Protein active site recognition method and device, computer equipment and storage medium

Abstract

The application relates to a protein active site recognition method and device, computer equipment and a storage medium. The method comprises the following steps: based on a protein amino acid sequence, a multi-modal protein graph combining residue level sequence representation and three-dimensional structure information is constructed; a four-layer isometric mapping neural network is used to jointly and iteratively update initial residue features and initial Cartesian coordinates in the multi-modal protein graph to determine geometrically perceived residue features; a protein function description text of the protein amino acid sequence is acquired, and a site perception fusion algorithm is used to determine target residue features according to the protein function description text and the geometrically perceived residue features; and a multi-layer perception machine is used to determine a protein active site recognition result of the protein amino acid sequence according to the target residue features. The above scheme improves the accuracy and reliability of protein active site recognition.

Description

Technical Field

[0001] This application relates to the field of protein structure analysis technology, and in particular to methods, apparatus, computer equipment and storage media for identifying protein active sites. Background Technology

[0002] The goal of protein active site identification is to accurately locate residues related to catalysis, ligand binding, or key functions within the amino acid sequence and three-dimensional structure of a protein, thereby providing fundamental support for enzyme function analysis, protein engineering, and structure-based drug design. Traditional active site discovery methods mainly rely on wet experimental techniques such as site-directed mutagenesis, biochemical experiments, and structural analysis. These methods suffer from long experimental cycles, high costs, and low throughput, making them unsuitable for large-scale protein functional analysis. With the development of deep learning and protein representation learning technologies, computational model-based active site identification has become an important technical approach to improve research efficiency and expand the scale of analysis. In recent years, protein language models have been able to learn contextualized residue representations from large-scale unlabeled protein sequences, demonstrating good capabilities in tasks such as protein structure prediction, function prediction, and variation effect analysis. However, protein activity is essentially determined by its three-dimensional structure. Many active residues are far apart in sequence but collectively form active pockets or catalytic microenvironments in space. Relying solely on one-dimensional sequences or only performing coarse-grained structural fusion often fails to fully utilize residue-level geometric relationships and long-range spatial dependencies. Meanwhile, research indicates that functional descriptions, protein names, species origins, and related attributes can provide supplementary semantic clues for active site identification. How to perform fine-grained co-modeling of sequence features, three-dimensional structural geometry, and functional semantics within a unified framework has become a key issue in the field of protein active site identification. Therefore, there is an urgent need for a protein active site identification method that can simultaneously model protein amino acid sequences, three-dimensional structures, and functional semantic information, while also considering both geometric sensitivity and predictive stability at the residue level. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, apparatus, and computer device for identifying protein active sites that can take into account both geometric sensitivity and predictive stability at the residue level, in order to address the above-mentioned technical problems.

[0004] In a first aspect, this application provides a method for identifying protein active sites, the method comprising:

[0005] Based on protein amino acid sequences, a multimodal protein map is constructed that jointly encodes residue-level sequence representation and three-dimensional structural information;

[0006] A four-layer isomorphic graph neural network is used to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein map to determine the geometrically aware residue features;

[0007] The protein functional description text of the protein amino acid sequence is obtained, and the target residue features are determined by the site-aware fusion algorithm based on the protein functional description text and the geometrically aware residue features.

[0008] The protein active site identification result of the protein amino acid sequence is determined by using a multilayer perceptron based on the characteristics of the target residues.

[0009] In one embodiment, a multimodal protein map jointly encoding residue-level sequence representation and three-dimensional structural information is constructed based on the protein amino acid sequence, including:

[0010] By using a pre-trained protein language model, features are extracted from the protein amino acid sequence to determine the initial residue features of each amino acid residue in the protein amino acid sequence; the pre-trained protein language model includes a protein sequence pre-training model and a protein sequence representation model.

[0011] By modeling a folding model at an evolutionary scale, the initial Cartesian coordinates of each amino acid residue in the protein's amino acid sequence are determined based on the protein's amino acid sequence.

[0012] Based on the initial residue features and the initial Cartesian coordinates, a multimodal protein map jointly encoding residue-level sequence representation and three-dimensional structural information is constructed.

[0013] In one embodiment, a pre-trained protein language model is used to extract features from the protein amino acid sequence to determine the initial residue features of each amino acid residue in the protein amino acid sequence, including:

[0014] By using a protein sequence pre-training model, features are extracted from the protein amino acid sequence to determine the first residue feature corresponding to each amino acid residue in the protein amino acid sequence.

[0015] By using a protein sequence representation model, features are extracted from the protein amino acid sequence to determine the second residue feature corresponding to each amino acid residue in the protein amino acid sequence.

[0016] Based on the first residue feature and the second residue feature, the splicing feature is determined, and the splicing feature is mapped to a unified feature space to determine the initial residue feature of each amino acid residue in the protein amino acid sequence.

[0017] In one embodiment, a four-layer isomorphic graph neural network is used to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein map to determine the geometrically aware residue features, including:

[0018] By embedding a mapping function, the initial residue features in the multimodal protein map are mapped to the hidden space to determine the hidden features;

[0019] If the undirected graph edges in the multimodal protein graph carry edge attributes, then the edge attributes are mapped to edge features based on a linear mapping function.

[0020] Using a four-layer isomorphic graph neural network, the initial residue features and initial Cartesian coordinates in the multimodal protein graph are jointly iteratively updated based on the hidden features, the edge features, and the initial Cartesian coordinates to determine the geometrically aware residue features.

[0021] In one embodiment, a four-layer isomorphic graph neural network is used to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein graph based on the hidden features, the edge features, and the initial Cartesian coordinates, to determine the geometry-aware residue features, including:

[0022] Based on the initial Cartesian coordinates, the relative geometric relationships between each amino acid residue are determined; the relative geometric relationships include relative displacement vectors and relative squared distances.

[0023] Using a four-layer isomorphic graph neural network, the initial residue features in the multimodal protein graph are jointly and iteratively updated layer by layer according to the hidden features, the edge features, and the relative geometric relationships to determine the geometrically aware residue features.

[0024] In one embodiment, a protein function description text of the protein's amino acid sequence is obtained, and a target residue feature is determined based on the protein function description text and the geometrically sensed residue features using a site-aware fusion algorithm, including:

[0025] The protein function description text is semantically encoded using a site-aware fusion algorithm to determine the text semantic vector.

[0026] Based on the geometrically perceived residue features and the text semantic vector, the enhanced residue features incorporating functional semantics are determined;

[0027] The target residue features are determined by splicing geometrically sensed residue features and enhanced residue features.

[0028] In one embodiment, determining enhanced residue features incorporating functional semantics based on the geometrically perceived residue features and the textual semantic vector includes:

[0029] The geometrically sensed residue features are pooled to determine the protein embedding vector, and the text semantic vector is pooled to determine the pooled text embedding vector.

[0030] Determine the text feature cosine similarity of the pooled text embedding vector and the protein feature cosine similarity of the protein embedding vector, and use the text feature cosine similarity and the protein feature cosine similarity as intramodal soft targets;

[0031] Based on the temperature coefficient, the pooled text embedding vector, and the protein embedding vector, a cross-modal soft target is determined;

[0032] The label loss data is determined based on the intra-modal soft target and the determined cross-modal soft target using the symmetric soft label loss function;

[0033] Based on the label loss data, the site-aware fusion algorithm is iteratively optimized through backpropagation to determine the target-aware fusion algorithm;

[0034] Using a target-aware fusion algorithm based on a cross-attention mechanism, geometrically aware residue features are used as query vectors and textual semantic vectors are used as key vectors to determine enhanced residue features incorporating functional semantics.

[0035] Secondly, this application also provides a protein active site identification device, the device comprising:

[0036] A multimodal protein map construction module is used to construct multimodal protein maps that jointly encode residue-level sequence representations and three-dimensional structural information based on protein amino acid sequences;

[0037] The geometry-aware feature determination module is used to perform joint iterative updates on the initial residue features and initial Cartesian coordinates in the multimodal protein map using a four-layer isovariant graph neural network to determine the geometry-aware residue features.

[0038] The target residue feature determination module is used to obtain the protein functional description text of the protein amino acid sequence and determine the target residue features based on the protein functional description text and the geometrically sensed residue features using a site-aware fusion algorithm.

[0039] The protein active site identification module is used to determine the protein active site identification result of the protein amino acid sequence based on the target residue characteristics using a multilayer sensor.

[0040] Thirdly, this application also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0041] Based on protein amino acid sequences, a multimodal protein map is constructed that jointly encodes residue-level sequence representation and three-dimensional structural information;

[0042] A four-layer isomorphic graph neural network is used to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein map to determine the geometrically aware residue features;

[0043] The protein functional description text of the protein amino acid sequence is obtained, and the target residue features are determined by the site-aware fusion algorithm based on the protein functional description text and the geometrically aware residue features.

[0044] The protein active site identification result of the protein amino acid sequence is determined by using a multilayer perceptron based on the characteristics of the target residues.

[0045] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0046] Based on protein amino acid sequences, a multimodal protein map is constructed that jointly encodes residue-level sequence representation and three-dimensional structural information;

[0047] A four-layer isomorphic graph neural network is used to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein map to determine the geometrically aware residue features;

[0048] The protein functional description text of the protein amino acid sequence is obtained, and the target residue features are determined by the site-aware fusion algorithm based on the protein functional description text and the geometrically aware residue features.

[0049] The protein active site identification result of the protein amino acid sequence is determined by using a multilayer perceptron based on the characteristics of the target residues.

[0050] The aforementioned protein active site identification method, apparatus, and computer equipment construct a multimodal protein map based on the protein amino acid sequence, jointly encoding residue-level sequence representation and three-dimensional structural information. A four-layer isovariant graph neural network is used to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein map to determine geometrically sensed residue features. Protein functional description text of the protein amino acid sequence is obtained, and target residue features are determined based on the protein functional description text and the geometrically sensed residue features using a site-aware fusion algorithm. Finally, the protein active site identification result of the protein amino acid sequence is determined based on the target residue features using a multilayer perceptron. This solves the problem that existing protein active site identification technologies rely solely on one-dimensional sequence or coarse-grained structural fusion, failing to fully utilize residue geometric relationships and long-range spatial dependencies, and struggling to complete fine-grained collaborative modeling of sequence, three-dimensional structure, and functional semantic information within a unified framework. The above scheme combines amino acid sequences to construct a multimodal protein map with fusion sequence and three-dimensional structural information. Then, it uses a four-layer isomorphic graph neural network to iteratively update features and coordinates to obtain geometrically perceptual residue features. The target residue features are obtained by fusing functional textual semantic information through a site-aware fusion algorithm. The active site is identified by a multilayer perceptron. This scheme can fully integrate multi-dimensional information of sequence, spatial geometry and functional semantics, accurately mine the spatial and functional associations of residues, and effectively improve the accuracy and reliability of protein active site identification. Attached Figure Description

[0051] Figure 1 This is a diagram illustrating the application environment of a protein active site identification method in one embodiment.

[0052] Figure 2 This is a flowchart illustrating a protein active site identification method in one embodiment;

[0053] Figure 3 This is a flowchart illustrating a method for constructing a multimodal protein map in one embodiment;

[0054] Figure 4 This is a flowchart illustrating a geometrically sensed residue feature determination method in one embodiment;

[0055] Figure 5 This is a flowchart illustrating a method for determining target residue features in one embodiment;

[0056] Figure 6 This is a structural block diagram of a protein active site recognition device in one embodiment;

[0057] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] The protein active site identification method provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed in the cloud or on another network server. Server 104 constructs a multimodal protein map based on the protein amino acid sequence, jointly encoding residue-level sequence representation and three-dimensional structural information; it uses a four-layer isovariant graph neural network to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein map to determine geometrically sensed residue features; it acquires the protein functional description text of the protein amino acid sequence, and uses a site-aware fusion algorithm to determine target residue features based on the protein functional description text and the geometrically sensed residue features; it uses a multilayer perceptron to determine the protein active site identification result of the protein amino acid sequence based on the target residue features, and sends the protein active site identification result to terminal 102 via a communication network.

[0060] In one embodiment, such as Figure 2 As shown, a method for identifying protein active sites is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0061] S210. Based on the protein amino acid sequence, construct a multimodal protein map that jointly encodes residue-level sequence representation and three-dimensional structural information.

[0062] Proteins are long chains of amino acids strung together in a fixed sequence. The amino acid sequence is the string of strings formed by the amino acids in that order. A multimodal protein diagram is a structured representation that combines the three-dimensional structure of a protein with the residue features of its amino acid sequence, thus incorporating both spatial topological and sequence characteristic information. Residues are the basic structural units that make up proteins.

[0063] Specifically, the protein amino acid sequence is obtained, and features are extracted from the protein amino acid sequence through a pre-trained protein language model to determine residue-level sequence features. The three-dimensional coordinates of each amino acid residue in the protein amino acid sequence are determined through an evolutionary scale modeling folding model, and topological edges are constructed based on the three-dimensional coordinates, thereby forming a multimodal protein graph that carries both residue sequence features and three-dimensional coordinates.

[0064] For example, such as Figure 3 As shown, a multimodal protein map jointly encoding residue-level sequence representation and three-dimensional structural information is constructed based on the protein amino acid sequence, including:

[0065] S2101. By using a pre-trained protein language model, feature extraction is performed on the protein amino acid sequence to determine the initial residue features of each amino acid residue in the protein amino acid sequence.

[0066] Pre-trained protein language models include protein sequence pre-trained models and protein sequence representation models.

[0067] The protein sequence pre-training model is the ESM2 model, and the protein sequence representation model is the Protein Sequence Representation Model based on the ProtTrans system, abbreviated as the ProtT5 model. The protein sequence pre-training model can extract deep sequence semantic representations residue-by-residue based on the protein amino acid sequence, outputting a high-dimensional embedding vector for each amino acid residue. The protein sequence representation model is used to perform residue-level feature encoding on the protein amino acid sequence, outputting a high-dimensional semantic embedding representation for each amino acid residue.

[0068] For example, by using a pre-trained protein language model, feature extraction is performed on the protein amino acid sequence to determine the initial residue features of each amino acid residue in the protein amino acid sequence, including:

[0069] By using a protein sequence pre-training model, features are extracted from the protein amino acid sequence to determine the first residue feature corresponding to each amino acid residue in the protein amino acid sequence. By using a protein sequence representation model, features are extracted from the protein amino acid sequence to determine the second residue feature corresponding to each amino acid residue in the protein amino acid sequence. Based on the first residue feature and the second residue feature, splicing features are determined and mapped to a unified feature space to determine the initial residue feature of each amino acid residue in the protein amino acid sequence.

[0070] Specifically, let the length of the protein amino acid sequence be N, where N is a positive integer. The protein amino acid sequence is input into a pre-trained ESM2 model, which extracts features from the sequence to determine the first residue feature corresponding to each amino acid residue. The protein amino acid sequence is then input into a pre-trained ProtT5 model, which extracts features from the sequence to determine the second residue feature corresponding to each amino acid residue. The first and second residue features are concatenated and mapped to a unified feature space to determine the initial residue feature of each amino acid residue in the protein amino acid sequence.

[0071] For example, when a protein amino acid sequence is input into a pre-trained ESM2 model, for the i-th amino acid residue in the protein amino acid sequence, the ESM2 model outputs the i-th amino acid residue. The first residue characteristic of each amino acid residue is The second residue feature of the i-th amino acid residue output by the ProtT5 model is: The first residue feature and the second residue feature of the i-th amino acid residue are concatenated and mapped to a unified feature space to determine the initial residue feature of the i-th amino acid residue in the protein amino acid sequence. The expression for the initial residue feature of the i-th amino acid residue in the protein amino acid sequence is shown in formula (1):

[0072] (1).

[0073] in, The vector obtained by representing a learnable linear projection. This serves as the initial residue feature of the node corresponding to the i-th amino acid residue.

[0074] The above scheme extracts two types of residue features by using two different pre-trained protein language models. The two types of residue features are then concatenated and mapped to a unified feature space to obtain the initial residue features. This approach can fully explore the multi-dimensional evolution and semantic information of amino acid sequences, enrich the feature expression ability of the initial residue features, and improve the recognition accuracy and robustness of subsequent active site identification.

[0075] S2102. By modeling a folding model at an evolutionary scale, the initial Cartesian coordinates of each amino acid residue in the protein amino acid sequence are determined based on the protein amino acid sequence.

[0076] Among them, the Evolutionary Scale Modeling Fold (ESMFold) can predict the three-dimensional Cartesian coordinates of each residue based on the protein amino acid sequence, providing basic geometric information for the structural topology construction of multimodal protein maps.

[0077] Specifically, the protein amino acid sequence is input into the evolutionary scale modeling folding model, which outputs the three-dimensional Cartesian coordinates of each amino acid residue in the protein amino acid sequence. This determines the spatial location information of each amino acid residue, providing basic geometric coordinate data for the subsequent construction of the multimodal protein map topology.

[0078] For example, the initial Cartesian coordinates corresponding to protein residue i are shown in Equation (2):

[0079] (2)

[0080] Where, p i Let be the Cartesian coordinates corresponding to protein residue i.

[0081] S2103. Based on the initial residue features and initial Cartesian coordinates, construct a multimodal protein map that jointly encodes residue-level sequence representation and three-dimensional structural information.

[0082] Specifically, based on the initial Cartesian coordinates of each amino acid residue, the Euclidean distance between every two amino acid residues is determined, and undirected graph edges are determined based on the Euclidean distances between amino acid residues. Based on the initial residue features, undirected graph edges, and initial Cartesian coordinates, an initial multimodal protein graph jointly encoding residue-level sequence representation and three-dimensional structural information is constructed. For example, when the Euclidean distance between two amino acid residues is less than a preset distance threshold, undirected graph edges are established between the nodes corresponding to these two amino acid residues.

[0083] For example, when the Euclidean distance between amino acid residue i and amino acid residue j satisfies formula (3), an undirected graph edge is added between amino acid residue i and amino acid residue j.

[0084] (3).

[0085] Where, p j Let j be the initial Cartesian coordinates of amino acid residue j. This represents the Euclidean distance between amino acid residue i and amino acid residue j. r is a positive number and can be set according to actual needs.

[0086] For example, the expression for a multimodal protein map is shown in Equation (4):

[0087] (4)

[0088] Each node It corresponds to a single amino acid residue and carries the initial residue characteristic x. i and initial Cartesian coordinates p i ; Each edge This represents the edges of an undirected graph determined by structural proximity relationships.

[0089] The above scheme extracts and determines the initial residue features of each residue by complementary extraction of two pre-trained protein language models. At the same time, it uses an evolutionary scale modeling folding model to predict the initial Cartesian coordinates of each residue from the amino acid sequence. Then, it integrates residue features and three-dimensional coordinates to construct a multimodal protein map. This approach can fully explore the multi-dimensional semantic information and three-dimensional geometric structure information of protein sequences, realize the joint encoding of sequence characterization and spatial structure, and lay the foundation for the accurate identification of protein active sites.

[0090] S220 employs a four-layer isomorphic graph neural network to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein map, thereby determining the geometrically aware residue features.

[0091] Among them, the four-layer isomorphic graph neural network EGNN is a backbone feature extraction network composed of four layers of isomorphic graph neural network stacked together. It maintains feature stability for rigid body transformations such as translation and rotation in the three-dimensional space of proteins. Taking the constructed multimodal protein map as input, it aggregates residue node features and spatial topological neighborhood information layer by layer, mines the three-dimensional structural associations and deep semantics of amino acid residues, and finally outputs geometrically perceived residue features that have both sequence semantics and geometric perception capabilities.

[0092] For example, such as Figure 4 As shown, a four-layer isomorphic graph neural network is used to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein map to determine the geometrically aware residue features, including:

[0093] S2201. By embedding a mapping function, the initial residue features in the multimodal protein map are mapped to the hidden space to determine the hidden features.

[0094] Specifically, after determining the multimodal protein map G=(V,E), the initial residue features in the multimodal protein map are embedded into the hidden space using an embedding mapping function to determine the hidden features. For example, the expression for embedding the initial residue features in the multimodal protein map into the hidden space is shown in formula (5):

[0095] (5).

[0096] in, The hidden feature is determined after the initial residue feature corresponding to the i-th amino acid residue is mapped to the hidden space. For predefined embedding mapping functions.

[0097] S2202. If the undirected graph edges in the multimodal protein graph carry edge attributes, then the edge attributes are mapped to edge features based on the linear mapping function.

[0098] Among them, edge attributes refer to the feature information attached to the edges of an undirected graph, which are used to describe the structured feature information of the relative position, spatial distance and spatial angle between the two amino acid residues corresponding to the edge of the undirected graph.

[0099] Specifically, if the undirected graph edges in the multimodal protein graph carry edge attributes e ij Then, based on a preset linear mapping function, the edge attributes carried by the edges of the undirected graph are mapped to edge features. For example, edge attribute e is mapped to edge features. ij The expression for mapping to edge features is shown in formula (6):

[0100] (6)

[0101] in, It is a linear mapping function.

[0102] S2203. Using a four-layer isomorphic graph neural network, the initial residue features and initial Cartesian coordinates in the multimodal protein graph are jointly iteratively updated based on hidden features, edge features, and initial Cartesian coordinates to determine geometrically aware residue features.

[0103] Examples of methods for determining geometry-aware residue features include:

[0104] Based on the initial Cartesian coordinates, the relative geometric relationships between amino acid residues are determined. The relative geometric relationships include the relative displacement vector and the relative squared distance. Through a four-layer isomorphic graph neural network, the initial residue features in the multimodal protein map are jointly updated layer by layer according to the hidden features, edge features and relative geometric relationships to determine the geometrically aware residue features.

[0105] Specifically, based on the initial Cartesian coordinates, the relative geometric relationships between amino acid residues are determined. Then, through each layer of the four-layer isovariant graph neural network, the initial residue features and initial Cartesian coordinates are updated layer-by-layer based on the hidden features, edge features, and relative geometric relationships output by the previous layer, thus determining the geometry-aware residue features and updating the Cartesian coordinates. It should be noted that the four-layer isovariant graph neural network consists of four layers. During the layer-by-layer joint iterative update of the initial residue features, each layer further optimizes and updates the updated residue features and updated Cartesian coordinates output by the previous layer based on the hidden features, edge features, and relative geometric relationships output by the previous layer. The updated residue features output by the last layer in the four-layer isovariant graph neural network are the geometry-aware residue features.

[0106] For example, in a four-layer isovariant graphical neural network, the determination of geometrically aware residue features includes:

[0107] Let the relative displacement vector between the i-th amino acid residue and the j-th amino acid residue be... The expression is: The square of the relative distance between the i-th amino acid residue and the j-th amino acid residue The expression is: The edge message between the i-th amino acid residue and the j-th amino acid residue. The expression is shown in formula (7):

[0108] (7)

[0109] in, This represents a multilayer perceptron.

[0110] To maintain the geometric equivariance of the model, the four-layer equivariant graph neural network employs a layered computation mechanism of weight adaptive prediction and relative displacement iterative coordinate update. For example, the update method for Cartesian coordinates in the l-th layer of the four-layer equivariant graph neural network is shown in equations (8), (9), and (10):

[0111] (8);

[0112] (9);

[0113] (10).

[0114] in, It is the indicator weight. Let N(i) be a learnable mapping function, and let N(i) be the set of neighborhood residues of the i-th amino acid residue. Update the offset for the coordinates. Let be the updated Cartesian coordinates of the i-th amino acid residue used as input to the (l+1)-th network layer. The input is the Cartesian coordinate of the i-th amino acid residue in the l-th network layer. It can be understood that 0 ≤ l ≤ 3, and l is an integer. If l = 0, then... Let be the initial Cartesian coordinates of the i-th amino acid residue.

[0115] When updating the initial residue features corresponding to the i-th amino acid residue, the edge messages from all neighboring residues of the amino acid residue to be updated are summed and aggregated to obtain the overall neighborhood message of the node. For example, the overall neighborhood message... The determination method is shown in formula (11):

[0116] (11).

[0117] The expression for updating the residue features corresponding to the i-th amino acid residue in the l-th network layer is shown in Equation (12):

[0118] (12).

[0119] in, The updated residue features are obtained by updating the residue features corresponding to the i-th amino acid residue in the l-th network layer. To input the residue feature corresponding to the i-th amino acid residue of the l-th network layer, it can be understood that if l=0, then This represents the initial residue feature corresponding to the i-th amino acid residue. This indicates that learnable features update the network, and residual connections are used for stability optimization.

[0120] It is understandable that by constructing the relative geometric relationship between residues using initial Cartesian coordinates, and then using a four-layer isovariant graph neural network to iteratively update the initial residue features layer by layer with hidden features, edge features and relative geometric relationships, it is possible to generate geometrically perceptual residue features with stronger expressive power and adapted to the active site identification task.

[0121] By stacking four layers of isotropic graphical neural networks layer by layer and iteratively updating features and coordinates, a geometrically perceptual residue representation that integrates sequence semantics and three-dimensional spatial structural information can be output. This geometrically perceptual residue representation is unaffected by rigid body transformations such as protein translation and rotation, possessing strict geometric isotropic properties. Simultaneously, it can accurately mine and capture the structural dependencies and spatial interactions between residues closely related to the formation of protein active sites.

[0122] S230. Obtain the protein functional description text of the protein amino acid sequence, and determine the target residue features based on the protein functional description text and geometrically sensed residue features using a site-aware fusion algorithm.

[0123] Among them, protein function description text is a natural language description text that records the biological functions, biochemical effects, and physiological regulatory characteristics of proteins.

[0124] Specifically, protein functional description texts of K protein amino acid sequences are obtained, where K is a positive integer. A site-aware fusion algorithm is used to perform cross-modal deep fusion of the protein functional semantic information corresponding to the protein functional description texts with geometrically sensed residue features to determine the target residue features.

[0125] For example, such as Figure 5 As shown, a protein functional description text of the protein's amino acid sequence is obtained, and target residue features are determined based on the protein functional description text and the geometrically sensed residue features using a site-aware fusion algorithm, including:

[0126] S2301. Semantically encode the protein function description text using a site-aware fusion algorithm to determine the text semantic vector.

[0127] Specifically, the semantic vector of the text is determined by semantically encoding the functional description text of K proteins through the MAcross (Multi-Attribute Cross-attention) mechanism in the site-aware fusion algorithm.

[0128] S2302. Based on geometrically perceived residue features and textual semantic vectors, determine the enhanced residue features that incorporate functional semantics.

[0129] An exemplary method for determining enhanced residue features incorporating functional semantics includes: pooling geometrically sensed residue features to determine protein embedding vectors, and pooling text semantic vectors to determine pooled text embedding vectors; determining the text feature cosine similarity of the pooled text embedding vectors and the protein feature cosine similarity of the protein embedding vectors, and using the text feature cosine similarity and protein feature cosine similarity as intramodal soft targets; determining cross-modal soft targets based on temperature coefficients, pooled text embedding vectors, and protein embedding vectors; determining label loss data using a symmetric soft label loss function, based on the intramodal soft targets and the determined cross-modal soft targets; determining a target-aware fusion algorithm by iteratively optimizing a site-aware fusion algorithm through backpropagation based on the label loss data; and determining enhanced residue features incorporating functional semantics using the target-aware fusion algorithm, based on a cross-attention mechanism, with geometrically sensed residue features as query vectors and text semantic vectors as key-value vectors.

[0130] For example, pooling is performed on geometry-aware residue features to determine the protein embedding vector. The text semantic vector is pooled to determine the pooled text embedding vector.

[0131] The expression for the pooled text embedding vector T is shown in formula (13):

[0132] (13).

[0133] in, .

[0134] Cosine similarity of text features from pooled text embedding vectors The expression is: Cosine similarity of protein features in protein embedding vectors The expression is: ;

[0135] Cross-modal soft targets The determination method is shown in formula (14):

[0136] (14).

[0137] in, Indicates the temperature coefficient. This represents the pairwise cosine similarity within a batch.

[0138] Furthermore, a symmetric soft-label loss function is adopted. The loss performs cross-modal alignment of geometrically perceived residue features and textual semantic vectors, and the label loss data. The method for determining is shown in formula (15):

[0139] (15).

[0140] Through the loss-driven cross-modal alignment process described above, the feature representation spatial distribution of protein structure embedding and text semantic embedding can be made consistent, and the similarity logic can be matched with each other. The feature differences between modalities are effectively reduced, providing a more compatible multimodal feature representation for subsequent site-level semantic injection, residue feature fusion and active site prediction.

[0141] Enhanced residue features incorporating functional semantics The method for determining is shown in formula (16):

[0142] (16).

[0143] in, This represents the text after it has been processed by the shared encoder.

[0144] S2303. The geometrically sensed residue features and the enhanced residue features are spliced ​​together to determine the target residue features.

[0145] For example, target residue features The expression is shown in formula (17):

[0146] (17).

[0147] S240. Using a multilayer perceptron, the protein active site identification result of the protein amino acid sequence is determined based on the characteristics of the target residues.

[0148] Specifically, the target residue features are input into a multilayer perceptron, and the protein active site identification result of the protein amino acid sequence is determined based on the output of the multilayer perceptron.

[0149] For example, the determination method for protein active site identification results is shown in formula (18):

[0150] (18).

[0151] MLP stands for Multilayer Perceptron.

[0152] The above scheme first performs semantic encoding on the protein function description text to obtain a text semantic vector, then combines geometrically perceived residue features to generate enhanced residue features that incorporate functional semantics, and finally splices the two types of features to obtain the target residue features. This achieves cross-modal deep fusion of protein three-dimensional structural information and functional semantic information, and improves the site association expression ability of residue features.

[0153] The aforementioned protein active site identification method constructs a multimodal protein map based on the protein amino acid sequence, jointly encoding residue-level sequence representation and three-dimensional structural information. A four-layer isovariant graph neural network is used to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein map to determine geometrically sensed residue features. Protein functional description text of the protein amino acid sequence is obtained, and target residue features are determined based on the protein functional description text and the geometrically sensed residue features using a site-aware fusion algorithm. Finally, a multilayer perceptron is used to determine the protein active site identification result based on the target residue features. This method solves the problem that existing protein active site identification technologies rely solely on one-dimensional sequence or coarse-grained structural fusion, failing to fully utilize residue geometric relationships and long-range spatial dependencies, and struggling to complete fine-grained collaborative modeling of sequence, three-dimensional structure, and functional semantic information within a unified framework. The above scheme combines amino acid sequences to construct a multimodal protein map with fusion sequence and three-dimensional structural information. Then, it uses a four-layer isomorphic graph neural network to iteratively update features and coordinates to obtain geometrically perceptual residue features. The target residue features are obtained by fusing functional textual semantic information through a site-aware fusion algorithm. The active site is identified by a multilayer perceptron. This scheme can fully integrate multi-dimensional information of sequence, spatial geometry and functional semantics, accurately mine the spatial and functional associations of residues, and effectively improve the accuracy and reliability of protein active site identification.

[0154] For example, the above-mentioned protein active site identification method is implemented using the multimodal isomorphic graph learning framework MEG-Site. MEG-Site is an end-to-end protein residue modeling and active site prediction framework built on isomorphic graph neural networks, designed for protein active site identification tasks. It integrates multimodal information such as protein amino acid sequence, three-dimensional structure, and functional textual semantics. The multimodal isomorphic graph learning framework includes a multimodal input and preprocessing component, a multimodal protein graph construction component, a four-layer isomorphic graph neural network, and a site-aware fusion algorithm.

[0155] The multimodal input and preprocessing component can receive and process protein amino acid sequences and protein functional description texts. The multimodal protein map construction component can be used to construct multimodal protein maps that jointly encode residue-level sequence representations and three-dimensional structural information. A four-layer isovariant graph neural network is used to determine geometrically aware residue features based on the multimodal protein map. A site-aware fusion algorithm is used to fuse protein functional description texts and geometrically aware residue features to determine target residue features.

[0156] For example, the training methods of the multimodal isomorphic graph learning framework MEG-Site include:

[0157] MEG-Site was trained using residue-level supervision information. Each amino acid residue was treated as an independent prediction unit, and the prediction logit for the i-th amino acid residue was set as follows: The real label is ,and The definition of the active site recognition loss of MEG-Site is shown in formula (19):

[0158] (19)

[0159] in, Loss due to active site recognition This represents the binary cross-entropy loss. `logit` refers to the unnormalized raw prediction score output by the MEG-Site framework for whether each amino acid residue is an active site.

[0160] Furthermore, the final training objective of the model is shown in equation (20):

[0161] (20)

[0162] in, This indicates a cross-modal alignment regularization term. This represents the weight coefficient of the alignment term, and L is the final training objective of the model.

[0163] Throughout the training process of the MEG-Site framework, the pre-trained encoder parameters remain frozen, and the fusion and prediction processes are optimized only through the aforementioned residue-level classification loss and cross-modal alignment loss. Through the synergistic effect of multimodal protein map construction, isomorphic map learning, and site-aware semantic fusion, fine-grained, geometrically sensitive, and semantically enhanced recognition of protein active sites can be achieved.

[0164] Based on the above embodiments, the protein active site identification method includes:

[0165] Obtain the protein amino acid sequence, with a sequence length of N (a positive integer). Input the protein amino acid sequence into a pre-trained ESM2 model. The ESM2 model extracts features from the protein amino acid sequence to determine the first residue feature corresponding to each amino acid residue. Input the protein amino acid sequence into a pre-trained ProtT5 model. The ProtT5 model extracts features from the protein amino acid sequence to determine the second residue feature corresponding to each amino acid residue. Concatenate the first and second residue features and map them to a unified feature space to determine the initial residue feature of each amino acid residue in the protein amino acid sequence.

[0166] The protein amino acid sequence is input into the evolutionary scale modeling folding model, which outputs the three-dimensional Cartesian coordinates of each amino acid residue in the protein amino acid sequence. This determines the spatial location information of each amino acid residue and provides basic geometric coordinate data for the subsequent construction of the multimodal protein map topology.

[0167] Based on the initial Cartesian coordinates of each amino acid residue, the Euclidean distance between every two amino acid residues is determined, and undirected graph edges are determined based on the Euclidean distances between amino acid residues. An initial multimodal protein graph jointly encoding residue-level sequence representation and three-dimensional structural information is constructed based on the initial residue features, undirected graph edges, and initial Cartesian coordinates. For example, when the Euclidean distance between two amino acid residues is less than a preset distance threshold, undirected graph edges are established between the nodes corresponding to these two amino acid residues.

[0168] A multimodal protein graph is defined. Initial residue features from the multimodal protein graph are embedded into the hidden space using an embedding mapping function to determine the hidden features. If undirected edges in the multimodal protein graph carry edge attributes, these attributes are mapped to edge features based on a predefined linear mapping function.

[0169] Based on the initial Cartesian coordinates, the relative geometric relationships between amino acid residues are determined. Then, through each layer of a four-layer isovariant graph neural network, the initial residue features and initial Cartesian coordinates are updated layer-by-layer based on the hidden features, edge features, and relative geometric relationships output by the previous layer, thus determining the geometry-aware residue features and updating the Cartesian coordinates. It should be noted that the four-layer isovariant graph neural network consists of four layers. During the layer-by-layer joint iterative update of the initial residue features, each layer further optimizes and updates the updated residue features and updated Cartesian coordinates output by the previous layer based on the hidden features, edge features, and relative geometric relationships output by the previous layer. The updated residue features output by the last layer in the four-layer isovariant graph neural network are the geometry-aware residue features.

[0170] The process involves obtaining functional description texts of K protein amino acid sequences, semantically encoding these texts using the MAcross mechanism in a site-aware fusion algorithm, and determining text semantic vectors. Geometrically sensed residue features are pooled to determine protein embedding vectors, and the text semantic vectors are pooled to determine pooled text embedding vectors. The cosine similarity of text features in the pooled text embedding vectors and the cosine similarity of protein features in the protein embedding vectors are determined, and these cosine similarities are used as intramodal soft targets. Cross-modal soft targets are determined based on temperature coefficients, pooled text embedding vectors, and protein embedding vectors. Label loss data is determined using a symmetric soft label loss function, based on the intra-modal soft targets and the determined cross-modal soft targets. Based on the label loss data, the site-aware fusion algorithm is iteratively optimized through backpropagation to determine the target-aware fusion algorithm. Using the target-aware fusion algorithm and a cross-attention mechanism, enhanced residue features incorporating functional semantics are determined, with geometrically sensed residue features as query vectors and text semantic vectors as key-value vectors. Through the loss-driven cross-modal alignment process described above, the feature representation spatial distribution of protein structure embedding and text semantic embedding can be made consistent, and the similarity logic can be matched with each other. The feature differences between modalities are effectively reduced, providing a more compatible multimodal feature representation for subsequent site-level semantic injection, residue feature fusion and active site prediction.

[0171] The target residue features are input into a multilayer perceptron, and the protein active site identification result is determined based on the output of the multilayer perceptron.

[0172] The aforementioned protein active site identification method constructs a multimodal protein map based on the protein amino acid sequence, jointly encoding residue-level sequence representation and three-dimensional structural information. A four-layer isovariant graph neural network is used to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein map to determine geometrically sensed residue features. Protein functional description text of the protein amino acid sequence is obtained, and target residue features are determined based on the protein functional description text and the geometrically sensed residue features using a site-aware fusion algorithm. Finally, a multilayer perceptron is used to determine the protein active site identification result based on the target residue features. This method solves the problem that existing protein active site identification technologies rely solely on one-dimensional sequence or coarse-grained structural fusion, failing to fully utilize residue geometric relationships and long-range spatial dependencies, and struggling to complete fine-grained collaborative modeling of sequence, three-dimensional structure, and functional semantic information within a unified framework. The above scheme combines amino acid sequences to construct a multimodal protein map with fusion sequence and three-dimensional structural information. Then, it uses a four-layer isomorphic graph neural network to iteratively update features and coordinates to obtain geometrically perceptual residue features. The target residue features are obtained by fusing functional textual semantic information through a site-aware fusion algorithm. The active site is identified by a multilayer perceptron. This scheme can fully integrate multi-dimensional information of sequence, spatial geometry and functional semantics, accurately mine the spatial and functional associations of residues, and effectively improve the accuracy and reliability of protein active site identification.

[0173] For example, the ProTAD protein active site identification benchmark dataset was used to evaluate the performance of the above-mentioned protein active site identification method. The ProTAD protein active site identification benchmark dataset contains 570,830 protein samples, each of which includes an amino acid sequence and a multi-attribute functional text description. When available, the 3D structure from the AlphaFold Protein Structure Database was incorporated; for samples without readily available structures, ESMFold was used to predict the 3D structure. The experiment selected multiple evaluation metrics, including Fmax, AUPRC, MCC, OS, and FPR, to comprehensively evaluate the model's performance in identifying active site residues from multiple dimensions, such as recognition accuracy, ranking effect, matching degree, and false positive probability.

[0174] Experimental results show that the proposed MEG-Site framework for protein active site identification achieves state-of-the-art performance on the benchmark dataset. When ESM-2 and ProtT5 are used together as the sequence characterization backbone, the model exhibits excellent performance across various metrics: Fmax (Maximum F-score) of 0.8623, AUPRC (Area Under Precision-Recall Curve) of 0.9189, MCC (Matthews Correlation Coefficient) of 0.8666, OS (Overall Score) of 0.8752, and FPR (False Positive Rate) as low as 0.1290. Compared to the existing mainstream multimodal identification method MMSite (Multimodal Protein Active Site Identification Benchmark Model), the above method significantly improves Fmax and AUPRC values ​​while effectively reducing the false positive rate, fully demonstrating that the above method can significantly improve the accuracy and reliability of active site residue localization.

[0175] Compared to traditional identification methods that rely solely on sequence information and simple structural fusion, the above methods have more prominent advantages. The Fmax values ​​of traditional methods such as ESM-1b, S-PLM, and PST t33_so are all much lower than those of the above methods. This proves that it is difficult to accurately identify active sites affected by the three-dimensional microenvironment by relying solely on sequence evolution information or shallow structural fusion. However, the above methods can accurately locate key residues with functional roles by fusing sequence characterization, three-dimensional geometric structure, and functional semantic information.

[0176] Furthermore, the above methods still possess stable and excellent recognition performance under a single sequence representation architecture. Good recognition results can be achieved by using ESM-2 or ProtT5 as the sequence backbone alone. After fusing and complementing the two sequence representations, the overall performance of the model can be further optimized and upgraded, demonstrating that the sequence, structure and semantic integrated modeling framework built by this invention can fully explore the complementary advantages between different sequence features.

[0177] The aforementioned protein active site identification method effectively improves the detection accuracy, feature sorting ability, and overall matching effect of protein active site identification, while significantly reducing the occurrence of misjudgments. The overall technical effect is excellent and has high practical application value.

[0178] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0179] Based on the same inventive concept, this application also provides a protein active site identification device for implementing the protein active site identification method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more protein active site identification device embodiments provided below can be found in the limitations of the protein active site identification method described above, and will not be repeated here.

[0180] In one embodiment, such as Figure 6 As shown, a protein active site identification device is provided, comprising: a multimodal protein map construction module 601, a geometric sensing feature determination module 602, a target residue feature determination module 603, and a protein active site identification module 604, wherein:

[0181] The multimodal protein map construction module 601 is used to construct multimodal protein maps that jointly encode residue-level sequence representations and three-dimensional structural information based on protein amino acid sequences;

[0182] The geometric perception feature determination module 602 is used to perform joint iterative updates on the initial residue features and initial Cartesian coordinates in the multimodal protein map using a four-layer isovariant graph neural network to determine the geometric perception residue features.

[0183] The target residue feature determination module 603 is used to obtain the protein functional description text of the protein amino acid sequence and determine the target residue features based on the protein functional description text and the geometrically sensed residue features using a site-aware fusion algorithm.

[0184] The protein active site identification module 604 is used to determine the protein active site identification result of the protein amino acid sequence based on the target residue characteristics using a multilayer sensor.

[0185] For example, the multimodal protein map construction module 601 is specifically used for:

[0186] By using a pre-trained protein language model, features are extracted from the protein amino acid sequence to determine the initial residue features of each amino acid residue in the protein amino acid sequence; the pre-trained protein language model includes a protein sequence pre-training model and a protein sequence representation model.

[0187] By modeling a folding model at an evolutionary scale, the initial Cartesian coordinates of each amino acid residue in the protein's amino acid sequence are determined based on the protein's amino acid sequence.

[0188] Based on the initial residue features and the initial Cartesian coordinates, a multimodal protein map jointly encoding residue-level sequence representation and three-dimensional structural information is constructed.

[0189] For example, the multimodal protein map construction module 601 is also specifically used for:

[0190] By using a protein sequence pre-training model, features are extracted from the protein amino acid sequence to determine the first residue feature corresponding to each amino acid residue in the protein amino acid sequence.

[0191] By using a protein sequence representation model, features are extracted from the protein amino acid sequence to determine the second residue feature corresponding to each amino acid residue in the protein amino acid sequence.

[0192] Based on the first residue feature and the second residue feature, the splicing feature is determined, and the splicing feature is mapped to a unified feature space to determine the initial residue feature of each amino acid residue in the protein amino acid sequence.

[0193] For example, the geometric perception feature determination module 602 is specifically used for:

[0194] By embedding a mapping function, the initial residue features in the multimodal protein map are mapped to the hidden space to determine the hidden features;

[0195] If the undirected graph edges in the multimodal protein graph carry edge attributes, then the edge attributes are mapped to edge features based on a linear mapping function.

[0196] Using a four-layer isomorphic graph neural network, the initial residue features and initial Cartesian coordinates in the multimodal protein graph are jointly iteratively updated based on the hidden features, the edge features, and the initial Cartesian coordinates to determine the geometrically aware residue features.

[0197] For example, the geometric perception feature determination module 602 is also specifically used for:

[0198] Based on the initial Cartesian coordinates, the relative geometric relationships between each amino acid residue are determined; the relative geometric relationships include relative displacement vectors and relative squared distances.

[0199] Using a four-layer isomorphic graph neural network, the initial residue features in the multimodal protein graph are jointly and iteratively updated layer by layer according to the hidden features, the edge features, and the relative geometric relationships to determine the geometrically aware residue features.

[0200] For example, the target residue feature determination module 603 is specifically used for:

[0201] The protein function description text is semantically encoded using a site-aware fusion algorithm to determine the text semantic vector.

[0202] Based on the geometrically perceived residue features and the text semantic vector, the enhanced residue features incorporating functional semantics are determined;

[0203] The target residue features are determined by splicing geometrically sensed residue features and enhanced residue features.

[0204] For example, the target residue feature determination module 603 is further specifically used for:

[0205] The geometrically sensed residue features are pooled to determine the protein embedding vector, and the text semantic vector is pooled to determine the pooled text embedding vector.

[0206] Determine the text feature cosine similarity of the pooled text embedding vector and the protein feature cosine similarity of the protein embedding vector, and use the text feature cosine similarity and the protein feature cosine similarity as intramodal soft targets;

[0207] Based on the temperature coefficient, the pooled text embedding vector, and the protein embedding vector, a cross-modal soft target is determined;

[0208] The label loss data is determined based on the intra-modal soft target and the determined cross-modal soft target using the symmetric soft label loss function;

[0209] Based on the label loss data, the site-aware fusion algorithm is iteratively optimized through backpropagation to determine the target-aware fusion algorithm;

[0210] Using a target-aware fusion algorithm based on a cross-attention mechanism, geometrically aware residue features are used as query vectors and textual semantic vectors are used as key vectors to determine enhanced residue features incorporating functional semantics.

[0211] Each module in the aforementioned protein active site identification device 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, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0212] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a method for identifying protein active sites. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or a key vector, trackball, or touchpad set on the computer device casing, or an external key vector disk, touchpad, or mouse, etc.

[0213] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0214] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0215] Step 1: Based on the protein amino acid sequence, construct a multimodal protein map that jointly encodes residue-level sequence representation and three-dimensional structural information;

[0216] Step 2: Using a four-layer isomorphic graph neural network, the initial residue features and initial Cartesian coordinates in the multimodal protein map are jointly iteratively updated to determine the geometrically aware residue features;

[0217] Step 3: Obtain the protein functional description text of the protein amino acid sequence, and determine the target residue features based on the protein functional description text and the geometrically sensed residue features using a site-aware fusion algorithm.

[0218] Step 4: Using a multilayer sensor, determine the protein active site identification result of the protein amino acid sequence based on the target residue characteristics.

[0219] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0220] Step 1: Based on the protein amino acid sequence, construct a multimodal protein map that jointly encodes residue-level sequence representation and three-dimensional structural information;

[0221] Step 2: Using a four-layer isomorphic graph neural network, the initial residue features and initial Cartesian coordinates in the multimodal protein map are jointly iteratively updated to determine the geometrically aware residue features;

[0222] Step 3: Obtain the protein functional description text of the protein amino acid sequence, and determine the target residue features based on the protein functional description text and the geometrically sensed residue features using a site-aware fusion algorithm.

[0223] Step 4: Using a multilayer sensor, determine the protein active site identification result of the protein amino acid sequence based on the target residue characteristics.

[0224] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0225] Step 1: Based on the protein amino acid sequence, construct a multimodal protein map that jointly encodes residue-level sequence representation and three-dimensional structural information;

[0226] Step 2: Using a four-layer isomorphic graph neural network, the initial residue features and initial Cartesian coordinates in the multimodal protein map are jointly iteratively updated to determine the geometrically aware residue features;

[0227] Step 3: Obtain the protein functional description text of the protein amino acid sequence, and determine the target residue features based on the protein functional description text and the geometrically sensed residue features using a site-aware fusion algorithm.

[0228] Step 4: Using a multilayer sensor, determine the protein active site identification result of the protein amino acid sequence based on the target residue characteristics.

[0229] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0230] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0231] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0232] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for identifying protein active sites, characterized in that, include: Based on protein amino acid sequences, a multimodal protein map is constructed that jointly encodes residue-level sequence representation and three-dimensional structural information; A four-layer isomorphic graph neural network is used to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein map to determine the geometrically aware residue features; The protein functional description text of the protein amino acid sequence is obtained, and the target residue features are determined by the site-aware fusion algorithm based on the protein functional description text and the geometrically aware residue features. The protein active site identification result of the protein amino acid sequence is determined by using a multilayer perceptron based on the characteristics of the target residues.

2. The method according to claim 1, characterized in that, Based on protein amino acid sequences, a multimodal protein map jointly encoding residue-level sequence representation and three-dimensional structural information is constructed, including: By using a pre-trained protein language model, features are extracted from the protein amino acid sequence to determine the initial residue features of each amino acid residue in the protein amino acid sequence; the pre-trained protein language model includes a protein sequence pre-training model and a protein sequence representation model. By modeling a folding model at an evolutionary scale, the initial Cartesian coordinates of each amino acid residue in the protein's amino acid sequence are determined based on the protein's amino acid sequence. Based on the initial residue features and the initial Cartesian coordinates, a multimodal protein map jointly encoding residue-level sequence representation and three-dimensional structural information is constructed.

3. The method according to claim 2, characterized in that, By using a pre-trained protein language model, feature extraction is performed on the protein amino acid sequence to determine the initial residue features of each amino acid residue in the protein amino acid sequence, including: By using a protein sequence pre-training model, features are extracted from the protein amino acid sequence to determine the first residue feature corresponding to each amino acid residue in the protein amino acid sequence. By using a protein sequence representation model, features are extracted from the protein amino acid sequence to determine the second residue feature corresponding to each amino acid residue in the protein amino acid sequence. Based on the first residue feature and the second residue feature, the splicing feature is determined, and the splicing feature is mapped to a unified feature space to determine the initial residue feature of each amino acid residue in the protein amino acid sequence.

4. The method according to claim 1, characterized in that, A four-layer isomorphic graph neural network is used to jointly iteratively update the initial residue features and initial Cartesian coordinates in the multimodal protein map to determine the geometrically aware residue features, including: By embedding a mapping function, the initial residue features in the multimodal protein map are mapped to the hidden space to determine the hidden features; If the undirected graph edges in the multimodal protein graph carry edge attributes, then the edge attributes are mapped to edge features based on a linear mapping function. Using a four-layer isomorphic graph neural network, the initial residue features and initial Cartesian coordinates in the multimodal protein graph are jointly iteratively updated based on the hidden features, the edge features, and the initial Cartesian coordinates to determine the geometrically aware residue features.

5. The method according to claim 4, characterized in that, Using a four-layer isomorphic graph neural network, based on the hidden features, edge features, and initial Cartesian coordinates, the initial residue features and initial Cartesian coordinates in the multimodal protein graph are jointly iteratively updated to determine the geometrically aware residue features, including: Based on the initial Cartesian coordinates, the relative geometric relationships between each amino acid residue are determined; the relative geometric relationships include relative displacement vectors and relative squared distances. Using a four-layer isomorphic graph neural network, the initial residue features in the multimodal protein graph are jointly and iteratively updated layer by layer according to the hidden features, the edge features, and the relative geometric relationships to determine the geometrically aware residue features.

6. The method according to claim 1, characterized in that, The protein functional description text of the protein's amino acid sequence is obtained, and target residue features are determined based on the protein functional description text and the geometrically sensed residue features using a site-aware fusion algorithm, including: The protein function description text is semantically encoded using a site-aware fusion algorithm to determine the text semantic vector. Based on the geometrically perceived residue features and the text semantic vector, the enhanced residue features incorporating functional semantics are determined; The target residue features are determined by splicing geometrically sensed residue features and enhanced residue features.

7. The method according to claim 6, characterized in that, Based on the geometrically perceived residue features and the textual semantic vector, enhanced residue features incorporating functional semantics are determined, including: The geometrically sensed residue features are pooled to determine the protein embedding vector, and the text semantic vector is pooled to determine the pooled text embedding vector. Determine the text feature cosine similarity of the pooled text embedding vector and the protein feature cosine similarity of the protein embedding vector, and use the text feature cosine similarity and the protein feature cosine similarity as intramodal soft targets; Based on the temperature coefficient, the pooled text embedding vector, and the protein embedding vector, a cross-modal soft target is determined; The label loss data is determined based on the intra-modal soft target and the determined cross-modal soft target using the symmetric soft label loss function; Based on the label loss data, the site-aware fusion algorithm is iteratively optimized through backpropagation to determine the target-aware fusion algorithm; Using a target-aware fusion algorithm based on a cross-attention mechanism, geometrically aware residue features are used as query vectors and textual semantic vectors are used as key vectors to determine enhanced residue features incorporating functional semantics.

8. A protein active site identification device, characterized in that, The protein active site identification device includes: A multimodal protein map construction module is used to construct multimodal protein maps that jointly encode residue-level sequence representations and three-dimensional structural information based on protein amino acid sequences; The geometry-aware feature determination module is used to perform joint iterative updates on the initial residue features and initial Cartesian coordinates in the multimodal protein map using a four-layer isovariant graph neural network to determine the geometry-aware residue features. The target residue feature determination module is used to obtain the protein functional description text of the protein amino acid sequence and determine the target residue features based on the protein functional description text and the geometrically sensed residue features using a site-aware fusion algorithm. The protein active site identification module is used to determine the protein active site identification result of the protein amino acid sequence based on the target residue characteristics using a multilayer sensor.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images