Multi-modal protein function prediction method and prediction system

Abstract

The invention provides a multi-modal protein function prediction method and system, and the method comprises the following steps: carrying out the processing of a one-dimensional convolution layer and a pooling layer of protein sequence data, extracting the local dependence features in a protein sequence, and capturing the mode and key sequence features in the protein; using protein sequence data as input, and using a large-scale pre-trained protein language model to extract residue-level embedded information in a protein sequence as node features to be applied to the contact graph; the method comprises the following steps: taking protein sequence data as input, predicting the probability of contact between protein residues based on a Cb-Cb distance, and constructing a contact graph; and performing joint processing on features of different modes by using a multi-head attention mechanism and a residual structure, integrating data from different modes through feature fusion, and fusing the extracted key sequence features, the residue level embedded information and the contact graph. The prediction method is accurate in prediction and good in performance.

Description

Technical Field

[0001] The present invention relates to the field of protein function prediction, and in particular to a multimodal protein function prediction method and prediction system. Background Art

[0002] With the development of high-throughput technology, the number of protein sequences in public databases has increased exponentially, but only a small number of protein sequences have been experimentally annotated. As the gap between the number of newly discovered proteins and the number of proteins with known functions continues to widen, reliable computational prediction methods are becoming increasingly important. Accurately predicting protein function not only helps to understand its role in disease pathogenesis and the function of the metagenome, but also helps to discover potential drug targets.

[0003] Despite differences in protein sequences, proteins with similar structures may have similar functions. Therefore, purely sequence-based unimodal methods perform poorly when dealing with completely novel protein sequences without any annotations, and are limited in the length of protein sequences that can be processed. Although existing multimodal methods have improved the performance of protein function prediction to a certain extent, they usually rely on other modal data that are difficult to obtain, such as protein-protein interaction networks or protein structure data. Compared with protein sequence databases with tens of millions of data, these other modal data are very scarce. For example, protein-protein interaction network data usually has only tens of thousands or even fewer. This data scarcity severely limits the breadth and scope of application of existing multimodal methods.

[0004] In view of this, the present invention is proposed. Summary of the invention

[0005] In view of this, the present invention proposes a multimodal protein function prediction method and prediction system, which only relies on protein sequence data to generate multimodal features, and combines the protein function prediction method of graph attention network and large-scale pre-trained model, effectively integrating structural information, sequence features and residue level embedding, reducing the dependence on multimodal data that is difficult to obtain or does not exist, while capturing the performance improvement provided by multimodal data.

[0006] Specifically, the present invention is achieved through the following technical solutions:

[0007] The present invention provides a multimodal protein function prediction method, comprising the following steps:

[0008] The protein sequence data is processed through one-dimensional convolutional layers and pooling layers to extract local dependency features in the protein sequence and capture patterns and key sequence features in the protein;

[0009] Taking protein sequence data as input, a large-scale pre-trained protein language model is used to extract residue-level embedding information in the protein sequence as node features and apply it to the contact graph.

[0010] Using protein sequence data as input, the contact probability between protein residues is predicted based on Cb-Cb distance to construct a contact map;

[0011] Using a multi-head attention mechanism and a residual structure to jointly process features of different modalities, integrating data from different modalities through feature fusion, and fusing the extracted key sequence features, the residue-level embedding information, and the contact map;

[0012] In the fused feature representation, a graph attention network is used for processing. The output of the graph attention network generates a new feature representation for each node, and the information of neighboring nodes is fused by weighted summation. Then, a global feature vector of fixed length is obtained through a global pooling layer. Finally, the probability of each Gene Ontology term is calculated through a fully connected layer and a SoftMax activation function to predict protein function.

[0013] In addition to providing a multimodal protein function prediction method, the present invention also provides a multimodal protein function prediction system, comprising:

[0014] Feature extraction module: used to process protein sequence data through one-dimensional convolutional layers and pooling layers to extract local dependency features in protein sequences and capture patterns and key sequence features in proteins;

[0015] Embedding information extraction module: It is used to take protein sequence data as input and use a large-scale pre-trained protein language model to extract residue-level embedding information in the protein sequence as node features for application to the contact graph;

[0016] Contact map construction module: It is used to take protein sequence data as input, predict the probability of contact between protein residues based on Cb-Cb distance, and construct a contact map;

[0017] Fusion module: used to jointly process features of different modalities using a multi-head attention mechanism and a residual structure, integrate data from different modalities through feature fusion, and fuse the extracted key sequence features, the residue-level embedding information, and the contact map;

[0018] Feature representation module: used to process the fused feature representation using a graph attention network. The output of the graph attention network generates a new feature representation for each node, fuses the information of neighboring nodes in a weighted summation manner, and then obtains a global feature vector of a fixed length through a global pooling layer. Finally, the probability of each Gene Ontology term is calculated through a fully connected layer and a SoftMax activation function to predict protein function.

[0019] In summary, the solution of the present invention has the following beneficial effects:

[0020] 1) The prediction method of the present invention reduces the reliance on multimodal data that is difficult to obtain or does not exist: Existing multimodal protein function prediction methods often rely on protein-protein interaction network data or protein structural data, which are expensive to obtain and are missing in a large number of proteins. In contrast, the present invention relies only on protein sequence data, which is huge and easy to obtain in public databases. Therefore, this solution greatly broadens the scope of application of the prediction model, making it applicable to more unannotated protein sequences.

[0021] 2) The prediction method of the present invention effectively integrates multimodal features to improve prediction accuracy: by extracting local dependency features of protein sequences through a one-dimensional convolutional neural network, combined with the residue-level embedding generated by the ESM2 large model, it is possible to capture deep evolutionary information in protein sequences. At the same time, the contact map generated by RaptorX provides sequence-derived structural information. These features effectively supplement the limitations of traditional single sequence prediction methods. In the graph attention network, modal fusion further enhances the expressiveness of the model, enabling the model to capture complex interactions between residues and enhance prediction accuracy.

[0022] 3) The prediction method of the present invention takes into account both the extensiveness and performance of the model: the entire method only needs to rely on sequence data to generate multimodal feature data. Compared with the traditional method that only uses sequence information, this technical solution can obtain a more comprehensive feature representation when capturing protein function, thereby improving the prediction performance. At the same time, since it does not rely on other scarce modal data, this method has higher applicability in practical applications.

[0023] 4) The prediction method of the present invention can process proteins with up to 2000 amino acids, covering more than 99% of the sequences in the UniProt database. This enhancement makes our model more flexible and practical in practical applications, capable of capturing more functional domains and structural information, while avoiding information loss caused by sequence segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Various other advantages and benefits will become apparent to those skilled in the art by reading the following detailed description of the preferred embodiments. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. The same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0025] Figure 1 The overall framework diagram of the multimodal protein function prediction method of the present invention;

[0026] Figure 2 A flowchart of a computer device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0027] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are only examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0028] The terms used in the present disclosure are only for the purpose of describing specific embodiments and are not intended to limit the present disclosure. The singular forms "a", "said" and "the" used in the present disclosure and the appended claims are also intended to include plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0029] It should be understood that although the terms first, second, third, etc. may be used in the present disclosure to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present disclosure, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" used herein may be interpreted as "at the time" or "when" or "in response to determining",

[0030] Example

[0031] like Figure 1 As shown, the multimodal protein function prediction method provided by the present invention comprises the following steps:

[0032] S1: The protein sequence data is used as the input of the one-dimensional convolutional neural network. The input protein sequence is first encoded by One-Hot to generate a 21×2000 matrix that meets the encoding requirements of all standard and non-standard amino acids. After processing by the one-dimensional convolutional layer and pooling layer, the local dependency features in the protein sequence are extracted to capture the patterns and key sequence features in the protein.

[0033] S2: Using a large-scale pre-trained protein language model, we take protein sequence data as input to extract residue-level embedding information in protein sequences. These embedding information captures the evolutionary information in protein sequences. We apply the residue-level embedding as node features to the contact graph to enrich the information represented by the nodes.

[0034] S3: Using protein sequence data as input, the contact probability between protein residues is predicted based on the Cb-Cb distance. When the contact probability between residues exceeds a certain threshold, edges are established in the graph to construct a protein contact graph for subsequent graph neural network processing.

[0035] S4: The contact map provides high-resolution information about the three-dimensional structure of the protein, while the residue-level features and sequence features reveal deep semantic information. The multi-head attention mechanism and residual structure are used to jointly process the features of different modalities, and the data from different modalities are integrated through the feature fusion layer. The extracted sequence features, residue-level embeddings and contact maps are fused to generate richer feature representations.

[0036] S5: In the fused feature representation, the graph attention network is used for processing. The graph attention network dynamically assigns different weights to each residue node through the self-attention mechanism to capture the complex relationship between nodes. The output of the graph attention network generates a new feature representation for each node, and the information of neighboring nodes is fused by weighted summation. Finally, a global feature vector of fixed length is obtained through the global pooling layer. Finally, the probability of each Gene Ontology term is calculated through the fully connected layer and the SoftMax activation function to predict protein function.

[0037] Furthermore, in step S1, the One-Hot encoded sequence passes through a one-dimensional convolutional layer and a pooling layer to capture local dependencies in the sequence. The feature representation of the sequence modality is obtained by extracting and regularizing features through a one-dimensional convolutional neural network and a maximum pooling layer.

[0038] The final feature representation of the sequence modality is: F1 is a deep neural network model that processes sequence modalities. Function f conv and f maxp Represent the operations of the convolutional layer and the maximum pooling layer respectively.

[0039] Furthermore, in step S2, a large-scale pre-trained model based on ESM2-15 B is used to obtain the residue-level embedding of each protein. The ESM2 large-scale model captures evolutionary information in amino acid sequences related to protein properties and learns the interaction patterns between residues in amino acid sequences. These residue-level feature representations are then embedded into the corresponding nodes in the contact graph. ESM2 is trained to predict the positions of amino acids in randomly masked protein sequences:

[0040] L MLM =-∑ i∈M log p(x i |x M )

[0041] Here, for a randomly generated mask set M that contains 15% of positions i in sequence x, the model’s task is to M (excluding masked positions) Predict masked amino acid x i identity.

[0042] This masked language modeling objective forces the model to learn dependencies between amino acids. While the training objective itself is simple and unsupervised, solving this problem on millions of evolutionarily diverse protein sequences requires the model to internalize evolutionary patterns in the sequences. We expect that this training will lead to biological structure emerging in the language model because it is closely related to sequence patterns.

[0043] The ESM2-15B large-scale pre-trained model is trained on sequences in the UniRef protein sequence database, the language model of protein sequences is expanded to 15 billion parameters, and atomic resolution images of protein structures appear in the learned representation. This leads to an order of magnitude acceleration of high-resolution structure prediction, thus enabling large-scale structural characterization of metagenomic proteins.

[0044] During training, sequences are drawn with uniform weights from the approximately 43 million UniRef 50 training clusters, which are drawn from the approximately 138 million UniRef 90 sequences, so that the model has seen approximately 65 million unique sequences throughout training. To obtain protein-level embeddings, we average these residue embeddings at the amino acid level and combine them with the other learned representations.

[0045] Furthermore, in step S3, in order to obtain structural information, RaptorX adopts a multi-task learning strategy to simultaneously predict residue distances and orientation matrices in the same model. A residual network is used to predict the discrete distance distribution between key backbone atom pairs. In this way, the model can not only predict whether a pair of residues are in contact, but also predict the actual distance between them.

[0046] These predictions are converted into probability distributions that reflect the possible contact distances for each pair of residues. RaptorX uses a deep residual network to predict the distances and contact relationships between protein residues. The residual network processes input features through a large number of convolutional layers, extracts local and global features of the protein sequence, and then predicts the distance distribution between residue pairs.

[0047] Once the distance and contact information is predicted, an enhanced gradient-based energy minimization method is used to convert these predicted distance and orientation information into a three-dimensional structural model of the protein. PyRosetta is used to pack the side chains and reduce steric clashes to optimize the final protein model.

[0048] This strategy reduces training time and the number of models required. This method can predict long-range contacts, i.e. contacts between residue pairs that are far apart in the protein sequence, which are essential for the correct folding of the protein. This method determines contacts between residues by predicting the Cb-Cb distance between the residues, with the input being the protein sequence. The contact probability is defined as the probability that the distance between the residues is less than When the contact probability between two residues exceeds a certain threshold, they are considered to be in contact and an edge is established between them.

[0049] Furthermore, in step S4, the feature fusion structure includes two parts, namely, a joint attention block and a feature fusion layer. For the joint attention block, it includes a multi-head joint attention module and two fully connected feed-forward layers, and its formula is as follows:

[0050] FFL(x)=max(0,xW1+b1)W2+b2

[0051] First, input the training data (X, Y), where X represents the input features and Y represents the corresponding labels. At the same time, input the hyperparameters, including the learning rate η, the number of attention heads h, and the maximum number of training rounds τ. Sample a mini-batch (x, y) from the dataset and extract features from the two modalities, denoted by ox1 and ox2 respectively.

[0052] Next, the model is calculated through a multi-head attention mechanism. For each co-attention block, we first calculate the attention matrices Q, K, and V, and divide these matrices into h attention heads. For each head, the attention score is calculated and the features are weighted according to the score to generate a weighted feature representation.

[0053] The calculation of attention scores is based on QK T / √d k, and then the attention weights are obtained through the softmax function. Subsequently, these weighted features are processed through residual connections and normalization layers to generate the output features of each layer. The output features of all layers will be concatenated to form the fused feature representation O.

[0054] Finally, the fused feature O is input into a fully connected layer and the probability value of each class is calculated through the softmax activation function. The model uses the cross entropy loss function to calculate the loss value, and the loss function is The back propagation algorithm is then used to update the model parameters based on the loss value. This process is repeated in τ iterations until the model converges or the maximum number of training rounds is reached.

[0055] Furthermore, in step S5, the graph attention network layer dynamically assigns different weights to each residue node through the self-attention mechanism, which allows the model to integrate key features from neighboring nodes while maintaining the feature information of the node itself.

[0056] In this paper, a graph attention network layer is used to capture the complex interactions between protein residues. The protein graph can be represented as a contact graph A∈R L×L , where L is the number of residues, X∈R L×c is a residue-level feature matrix containing the representation of each residue. GAT receives the adjacency matrix A and the residue-level embedding of the previous layer As input, it outputs the residue-level embedding of the next layer

[0057] H (l) ∈GAT(A,H (l) )

[0058] In the graph attention network layer, the pairwise importance score e between the residue nodes is ij =a(θh i ,θh i ) based on the hidden states h of nodes i and j in layer l i and h i Calculate these hidden states through the weight matrix W (l) The model uses masked attention scores to convert To update the new feature representation of the graph structure information. ij It reflects the relative importance of node j to node i.

[0059] New feature representation of nodes It is updated by the weighted sum of neighbor node features. The ReLU activation function is used to introduce nonlinearity, allowing the model to capture more complex patterns. In order to integrate the information of all nodes in the graph, a global pooling layer is used to extract a global feature vector of fixed length, thereby enhancing the model's perception of overall structural information. Subsequently, the potential representation of each GO term is calculated through a fully connected layer combined with the ReLU activation function.

[0060] Finally, the SoftMax layer outputs the probability prediction of each GO term, and the model can intuitively express the prediction results of various functions through probability vectors.

[0061] In addition, in addition to providing a multimodal protein function prediction method, the present invention also provides a multimodal protein function prediction system, which specifically includes:

[0062] Feature extraction module: used to process protein sequence data through one-dimensional convolutional layers and pooling layers to extract local dependency features in protein sequences and capture patterns and key sequence features in proteins;

[0063] Embedding information extraction module: It is used to take protein sequence data as input and use a large-scale pre-trained protein language model to extract residue-level embedding information in the protein sequence as node features for application to the contact graph;

[0064] Contact map construction module: It is used to take protein sequence data as input, predict the probability of contact between protein residues based on Cb-Cb distance, and construct a contact map;

[0065] Fusion module: used to jointly process features of different modalities using a multi-head attention mechanism and a residual structure, integrate data from different modalities through feature fusion, and fuse the extracted key sequence features, the residue-level embedding information, and the contact map;

[0066] Feature representation module: used to process the fused feature representation using a graph attention network. The output of the graph attention network generates a new feature representation for each node, fuses the information of neighboring nodes in a weighted summation manner, and then obtains a global feature vector of a fixed length through a global pooling layer. Finally, the probability of each Gene Ontology term is calculated through a fully connected layer and a SoftMax activation function to predict protein function.

[0067] In specific implementation, the above modules can be implemented as independent entities, or can be arbitrarily combined and implemented as the same or several entities. The specific implementation of the above units can refer to the previous method embodiments, which will not be repeated here.

[0068] The multimodal protein prediction method of the present invention can be applied in many fields:

[0069] Although the patent focuses on function prediction, the contact map information used can potentially be used to derive protein structure. Through contact map and graph neural network technology, it can be expanded to sequence-based structure prediction. Pharmaceutical companies and research institutes can combine structure prediction tools to predict the three-dimensional structure of unknown proteins, thereby providing support for drug design.

[0070] Personalized medicine: With the development of genomics and proteomics technologies, personalized medicine requires functional annotation of patient-specific proteins. The technology in this patent can be used for large-scale protein function prediction to help develop personalized disease treatment plans. This technology can be used to analyze protein sequences in patient samples, predict their functions, and associate them with specific diseases, thereby helping to develop personalized treatments.

[0071] Functional food development: In the field of food science, predicting the function of specific proteins can provide a basis for the development of functional foods, such as identifying proteins related to immunity, antioxidant and other functions. Food companies can use this technology to identify functional components in proteins from plants or animals, design foods containing specific functional proteins, and enhance their health value.

[0072] Agricultural biotechnology: In agriculture, predicting protein functions related to biological resistance, nutrient transport or growth regulation can provide new directions for crop improvement. Agricultural biotechnology companies can use this technology to identify functional proteins related to disease resistance or yield in crops and improve them through gene editing or breeding technology.

[0073] Other potential application areas:

[0074] In biomanufacturing, through function prediction, new functional proteins in microorganisms or enzymes can be discovered for catalytic reactions or metabolic engineering in biomanufacturing. This technology can be used by industrial biotechnology companies to discover new enzymes through protein function prediction, optimize biological reaction processes, and improve production efficiency.

[0075] Environmental monitoring: By analyzing protein sequences in environmental samples and predicting their functions, functional proteins related to the degradation of environmental pollutants can be identified. Environmental monitoring agencies can use this technology to predict functional proteins in specific environments and help develop new bioremediation strategies.

[0076] Figure 2 This is a schematic diagram of the structure of a computer device disclosed in the present invention. Figure 2As shown, the computer device 400 includes at least a memory 402 and a processor 401; the memory 402 is connected to the processor via a communication bus 403, and is used to store computer instructions executable by the processor 401, and the processor 401 is used to read computer instructions from the memory 402 to implement the steps of the method described in any of the above embodiments.

[0077] For the above-mentioned device embodiments, since they basically correspond to the method embodiments, the relevant parts can refer to the partial description of the method embodiments. The device embodiments described above are only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without paying any creative work.

[0078] Computer readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media and memory devices, including, for example, semiconductor memory devices (e.g., EPROM, EEPROM and flash memory devices), magnetic disks (e.g., internal disks or removable disks), magneto-optical disks, and CD ROM and DVD-ROM disks. The processor and memory may be supplemented by, or incorporated in, special purpose logic circuits.

[0079] Finally, it should be noted that although this specification contains many specific implementation details, these should not be interpreted as limiting the scope of any invention or the scope of protection claimed, but are mainly used to describe the features of specific embodiments of specific inventions. Certain features described in multiple embodiments within this specification may also be implemented in combination in a single embodiment. On the other hand, the various features described in a single embodiment may also be implemented separately in multiple embodiments or in any suitable sub-combination. In addition, although features may work in certain combinations as described above and even initially claimed as such, one or more features from the claimed combination may be removed from the combination in some cases, and the claimed combination may point to a sub-combination or a variation of a sub-combination.

[0080] Similarly, although operations are depicted in a particular order in the accompanying drawings, this should not be understood as requiring that these operations be performed in the particular order shown or performed sequentially, or requiring that all illustrated operations be performed to achieve the desired results. In some cases, multitasking and parallel processing may be advantageous. In addition, the separation of various system modules and components in the above-described embodiments should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product, or packaged into multiple software products.

[0081] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims can be performed in a different order and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the particular order or sequential order shown to achieve the desired results. In some implementations, multitasking and parallel processing may be advantageous.

[0082] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present disclosure shall be included in the scope of protection of the present disclosure.

Claims

1. A multimodal protein function prediction method, characterized in that: The steps include: The protein sequence data is processed through one-dimensional convolutional layers and pooling layers to extract local dependency features in the protein sequence and capture patterns and key sequence features in the protein; Taking protein sequence data as input, a large-scale pre-trained protein language model is used to extract residue-level embedding information in the protein sequence as node features and apply it to the contact graph. Using protein sequence data as input, the contact probability between protein residues is predicted based on Cb-Cb distance to construct a contact map; Using a multi-head attention mechanism and a residual structure to jointly process features of different modalities, integrating data from different modalities through feature fusion, and fusing the extracted key sequence features, the residue-level embedding information, and the contact map; In the fused feature representation, a graph attention network is used for processing. The output of the graph attention network generates a new feature representation for each node, and the information of neighboring nodes is fused by weighted summation. Then, a global feature vector of fixed length is obtained through a global pooling layer. Finally, the probability of each GeneOntology term is calculated through a fully connected layer and a SoftMax activation function to predict protein function.

2. The multimodal protein function prediction method according to claim 1, characterized in that: The key sequence feature is expressed as: Where F1 is a deep neural network model that processes sequence modalities, and function f conv and f maxp Represent the operations of the convolutional layer and the maximum pooling layer respectively.

3. The multimodal protein function prediction method according to claim 1, characterized in that: The following formula is used to extract residue-level embedding information in protein sequences using a large-scale pre-trained protein language model: L MLM =-∑ i∈M logp(x i |x M ) The mask set M contains 15% of the positions i in the sequence x, according to the surrounding context x M Predict masked amino acids x i identity.

4. The multimodal protein function prediction method according to claim 1, characterized in that: The method for constructing a contact map includes the following steps: RaptorX uses a deep residual network to predict the distances and contact relationships between protein residues, and then uses a gradient-based energy minimization method to convert these predicted distance and direction information into a three-dimensional structural model of the protein.

5. The multimodal protein function prediction method according to claim 2, characterized in that: The method for integrating data from different modalities by feature fusion includes: a joint attention step and a feature fusion step, wherein the joint attention step includes a multi-head joint attention module and two fully connected feed-forward layers, and the formula is as follows: FFL(x)=max(0,xW1+b1)W2+b2 Input training data (X, Y), where X represents the input features and Y represents the corresponding labels. At the same time, input hyperparameters, including learning rate η, number of attention heads h, and maximum number of training rounds τ. Sample a mini-batch (x, y) from the dataset and extract features from the two modalities, denoted by ox1 and ox2 respectively. Calculate the attention matrices Q, K, and V, and divide these matrices into h attention heads. For each head, calculate the attention score and weight the features according to the score to generate a weighted feature representation.

6. The multimodal protein function prediction method according to claim 5, characterized in that: The method of the feature fusion step includes: The weighted features are processed through residual connections and normalization layers to generate output features of each layer, and the output features of all layers are concatenated to form a fused feature representation O; The fused feature O is input into the fully connected layer, and the probability value of each class is calculated through the softmax activation function. The model uses the cross entropy loss function to calculate the loss value. The loss function is The back-propagation algorithm is then used to update the model parameters according to the loss value, and this process is repeated in τ iterations until the model converges or the maximum number of training rounds is reached.

7. The multimodal protein function prediction method according to claim 5, characterized in that: Methods for processing using graph attention networks include: The protein graph is represented as a contact graph A∈R L×L , where L is the number of residues, X∈R L×c is a residue-level feature matrix containing the representation of each residue. GAT receives the adjacency matrix A and the residue-level embedding of the previous layer. As input, it outputs the residue-level embedding of the next layer H (l) ∈GAT(A,H (l) ) In the graph attention network layer, the pairwise importance score between residue nodes is e ij =a(θh i ,θh i ), based on the hidden states h of nodes i and j in layer l i and h i Calculate the hidden state through the weight matrix W (l) Convert and calculate through a single-layer feedforward neural network, through the mask attention score To update the new feature representation of the graph structure information, the ae in the result ij Reflects the relative importance of node j to node i; Then each node generates a new feature representation 8. A prediction system using the multimodal protein function prediction method according to any one of claims 1 to 7, characterized in that: include: Feature extraction module: used to process protein sequence data through one-dimensional convolutional layers and pooling layers to extract local dependency features in protein sequences and capture patterns and key sequence features in proteins; Embedding information extraction module: It is used to take protein sequence data as input and use a large-scale pre-trained protein language model to extract residue-level embedding information in the protein sequence as node features for application to the contact graph; Contact map construction module: It is used to take protein sequence data as input, predict the probability of contact between protein residues based on Cb-Cb distance, and construct a contact map; Fusion module: used to jointly process features of different modalities using a multi-head attention mechanism and a residual structure, integrate data from different modalities through feature fusion, and fuse the extracted key sequence features, the residue-level embedding information, and the contact map; Feature representation module: used to process the fused feature representation using a graph attention network. The output of the graph attention network generates a new feature representation for each node, fuses the information of neighboring nodes in a weighted summation manner, and then obtains a global feature vector of a fixed length through a global pooling layer. Finally, the probability of each Gene Ontology term is calculated through a fully connected layer and a SoftMax activation function to predict protein function.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the multimodal protein function prediction method according to any one of claims 1 to 7 are executed.

10. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the multimodal protein function prediction method according to any one of claims 1 to 7 are implemented.

Images