Protein function prediction method and device based on topology-aware attention network

Abstract

This invention proposes a method and apparatus for protein function prediction based on a topology-aware attention network. The method includes the following steps: S1: acquiring a publicly available protein function annotation dataset, a publicly available gene ontology description, and relational .obo files; S2: constructing a protein graph input and a gene ontology graph input; S3: building a topology-aware protein-gene ontology attention network, including a topology-aware attention module and a multi-head aggregator; S4: training the topology-aware protein-gene ontology attention network from step S3 using the data from step S1 to obtain a trained protein function prediction model; S5: for a protein whose function is to be predicted, using the trained model from step S5, generating the probability that it is annotated by each GO term. This method can improve the performance of protein function prediction.

Description

Technical Field

[0001] This invention relates to the field of protein function prediction and analysis, and in particular to a method and apparatus for protein function prediction based on a topologically perceptive attention network. Background Technology

[0002] Proteins play an indispensable role in biological systems, participating in a variety of important functions, such as catalyzing chemical reactions, regulating gene expression, and transmitting molecular signals. Annotating protein functions allows us to gain a deeper understanding of cellular mechanisms, identify disease-related changes in gene or protein function, and discover new tools for disease prevention, diagnosis, and treatment. However, due to the high cost of biochemical experiments, the amount of experimental functional annotation data is relatively limited. Although UniProtKB has recorded a vast number of protein sequences, exceeding 115 million as of April 2018, only a small fraction (approximately 150,000, less than 0.15%) have been experimentally annotated. Therefore, using computational methods to predict protein functions and compensate for the lack of experimental annotation data is becoming increasingly important.

[0003] Gene ontology (GO) is widely used to systematically describe protein function, providing a standardized descriptive framework for protein function. GO contains over 40,000 terms, covering three domains: molecular function (MF), biological processes (BP), and cellular components (CC). The hierarchical relationships between these terms can be described using a directed acyclic graph (DAG). Proteins are typically annotated with multiple GO terms. Furthermore, if a protein is annotated with a particular GO term, it is automatically annotated with all its corresponding ancestor GO terms.

[0004] Protein function prediction faces significant challenges due to the substantial semantic gap between the low-dimensional semantic space of proteins and the vast, hierarchical, and highly imbalanced semantic space of GO terms. Current methods primarily rely on constructing various feature extraction models to derive protein-level embeddings from protein sequences or structures. These embeddings are then fed into fully connected layers to project them into the semantic space of GO terms to predict protein function. While these methods have made some recent progress, several limitations remain to be addressed. First, these methods use fully connected layers to compute the semantic similarity between protein-level embeddings and GO terms. However, they often fail to effectively capture the topological hierarchy information of GO terms. This limitation affects their ability to adequately model the topological semantic similarity between GO terms and protein residues, leading to a significant semantic gap. Consequently, these methods struggle to predict multiple functions simultaneously and may perform poorly when dealing with less common functions. Second, protein functions are often associated with specific structural regions composed of specific residues, rather than the entire protein sequence. Therefore, relying on protein-level embeddings may not accurately capture the relationship between these functions and specific structural regions.

[0005] Therefore, it can be seen that the existing methods have the technical problem of poor prediction performance. Summary of the Invention

[0006] In view of the shortcomings of the prior art mentioned in the background section, the present invention proposes a protein function prediction method and device based on topological sensing attention network.

[0007] The technical solution adopted in this invention is as follows:

[0008] The first aspect discloses a protein function prediction method based on topology-aware attention networks, including:

[0009] S1: Obtain publicly available protein function annotation datasets, publicly available gene ontology descriptions, and relational OBO files, where each protein sequence in the publicly available protein function annotation dataset is annotated with multiple GO terms.

[0010] S2: Construct a protein graph input based on publicly available protein function annotation datasets, and construct a basic graph input based on publicly available gene ontology and relation obo files;

[0011] S3: Construct a topology-aware protein-gene ontology attention network. The network includes a topology-aware attention module and a multi-head aggregator. The topology-aware attention module is used to extract topology-aware residue-level embeddings and GO term-level embeddings from the input data, project them into the joint semantic space, and then use attention scores to model the topological semantic similarity between them. The multi-head aggregator is used to aggregate the similarity calculated by the topology-aware attention module and calculate the probability that the protein is annotated by GO terms.

[0012] S4: Using the protein graph input and the basic graph input as training data input, train the protein-gene ontology attention network based on topological awareness to obtain a trained protein function prediction model.

[0013] S5: For a protein whose function is to be predicted, use the trained model to generate the probability that it is annotated by each GO term.

[0014] In one implementation, step S2 includes:

[0015] The 3D structure of a protein is obtained using a protein structure prediction tool, and the coordinates of the central carbon atom of each residue are extracted. The distance between two residues is calculated based on the coordinates of the central carbon atoms of each pair of residues. Then, a contact threshold is used to determine whether two residues are in contact, resulting in a protein contact map, which is used as the adjacency matrix of the protein graph input. The protein-level embedding and residue-level embedding are calculated using an ESM tool, and the obtained residue-level embedding is used as the node feature of the protein graph input. The adjacency matrix and the node feature of the protein graph input together constitute the protein graph input.

[0016] The topological hierarchy relationships between GO terms in the obo file containing the gene ontology and its relationships are read using the obonet tool and used as the adjacency matrix of the gene ontology graph input. Then, one-hot encoding is used to encode the GO terms, and the encoded GO terms are used as the node features of the gene ontology graph input. The adjacency matrix and the node features of the gene ontology graph input together constitute the gene ontology graph input.

[0017] In one implementation, the processing procedure of the topology-aware attention module in step S3 includes:

[0018] Soft-masked graph neural networks are used to extract topology-aware residue-level embeddings from protein graph inputs;

[0019] Graph attention networks are used to extract topology-aware GO term-level embeddings from gene ontology graph inputs.

[0020] We use an attention mechanism to project topology-aware residue-level embeddings and topology-aware GO term-level embeddings into a joint semantic space, and use attention score modeling to calculate the topological semantic similarity between residue-level embeddings and topology-aware GO term-level embeddings.

[0021] The attention score is activated using the Softmax function, which is defined as follows:

[0022]

[0023] Where e is the natural constant, representing the sum of powers of the similarity to the base e, m is the number of GO terms, and S... i,j The soft max(S) represents the topological semantic similarity between residues i and j of the i-th GO technique. i,j ) represents S i,j The result after applying Softmax.

[0024] In one implementation, when extracting topology-aware residue-level embeddings from a protein map input using a soft-masked graph neural network, the convolutions employed include mask convolution and embedding convolution, wherein the formula for embedding convolution is:

[0025]

[0026] in Let || represent the soft mask of node v in the k-th layer, and || represent the concatenation operation. Let d represent the trainable matrix, d represent the dimension of the node representation, and k represent the number of layers. This represents the embedding of node v at the k-th layer. Let u represent the embedding of node v at level k-1, and let u∈N(v) represent all neighbors of node v. This indicates the embedding of node u at layer (k-1).

[0027] The formula for masked convolution is:

[0028]

[0029] in and Representing an affine mapping, the MLP consists of a two-layer perceptron with an output feature dimension of 1. A sigmoid activation layer is applied after each MLP layer to... The mapping is from 0 to 1, where u∈N(v) represents all neighbors of node v. These represent the soft masks for nodes v and u at layer k-1, respectively.

[0030] In one implementation, the multi-head aggregator uses a trainable weight matrix to aggregate the semantic similarity between each GO term and all residues of the protein, and the aggregation result serves as the probability that the protein is annotated by each GO term. Specifically:

[0031]

[0032] Where S i ∈R 1×n Let represent the semantic similarity between the i-th row of the semantic similarity matrix, i.e., the i-th GO term, and all residues of the protein, where i = 1, 2, ..., m, b represents the bias, m is the number of GO terms, n is the length of the protein sequence, and W i f Let f represent the f-th aggregator head of the i-th GO term, where f = 1, 2, ..., F, and F is the number of aggregator heads; then, a max-pooling layer is applied to obtain the final probability output C ∈ R of the protein being annotated by each GO term. m×1 .

[0033] In one embodiment, the method further includes inputting protein-level embeddings into a fully connected layer and a sigmoid layer to obtain a protein prediction output C′∈R. m×1 Then, add C′ to C to obtain the final function prediction probability.

[0034] In one implementation, during the training process in step S4, the model is trained using a binary cross-entropy loss function and an Adam optimizer.

[0035] Based on the same inventive concept, a second aspect of the present invention provides a protein function prediction device based on a topology-aware attention network, comprising:

[0036] The data acquisition module is used to acquire publicly available protein function annotation datasets, publicly available gene ontology descriptions, and relational OBO files. Each protein sequence in the publicly available protein function annotation dataset is annotated with multiple GO terms.

[0037] The input building module is used to construct protein graph inputs based on publicly available protein function annotation datasets and basic body graph inputs based on publicly available descriptive gene ontology and relational obo files.

[0038] The model building module is used to build a topology-aware protein-gene ontology attention network. The network includes a topology-aware attention module and a multi-head aggregator. The topology-aware attention module is used to extract topology-aware residue-level embeddings and GO term-level embeddings from the input data, project them into the joint semantic space, and then use attention scores to model the topological semantic similarity between them. The multi-head aggregator is used to aggregate the similarity calculated by the topology-aware attention module and calculate the probability that the protein is annotated by GO terms.

[0039] The model training module is used to train a topology-aware protein-gene ontology attention network by taking the protein graph input and the basic graph input as training data input to obtain a trained protein function prediction model.

[0040] The protein function prediction module is used to generate the probability of a protein with a function to be predicted being annotated by each GO term using the trained model.

[0041] Based on the same inventive concept, a third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the protein function prediction method based on a topology-aware attention network as described in the first aspect.

[0042] Based on the same inventive concept, a fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the protein function prediction method based on a topology-aware attention network as described in the first aspect.

[0043] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows:

[0044] This invention provides a protein function prediction method based on a topology-aware attention network. A protein function prediction model is constructed and trained. This model introduces a topology-aware attention module to extract topology-aware residue-level embeddings and topology-aware GO term-level embeddings from protein and gene ontology graphs, and injects them into a joint semantic space. Using attention scores within the joint semantic space, fine-grained semantic similarity between the topology-aware residue-level embeddings and topology-aware GO term-level embeddings is modeled, thereby reducing the semantic gap. Furthermore, the model includes a multi-head aggregator to aggregate the fine-grained semantic similarity between the topology-aware residue-level embeddings and topology-aware GO term-level embeddings to predict protein function. This aggregator cleverly captures function-related fine-grained semantic similarities and filters out task-irrelevant parts, effectively revealing the protein-function relationship, thus enhancing the universality and robustness of the model and improving the prediction performance of protein function. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram illustrating the hierarchical relationship between GO terms provided in the embodiments of the present invention, taking molecular function as an example;

[0047] Figure 2 This is a flowchart illustrating the internal processing of the protein-gene ontology attention network based on topology awareness in an embodiment of the present invention.

[0048] Figure 3 This is a flowchart illustrating the workflow of the model in this embodiment of the invention. Detailed Implementation

[0049] This invention provides a method and apparatus for protein function prediction based on a topology-aware attention network, which improves the effectiveness of protein function prediction. To achieve this effect, the main concepts of this invention are as follows:

[0050] 1. This invention introduces a topology-aware attention module to extract topology-aware residue-level embeddings and topology-aware GO term-level embeddings from protein maps and gene ontology maps, and injects them into a joint semantic space. By utilizing attention scores within the joint semantic space, fine-grained semantic similarity between the topology-aware residue-level embeddings and the topology-aware GO term-level embeddings is modeled, thereby reducing the semantic gap.

[0051] 2. Furthermore, this invention proposes a multi-head aggregator for aggregating fine-grained semantic similarities between topology-aware residue-level embeddings and topology-aware GO term-level embeddings to predict protein function. This aggregator cleverly captures function-related fine-grained semantic similarities and filters out task-irrelevant parts, effectively revealing protein-function relationships, thereby enhancing the universality and robustness of this invention.

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1

[0054] This invention discloses a protein function prediction method based on a topology-aware attention network, comprising:

[0055] S1: Obtain publicly available protein function annotation datasets, publicly available gene ontology descriptions, and relational OBO files, where each protein sequence in the publicly available protein function annotation dataset is annotated with multiple GO terms.

[0056] S2: Construct a protein graph input based on publicly available protein function annotation datasets, and construct a basic graph input based on publicly available gene ontology and relation obo files;

[0057] S3: Construct a topology-aware protein-gene ontology attention network. The network includes a topology-aware attention module and a multi-head aggregator. The topology-aware attention module is used to extract topology-aware residue-level embeddings and GO term-level embeddings from the input data, project them into the joint semantic space, and then use attention scores to model the topological semantic similarity between them. The multi-head aggregator is used to aggregate the similarity calculated by the topology-aware attention module and calculate the probability that the protein is annotated by GO terms.

[0058] S4: Using the protein graph input and the basic graph input as training data input, train the protein-gene ontology attention network based on topological awareness to obtain a trained protein function prediction model.

[0059] S5: For a protein whose function is to be predicted, use the trained model to generate the probability that it is annotated by each GO term.

[0060] Please see Figure 1 The hierarchical relationship between GO terms is illustrated using molecular function as an example.

[0061] The specific implementation methods of each step of the present invention are described in detail below.

[0062] S2 can be achieved through the following steps:

[0063] S2.1: The 3D structure of the protein is obtained using a protein structure prediction tool, and the coordinates of the central carbon atom of each residue are extracted. The distance between two residues is calculated based on the coordinates of the central carbon atoms of each pair of residues. Then, a contact threshold is used to determine whether two residues are in contact, thereby deriving the protein's contact map, which is used as the adjacency matrix of the protein graph input. In addition, the ESM tool is used to calculate protein-level embeddings and residue-level embeddings, and the residue-level embeddings obtained by the ESM tool are used as node features in the protein's graph input. The aforementioned adjacency matrix and node features together constitute the protein's graph input.

[0064] S2.2: The obonet tool is used to read the topological hierarchy relationships between GO terms in the ontology and relation obo files as the adjacency matrix in the gene ontology graph. Then, one-hot encoding is used to encode the GO terms, which are used as node features in the gene ontology graph. The adjacency matrix and node features together constitute the gene ontology graph input.

[0065] The processing steps of the topology-aware attention module in step S3 include:

[0066] S3.1: Extract topology-aware residue-level embeddings P∈R from the protein graph input using a soft-mask graph neural network (soft-mask GNN). n×d , where n is the length of the protein sequence. This process includes two parts: mask convolution and embedding convolution.

[0067] The formula for embedding convolution is:

[0068]

[0069] in Let || denote the soft mask of node v in the (k-th)th layer, and || denote the concatenation operation. Furthermore… Let represent a trainable matrix, where d represents the dimension of the node representation and k represents the number of layers. This represents the embedding of node v at the k-th layer. Let u represent the embedding of node v at level k-1, and let u∈N(v) represent all the neighbors of node v.

[0070] The formula for masked convolution is:

[0071]

[0072] in and This represents an affine mapping, and the MLP consists of a two-layer perceptron with an output feature dimension of 1. A sigmoid activation layer is applied after each MLP layer to achieve... Represents all neighbors of node v. These represent the masks for nodes v and u at layer k-1, respectively.

[0073] S3.2: Extracting topology-aware GO term-level embeddings Q∈R from gene ontology graph terms using a Graph Attention Network (GAT). m×d , where m is the number of GO terms.

[0074] S3.3: Utilizing attention to embed topologically aware residue-level structures P∈R n×d And topology-aware GO term-level embeddings Q∈R m×d Projecting onto the joint semantic space, note that the fractional matrix is ​​calculated as follows:

[0075] S = dot(Q, P) T ),

[0076] Where S i,j The topological semantic similarity between GO term i and residue j is represented by dot(·), where dot(·) represents dot product.

[0077] Then, the attention score is activated using the Softmax function. The Softmax function is defined as follows:

[0078]

[0079] Where e is the natural constant, representing the sum of powers raised to the power of e with the similarity as the exponent, and m is the number of GO terms. softmax(S i,j) represents the result of softmax processing of the topological semantic similarity between the i-th GO term and the j-th residue.

[0080] The processing steps of the multi-head aggregator in step S3 include:

[0081] S3.4: The multi-head aggregator uses a trainable weight matrix to aggregate the semantic similarity between each GO term and all residues of the protein, as the probability that the protein is annotated by each GO term. This process is described as follows:

[0082]

[0083] Where S i ∈R 1×n (i = 1, 2, ..., m) represents the i-th row in the semantic similarity matrix, b represents the bias, m is the number of GO terms, and n is the length of the protein sequence. W i f ∈R n×1 (i = 1, 2, ..., m)(f = 1, 2, ..., F) represents the aggregator for the i-th GO term. F is the number of aggregator heads. Then we apply a max-pooling layer, and finally obtain the output C ∈ R of the probability that the protein is annotated by each GO term. m×1 .

[0084] To further improve prediction performance, step S3.5 is included: inputting the protein-level embedding into the fully connected layer and the Sigmoid layer to obtain the output C′∈R. m×1 and with C∈R m×1 Add them together to get the final function prediction probability.

[0085] Please see Figure 2 and Figure 3 , Figure 2 This is a flowchart of the internal processing of a topology-aware protein-gene ontology attention network. The GO directed acyclic graph is the input form of GO terminology. Figure 3 This is a flowchart of the model's workflow.

[0086] The training process in step S4 includes training the model using the Binary CrossEntropy Loss (BCE) function and the Adam optimizer.

[0087] The training process requires finding the optimal model value based on the loss function and gradient descent. The gradient descent method employs the adaptive momentum estimation algorithm, Adam, which combines momentum and RMSprop. It not only uses momentum as the parameter update direction but also adaptively adjusts the learning rate. The main idea of ​​momentum is to use accumulated momentum instead of the actual gradient at each iteration, effectively mitigating the severe impact of oscillations on optimization speed during gradient descent. In the t-th iteration, the parameter update direction is:

[0088] Δθ t =ρΔθ t-1 -αg t ;

[0089] Where ρ is the momentum factor, α is the learning rate, and g t Let represent the updated gradient, and t represent the number of update steps; the actual update difference for each parameter depends on the weighted average of the gradients over a recent period; Adam not only uses momentum as the parameter update direction, but also adaptively adjusts the learning rate, specifically by calculating the squared gradient g. t 2 The exponentially weighted average is then used to calculate the gradient g. t The exponentially weighted average, the relevant formula is:

[0090] M t =β1M t-1 +(1-β1)g t ;

[0091] G t =β2G t-1 +(1-β2)g t ⊙g t ;

[0092] Where β1 and β2 are the decay rates of the two moving averages, g t M represents the updated gradient. t and G t They represent g respectively t The first and second moments are given, and t represents the number of update steps. Calculate:

[0093]

[0094] and G t and M t Bias correction.

[0095] Adam's parameter update difference is:

[0096]

[0097] Here, ε represents a very small number to avoid a denominator of 0. The number of training epochs is set to 100, with multiple iterations in each epoch until all training samples have participated in one training iteration. The batch size for each iteration is set to 32. The initial learning rate is set to 0.0005.

[0098] The model training epochs are set to 100, with multiple iterations in each epoch until all training samples have participated in one training iteration. The batch size for each iteration is set to 32. The initial learning rate is set to 0.0005.

[0099] The method proposed in this invention will be illustrated and verified through specific examples below.

[0100] Example 1: Preprocess the publicly available protein function prediction dataset PDB-cdhit according to step S1; construct the protein graph input and gene ontology input according to step S2; build the model according to step S3; train according to step S4, and randomly divide the dataset into training, validation, and test sets in an 8:1:1 ratio. The training and validation sets are used to train and select the model, and the test set is used to test the performance of the trained model.

[0101] Table 1 shows the predictive performance of the protein function prediction method based on topology-aware protein-gene ontology attention network proposed in this invention on the problem of protein function prediction.

[0102] Table 1. Performance of protein function prediction based on topology-aware protein-gene ontology attention network.

[0103]

[0104] In Table 1, MF represents molecular function, BP represents biological processes, and CC represents cellular components. AUPRC represents the area under the precision-recall curve. Fmax represents the maximum F1 score achieved across all prediction thresholds from 0 to 1 (with a step size of 0.01).

[0105] Example 2

[0106] Based on the same inventive concept, this embodiment discloses a protein function prediction device based on a topology-aware attention network, comprising:

[0107] The data acquisition module is used to acquire publicly available protein function annotation datasets, publicly available gene ontology descriptions, and relational OBO files. Each protein sequence in the publicly available protein function annotation dataset is annotated with multiple GO terms.

[0108] The input building module is used to construct protein graph inputs based on publicly available protein function annotation datasets and basic body graph inputs based on publicly available descriptive gene ontology and relational obo files.

[0109] The model building module is used to build a topology-aware protein-gene ontology attention network. The network includes a topology-aware attention module and a multi-head aggregator. The topology-aware attention module is used to extract topology-aware residue-level embeddings and GO term-level embeddings from the input data, project them into the joint semantic space, and then use attention scores to model the topological semantic similarity between them. The multi-head aggregator is used to aggregate the similarity calculated by the topology-aware attention module and calculate the probability that the protein is annotated by GO terms.

[0110] The model training module is used to train a topology-aware protein-gene ontology attention network by taking the protein graph input and the basic graph input as training data input to obtain a trained protein function prediction model.

[0111] The protein function prediction module is used to generate the probability of a protein with a function to be predicted being annotated by each GO term using the trained model.

[0112] Since the device described in Embodiment 2 of this invention is the same device used to implement the protein function prediction method based on topological perceptual attention network in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this device based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All devices used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0113] Example 3

[0114] Based on the same inventive concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in Embodiment 1.

[0115] Since the computer-readable storage medium described in Embodiment 3 of this invention is the same computer-readable storage medium used in implementing the protein function prediction method based on a topology-aware attention network in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer-readable storage medium based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All computer-readable storage media used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0116] Example 4

[0117] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in Embodiment 1.

[0118] Since the computer device described in Embodiment 4 of this invention is the same computer device used to implement the protein function prediction method based on topological perceptual attention network in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer device based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All computer devices used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0119] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0121] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various modifications and variations to the embodiments of the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the embodiments of the invention fall within the scope of the claims of the invention and their equivalents, the invention also intends to include these modifications and variations.

Claims

1. A protein function prediction method based on topologically perceptive attention networks, characterized in that, include: S1: Obtain publicly available protein function annotation datasets, publicly available gene ontology descriptions, and relational OBO files, where each protein sequence in the publicly available protein function annotation dataset is annotated with multiple GO terms. S2: Construct a protein graph input based on publicly available protein function annotation datasets, and construct a basic graph input based on publicly available gene ontology and relation obo files; S3: Construct a topology-aware protein-gene ontology attention network. The network includes a topology-aware attention module and a multi-head aggregator. The topology-aware attention module is used to extract topology-aware residue-level embeddings and GO term-level embeddings from the input data, project them into the joint semantic space, and then use attention scores to model the topological semantic similarity between them. The multi-head aggregator is used to aggregate the similarity calculated by the topology-aware attention module and calculate the probability that the protein is annotated by GO terms. S4: Using the protein graph input and the basic graph input as training data input, train the protein-gene ontology attention network based on topological awareness to obtain a trained protein function prediction model. S5: For a protein whose function is to be predicted, use the trained model to generate the probability that it is annotated by each GO term. The processing steps of the topology-aware attention module in step S3 include: Soft-masked graph neural networks are used to extract topology-aware residue-level embeddings from protein graph inputs; Graph attention networks are used to extract topology-aware GO term-level embeddings from gene ontology graph inputs. We use an attention mechanism to project topology-aware residue-level embeddings and topology-aware GO term-level embeddings into a joint semantic space, and use attention score modeling to calculate the topological semantic similarity between residue-level embeddings and topology-aware GO term-level embeddings. The attention score is activated using the Softmax function, which is defined as follows: in, The natural constant is represented by . The base is the sum of powers raised to the power of that similarity. The number of GO terms. Represents the i-th GO technique Topological semantic similarity between j residues express The result after Softmax; The multi-head aggregator uses a trainable weight matrix to aggregate the semantic similarity between each GO term and all residues of the protein. The aggregation result serves as the probability that the protein is annotated by each GO term. Specifically: in Represents the first in the semantic similarity matrix Line, i.e., the semantic similarity between the i-th GO term and all residues of the protein, , Indicates bias. The number of GO terms. The length of the protein sequence, Indicates the first The first GO term One aggregator head, , The number of aggregator heads is used; then a max-pooling layer is applied to obtain the final probability output of the protein being annotated by each GO term. .

2. The protein function prediction method based on topologically aware attention network as described in claim 1, characterized in that, Step S2 includes: The 3D structure of a protein is obtained using a protein structure prediction tool, and the coordinates of the central carbon atom of each residue are extracted. The distance between two residues is calculated based on the coordinates of the central carbon atoms of each pair of residues. Then, a contact threshold is used to determine whether two residues are in contact, resulting in a protein contact map, which is used as the adjacency matrix of the protein graph input. The protein-level embedding and residue-level embedding are calculated using an ESM tool, and the obtained residue-level embedding is used as the node feature of the protein graph input. The adjacency matrix and the node feature of the protein graph input together constitute the protein graph input. The topological hierarchy relationships between GO terms in the obo file containing the gene ontology and its relationships are read using the obonet tool and used as the adjacency matrix of the gene ontology graph input. Then, one-hot encoding is used to encode the GO terms, and the encoded GO terms are used as the node features of the gene ontology graph input. The adjacency matrix and the node features of the gene ontology graph input together constitute the gene ontology graph input.

3. The protein function prediction method based on topologically perceptive attention network as described in claim 1, characterized in that, When extracting topology-aware residue-level embeddings from protein map input using a soft-masked graph neural network, the convolutions used include mask convolution and embedding convolution. The formula for embedding convolution is as follows: , in Indicates the first Layer nodes The soft mask, where || represents the concatenation operation. Let d represent the trainable matrix, d represent the dimension of the node representation, and k represent the number of layers. Represents a node Embedding at layer k, Represents a node Embedding at layer k-1, Represents all neighbors of node v. Represents a node Embedding at the (k-1)th layer; The formula for masked convolution is: ， in and Representing an affine mapping, the MLP consists of a two-layer perceptron with an output feature dimension of 1. A sigmoid activation layer is applied after each MLP layer to... Mapped to 0 to 1, Represents all neighbors of node v. , Representing nodes respectively and nodes In the Soft mask for the layer.

4. The protein function prediction method based on topologically aware attention network as described in claim 1, characterized in that, The method further includes inputting protein-level embeddings into a fully connected layer and a sigmoid layer to obtain protein prediction outputs. and will and Add them together to get the final function prediction probability.

5. The protein function prediction method based on topologically perceptive attention network as described in claim 1, characterized in that, During the training process in step S4, the model is trained using the binary cross-entropy loss function and the Adam optimizer.

6. A protein function prediction device based on a topologically perceptive attention network, characterized in that, Based on the method described in claim 1, the apparatus includes: The data acquisition module is used to acquire publicly available protein function annotation datasets, publicly available gene ontology descriptions, and relational OBO files. Each protein sequence in the publicly available protein function annotation dataset is annotated with multiple GO terms. The input building module is used to construct protein graph inputs based on publicly available protein function annotation datasets and basic body graph inputs based on publicly available descriptive gene ontology and relational obo files. The model building module is used to build a topology-aware protein-gene ontology attention network. The network includes a topology-aware attention module and a multi-head aggregator. The topology-aware attention module is used to extract topology-aware residue-level embeddings and GO term-level embeddings from the input data, project them into the joint semantic space, and then use attention scores to model the topological semantic similarity between them. The multi-head aggregator is used to aggregate the similarity calculated by the topology-aware attention module and calculate the probability that the protein is annotated by GO terms. The model training module is used to train a topology-aware protein-gene ontology attention network by taking the protein graph input and the basic graph input as training data input to obtain a trained protein function prediction model. The protein function prediction module is used to generate the probability of a protein with a function to be predicted being annotated by each GO term using the trained model.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the protein function prediction method based on a topology-aware attention network as described in any one of claims 1 to 5.

8. 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, it implements the protein function prediction method based on a topology-aware attention network as described in any one of claims 1 to 5.

Images