Metal binding site mutation effect prediction method based on panoramic attention enhancement

Abstract

The invention relates to the technical field of biological information, and particularly discloses a metal binding site mutation effect prediction method based on panoramic attention enhancement, and the method comprises the steps: S1, obtaining disease-related mutation data of a metal binding site from a database; s2, dividing the data into training, verification and test sample sets; s3, performing data preprocessing operation on the sample set; s4, training and verifying the neural network framework by using the preprocessed training and verification sample set, and storing the trained neural network framework; and S5, inputting the preprocessed test sample set into the trained neural network framework to obtain a metal binding site mutation effect prediction result. According to the method, feature semantic information implied by a protein language model is deeply excavated, the panoramic attention network is driven to be constructed, the perception ability of the network to a metal protein structure is remarkably enhanced, finally, effective fusion of structure and sequence representation is achieved, and the prediction precision of the metal binding site mutation effect is improved.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method for predicting metal binding site mutation effects based on image panoramic attention enhancement. Background Art

[0002] It is estimated that more than 25% of proteins need to interact with one or more metal ions to perform their biological functions within cells, such as electron transfer, molecular recognition, gene transcription and expression regulation. Missense mutations in metal binding sites caused by single nucleotide changes can hinder protein-metal interactions and destroy metalloprotein structures, thereby impairing cellular protein network function, which may lead to disease development and affect drug resistance. Therefore, accurate prediction of disease-associated mutations in metal binding sites is crucial for understanding the molecular mechanisms of many diseases and developing effective treatment strategies.

[0003] Currently, existing methods for predicting the effects of metal binding site mutations are mainly divided into two categories: traditional methods and deep learning methods. Traditional methods combine various machine learning algorithms, including naive Bayes, support vector machines, and random forests, etc. However, these methods rely heavily on manually extracted features, and their limited ability to learn from given data makes further improvement challenging.

[0004] In recent years, deep learning methods have been widely used in the research direction of mutation effect prediction due to their efficient feature extraction and good feature expression capabilities. They mainly include extracting sequence information embedded in protein language models through simple CNN or Transformer encoders for prediction, and learning the structural context around the mutation site through neural networks for prediction.

[0005] The above two deep learning models have the following main problems: 1. The first model does not fully explore the deeper semantic information in the protein language model, and lacks the learning of metal protein structure, while the structure of protein is an important factor affecting the realization of protein biological function. 2. The second model lacks comprehensive information aggregation of the environment around the metal binding site, and it is difficult to capture the inherent complexity of different metal coordination structure patterns.

[0006] Therefore, how to better integrate the sequence information embedded in the protein language model and the complex structural information of metalloprotein coordination to give the deep learning model more powerful feature expression capabilities to improve the accuracy of mutation effect prediction is still a difficult problem before us. Summary of the invention

[0007] In order to solve the above-mentioned technical problems existing in the prior art, the present invention provides a method for predicting the effects of metal binding site mutations based on graph panoramic attention enhancement. The method deeply mines the rich semantic information contained in the protein language model embedding, drives the construction of a graph panoramic attention network, and deeply perceives the structure of metal proteins. Through the guided fusion mechanism of bidirectional feedback, it can effectively fuse structure and sequence representations to achieve comprehensive information utilization, making the final feature representation powerful and generalizable, thereby improving the accuracy of metal binding site mutation effect prediction.

[0008] The technical solution of the present invention to solve the above technical problems is: a method for predicting the effect of metal binding site mutations based on image panoramic attention enhancement, comprising the following steps:

[0009] S1: Obtain disease-related mutation data of metal binding sites from the database;

[0010] S2: Divide the acquired disease-related mutation dataset on metal binding sites into a training sample set, a validation sample set, and a test sample set;

[0011] S3: performing data preprocessing operations on the sample set divided in step S2; the data preprocessing operations include constructing a metal coordination network diagram, encoding biological feature information, and obtaining deep semantic embedding of a protein language model;

[0012] S4: Use the preprocessed training sample set and verification sample set obtained in step S3 to train and verify the neural network framework, and save the trained neural network framework;

[0013] S5: Input the preprocessed test sample set obtained in step S3 into the neural network framework trained in step S4 to obtain the prediction result of the metal binding site mutation effect.

[0014] Furthermore, in step S3, when constructing the metal coordination network diagram, the sample set divided in step S2 is constructed according to the composition mode of metal ions, ligands and ligand neighbors to construct an amino acid diagram of the metal binding site network, and an atomic diagram is constructed according to a threshold value of 3.

[0015] In step S3, when encoding the biometric information, the biometric information including residue depth, secondary structure, relative solvent accessible surface area, coordination bonds, and hydrogen bonds is encoded;

[0016] In step S3, when obtaining the deep semantic embedding of the protein language model, the embedding of the 33rd layer of the ESM2 protein language model esm2_t33_650M_UR50D() on the long sequence is obtained and mapped to the corresponding amino acids on the metalpdb structure to obtain the ESM2 semantic embedding.

[0017] Furthermore, the neural network framework in step S4 is a protein language model embedding enhanced graph panoramic attention learning framework, specifically:

[0018] In the atomic view attention aggregation module of the framework, the pre-processed atomic graph and amino acid graph are input to extract the structural features of the neighborhood around the metal binding site from both atomic and amino acid perspectives, and the feature embedding of atoms and amino acids is integrated through the self-attention mechanism of atomic clusters.

[0019] In the graph-wide attention perception enhancement module of the framework, the network weaves four attention-driven modules, namely node-to-node, node-to-edge, edge-to-edge, and edge-to-node, to build a panoramic message delivery path. At the same time, each driver module injects deep semantic information from the protein language model to enhance the learning of graph structure, and iteratively feeds back structural information to the ESM encoding module to continuously optimize the representation of sequence information.

[0020] In the self-distillation knowledge part of the framework, the ESM2 embedding obtained from the masked graph structure features of the mutation site and the amino acid sequence after the amino acid at the mutation site is replaced with "X" is input into the atomic view attention aggregation module and the graph panoramic attention perception enhancement module of the normal gradient feedback; and a target feature generator is constructed, the input of the target feature generator is the graph structure features and the ESM2 embedding obtained from the normal sequence; the target feature generator performs parameter updates through exponential moving average;

[0021] In the mutation effect prediction part of the framework, the feature representation of the fusion of the sequence and structure obtained in the coding part is passed through a linear layer (D, 21) to obtain the predicted probability of the amino acid type before and after the mutation, and its maximum likelihood value is calculated. Finally, it passes through a linear layer and a sigmoid function to obtain the probability of the pathogenicity of the mutation.

[0022] Furthermore, the protein language model is embedded in the graph panoramic attention perception enhancement module of the enhanced graph panoramic attention learning framework:

[0023] In the node-to-node module, use the formula The node feature matrix updated by the atomic view attention aggregation module and the ESM2 semantic embedding after dimensionality reduction are used for preliminary feature fusion; It means adding the elements at corresponding positions. Represents the node feature matrix that has been initially updated by the atomic view attention aggregation module; represents the semantic embedding of the ESM2 protein language model after dimensionality reduction;

[0024] Using the formula The three-hop information of the splicing node matrix, the three-hop information are zero-hop, one-hop and two-hop information; among them, Represents the node matrix The first-order adjacency matrix of Represents the node matrix The second-order adjacency matrix of Represents a fully connected layer with a ReLU linear rectification activation function;

[0025] The three-hop information of all nodes constructs a fully connected graph, using the formula Obtain the updated three-hop feature representation of node i; where, represents the three-hop feature representation of node i that has not been updated, represents the dimension of the node feature vector, represents the learnable weight matrix, Represents the transpose of a matrix;

[0026] Using the formula ; Get the node feature matrix after attention aggregation; among them, Indicates the zeroth hop feature matrix of multi-hop information interaction, It represents the first-hop and second-hop feature matrices of multi-hop information interaction. represents the learnable weight matrix;

[0027] Furthermore, the protein language model is embedded in the graph panoramic attention perception enhancement module of the enhanced graph panoramic attention learning framework:

[0028] In ESM coding module 1, using the formula Get the Q, K, V matrices of the initial semantic embedding of ESM2; where E represents the semantic embedding of ESM2 after dimensionality reduction by the fully connected layer; represents the learnable weight matrix;

[0029] Using the formula Inductive structural bias and update ESM2 feature representation; represents the node structure feature matrix updated by the node-to-node module, express The dimensions of the eigenvectors in the matrix, Represents the transpose of a matrix;

[0030] Furthermore, the protein language model is embedded in the graph panoramic attention perception enhancement module of the enhanced graph panoramic attention learning framework:

[0031] In the node-to-edge module, using the formula Get the self-attention weight coefficient after the concatenation of node i and outgoing edge ij features; represents the feature concatenation operation, represents the feature representation of node i, Represents edge The characteristic representation of represents the learnable weight matrix;

[0032] Using the formula Get the updated edge Feature representation; represents the learnable weight matrix;

[0033] Using the formula The semantic information in the ESM2 embedding is weighted to guide the encoding of edge features in the graph structure; represents the feature representation of node i,j after being updated by ESM encoding module 1, represents the vector outer product, Indicates row-wise averaging;

[0034] Using the formula Get the edge updated by the node-to-edge module Feature representation, the feature matrix after all edges are updated is expressed as .

[0035] Furthermore, the protein language model is embedded in the graph panoramic attention perception enhancement module of the enhanced graph panoramic attention learning framework:

[0036] In ESM coding module 2, using the formula Get the ESM feature representation after the first update Q, K, V matrices; where represents the learnable weight matrix;

[0037] Using the formula Summarize the structural information of all incoming edges of the corresponding node; among them, Represents the edge after the node-to-edge module updates Feature representation, represents the set of all incoming edges of node i, Indicates row concatenation. represents the learnable weight matrix;

[0038] Using the formula The summarized edge structure information is fed back to the sequence encoding branch for attention update of the ESM2 feature representation to obtain the second updated ESM feature representation; express The dimensions of the eigenvectors in the matrix, Represents the transpose of a matrix.

[0039] Furthermore, the protein language model is embedded in the graph panoramic attention perception enhancement module of the enhanced graph panoramic attention learning framework:

[0040] In the edge-to-edge module, using the formula , , Get all incoming edges of node i The K, V matrix of an outgoing edge The Q matrix of Represents adjacent edges and The relative angle between The relative angle range is divided into 6 small intervals, corresponding to the 6-dimensional onehot encoding. represents the learnable weight matrix, Represents the edge after the node-to-edge module updates Feature representation;

[0041] Using the formula Planning ESM2 Embedded Virtual Edge features to construct , and The triangular perception formed; among them, and They respectively represent the feature representations of nodes x and j after being updated by ESM encoding module 2;

[0042] Using the formula Injecting triangle awareness into edges and Calculation of attention coefficient between; express The dimensions of the matrix eigenvectors;

[0043] Using the formula and Get the updated edge after aggregation Feature representation; among them, Represents the set of all incoming edges of a node. The updated feature matrix of all edges is expressed as .

[0044] Furthermore, the protein language model is embedded in the graph panoramic attention perception enhancement module of the enhanced graph panoramic attention learning framework:

[0045] In the edge-to-node module, use the formula Get the Q matrix of node j, the K matrix of node i, and the edge The V matrix of represents the feature representation of node j and node i after being updated by the node-to-node module; Represents the edge updated by the edge-to-edge module The feature representation of represents the learnable weight matrix;

[0046] Using the formula and Get the edge The importance of node j; express The dimensions of the eigenvectors in the matrix, Represents the set of all neighbor nodes of node i.

[0047] Using the formula The updated semantic features of ESM encoding module 3 To guide the convergence update of node j features; where, represents the set of all incoming edges of node j, and the updated feature matrix of all nodes is expressed as ;

[0048] In the bidirectional cross attention fusion module, using formula and Get the updated structural features after cross attention and sequence characteristics ;in, represents the node features updated by the edge-to-node module, represents the updated sequence semantic features after ESM encoding module 3, represents the learnable weight matrix, represents the transpose of a matrix, express The dimension of the eigenvector of the matrix after linear transformation;

[0049] Using the formula Enhance the nonlinear expression of sequence features. Similarly, use the formula Enhance the nonlinearity of structural representation; using the formula The final structure and sequence feature fusion representation is obtained; represents the learnable weight matrix; Represents the bias term.

[0050] Furthermore, in the prediction part of the protein language model embedding-enhanced graph panoramic attention learning framework, the loss formula is used Maximize the probability that the amino acid predicted by the masked mutation site is the wild-type amino acid type, and minimize the probability that the amino acid predicted by the mutation site is the pathogenic amino acid type; represents the probability that the amino acid type is predicted to be a wild-type amino acid, It indicates the probability that the amino acid at the same mutation site in the training set is predicted to be the pathogenic amino acid type. Indicates the number of mutation samples located at the same site in the same Metalpdb structure in the training set.

[0051] The beneficial effects of the present invention are:

[0052] 1. The present invention creatively constructs a panoramic attention perception enhancement module based on the protein language model and the panoramic attention mechanism, which fully exploits the rich semantic features contained in the protein language model embedding to enhance the learning of metal protein structures.

[0053] 2. In the present invention, multiple attention-driven units are combined to construct a panoramic message transmission path, so that the network can capture the complex structural patterns under different metal binding conformations. The framework seamlessly integrates the structural information at the residue and atomic levels, deeply integrates the structure and sequence representation, and the comprehensive information induction better solves the problem of the difficulty in characterizing the complex coordination patterns of metal ions and proteins in different metal coordination environments.

[0054] 3. The present invention maximizes the probability of predicting the mutation site as the wild-type amino acid type and minimizes the probability of predicting the site as the pathogenic amino acid type through the design of a new loss, thereby enhancing the network's ability to reason about the true amino acid identity of the mutation site, suppressing prediction results that are contrary to biological common sense, and better solving the problem of inaccurate prediction of the mutation effect of metal binding sites. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is the overall flow chart of the present invention; Figure 2 It is a schematic diagram of the structure of the protein language model embedding-enhanced graph panoramic attention learning framework in the present invention; Figure 3 A schematic diagram of a graph panoramic attention perception enhancement module based on a protein language model embedding enhanced graph panoramic attention learning framework in the present invention; Figure 4 A schematic diagram of a panoramic attention module of a panoramic attention perception enhancement module in a panoramic attention learning framework based on protein language model embedding enhancement in the present invention; Figure 5 It is a schematic diagram of a bidirectional cross-attention fusion module of a panoramic attention perception enhancement module in a panoramic attention learning framework based on protein language model embedding enhancement in the present invention; Figure 6 for Figure 2 A magnified view of the structural information after the mask; Figure 7 for Figure 2 A zoomed-in view of the unmasked structural information in . DETAILED DESCRIPTION

[0056] The following will be combined with the attached Figure 1-7 The embodiments of the present invention are described in further detail with reference to the accompanying drawings and specific examples.

[0057] A method for predicting the effects of metal binding site mutations based on graph panoramic attention enhancement, specifically a method for predicting the effects of metal binding site mutations based on protein language model-guided graph structure learning, such as Figure 1 As shown in the flowchart in, the overall process includes:

[0058] S1: Obtain metal binding site disease-related mutation data from Clinvar and Uniprot databases, specifically:

[0059] Download the variant summary.txt.gz file from the Clinvar database, download humsavar.txt from the Uniprot Humsavar database, obtain all the single missense variants therein, download all the XML files that record the correspondence between uniprot and pdb positions from the sifts website, and obtain the complete mapping file after processing. Use the mapping file to screen out samples with a determined metalpdb structure from the obtained disease-related mutation data as the final metal binding site disease-related mutation data set. All single missense variants obtain the corresponding uniprotID and pdbID through the ID mapping online conversion function of the uniprot website.

[0060] Then, the corresponding position of the mutation position on uniprot in pdb is found through the total mapping file. In this step, multiple pdb files may be found, and each pdb file may correspond to multiple metalpdb files. Finally, according to the pdb resolution (the lower the value, the higher the priority) and metalpdb coverage (the higher the value, the higher the priority), the metalpdb with the highest priority is found, that is, the local metal protein structure of the mutant sample is found.

[0061] S2: Divide the acquired disease-related mutation dataset on metal binding sites into a training sample set, a validation sample set, and a test sample set. Specifically:

[0062] After balancing the positive and negative samples of the metal binding site mutation data set obtained in step S1, it is randomly divided into a training sample set, a validation sample set, and a test sample set in a ratio of 0.8:0.1:0.1.

[0063] S3: Perform data preprocessing operations on the sample set divided in step S2. The data preprocessing operations include constructing a metal coordination network diagram, encoding biological feature information, and obtaining deep semantic embedding of a protein language model.

[0064] When constructing the metal coordination network diagram, the sample set divided in step S2 is constructed according to the composition of metal ions, ligands and ligand neighbors to construct an amino acid map of the metal binding site network, and an atomic map is constructed according to a threshold of 3 distances. Specifically, in the amino acid map, the metal ion is represented as a central node, and edges are added between the central node and all metal ligands, where the metal ligand is defined as a residue or cofactor containing at least one non-hydrogen atom within 3 Å of the metal ion. In addition, edges are added between the metal ligand and the ligand neighbor, and the ligand neighbor is defined as a residue or cofactor containing at least one non-hydrogen atom within 5 Å of the metal ligand. In the atomic map, all atoms are used as nodes, and two atoms within a distance of 3 Å are connected by edges to simulate the interaction between atoms.

[0065] When encoding biological feature information, biological feature information including residue depth, secondary structure, relative solvent accessible surface area, coordination bond, and hydrogen bond is encoded. Specifically, the node features of the amino acid graph include residue type, residue depth, secondary structure, relative solvent accessible surface area, whether it is a coordination or receptor of hydrogen bond, whether it is a ligand or receptor of coordination bond; edge features include the length and angle of the edge (composed of polar angle and azimuthal angle in spherical coordinates); node features of the atomic graph include atom type, number of covalent bonds formed by atoms, number of electronic charges assigned to atoms, chiral state, hybridization state, van der Waals radius, and atomic mass; the encoding of edge features and amino acid Figure 1 To.

[0066] When obtaining the deep semantic embedding of the protein language model, the embedding of the 33rd layer of the ESM2 protein language model esm2_t33_650M_UR50D() on the long sequence is obtained and mapped to the corresponding amino acids on the metalpdb structure to obtain the ESM2 semantic embedding. Specifically, first traverse the metalpdb to obtain several pdb chains contained therein, and then use the mapping file constructed in step S1 to find the corresponding complete uniprot single chain; according to the pre-trained esm2_t33_650M_UR50D() model weights provided by facebookresearch, obtain the semantic feature embedding of all unipro chains at the 33rd layer of the network, and finally map back to the amino acids on the incomplete pdb chain in metalpdb to obtain the ESM2 semantic embedding.

[0067] S4: Use the preprocessed training sample set and verification sample set obtained in step S3 to train and verify the neural network model, and save the neural network model framework after training. Input the preprocessed training sample set obtained in step S3 into the protein language model embedding enhanced graph panoramic attention learning framework for training. Use an early stopping strategy. When the model loss does not decrease after 30 consecutive iterations on the verification set, or the number of iterations reaches the set value of 300, stop the training process, and save the protein language model embedding enhanced graph panoramic attention learning framework when the mutation effect prediction performance on the verification sample set is the best.

[0068] The neural network framework in step S4 is a protein language model embedding enhanced graph panoramic attention learning framework, specifically:

[0069] In the atomic view attention aggregation module of the framework, the pre-processed atomic graph and amino acid graph are input to extract detailed structural features of the neighborhood around the metal binding site from both atomic and amino acid perspectives. The self-attention mechanism of atomic clusters can make up for the inconsistency of atomic and amino acid dimensions, thereby seamlessly integrating the feature embedding of the two.

[0070] In the graph panoramic attention perception enhancement module part of the framework, the network will weave four attention-driven modules, node-to-node, node-to-edge, edge-to-edge, and edge-to-node, to build a panoramic message transmission path to fully characterize the complex structure of the neighborhood around the metal binding site. During this period, deep semantic sequence information from the protein language model will be injected into each driving module to guide the learning of the graph structure. At the same time, the iteration of structural information will also be fed back to the ESM encoding module to continuously optimize the representation of sequence information. The two-way guided fusion mechanism can help the network better learn the graph structure and deeply fuse the sequence and structure representations to ensure that more comprehensive information is used for more accurate predictions.

[0071] In the self-distillation knowledge part of the framework, the masked graph structure features of the mutation site and the ESM2 embedding obtained from the amino acid sequence after the mutation site amino acid is replaced with "X" are input into the atomic view attention aggregation module and the graph panoramic attention perception enhancement module of the normal gradient feedback. In addition, an additional target feature generator is constructed, and its input is the complete graph structure features and the ESM2 embedding obtained from the normal sequence. After message passing, it will be aligned in the feature space in order to enable the encoder to obtain stronger information reasoning capabilities. The target feature generator updates its parameters through the exponential moving average (EMA).

[0072] In the mutation effect prediction part of the framework, the feature representation of the fusion of the sequence and structure obtained in the coding part is passed through a linear layer (D, 21) to obtain the predicted probability of the amino acid type before and after the mutation, and its maximum likelihood value is calculated. Finally, it passes through a linear layer and a sigmoid function to obtain the probability of the pathogenicity of the mutation.

[0073] See also Figure 2 In the atomic view attention aggregation module of the protein language model embedding enhanced graph panoramic attention learning framework, the preprocessed training sample set obtained in step S3 is first input into the atomic view attention aggregation module, and the formula is used To obtain the attention weights between adjacent atoms, represents the initial feature representation of atoms i, j, Represents atomic graph edges The initial feature representation of represents the dimension of atomic feature representation, represents the attention coefficient between atoms i and j, represents the matrix transpose, represents the learnable weight matrix.

[0074] For atom i, the feature representation after attention aggregation is expressed by the formula obtain; among them, Indicates the number of attention heads; represents the learnable weight matrix.

[0075] All atomic features are updated with express.

[0076] The characteristic matrix of amino acids is calculated using the formula Perform an initial update, where represents the initial amino acid biosignature encoding matrix; is a learnable weight matrix.

[0077] All atoms belonging to the same amino acid are related by the formula Get the self-attention weight coefficient of the atom; is a learnable weight matrix.

[0078] Using the formula Align the atomic scale with the amino acid scale to generate an atomic graph feature representation optimized by attention guidance;

[0079] Using the formula Integrate structural information at different scales of amino acids and atoms.

[0080] See also Figure 3The present invention is based on the protein language model embedding enhancement graph panoramic attention learning framework encoding part of the graph panoramic attention perception enhancement module, specifically: see Figure 4 (a) The node-to-node module in the graph panoptic attention-aware enhancement module based on protein language model embedding enhanced graph panoptic attention learning framework, using formula The node feature matrix preliminarily updated by the atomic view attention aggregation module and the ESM2 semantic embedding after dimensionality reduction are preliminarily fused to enhance the overall information representation. It means adding the elements at corresponding positions. Represents the node feature matrix that has been initially updated by the atomic view attention aggregation module; Represents the semantic embedding of the ESM2 protein language model after dimensionality reduction.

[0081] Using the formula The three-hop information of the splicing node matrix, the three-hop information are zero-hop, one-hop and two-hop information; among them, Represents the node matrix The first-order adjacency matrix of ; Represents the node matrix The second-order adjacency matrix of ; Represents a fully connected layer with a ReLU linear rectification activation function.

[0082] The three-hop information of all nodes constructs a fully connected graph, using the formula Obtain the updated three-hop feature representation of node i; where, represents the three-hop feature representation of node i that has not been updated, represents the dimension of the node feature vector, represents the learnable weight matrix, Represents the transpose of the matrix. By replacing the interaction between nodes with the interaction between multi-hop information, the defect of graph neural network that the node features are over-smoothed as the number of network layers increases is alleviated.

[0083] Using the formula ; Get the node feature matrix after attention aggregation; among them, Indicates the zeroth hop feature matrix of multi-hop information interaction, It represents the first-hop and second-hop feature matrices of multi-hop information interaction. represents a learnable weight matrix. By exploiting nonlinear interactions between multi-hop neighbor representations to drive information transfer between nodes, the network is able to acquire representative embeddings of complex patterns around metal ions and metal binding sites.

[0084] In ESM coding module 1, using the formula Get the Q, K, V matrices of the initial semantic embedding of ESM2; where E represents the semantic embedding of ESM2 after dimensionality reduction by the fully connected layer; represents the learnable weight matrix.

[0085] Using the formula Inductive structural bias and update ESM2 feature representation; represents the node structure feature matrix updated by the node-to-node module, express The dimensions of the eigenvectors in the matrix, Represents the transpose of a matrix.

[0086] See also Figure 4 (b) The node-to-edge module in the panoramic attention perception enhancement module uses the formula Get the self-attention weight coefficient after the concatenation of node i and outgoing edge ij features; represents the feature concatenation operation, represents the feature representation of node i, Represents edge The characteristic representation of represents the learnable weight matrix.

[0087] Using the formula Get the updated edge Feature representation; represents the learnable weight matrix.

[0088] Using the formula The rich semantic information in ESM2 embedding is weighted to guide the encoding of edge features in the optimized graph structure; represents the feature representation of node i,j after being updated by ESM encoding module 1, represents the vector outer product, Indicates row-wise averaging.

[0089] Using the formula Get the edge updated by the node-to-edge module Feature representation, this module effectively models the information interaction between nodes and edges without losing the inherent information of the edges. The updated feature matrix of all edges is represented as .

[0090] In ESM coding module 2, using the formula Get the ESM feature representation after the first update Q, K, V matrices; where represents the learnable weight matrix.

[0091] Using the formula Summarize the structural information of all incoming edges of the corresponding node; among them, Represents the edge after the node-to-edge module updates Feature representation, represents the set of all incoming edges of node i, Indicates row concatenation. represents the learnable weight matrix.

[0092] Using the formula The summarized edge structure information is fed back to the sequence encoding branch for attention update of the ESM2 feature representation to obtain the second updated ESM feature representation; express The dimensions of the eigenvectors in the matrix, Represents the transpose of a matrix.

[0093] See also Figure 4 (c) Edge-to-edge module in the panoramic attention-aware enhancement module. Using the formula , , Get all incoming edges of node i The K, V matrix of an outgoing edge The Q matrix of Represents adjacent edges and The relative angle between The relative angle range is divided into 6 small intervals, corresponding to the 6-dimensional onehot encoding. represents the learnable weight matrix, Represents the edge after the node-to-edge module updates Feature representation.

[0094] Using the formula Planning ESM2 Embedded Virtual Edge features to construct , and The triangular perception formed; among them, and They respectively represent the feature representations of nodes x and j after being updated by ESM encoding module 2.

[0095] Using the formula Injecting triangle-aware bias into edges and Calculation of attention coefficient between; express Dimensions of the matrix eigenvectors.

[0096] Using the formula and Get the updated edge after aggregation Feature representation; Represents the set of all incoming edges of a node. The updated feature matrix of all edges is expressed as This module performs message passing and updates from the perspective of edges, which helps to identify structural details that may be overlooked when learning graph representations only from the perspective of nodes.

[0097] See also Figure 4 (d) The edge-to-node module in the panoramic attention perception enhancement module. In the edge-to-node module, the formula is used Get the Q matrix of node j, the K matrix of node i, and the edge The V matrix of represents the feature representation of node j and node i after being updated by the node-to-node module; Represents the edge updated by the edge-to-edge module The feature representation of represents the learnable weight matrix.

[0098] Using the formula and Get the edge The importance of node j; express The dimensions of the eigenvectors in the matrix; Represents the set of all neighbor nodes of node i.

[0099] Using the formula The updated semantic features of ESM encoding module 3 To guide the convergence update of node j features; where, represents the set of all incoming edges of node j, and the updated feature matrix of all nodes is expressed as This module effectively aggregates key features from different edge structure representations to generate more robust graph embeddings.

[0100] See also Figure 5 , Figure 2. The bidirectional cross attention fusion module in the panoramic attention perception enhancement module is as follows: Using the formula and Get the updated structural features after cross attention and sequence characteristics ;in, represents the node features updated by the edge-to-node module, represents the updated sequence semantic features after ESM encoding module 3, represents the learnable weight matrix, represents the transpose of a matrix, express The dimension of the eigenvector of the matrix after linear transformation;

[0101] Using the formula Enhance the nonlinear expression of sequence features. Similarly, use the formula Enhance the nonlinearity of structural representation; using the formula The final structure and sequence feature fusion representation is obtained; represents the learnable weight matrix; Represents the bias term, which further enhances the nonlinearity of structural and sequence features and better fits the complex correlations between different types of features.

[0102] The self-distillation module in the framework of enhanced graph panoramic attention learning based on protein language model embedding is specifically as follows: the ESM2 embedding obtained from the masked graph structure features of the mutation site and the amino acid sequence where the amino acid at the mutation site is replaced by 'X' is input into the framework encoding block composed of the atomic view attention aggregation module and the graph panoramic attention perception enhancement module. The ESM2 embedding obtained from the unmasked graph structure features and the normal sequence is input into the target feature generator network, which shares the same architecture as the encoding block.

[0103] Using the formula The feature outputs of the encoding block and the target feature generator network are aligned in the feature space to enhance the network's information reasoning ability and its ability to learn about the global feature distribution; represents the feature output of the encoding block, represents the feature output of the target feature generator; N represents the batch size; represents the sensitivity hyperparameter. Using the formula To update the parameters of the target feature generator network; where, are the parameters of the frame encoder atomic view attention aggregation module and the guide graph panoramic attention learning module; is the weight decay coefficient, and the weight of the encoding block is updated by the normal gradient feedback.

[0104] In the mutation effect prediction part of the protein language model embedding enhanced graph panoramic attention learning framework, the sequence and structure fusion features obtained in the encoding block are input into the (d, 21) dimension fully connected layer for subsequent prediction. The prediction process is as follows:

[0105] Using the formula The probability of the amino acid type at mutation position i being predicted as the wild-type amino acid and the mutated amino acid is obtained, and the log-likelihood ratio between the two is used for variant effect prediction, where, represents the probability that the amino acid type at position i is predicted to be a wild-type amino acid, Represents the probability that the amino acid type at position i is predicted to be the amino acid type a after mutation.

[0106] The log-likelihood ratio score calculated for A fully connected layer and a sigmoid activation layer are used to obtain the probability that the amino acid at position i will cause the disease after mutating from the wild-type amino acid to the amino acid a.

[0107] Utilizing the loss formula Maximize the probability that the amino acid type predicted by the masked mutation site is the wild-type amino acid, and minimize the probability that the amino acid type predicted by the mutation site is the pathogenic amino acid type, where represents the probability that the amino acid type is predicted to be a wild-type amino acid, It indicates the probability that the amino acid at the same mutation site in the training set is predicted to be the pathogenic amino acid type. It represents the number of mutation samples at the same site in the same metalpdb structure in the training set. This is consistent with the evolutionary conservation of amino acid sequences. Long-term natural selection determines that protein sequence distribution has a certain statistical law. By designing this loss function, the network can be optimized in a direction that is more in line with the biological background setting.

[0108] Get the weighted binary cross entropy loss of the predicted probability value , combined with the weighted scaled cosine error loss and weighting for design Loss, get the overall loss value of the neural network model, and optimize the model parameters through the adaptive gradient method (AdamW optimizer) to make the loss continuously approach the minimum value to train the neural network model.

[0109] The preprocessed verification sample set obtained in step S3 is input into the above-mentioned trained protein language model embedding-enhanced graph panoramic attention learning framework, and the classification performance of the verification sample set is calculated using the commonly used classification indicators AUROC, AUPRC, ACC, F1-score, MCC, PPV, and Recall, and the relevant information of the parameters of the protein language model embedding-enhanced graph panoramic attention learning framework when the classification performance on the verification sample set is the best is saved.

[0110] S5: Input the preprocessed test sample set obtained in step S3 into the neural network model trained in step S4 to obtain the prediction result of the metal binding site mutation effect.

[0111] The above description is only used to illustrate the design ideas and implementation plans of the present invention, rather than to limit them. Those skilled in the art should understand that other solutions that modify or equivalently replace the technical solutions of the present invention are still included in the scope defined by the claims of this application.

Claims

1. A method for predicting the effect of metal binding site mutations based on enhanced image panorama attention, characterized in that: The following steps are involved: S1: Obtain disease-related mutation data of metal binding sites from the database; S2: Divide the acquired disease-related mutation dataset on metal binding sites into a training sample set, a validation sample set, and a test sample set; S3: performing data preprocessing operations on the sample set divided in step S2; the data preprocessing operations include constructing a metal coordination network diagram, encoding biological feature information, and obtaining deep semantic embedding of a protein language model; S4: Use the preprocessed training sample set and verification sample set obtained in step S3 to train and verify the neural network framework, and save the trained neural network framework; S5: Input the preprocessed test sample set obtained in step S3 into the neural network framework trained in step S4 to obtain the prediction result of the metal binding site mutation effect.

2. The method for predicting the effect of metal binding site mutations based on enhanced image panorama attention according to claim 1, characterized in that: In step S3, when constructing the metal coordination network diagram, the sample set divided in step S2 is constructed according to the composition mode of metal ions, ligands and ligand neighbors to construct an amino acid diagram of the metal binding site network, and an atomic diagram is constructed according to a threshold value of 3; In step S3, when encoding the biometric information, the biometric information including residue depth, secondary structure, relative solvent accessible surface area, coordination bonds, and hydrogen bonds is encoded; In step S3, when obtaining the deep semantic embedding of the protein language model, the embedding of the 33rd layer of the ESM2 protein language model esm2_t33_650M_UR50D() on the long sequence is obtained and mapped to the corresponding amino acids on the metalpdb structure to obtain the ESM2 semantic embedding.

3. The method for predicting the effect of metal binding site mutations based on enhanced image panorama attention according to claim 1, characterized in that: The neural network framework in step S4 is based on the protein language model embedding enhanced graph panoramic attention learning framework, specifically: In the atomic view attention aggregation module of the framework, the pre-processed atomic graph and amino acid graph are input to extract the structural features of the neighborhood around the metal binding site from both atomic and amino acid perspectives, and the feature embedding of atoms and amino acids is integrated through the self-attention mechanism of atomic clusters. In the graph-wide attention perception enhancement module of the framework, the network weaves four attention-driven modules, namely node-to-node, node-to-edge, edge-to-edge, and edge-to-node, to build a panoramic message delivery path. At the same time, each driver module injects deep semantic information from the protein language model to enhance the learning of graph structure, and iteratively feeds back structural information to the ESM encoding module to continuously optimize the representation of sequence information. In the self-distillation knowledge part of the framework, the ESM2 embedding obtained from the masked graph structure features of the mutation site and the amino acid sequence after the amino acid at the mutation site is replaced with "X" is input into the atomic view attention aggregation module and the graph panoramic attention perception enhancement module of the normal gradient feedback; and a target feature generator is constructed, and the input of the target feature generator is the graph structure features and the ESM2 embedding obtained from the normal sequence; The target feature generator updates parameters through exponential moving average; In the mutation effect prediction part of the framework, the feature representation of the fusion of the sequence and structure obtained in the coding part is passed through a linear layer (D, 21) to obtain the predicted probability of the amino acid type before and after the mutation, and its maximum likelihood value is calculated. Finally, it passes through a linear layer and a sigmoid function to obtain the probability of the pathogenicity of the mutation.

4. The method for predicting metal binding site mutation effects based on enhanced image panorama attention according to claim 3, characterized in that: In the graph panoramic attention perception enhancement module of the enhanced graph panoramic attention learning framework based on protein language model embedding: In the node-to-node module, use the formula The node feature matrix updated by the atomic view attention aggregation module and the ESM2 semantic embedding after dimensionality reduction are used for preliminary feature fusion; It means adding the elements at corresponding positions. Represents the node feature matrix that has been initially updated by the atomic view attention aggregation module; represents the semantic embedding of the ESM2 protein language model after dimensionality reduction; Using the formula The three-hop information of the splicing node matrix, the three-hop information are zero-hop, one-hop and two-hop information; among them, Represents the node matrix The first-order adjacency matrix of Represents the node matrix The second-order adjacency matrix of Represents a fully connected layer with a ReLU linear rectification activation function; The three-hop information of all nodes constructs a fully connected graph, using the formula Obtain the updated three-hop feature representation of node i; where, represents the three-hop feature representation of node i that has not been updated, represents the dimension of the node feature vector, represents the learnable weight matrix, Represents the transpose of a matrix; Using the formula ; Get the node feature matrix after attention aggregation; among them, Indicates the zeroth hop feature matrix of multi-hop information interaction, It represents the first-hop and second-hop feature matrices of multi-hop information interaction. represents the learnable weight matrix.

5. The method for predicting the effect of metal binding site mutations based on enhanced image panorama attention according to claim 4, characterized in that: In the graph panoramic attention perception enhancement module of the enhanced graph panoramic attention learning framework based on protein language model embedding: In ESM coding module 1, using the formula Get the Q, K, V matrices of the initial semantic embedding of ESM2; where E represents the semantic embedding of ESM2 after dimensionality reduction by the fully connected layer; represents the learnable weight matrix; Using the formula Inductive structural bias and update ESM2 feature representation; represents the node structure feature matrix updated by the node-to-node module, express The dimensions of the eigenvectors in the matrix, Represents the transpose of a matrix.

6. The method for predicting the effect of metal binding site mutations based on enhanced image panorama attention according to claim 5, characterized in that: In the graph panoramic attention perception enhancement module of the enhanced graph panoramic attention learning framework based on protein language model embedding: In the node-to-edge module, using the formula Get the self-attention weight coefficient after the concatenation of node i and outgoing edge ij features; represents the feature concatenation operation, represents the feature representation of node i, Represents edge The characteristic representation of represents the learnable weight matrix; Using the formula Get the updated edge Feature representation; represents the learnable weight matrix; Using the formula The semantic information in the ESM2 embedding is weighted to guide the encoding of edge features in the graph structure; represents the feature representation of node i,j after being updated by ESM encoding module 1, represents the vector outer product, Indicates row-wise averaging; Using the formula Get the edge updated by the node-to-edge module Feature representation, the feature matrix after all edges are updated is expressed as .

7. The method for predicting the effect of metal binding site mutations based on enhanced image panorama attention according to claim 6, characterized in that: In the graph panoramic attention perception enhancement module of the enhanced graph panoramic attention learning framework based on protein language model embedding: In ESM coding module 2, using the formula Get the ESM feature representation after the first update Q, K, V matrices; where represents the learnable weight matrix; Using the formula Summarize the structural information of all incoming edges of the corresponding node; among them, Represents the edge after the node-to-edge module updates Feature representation, represents the set of all incoming edges of node i, Indicates row concatenation. represents the learnable weight matrix; Using the formula The summarized edge structure information is fed back to the sequence encoding branch for attention update of the ESM2 feature representation to obtain the second updated ESM feature representation; express The dimensions of the eigenvectors in the matrix, Represents the transpose of a matrix.

8. The method for predicting the effect of metal binding site mutations based on enhanced image panorama attention according to claim 7, characterized in that: In the graph panoramic attention perception enhancement module of the enhanced graph panoramic attention learning framework based on protein language model embedding: In the edge-to-edge module, Using the formula , , Get all incoming edges of node i The K, V matrix of an outgoing edge The Q matrix of Represents adjacent edges and The relative angle between The relative angle range is divided into 6 small intervals, corresponding to the 6-dimensional onehot encoding. represents the learnable weight matrix, Represents the edge after the node-to-edge module updates Feature representation; Using the formula Planning ESM2 Embedded Virtual Edge features to construct , and The triangular perception formed; among them, and They respectively represent the feature representations of nodes x and j after being updated by ESM encoding module 2; Using the formula Injecting triangle-aware bias into edges and Calculation of attention coefficient between; express The dimensions of the matrix eigenvectors; Using the formula and Get the updated edge after aggregation Feature representation; among them, Represents the set of all incoming edges of a node. The updated feature matrix of all edges is expressed as .

9. The method for predicting the effect of metal binding site mutations based on enhanced image panorama attention according to claim 8, characterized in that: In the graph panoramic attention perception enhancement module of the enhanced graph panoramic attention learning framework based on protein language model embedding: In the edge-to-node module, use the formula Get the Q matrix of node j, the K matrix of node i, and the edge The V matrix of represents the feature representation of node j and node i after being updated by the node-to-node module; Represents the edge updated by the edge-to-edge module The feature representation of represents the learnable weight matrix; Using the formula and Get the edge The importance of node j; express The dimensions of the eigenvectors in the matrix, Represents the set of all neighbor nodes of node i; Using the formula The updated semantic features of ESM encoding module 3 To guide the convergence update of node j features; where, represents the set of all incoming edges of node j, and the updated feature matrix of all nodes is expressed as ; In the bidirectional cross attention fusion module, using formula and Get the updated structural features after cross attention and sequence characteristics ;in, represents the node features updated by the edge-to-node module, represents the updated sequence semantic features after ESM encoding module 3, represents the learnable weight matrix, represents the transpose of a matrix, express The dimension of the eigenvector of the matrix after linear transformation; Using the formula Enhance the nonlinear expression of sequence features. Similarly, Using the formula Enhance the nonlinearity of structural representation; using the formula The final structure and sequence feature fusion representation is obtained; represents the learnable weight matrix; Represents the bias term.

10. The method for predicting metal binding site mutation effects based on enhanced image panorama attention according to claim 3, characterized in that: In the prediction part of the protein language model embedding-enhanced graph panoramic attention learning framework, the loss formula is used Maximize the probability that the amino acid predicted by the masked mutation site is the wild-type amino acid type, and minimize the probability that the amino acid predicted by the mutation site is the pathogenic amino acid type; represents the probability that the amino acid type is predicted to be a wild-type amino acid, It indicates the probability that the amino acid at the same mutation site in the training set is predicted to be the pathogenic amino acid type. Indicates the number of mutation samples located at the same site in the same Metalpdb structure in the training set.

Images