein graph, the InterProScan tool is used to extract protein domain features as node features, and four edge types are included: original PPI edges, original homology edges, homology edges supplemented by sequence similarity, and PPI edges.

[0120] For graph node construction, protein domains typically correspond to specific structural and functional modules. Compared to raw sequence information, domain-based features provide stronger biological interpretability and are an important basis for protein function prediction. Therefore, InterProScan is used to obtain protein domain information to construct node features in the graph, thereby supporting the effective learning of downstream models. InterProScan is a widely used protein annotation tool that integrates multiple mature databases (such as Pfam, COILS, PRINTS, and PANTHER) and can identify conserved domains, functional sites, and protein family affiliations in protein sequences.

[0121] Formally, for a group of proteins Each protein p i InterProScan is used to process and obtain the annotated InterPro entries for each protein p. Assuming there are m unique InterPro entries in the entire dataset, for each protein p... i Construct a binary vector x i ∈{0,1} m The j-th element x i,j =1 indicates protein p i It has the j-th InterPro annotation, otherwise x i,j =0.

[0122] To compress these high-dimensional sparse vectors and extract latent semantic features, a fully connected layer is used to perform a linear transformation on the Interpro sparse binary feature vector of each protein, thereby obtaining a new protein domain feature table.

[0123] h inter =σ(Wx i +b),

[0124] in, It is a learnable weight matrix. This is the bias term, where d represents the dimension of the hidden layer vector, and σ(·) represents the non-linear activation function (such as ReLU). The final embedded vector h is obtained. inter It captures information at the protein domain level, providing a powerful characterization for protein function prediction based on graph neural networks.

[0125] For primitive edge construction, homology information primarily reflects the global evolutionary relationships between proteins, while PPI networks emphasize local functional associations under specific biological conditions. The former provides global structural and functional evolutionary clues, while the latter reveals dynamic interactions under specific physiological conditions. Homologous proteins typically participate in similar biological processes and exhibit similar interaction patterns; these two types of information are naturally complementary in a biological sense. Given the close relationship between homology and interaction, integrating them into a heterogeneous graph can leverage their synergistic effect, reducing homology data errors caused by protein variation and noise due to inaccurate protein interaction information. Therefore, in this embodiment, homology and interaction relationships are modeled in a unified heterogeneous graph to obtain a more comprehensive representation of protein features, thereby improving the accuracy and robustness of functional prediction.

[0126] First, initial edges are constructed directly from the database. The PPI network edges are built using the STRING database, which integrates large-scale protein-protein interaction data and assigns a confidence score to each pair of interacting proteins to represent the confidence of their interaction association. To reduce potential false positives and computational overhead, only the top k interacting proteins for each protein are retained, and an edge is added between each retained pair of interacting proteins. This strategy helps control the density of the graph, resulting in a high-quality, compact graph. In this embodiment, k is set to 50. Then, a homology network is constructed based on the eggNOG database, which provides large-scale orthologous and paralogous group annotations across species. If two proteins belong to the same orthologous group, a homology edge is added between them.

[0127] For the construction of supplementary edges, since the information of most proteins in the protein database is incomplete, and mutations may lead to inaccurate homology information recorded in the homology database and PPI networks often have a lot of noise, supplementary edge information based on sequence similarity is introduced to enhance the robustness and generalization ability of the constructed heterogeneous graph.

[0128] Specifically, DIAMOND is used to calculate the sequence similarity between pairs of proteins, and two protein sequences with an e-value less than 0.0001 are defined as similar. For homology edges based on sequence similarity, the top 10 interacting proteins of the target protein are first selected. Then, for each interacting protein, its top 10 interacting proteins are retrieved, resulting in a total of 100 candidate proteins. Among these candidate proteins, the 10 proteins with the highest sequence similarity to the target protein are selected, and a homology edge is defined between them and the target protein. For interaction edges based on sequence similarity, the top 10 proteins with the highest sequence similarity to the target protein are first identified, and for each protein, its top 10 interacting proteins based on interaction score are retrieved. Additionally, the top 10 interacting proteins based on the target protein's interaction score are selected, and for each interacting protein, its top 10 proteins with the highest sequence similarity are retrieved. This process yields a total of 200 candidate proteins. A comprehensive ranking (similarity ranking + interaction ranking) is calculated, and the top 10 proteins in the comprehensive ranking are defined as having interaction edges with the target protein.

[0129] Ultimately, the constructed heterogeneous graph integrates node features based on structural domains, topological features based on interaction information and homology information, providing rich multi-level structural information for learning protein features.

[0130] Furthermore, in some embodiments, step S3, "using a graph attention mechanism to encode the features of the heterogeneous graph to obtain protein graph features," may specifically include:

[0131] S31. Input the node features of the heterogeneous graph into the input layer of the graph attention network to initialize the features;

[0132] Specifically, for step S31, the input layer of the Graph Attention Network (GAT) needs to perform standardized preprocessing and dimensionality adaptation on the original node features of the heterogeneous graph to ensure that the feature format conforms to the computational logic of the subsequent hidden layers, while preserving the core information of the protein domain structure contained in the node features (such as the DNA binding domain features of TP53). The node features are initially low-dimensional dense vectors (e.g., 256-dimensional) generated based on the domains. After being input into the GAT input layer, the features are initially mapped through a trainable linear transformation matrix (the dimension can remain unchanged or be adjusted to adapt to the dimension of attention calculation), completing the feature initialization and laying the foundation for subsequent attention weight calculation.

[0133] S32. In the hidden layer of the graph attention network, the attention weights between nodes are calculated for each type of edge. The attention weights are calculated by using the node features through a learnable function and are used to characterize the strength of the association between nodes.

[0134] Specifically, for step S32, the heterogeneous graph contains various types of edges, including original interaction edges, original homologous edges, and complementary edges. Different edges have different biological meanings and information reliability (e.g., original edges are more reliable than complementary edges). Attention weights need to be calculated independently for each edge type to accurately distinguish the contribution of different associations to functional prediction. Attention weights are calculated from node features through a learnable function mapping and are used to characterize the strength of association between nodes. Higher association strength indicates a closer functional association between two proteins, corresponding to a larger attention weight. This is divided into three steps: ① Perform a linear transformation on the initial features of the current node and its neighboring nodes; ② Calculate the similarity of the transformed features to obtain an initial attention score; ③ Normalize the scores of all neighboring nodes of the same node (e.g., using the Softmax function) to obtain attention weights in the 0-1 interval.

[0135] S33. Perform weighted aggregation of the features of neighboring nodes based on attention weights, and update the feature representation of the current node;

[0136] Specifically, for step S33, unlike the average aggregation of traditional graph models, GAT performs differentiated aggregation of neighborhood node features through attention weights. Important neighborhood nodes (high weights) contribute more to the features, while secondary neighborhood nodes (low weights) contribute less, thereby highlighting key functional association information and suppressing noise interference. The node feature update method is based on the aggregation formula: the updated feature of the current node = the sum of all neighborhood node features × the corresponding attention weights. Then, a non-linear activation function (such as ReLU) is used to enhance the feature expressive power, ultimately obtaining the node features that integrate neighborhood association information.

[0137] S34. After iterative calculations by a multi-layer graph attention network, the final protein graph features are output. These protein graph features are used to fuse node self-information and neighborhood association information.

[0138] Specifically, for step S34, a single-layer GAT can only capture the information of the node's direct neighborhood (first-order neighbors), while protein functional associations often involve indirect cooperation (such as the indirect regulation of TP53→MDM2→E3 ubiquitin ligase). For example, if the GAT is set to two layers, multi-layer iteration can capture higher-order neighborhood information (second-order and above neighbors), achieving a deep fusion of "self-information + local direct association + global indirect association". The final output protein map features are low-dimensional dense vectors, which not only retain the node's own domain information (such as the DNA-binding domain of TP53), but also integrate the functional cooperation signals of direct neighbors (such as the interaction with MDM2) and the evolutionary association signals of indirect neighbors (such as the homologous protein network of Trp53), possessing comprehensive topological feature expression capabilities.

[0139] This embodiment achieves differentiated modeling of different associations by calculating attention weights based on edge type. The relevance-based weight allocation accurately describes the association strength between key functions. Weighted aggregation and multi-level iteration gradually fuse self-information with local-global neighborhood associations, effectively mining potential functional cooperation signals in the protein network. The final output protein graph features possess comprehensive and accurate topological information representation capabilities, providing high-quality network structure feature inputs for subsequent multimodal fusion, and improving the overall model's ability and robustness in modeling protein functional associations.

[0140] Furthermore, in some embodiments, step S32, "calculating the attention weights between nodes for each type of edge," may specifically include:

[0141] S321. Perform a linear transformation on the features of the current node and the features of the neighboring nodes to obtain the transformed feature vector;

[0142] Specifically, in step S321, the original features of the current node and its neighboring nodes are mapped to the same high-dimensional feature space, enhancing the expressive power of the features and providing a unified vector basis for subsequent similarity calculations. The transformation process is implemented using a trainable weight matrix, whose parameters are dynamically optimized during model training to capture the unique feature patterns of different edge types (such as original interaction edges and homologous edges). For each edge type, independent linear transformation matrices are assigned to the features of the current node and its neighboring nodes (e.g., matrix W1 for original interaction edges and matrix W2 for homologous edges). Matrix multiplication is used to perform feature dimensionality increase or decrease, preserving core biological information (such as domain associations and evolutionary conservation) while eliminating the interference of dimensionality differences on similarity calculations.

[0143] S322. The initial attention score is calculated based on the transformed node features using a learnable attention function;

[0144] Specifically, for step S322, unlike the fixed-weight approach, the learnable attention function dynamically learns the correlation patterns between features through neural network layers, more accurately capturing the nonlinear correlations of node features in different edge types (such as the combined effects of domain similarity and evolutionary conservation). The learnable attention function typically consists of a single-layer neural network (including weight matrix and bias) plus LeakyReLU activation, outputting the original attention score. The transformed current node features are concatenated with the features of neighboring nodes and input into the learnable attention function. A nonlinear transformation outputs a scalar value as the initial attention score; a higher score indicates a stronger functional correlation between the two features under that edge type.

[0145] S323. Normalize the initial attention scores of all neighboring nodes of the same node to obtain the attention weights;

[0146] Specifically, in step S323, the initial scores of different neighboring nodes are mapped to the 0-1 interval, and the sum of all scores is 1, forming attention weights that can be directly used for weighted aggregation. The Softmax function is used, which amplifies high scores and suppresses low scores exponentially, strengthening the contribution of important neighboring nodes while ensuring the probabilistic interpretability of the weight allocation. For the initial scores of all neighboring nodes of the current node, the exponential function value is calculated for each, and then divided by the sum of all exponential values ​​to obtain the normalized weight of each neighboring node.

[0147] This embodiment uses linear transformation to customize the feature space for each edge type, ensuring the targeting of similarity calculation; it uses a learnable attention function to dynamically capture nonlinear relationships between features, improving scoring accuracy; and it uses Softmax normalization to enhance the contribution of key neighborhood nodes. The resulting attention weights can adaptively distinguish the biological value of different types of edges, such as original interactions and homology relationships, making the graph attention network more focused on functional associations when aggregating neighborhood features. This improves the discriminative power and biological interpretability of protein graph features, laying a high-quality topological feature foundation for subsequent multimodal fusion and functional prediction.

[0148] In specific implementations, existing methods typically utilize graph neural networks (GNNs) to extract topological features from protein-protein interaction networks. PPI scores are scaled and used as edge weights, and for homology networks, a fixed and uniform weight is usually assigned to all edges. However, PPI data comes from diverse sources and is often noisy, making it difficult for edge weights to reliably reflect the true contribution of interacting proteins to the functional prediction of the target protein. Furthermore, due to factors such as protein variation, homology relationships themselves are uncertain, and homologous proteins may not have a beneficial impact on functional prediction. Therefore, GNNs based on fixed-weight aggregation struggle to distinguish the actual importance of neighboring nodes in a protein network, potentially obscuring crucial information or even introducing misleading features.

[0149] In this embodiment, Graph Attention Networks (GATs) introduce a learnable attention mechanism that enables the model to dynamically adjust the aggregation weights of neighbors based on the functional relevance between nodes. This mechanism not only improves the model's ability to capture information in heterogeneous graphs but also enhances its sensitivity to differences in local network structure, thereby enabling it to more effectively capture potential non-uniform dependencies.

[0150] Based on the above considerations, this embodiment uses GATs to extract features from the constructed heterogeneous protein network, so as to fully explore the complex topological features and functional relationships between proteins and provide a more accurate and discriminative representation for downstream functional prediction tasks.

[0151] Assume a heterogeneous network contains N protein nodes, each node p i Initialize as a feature vector Using a multi-layer GATs encoder, the representation of node i in the (l+1)th layer is calculated as follows:

[0152]

[0153] in, Let i represent the set of neighbors of node i. It is the trainable linear transformation matrix of the l-th layer, and d′ represents the dimension after transformation. σ(·) represents the normalized attention coefficients from node j to i, and σ(·) represents the nonlinear activation function (such as ReLU).

[0154] Attention coefficient The calculation is divided into two steps. First, the unnormalized attention score... The calculation is as follows:

[0155]

[0156] in, It is a learnable attention vector, where || denotes vector concatenation. Next, the normalized attention coefficients are expressed as:

[0157]

[0158] Since the constructed protein network contains four different edge types, including original PPI, original homology edge, supplementary PPI edge, and supplementary homology edge, attention is calculated for each edge type, and an adjustable edge type-specific weight γ is introduced for each. r ∈(0,1], to mitigate the uncertainty that reinforcing edges may introduce. The node representation is updated as follows:

[0159]

[0160] in, This represents the set of neighbors of node i under edge type r. It is the attention coefficient of edge type r. It is a transformation matrix specific to the edge type. During training, the weights of the original edges (i.e., the original PPI edges and the original homology edges) are initialized to γ. r =1.0, while the weights of the reinforcing edges are initialized to smaller values ​​(e.g., γ). r =0.4), to set a heuristic construction method and reduce the impact of potential noise. This design allows the model to carefully introduce supplementary information while preserving the reliability of the original interaction.

[0161] After multi-layer GATs computation, the final node representation is obtained. It integrates protein domain information, local protein-protein interaction topology information, and global evolutionary relationships. These representations will be used as input to the downstream protein function prediction module.

[0162] Furthermore, in some embodiments, step S4, "using a gated fusion mechanism to perform multimodal fusion of protein sequence features and protein map features to obtain fused features," may specifically include:

[0163] S41. Perform dimension alignment on protein sequence features and protein graph features;

[0164] Specifically, in step S41, protein sequence features (such as the 1280-dimensional vector output by the ESM2 model) and protein graph features (such as the 256-dimensional vector output by the attention network) have dimensional differences due to their different sources, and direct fusion would lead to an imbalance in feature contributions. A linear transformation is used to unify the dimensions, ensuring that the two types of features have an equal basis for weight adjustment in subsequent fusion. Independent trainable linear transformation matrices (including bias terms) are applied to the sequence features and graph features respectively, mapping them to a preset target dimension (such as 512 dimensions). The transformation process preserves the core biological information of the original features (such as evolutionary signals of the sequence and topological relationships of the graph) while eliminating the interference of dimensional differences on fusion.

[0165] S42. Concatenate the dimension-aligned protein sequence features and protein map features to obtain a joint feature vector;

[0166] Specifically, for step S42, the joint features must fully preserve the original information of both sequence and graph features, providing a global perspective for the calculation of the gating factor. The gating factor needs to dynamically adjust its weights based on the overall correlation between the two features, rather than the local information of a single feature. The concatenation method uses column-wise concatenation, connecting the dimension-aligned sequence features and graph features end-to-end to form a joint feature vector with dimensions of "sequence feature dimension + graph feature dimension" (e.g., 512 + 512 = 1024 dimensions). This vector contains both the protein's own sequence characteristics and its correlation patterns within the network.

[0167] S43. A multilayer perceptron is used to map the joint feature vector, and a gate factor is obtained through a nonlinear activation function. The gate factor is used to adaptively adjust the weights of the two modal features.

[0168] Specifically, for step S43, the contributions of sequence features and graph features are dynamically balanced. When graph features contain more critical functional associations (such as functional labels of known interacting proteins), the gating factor increases; when sequence features are more distinctive (such as unique domain sequences), the gating factor decreases, achieving "weighted weighting on demand." The multilayer perceptron (MLP) design includes one hidden layer (e.g., 256 neurons, ReLU activated) and one output layer (one neuron, Sigmoid activated), ensuring that the output gating factor ranges from 0 to 1. The MLP automatically optimizes the weight allocation logic by learning the complementarity of the two types of features in the training data.

[0169] S44. Based on the gating factor, protein sequence features and protein map features are weighted and fused to obtain fused features;

[0170] Specifically, in step S44, protein sequence features and protein graph features are weighted and fused according to a preset dynamic weighting formula. Weights are dynamically allocated through a gating factor, allowing the fused features to emphasize more discriminative modalities. Compared to fixed-weight fusion (such as simple averaging), gating fusion can adaptively address the feature differences of different proteins. For example, for newly discovered proteins (with sparse graph features), the gating factor automatically decreases, relying more on sequence features; for proteins with rich interaction networks, the gating factor increases, emphasizing graph features.

[0171] S45. Perform a linear transformation on the fusion features and obtain the predicted probabilities of each functional label through an activation function, and output the protein function annotation results;

[0172] Specifically, for step S45, the fused features are mapped to a functional label space (such as the set of functional terms in GeneOntology), and the output is converted into a probability distribution through an activation function to achieve quantitative prediction of protein function. The specific process includes: mapping the fused features (e.g., 512 dimensions) to the dimension of the number of functional labels (e.g., 1000 GO terms corresponding to 1000 dimensions) through a weight matrix; using Sigmoid activation (in multi-label classification scenarios) to output a probability value of 0-1 for each dimension, representing the probability that the protein has that function; and selecting functional labels with probability values ​​higher than a preset threshold (e.g., 0.5) as the final protein function annotation result.

[0173] This embodiment ensures feature comparability through dimension alignment, provides global correlation information through joint features, achieves adaptive weight allocation through the gating factor learned by MLP, and finally outputs accurate functional annotations through probabilistic prediction. This fully leverages the complementary value of sequence features and graph features, improves the accuracy, robustness, and biological interpretability of protein function prediction, and shows outstanding advantages, especially when dealing with proteins with sparse features or complex functions.

[0174] Furthermore, in some embodiments, step S5, "performing multi-label classification prediction based on fusion features to obtain protein function annotation results," may specifically include:

[0175] S51. Input the fused features into a multi-label classifier, which includes at least one fully connected layer;

[0176] Specifically, for step S51, protein function prediction is a typical multi-label classification task. A protein can have multiple functions, such as "DNA binding" and "regulation of apoptosis." The classifier needs to map the fused features (low-dimensional vectors integrating sequence and graph information) to the functional label space to achieve association modeling from features to functions. The classifier contains at least one fully connected layer (two layers in the example). Its role is to extract high-order abstract information from the fused features through linear transformation and nonlinear activation, gradually narrowing the gap between the feature dimension and the number of functional labels. The first fully connected layer is used for feature dimension compression and association mining (e.g., mapping 512-dimensional fused features to 256 dimensions), while the second layer directly maps to the functional label dimension (e.g., 1000 GO terms correspond to 1000-dimensional output).

[0177] S52. Calculate the predicted probability of each functional label using an activation function;

[0178] Specifically, in step S52, in multi-label classification, the prediction of each functional label is an independent event (a protein may or may not possess a certain function). The original output of the fully connected layer needs to be converted into a probability value in the 0-1 range. The Sigmoid function can be used to independently normalize the score of each label, avoiding interference between labels. The output probability value represents the confidence that the protein possesses the corresponding function; the closer the value is to 1, the higher the reliability of inferring the function based on the fusion features. For example, a probability of 0.92 indicates a high probability of possessing the function.

[0179] S53. Use the function labels with a predicted probability greater than a preset probability threshold as the predicted functions of the protein, and generate protein function annotation results containing the function labels;

[0180] Specifically, for step S53, the preset probability threshold needs to balance the accuracy and completeness of the prediction—a threshold that is too high will reduce false positives but may miss true functions; a threshold that is too low will cover more functions but may introduce noise. This embodiment uses 0.5 as the default threshold, which can be fine-tuned during training based on the performance of the validation set (such as the F1 score). All function labels with probabilities greater than the preset probability threshold are filtered, sorted from high to low probability, and finally output with annotation results including label name, GO number, and corresponding probability, which facilitates biological interpretation and experimental verification.

[0181] This embodiment effectively mines the deep association between fusion features and functional labels through fully connected layers; Sigmoid activation ensures that the probability of each functional label is independently interpretable; threshold screening balances the accuracy and completeness of prediction, so that the final functional annotation results cover both the core functions of proteins and retain secondary but potential functions, providing accurate, comprehensive and verifiable functional prediction basis for biological research.

[0182] In specific embodiments, after obtaining protein sequence features and graph features, existing methods typically perform functional prediction by simply splicing or linearly combining the two modalities. Such shallow fusion strategies lack the ability to capture their deep correlations, are difficult to effectively integrate heterogeneous information, and cannot fully capture their potential interdependencies, thereby limiting the model's ability to comprehensively and multidimensionally characterize proteins, ultimately affecting prediction performance.

[0183] To address this issue, this embodiment designs a collaborative fusion and prediction module that effectively integrates information from two modalities to achieve multi-label protein function prediction.

[0184] set up Represents a graph-based representation of protein node i, where d g The dimension of the graph feature vectors; Represents a sequence-based representation, where d s Let be the dimension of the sequence feature vector. First, concatenate the two to form a joint representation:

[0185]

[0186] Then, the gating factor α is calculated using a multilayer perceptron (MLP). i ∈(0,1):

[0187] α i =σ(W g z i +b g ),

[0188] in, The weight matrix is ​​a learnable matrix. This is a bias term.

[0189] The final fusion is represented as:

[0190]

[0191] This mechanism enables the fusion operation to be context-adaptive, i.e., the gating factor α i It is adaptively determined by the joint features of each sample, rather than being fixed, thus enabling more flexible and accurate integration of multimodal information.

[0192] Next, a multi-label function prediction module is designed. In this task, each protein may correspond to multiple GO terms. This is based on fusion representation. By performing a linear transformation and then applying the sigmoid activation function, the probability distribution over all labels is obtained:

[0193]

[0194] Where C represents the number of GO categories. The classifier weight matrix is... σ is the bias term, and σ(·) is the sigmoid activation function used to map the output to the (0,1) interval.

[0195] The model is trained using the Binary Cross-Entropy (BCE) loss function, which is suitable for multi-label classification tasks. The loss for the i-th sample on the j-th label is defined as follows:

[0196]

[0197] Among them, y i,j ∈{0,1} indicates whether sample i is annotated as the j-th GO term. To predict probabilities.

[0198] The final objective function is the average loss over all samples and all labels:

[0199]

[0200] Furthermore, the multimodal collaborative learning framework proposed in this embodiment is implemented using PyTorch. All hyperparameters were tuned based on the performance of the validation set, aiming to achieve a balance between accuracy and generalization. To control model complexity, the GATs module was configured as a two-layer module. The optimizer was heterogeneous AdamW with an initial learning rate of 0.001. During training, an early stopping strategy was adopted, terminating training prematurely when the Fmax score on the validation set no longer improved after 5 consecutive epochs. The maximum number of training epochs was set to 20, and the batch size was set to 40.

[0201] To further enhance the robustness and generalization ability of the model, a model ensemble strategy is adopted. Specifically, the top three models that perform best on the validation set are retained, denoted as f1(·), f2(·), and f3(·). Their prediction results are averaged and used as the final output.

[0202]

[0203] This ensemble approach helps reduce the uncertainty of individual model predictions and improves the overall stability and accuracy of the final results.

[0204] like Figure 3 As shown, Figure 3 The schematic diagram of the overall architecture of the protein function prediction method based on multimodal protein data provided in this embodiment includes four main components: (1) sequence encoding, (2) heterogeneous graph construction, (3) graph-based feature encoding, and (4) feature fusion and prediction. The specific implementation process is as follows: First, the amino acid sequence of the protein is input. On one hand, the sequence is fed into the ESM2 model to obtain the original sequence features, and then processed by a multilayer perceptron (MLP) to extract protein sequence features that reflect sequence evolution, structure, and other information. On the other hand, the sequence is analyzed by InterProScan to identify structural domains, and combined with the homology network constructed by EggNOG and the protein-protein interaction (PPI) network constructed by STRING to form a heterogeneous graph containing multiple types of relationships such as homology association and interaction association. Next, the heterogeneous graph is input into a graph attention network (GAT), and the node features of different types of edges are aggregated and updated through a multilayer (1stlayer, 2ndlayer) attention mechanism to extract protein network features that can capture the topological relationships of the protein network. Finally, the protein sequence features and network features are fused through a gated fusion mechanism (GMU), and the fused features are then input into a multilayer perceptron (MLP) for multi-label classification. Finally, multiple GeneOntology (GO) functional labels corresponding to the protein are output to generate protein functional annotation results.

[0205] This embodiment proposes a novel protein function prediction framework based on customized heterogeneous graph modeling and an attention mechanism. Specifically, homology and interaction information are integrated into the heterogeneous graph. To address the issues of incomplete PPI and inaccurate homology information caused by protein variation, supplementary relationships are constructed in the heterogeneous graph based on protein similarity to provide more comprehensive protein background information. Building upon this, Graph Attention Networks (GATs) with an attention mechanism are applied to automatically extract important features, enabling the model to capture local and global correlations. Simultaneously, the synergistic effect of the two data sources is utilized for information verification, reducing noise from potential errors. Furthermore, a Gated Multimodal Unit (GMU) is designed to provide more informative features for protein function prediction, thereby more effectively fusing sequence-based and graph-based features.

[0206] To verify the protein function prediction method based on multimodal protein data provided in this embodiment, the following verification is carried out in conjunction with specific experiments.

[0207] 1. Data preparation.

[0208] The experimental data used included protein sequences, PPI data, orthologous relationships, GO functional annotations, and GeneOntology files, sourced from the following sources:

[0209] GO annotation data comes from the GeneOntologyAnnotation (GOA) database. Annotation files (version dated January 17, 2024) were downloaded from the official GO website. To ensure the reliability of the annotations, all entries with the evidence code IEA (Inferred from Electronic Annotation) were removed, retaining only manually compiled or experimentally supported GO annotations.

[0210] The protein sequences were obtained from the UniProt database.

[0211] The GO file can be downloaded from the GeneOntology official website.

[0212] The direct kinship data comes from the EggNOGv5.0 database.

[0213] PPI data comes from the STRING database, version 12.

[0214] Three representative model organisms widely used in biological research were selected: HUMAN (human, Homosapiens), MOUSE (mouse, Musmusculus), and ARATH (Arabidopsis thaliana). Among them, HUMAN represents higher mammals and has the richest functional annotations; MOUSE is a commonly used model organism with abundant experimental data; and ARATH is a classic model species in plant biology research.

[0215] To better reflect real-world application scenarios, a time-based annotation partitioning strategy was adopted instead of random partitioning. This setting follows the CAFA evaluation framework. Specifically, proteins annotated before 2022 were used for training; proteins first annotated between January 1, 2022 and December 31, 2022 were used for the validation set; and proteins first annotated between January 1, 2023 and December 31, 2023 were used for the test set. This time-based partitioning more accurately simulates the task of predicting the function of previously unannotated proteins, thus enabling a more realistic evaluation of the model's performance in real-world applications. Table 1 shows detailed statistics for each GO ontology branch (BP, MF, and CC) under different dataset partitions.

[0216] Table 1 Detailed Statistical Information

[0217] Dataset BP CC MF ALL Training set 29777 30323 27757 37226 Validation set 628 696 703 912 Test set 568 565 490 732 Number of GO terms 15624 2008 5261 22916

[0218] 2. Evaluation indicators.

[0219] Two widely used evaluation metrics in this field were employed: maximum F1 score (Fmax) and area under the precision-recall curve (AUPR).

[0220] Fmax is a commonly used evaluation metric in multi-label classification tasks, particularly suitable for measuring the consistency between predicted labels and multiple ground truth annotations. It is defined as the maximum F1 score calculated over a series of probability thresholds. The F1 score is the harmonic mean of precision and recall.

[0221]

[0222] Among them, accuracy Recall rate TP, FP, and FN represent the number of true positives, false positives, and false negatives, respectively.

[0223] The definition of Fmax is:

[0224]

[0225] AUPR was also used as a supplementary metric to evaluate the model's overall predictive performance across all labels. AUPR is calculated by plotting precision-recall curves at different thresholds and then calculating the area under the curve. It reflects the model's robustness and consistency on positive samples (i.e., samples with functional annotations).

[0226] 3. Experimental results.

[0227] To evaluate the predictive performance of the proposed method, it was compared with nine state-of-the-art methods: Diamond, BlastKNN, DeepGO, DeepGOA, DeepGOCNN, DeepGOZero, DeepGraphGO, DeepFMB, and DeepFMB+.

[0228] Table 2. Experimental Comparison Results

[0229]

[0230] Table 2 presents the specific comparison results, where the best-performing model (MFPFP model) outperforms all other methods across all three branches. These results demonstrate that MFPFP is a superior method that can significantly improve prediction performance.

[0231] In summary, the protein function prediction method based on multimodal protein data provided in this embodiment firstly, by acquiring multi-source protein data and extracting sequence features, breaks the dependence of traditional methods on a single data source, providing basic functional information at the sequence level for the prediction task. Next, a heterogeneous graph containing domain-derived node features and various types of edges is constructed based on the multi-source data. The domain node features naturally associate with protein functional modules, while the various edge designs, including original edges and edge supplementation based on sequence similarity, effectively compensate for deficiencies such as incomplete PPI data and errors in homology information, enriching the context of protein associations. Then, a graph attention mechanism is used to encode features in the heterogeneous graph, adaptively identifying and aggregating effective information in nodes and their neighborhoods, filtering data noise interference, and generating more accurate protein graph features. Subsequently, a gating fusion mechanism dynamically adjusts the fusion weights of sequence features and graph features, rather than simply superimposing them, fully exploring the complementary value of the two modalities and forming a more comprehensive protein function representation. Finally, multi-label classification prediction based on high-quality fused features can more accurately match multiple functional annotations corresponding to proteins, improving the accuracy and robustness of protein function prediction results.

[0232] To facilitate better implementation of the protein function prediction method based on multimodal protein data in this application, this application also provides a protein function prediction device based on multimodal protein data, which is based on the aforementioned protein function prediction method based on multimodal protein data. The meanings of the terms used are the same as in the aforementioned protein function prediction method based on multimodal protein data, and specific implementation details can be found in the descriptions in the method embodiments.

[0233] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of a protein function prediction device based on multimodal protein data provided in an embodiment of this application. Specifically, this device may include a feature extraction module 201, a heterogeneous graph module 202, a feature encoding module 203, a feature fusion module 204, and a classification prediction module 205, as follows:

[0234] The feature extraction module 201 is used to acquire protein multi-source data and extract features from the protein sequences in the protein multi-source data to obtain protein sequence features.

[0235] Heterogeneous graph module 202 is used to construct heterogeneous graphs based on multi-source protein data. The heterogeneous graph includes node features and multiple types of edges. The node features are generated based on protein domain information. The multiple types of edges include at least one of original interaction edges, original homology edges, and edges supplemented based on sequence similarity.

[0236] Feature encoding module 203 is used to encode the features of heterogeneous graphs using a graph attention mechanism to obtain protein graph features;

[0237] The feature fusion module 204 is used to perform multimodal fusion of protein sequence features and protein map features using a gated fusion mechanism to obtain fused features;

[0238] The classification prediction module 205 is used to perform multi-label classification prediction based on fusion features to obtain protein function annotation results.

[0239] Specific limitations regarding the protein function prediction device based on multimodal protein data can be found in the limitations of the protein function prediction method based on multimodal protein data described above, and will not be repeated here. Each module in the aforementioned protein function prediction device based on multimodal protein data can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0240] The protein function prediction device based on multimodal protein data provided in this embodiment includes a feature extraction module 201 that acquires multi-source protein data and extracts features from the protein sequences in the multi-source protein data to obtain protein sequence features; a heterogeneous graph module 202 that constructs a heterogeneous graph based on the multi-source protein data, the heterogeneous graph including node features and various types of edges, the node features being generated based on protein domain information, and the various types of edges including at least one of original interaction edges, original homology edges, and edges supplemented based on sequence similarity; a feature encoding module 203 that uses a graph attention mechanism to encode the features of the heterogeneous graph to obtain protein graph features; a feature fusion module 204 that uses a gating fusion mechanism to perform multimodal fusion of protein sequence features and protein graph features to obtain fused features; and a classification prediction module 205 that performs multi-label classification prediction based on the fused features to obtain protein function annotation results. This embodiment achieves efficient synergistic utilization of multimodal protein data by constructing an enhanced heterogeneous graph and utilizing attention mechanisms and gating fusion, thereby effectively improving the accuracy and reliability of protein function prediction and solving the problems of existing technologies that are difficult to effectively integrate multi-source protein data and are sensitive to data noise.

[0241] Furthermore, embodiments of this application also provide an electronic device, such as... Figure 5 As shown, it illustrates a structural schematic diagram of the electronic device involved in the embodiments of this application, specifically:

[0242] The electronic device may include components such as a processor 301 with one or more processing cores, a memory 302 with one or more computer-readable storage media, a power supply 303, and an input unit 304. Those skilled in the art will understand that... Figure 5 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0243] The processor 301 is the control center of the electronic device. It connects various parts of the electronic device via various interfaces and lines, and performs various functions and processes data by running or executing software programs and / or modules stored in the memory 302, and by calling data stored in the memory 302, thereby providing overall monitoring of the electronic device. Optionally, the processor 301 may include one or more processing cores; preferably, the processor 301 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 301.

[0244] The memory 302 can be used to store software programs and modules. The processor 301 executes various functional applications and protein function prediction methods based on multimodal protein data by running the software programs and modules stored in the memory 302. The memory 302 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 302 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 302 may also include a memory controller to provide the processor 301 with access to the memory 302.

[0245] The electronic device also includes a power supply 303 that supplies power to various components. Preferably, the power supply 303 can be logically connected to the processor 301 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 303 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0246] The electronic device may also include an input unit 304, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0247] Although not shown, the electronic device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 301 in the electronic device loads the executable files corresponding to the processes of one or more applications into the memory 302 according to the following instructions, and the processor 301 runs the applications stored in the memory 302 to realize various functions, as follows:

[0248] This process involves acquiring multi-source protein data and extracting features from the protein sequences to obtain protein sequence features. A heterogeneous graph is constructed based on the multi-source protein data. This graph includes node features and various types of edges. Node features are generated based on protein domain information, and the various edge types include at least one of original interaction edges, original homology edges, and edges supplemented based on sequence similarity. A graph attention mechanism is used to encode the features of the heterogeneous graph to obtain protein graph features. A gating fusion mechanism is then used to perform multimodal fusion of the protein sequence features and protein graph features to obtain fused features. Finally, multi-label classification prediction is performed based on the fused features to obtain protein function annotation results.

[0249] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0250] By constructing an enhanced heterogeneous graph and utilizing attention mechanisms and gating fusion, efficient collaborative utilization of multimodal protein data is achieved, thereby effectively improving the accuracy and reliability of protein function prediction and solving the problem that existing technologies are difficult to effectively integrate multi-source protein data and are sensitive to data noise.

[0251] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0252] Therefore, embodiments of this application provide a storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the protein function prediction methods based on multimodal protein data provided in embodiments of this application. For example, the instructions can execute the following steps:

[0253] This process involves acquiring multi-source protein data and extracting features from the protein sequences to obtain protein sequence features. A heterogeneous graph is constructed based on the multi-source protein data. This graph includes node features and various types of edges. Node features are generated based on protein domain information, and the various edge types include at least one of original interaction edges, original homology edges, and edges supplemented based on sequence similarity. A graph attention mechanism is used to encode the features of the heterogeneous graph to obtain protein graph features. A gating fusion mechanism is then used to perform multimodal fusion of the protein sequence features and protein graph features to obtain fused features. Finally, multi-label classification prediction is performed based on the fused features to obtain protein function annotation results.

[0254] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0255] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0256] Since the instructions stored in the storage medium can execute the steps in any of the protein function prediction methods based on multimodal protein data provided in the embodiments of this application, the beneficial effects that any of the protein function prediction methods based on multimodal protein data provided in the embodiments of this application can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0257] The foregoing has provided a detailed description of a protein function prediction method and apparatus based on multimodal protein data provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting protein function based on multimodal protein data, characterized in that, include: Acquire multi-source protein data and extract features from the protein sequences in the multi-source protein data to obtain protein sequence features; Heterogeneous graphs are constructed based on the protein multi-source data. The heterogeneous graphs include node features and multiple types of edges. The node features are generated based on protein domain information. The multiple types of edges include at least one of original interaction edges, original homology edges, and edges supplemented based on sequence similarity. The heterogeneous graph is feature-encoded using a graph attention mechanism to obtain protein graph features; A gated fusion mechanism is used to perform multimodal fusion of the protein sequence features and the protein map features to obtain fused features; Based on the fusion features, multi-label classification prediction is performed to obtain protein function annotation results.

2. The protein function prediction method based on multimodal protein data according to claim 1, characterized in that, The process of acquiring multi-source protein data and extracting features from the protein sequences in the multi-source protein data to obtain protein sequence features includes: Protein multi-source data are collected from public databases, including protein sequence data, protein interaction data, homology data, and functional annotation data. The collected multi-source protein data were cleaned. A pre-trained protein language model was used to extract features from the cleaned protein sequence data; The output vector of the last layer of the pre-trained protein language model is extracted as the protein sequence feature, which includes at least one of the protein's evolutionary signal, structural cues, and functional information.

3. The protein function prediction method based on multimodal protein data according to claim 2, characterized in that, The process of using a pre-trained protein language model to extract features from the cleaned protein sequence data includes: The protein sequence is preprocessed by adding special markers at the beginning and end of the sequence and converting the amino acids into corresponding numerical codes; The preprocessed protein sequence is input into a pre-trained protein language model, and long-distance dependencies in the protein sequence are captured through a multi-head self-attention mechanism. The pre-trained protein language model outputs a latent vector representation of the protein sequence, which includes information on the protein's evolutionary conservation and structural features.

4. The protein function prediction method based on multimodal protein data according to claim 1, characterized in that, The construction of the heterogeneity map based on the protein multi-source data includes: Construct a set of nodes for a heterogeneous graph, using a single protein as a node in the heterogeneous graph; The domain composition of each protein is obtained by using a domain prediction tool, the domain composition is converted into a binary feature vector, and the binary feature vector is dimensionality reduced to obtain node features. Multiple types of edges are constructed from the aforementioned protein multi-source data to create the heterogeneous graph; Initialize the weights of edges of various types, assigning initial weight values ​​to edges of different types.

5. The protein function prediction method based on multimodal protein data according to claim 4, characterized in that, The various types of edges constructed based on the protein multi-source data for the heterogeneous graph include: Interactions with confidence scores greater than a preset confidence threshold are selected from a protein-protein interaction database to establish primal interaction edges between proteins; Based on homology grouping information, proteins belonging to the same homology family are connected by original homology edges; Calculate the sequence similarity between the protein sequences. When the similarity is greater than a preset similarity threshold, add supplementary edges between the corresponding protein nodes. The supplementary edges include supplementary interaction edges and supplementary homology edges.

6. The protein function prediction method based on multimodal protein data according to claim 1, characterized in that, The step of using a graph attention mechanism to encode features of the heterogeneous graph to obtain protein graph features includes: The node features of the heterogeneous graph are input into the input layer of the graph attention network for feature initialization. In the hidden layer of the graph attention network, attention weights between nodes are calculated for each type of edge. These attention weights are obtained by using the node features through a learnable function and are used to characterize the strength of the association between nodes. The features of neighboring nodes are weighted and aggregated according to the attention weights, and the feature representation of the current node is updated. After iterative calculations by a multi-layer graph attention network, the final protein graph features are output, which are used to fuse node self-information and neighborhood association information.

7. The protein function prediction method based on multimodal protein data according to claim 6, characterized in that, The calculation of attention weights between nodes for each type of edge includes: Perform a linear transformation on the current node features and the features of neighboring nodes to obtain the transformed node features; The initial attention score is calculated based on the transformed node features using the learnable attention function; The initial attention scores of all neighboring nodes of the same node are normalized to obtain the attention weights.

8. The protein function prediction method based on multimodal protein data according to claim 1, characterized in that, The method employs a gated fusion mechanism to perform multimodal fusion of the protein sequence features and the protein map features to obtain fused features, including: Perform dimensional alignment on the protein sequence features and the protein graph features; The dimension-aligned protein sequence features and protein map features are concatenated to obtain a joint feature vector. The joint feature vector is mapped using a multilayer perceptron, and a gating factor is obtained through a nonlinear activation function. The gating factor is used to adaptively adjust the weights of the two modal features. Based on the gating factor, the protein sequence features and protein map features are weighted and fused to obtain fused features; The fusion features are linearly transformed, and the predicted probabilities of each functional label are obtained through an activation function, outputting the protein function annotation results.

9. The protein function prediction method based on multimodal protein data according to claim 1, characterized in that, The multi-label classification prediction based on the fusion features yields protein function annotation results, including: The fused features are input into a multi-label classifier, which includes at least one fully connected layer. The predicted probability of each feature label is calculated using an activation function; Functional tags with predicted probabilities greater than a preset probability threshold are used as predicted functions of proteins, generating protein function annotation results that include functional tags.

10. A protein function prediction device based on multimodal protein data, characterized in that, include: The feature extraction module is used to acquire protein multi-source data and extract features from the protein sequences in the protein multi-source data to obtain protein sequence features. A heterogeneous graph module is used to construct a heterogeneous graph based on the protein multi-source data. The heterogeneous graph includes node features and multiple types of edges. The node features are generated based on protein domain information. The multiple types of edges include at least one of original interaction edges, original homology edges, and edges supplemented based on sequence similarity. The feature encoding module is used to encode the heterogeneous graph using a graph attention mechanism to obtain protein graph features; The feature fusion module is used to perform multimodal fusion of the protein sequence features and the protein map features using a gated fusion mechanism to obtain fused features; The classification prediction module is used to perform multi-label classification prediction based on the fused features to obtain protein function annotation results.

Images