Protein mutation prediction method and device

Abstract

The invention provides a protein mutation prediction method and device, and belongs to the crossing field of biology and artificial intelligence. The method comprises the following steps: acquiring amino acid sequence information and three-dimensional structure information of a to-be-predicted protein; constructing a first multi-modal feature map based on the amino acid sequence information and the three-dimensional structure information; the first multi-modal feature graph contains node features and edge features and does not contain explicit evolutionary energy features; inputting the first multi-modal feature map into a pre-trained prediction model to obtain a predicted mutation result; the prediction model is obtained by training the second multi-modal feature map; the second multi-modal feature map at least comprises amino acid sequence information of the training protein data set, three-dimensional structure information and an evolutionary energy matrix which is obtained based on multi-sequence alignment MSA calculation and through feature alignment processing. In this way, the accuracy and efficiency of protein mutation prediction can be improved, and the method can adapt to various application scenes.

Description

Technical Field

[0001] This disclosure relates to the intersection of biology and artificial intelligence, and in particular to a method and apparatus for predicting protein mutations. Background Technology

[0002] Currently, most existing protein prediction methods are either single sequence / structure-based models or simple combinations of sequence and structure models. They either rely solely on protein sequence information for modeling, ignoring the physical meaning supported by three-dimensional structure and evolutionary constraints, and failing to capture the profound impact of mutations on protein spatial conformation and function; or they are based solely on structural feature analysis, lacking evolutionary conservation information at the sequence level, resulting in insufficient model predictive ability. Even when using a sequence-structure combination model, it is often a simple feature splicing, failing to address the semantic misalignment problem of multimodal features, and failing to explore the intrinsic connections between sequence, structure, and evolutionary information. Furthermore, their statistical extrapolation costs are high, and they rely on a large number of labeled samples for model training, making efficient prediction difficult in low-quality data scenarios with scarce labeled samples and few homologous sequences. Thus, existing protein mutation prediction methods have low accuracy and efficiency, making them unsuitable for various application scenarios. Summary of the Invention

[0003] This disclosure provides a protein mutation prediction method and apparatus, which to some extent solves the problem that existing protein mutation prediction methods have low accuracy and efficiency and are difficult to adapt to various application scenarios.

[0004] According to one aspect of this disclosure, a method for predicting protein mutations is provided. The method includes: acquiring amino acid sequence information of the protein to be predicted; the amino acid sequence information includes the amino acid sequence order, amino acid type, and amino acid sequence length; acquiring three-dimensional structural information of the protein to be predicted; the three-dimensional structural information includes the spatial coordinates of the amino acids, the dihedral angles of the amino acids, and the relative spatial distances between the amino acids; constructing a first multimodal feature map based on the amino acid sequence information and the three-dimensional structural information; the first multimodal feature map includes node features and edge features, but does not include explicit evolutionary energy features; inputting the first multimodal feature map into a pre-trained prediction model to obtain the predicted mutation result; the prediction model is a multimodal fusion prediction model determined based on a graph neural network, and is obtained by training using a second multimodal feature map; the second multimodal feature map includes at least the amino acid sequence information, three-dimensional structural information, and an evolutionary energy matrix calculated based on multiple sequence alignment (MSA) and processed by feature alignment of the training protein dataset; the evolutionary energy matrix is ​​used as a regularization constraint in the training of the prediction model.

[0005] Furthermore, according to one aspect of the method disclosed herein, the predicted mutation results include: the mutation type of the amino acid and the mutation score of the amino acid.

[0006] Furthermore, according to one aspect of the method of this disclosure, obtaining the amino acid sequence information of a protein to be predicted includes: obtaining an identifier of the protein to be predicted; using the identifier, retrieving raw amino acid sequence data from a protein sequence database; preprocessing the raw amino acid sequence data and extracting the amino acid sequence order, amino acid type, and amino acid sequence length to obtain amino acid sequence information; the preprocessing includes at least one of the following: removing redundant sequence fragments, removing unknown amino acid residues, or invalid whitespace characters.

[0007] Furthermore, according to one aspect of the method disclosed herein, obtaining the three-dimensional structural information of a protein to be predicted includes: extracting the spatial coordinates of the backbone atoms of any amino acid of the protein to be predicted, and using the average spatial coordinates of the backbone atoms as the spatial coordinates of the amino acid; determining the torsion angle between adjacent peptide bonds of the amino acid based on the spatial coordinates of the amino acid to obtain the dihedral angle of the amino acid; calculating the Euclidean distance between the spatial coordinates of any two amino acids to obtain the relative spatial distance of the amino acids; and determining the spatial coordinates of the amino acid, the dihedral angle of the amino acid, and the relative spatial distance of the amino acid as the three-dimensional structural information.

[0008] Furthermore, according to one aspect of the method disclosed herein, a first multimodal feature map is constructed based on amino acid sequence information and three-dimensional structural information, including: performing feature encoding and dimension mapping processing on both the amino acid sequence information and the three-dimensional structural information to obtain initial node feature vectors and initial edge feature vectors of amino acids; constructing inference node features; the inference node features are obtained by concatenating the initial node feature vectors with a first zero vector in terms of dimension; wherein the dimension of the first zero vector is consistent with the dimension of the single-point potential energy vector in the evolutionary energy feature, and is used to shield the single-point energy feature; constructing inference edge features; the inference edge features are obtained by determining the nearest neighbor connection relationship based on the relative spatial distance of amino acids, and concatenating the geometric feature vector extracted based on the nearest neighbor connection relationship with a second zero vector in terms of dimension; wherein the dimension of the second zero vector is consistent with the dimension of the pairwise coupling potential energy matrix in the evolutionary energy feature, and is used to shield the coupling energy feature; and generating the first multimodal feature map based on the inference node features and the inference edge features.

[0009] Furthermore, according to one aspect of the method disclosed herein, the prediction model includes: a sequence encoding module for learning long-range dependencies between residues based on amino acid sequence information from a training protein dataset, thereby obtaining a sequence semantic embedding vector; a structure encoding module for extracting geometric features based on three-dimensional structural information from the training protein dataset, and learning a joint representation of geometric structure and evolutionary energy using the evolutionary energy matrix from the training protein dataset for normalization constraints, thereby obtaining a structure embedding vector; a multimodal fusion module for concatenating and fusing the sequence semantic embedding vector, the structure embedding vector, and the position-specific score matrix, thereby obtaining a fused embedding vector; and a predicted mutation decoding module for decoding the fused embedding vector and outputting the amino acid probability distribution at each residue position, thereby obtaining the predicted mutation result.

[0010] Furthermore, according to one aspect of the method of this disclosure, the evolutionary energy matrix includes: calculating a position-specific score matrix based on the MSA of the training protein dataset; calculating a single-point potential vector and a pairwise coupling potential matrix using a statistical physical model based on the MSA of the training protein dataset; obtaining the first vocabulary order output by the statistical model and the second vocabulary order of the multimodal fusion prediction model; and performing dimensional rearrangement of the single-point potential vector and the pairwise coupling potential matrix based on the mapping relationship between the first and second vocabulary orders and the position-specific score matrix to determine the semantically aligned evolutionary energy matrix.

[0011] Furthermore, according to one aspect of the method of this disclosure, the second multimodal feature map includes: acquiring amino acid sequence information, three-dimensional structure information, and evolutionary energy matrix of a training protein dataset; encoding the amino acid sequence information and three-dimensional structure information to obtain initial node feature vectors and initial edge feature vectors; constructing training node features; the training node features are obtained by dimensionally concatenating the initial node feature vectors with the single-point potential vectors in the evolutionary energy matrix; constructing training edge features; the training edge features are obtained by dimensionally concatenating the initial edge feature vectors with the paired coupling potential matrices in the evolutionary energy matrix; and determining the second multimodal feature map based on the training node features and training edge features.

[0012] Furthermore, according to one aspect of the method of this disclosure, the multimodal fusion prediction model further includes: an adaptive gating module for dynamically adjusting the weight ratio of evolutionary energy matrix features in the structure encoding module based on the effective sequence depth of the input MSA.

[0013] According to another aspect of this disclosure, a protein mutation prediction device is provided, comprising: a first acquisition unit for acquiring amino acid sequence information of the protein to be predicted; the amino acid sequence information includes the amino acid sequence order, amino acid type, and amino acid sequence length; a second acquisition unit for acquiring three-dimensional structural information of the protein to be predicted; the three-dimensional structural information includes the spatial coordinates of the amino acids, the dihedral angles of the amino acids, and the relative spatial distances between the amino acids; a construction unit for constructing a first multimodal feature map based on the amino acid sequence information and the three-dimensional structural information; the first multimodal feature map includes node features and edge features, but does not include explicit evolutionary energy features; and a prediction unit for inputting the first multimodal feature map into a pre-trained prediction model to obtain a predicted mutation result; the prediction model is a multimodal fusion prediction model determined based on a graph neural network, and is trained using a second multimodal feature map; the second multimodal feature map includes at least the amino acid sequence information, three-dimensional structural information, and an evolutionary energy matrix calculated based on multiple sequence alignment (MSA) and processed by feature alignment of the training protein dataset; the evolutionary energy matrix is ​​used as a regularization constraint in the training of the prediction model.

[0014] This disclosure provides a method and apparatus for predicting protein mutations. The method involves: acquiring the amino acid sequence information of the protein to be predicted; the amino acid sequence information includes the amino acid order, amino acid type, and amino acid sequence length; acquiring the three-dimensional structural information of the protein to be predicted; the three-dimensional structural information includes the spatial coordinates of the amino acids, the dihedral angles of the amino acids, and the relative spatial distances between the amino acids; constructing a first multimodal feature map for inference based on the amino acid sequence information and the three-dimensional structural information; inputting the first multimodal feature map into a pre-trained multimodal fusion prediction model to obtain the predicted mutation result; the multimodal fusion prediction model is trained and determined based on a second multimodal feature map; the second multimodal feature map, in addition to containing the amino acid sequence information and three-dimensional structural information based on the training protein dataset, also contains an evolutionary energy matrix calculated based on multiple sequence alignment (MSA) and processed by feature semantic alignment; the evolutionary energy matrix serves as a regularization constraint in model training. In this way, compared with existing single sequence / structure models or simple combination models, this disclosure can construct a multimodal fusion architecture through graph neural networks. This not only achieves deep fusion of sequence and structure but also introduces an evolutionary energy matrix based on statistical physics, compensating for the lack of explicit physical / evolutionary constraints in deep learning models. This fundamentally improves the model's ability to capture the correlation between mutations and protein conformation and function. Simultaneously, by adopting an asymmetric design that integrates energy features during training and only requires sequence and structural features during inference, the computationally time-consuming evolutionary energy features are shielded during inference, solving the problems of high inference cost and slow speed in traditional statistical physics models. Furthermore, through energy constraint guidance during training and the complementarity of multimodal information, the model reduces its dependence on high-quality multi-sequence alignment data during inference. Even in scenarios where homologous sequences are scarce or data quality is low, the model can still maintain high prediction accuracy. In summary, the technical solution provided by this disclosure can improve the accuracy and efficiency of protein mutation prediction and can be adapted to various application scenarios.

[0015] It should be understood that both the foregoing general description and the following detailed description are exemplary and intended to provide further illustration of the claimed technology. Attached Figure Description

[0016] The above and other objects, features, and advantages of this disclosure will become more apparent from the more detailed description of the embodiments thereof in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this disclosure and form part of the specification. They are used together with the embodiments of this disclosure to explain the disclosure and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0017] Figure 1 A schematic flowchart illustrating a protein mutation prediction method provided in this embodiment of the disclosure; Figure 2 A schematic flowchart illustrating a protein mutation prediction method provided in this embodiment of the disclosure; Figure 3 A schematic diagram of another protein mutation prediction process provided in this disclosure embodiment; Figure 4 This is a structural block diagram of a protein mutation prediction device provided in an embodiment of the present disclosure. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this disclosure more apparent, exemplary embodiments according to this disclosure will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments of this disclosure. It should be understood that this disclosure is not limited to the exemplary embodiments described herein.

[0019] Currently, most existing protein prediction methods are either single sequence / structure-based models or simple combinations of sequence and structure models. They either rely solely on protein sequence information for modeling, ignoring the physical meaning supported by three-dimensional structure and evolutionary constraints, and failing to capture the profound impact of mutations on protein spatial conformation and function; or they are based solely on structural feature analysis, lacking evolutionary conservation information at the sequence level, resulting in insufficient model predictive ability. Even when using a sequence-structure combination model, it is often a simple feature splicing, failing to address the semantic misalignment problem of multimodal features, and failing to explore the intrinsic connections between sequence, structure, and evolutionary information. Furthermore, their statistical extrapolation costs are high, and they rely on a large number of labeled samples for model training, making efficient prediction difficult in low-quality data scenarios with scarce labeled samples and few homologous sequences. Thus, existing protein mutation prediction methods have low accuracy and efficiency, making them unsuitable for various application scenarios.

[0020] Therefore, to address the aforementioned problems, this disclosure provides a protein mutation prediction method. Compared to existing single-sequence / structure models or simple combination models, it constructs a multimodal fusion architecture through graph neural networks. This not only achieves deep fusion of sequence and structure but also introduces an evolutionary energy matrix based on statistical physics, compensating for the lack of explicit physical / evolutionary constraints in deep learning models. This fundamentally improves the model's ability to capture the correlation between mutations and protein conformation and function. Simultaneously, it employs an asymmetric design that integrates energy features during training and requires only sequence and structural features during inference. This shields computationally time-consuming evolutionary energy features during inference, solving the problems of high inference cost and slow speed in traditional statistical physics models. Furthermore, through energy constraint guidance during training and multimodal information complementarity, the model reduces its dependence on high-quality multi-sequence alignment data during inference. Even in scenarios where homologous sequences are scarce or data quality is low, the model can still maintain high prediction accuracy. In summary, the technical solution provided by this disclosure can improve the accuracy and efficiency of protein mutation prediction and can be adapted to various application scenarios.

[0021] First, please refer to Figure 1 , Figure 1 This is a schematic flowchart illustrating a protein mutation prediction method provided in an embodiment of this disclosure. Figure 1 As shown, the method includes: In step S101, the amino acid sequence information of the protein to be predicted is obtained; the amino acid sequence information includes the amino acid sequence order, the type of amino acid, and the sequence length of the amino acid. In step S102, the three-dimensional structural information of the protein to be predicted is obtained; the three-dimensional structural information includes the spatial coordinates of amino acids, the dihedral angles of amino acids, and the relative spatial distances between amino acids. In step S103, a first multimodal feature map is constructed based on amino acid sequence information and three-dimensional structure information; the first multimodal feature map includes node features and edge features, but does not include explicit evolutionary energy features; In step S104, the first multimodal feature map is input into the pre-trained prediction model to obtain the predicted mutation result; the prediction model is a multimodal fusion prediction model determined based on a graph neural network, and is obtained by training using the second multimodal feature map; the second multimodal feature map contains at least the amino acid sequence information, three-dimensional structure information, and evolutionary energy matrix calculated based on multiple sequence alignment (MSA) and processed by feature alignment of the training protein dataset; the evolutionary energy matrix is ​​used as a regularization constraint in the training of the prediction model.

[0022] In this disclosure, amino acid sequence information can be understood as the sequence-related characterization information of amino acid residues in the protein to be predicted. Specifically, the amino acid sequence can be understood as the arrangement of amino acid residues in the protein according to the order in which they are linked by peptide bonds; this order is the core basis for determining the three-dimensional structure and function of the protein. The amino acid type can be understood as the specific category of the 20 naturally occurring amino acids contained in the protein to be predicted. The amino acid sequence length can be understood as the total number of amino acid residues in the protein to be predicted; it is a parameter characterizing the size of the protein molecule and defining the dimensionality of sequence features.

[0023] In this disclosure, three-dimensional structural information can be understood as the spatial correlation characterization information of amino acid residues and backbone atoms in the protein to be predicted. Specifically, the spatial coordinates of an amino acid can be understood as the specific position coordinates of key backbone atoms (nitrogen atom N, alpha carbon atom CA, carbon atom C, oxygen atom O) of each amino acid residue in the protein to be predicted in a three-dimensional Cartesian coordinate system. The dihedral angle of an amino acid can be understood as the torsion angle formed by the backbone atoms of adjacent amino acid residues in the peptide chain of the protein to be predicted, and is an indicator reflecting the folding state of the protein peptide chain. The relative spatial distance of amino acids can be understood as the linear spatial distance between any two key backbone atoms (preferably CA atoms) of any two amino acid residues in the protein to be predicted, and is used to measure the spatial proximity and interaction strength between residues.

[0024] In this disclosure, the multimodal fusion prediction model can be understood as a mutation prediction model built on graph neural networks. This model has the ability to deeply fuse sequence features and structural features, and has the characteristics of "enhancement during training and pruning during inference". That is, during the training phase, physical constraint rules are internalized by ingesting evolutionary energy features, while during the inference phase, efficient and accurate mutation effect prediction can be achieved based on implicit weights without inputting explicit evolutionary features. The core is used to receive the first multimodal feature map and output the mutation effect prediction result.

[0025] In this disclosure, the second multimodal feature map can be understood as a graph structure representation containing full information constructed for the proteins in the training set. In addition to basic sequence and structural information, the evolutionary energy matrix is ​​explicitly injected into the node and edge features. This feature map is specifically used during the model training phase, acting as a regularization constraint to train the multimodal fusion prediction model, forcing the model to learn the deep intrinsic relationships between sequence, structure, and evolutionary energy.

[0026] In this disclosure, the evolutionary energy matrix can be understood as an energy characterization matrix inferred from the multiple sequence alignment results of the training protein dataset using a statistical physical model (such as the Potts model). This matrix contains a single-point potential vector characterizing single-point amino acid preferences and a pairwise coupling potential matrix characterizing the strength of co-evolutionary coupling between residue pairs. It is the core physical feature for quantifying the strength of protein evolutionary constraints. It can be determined based on multiple sequence alignment (MSA). MSA can be understood as the sequence analysis result of aligning and arranging the amino acid sequences of homologous proteins in the training protein dataset, identifying conserved and variable sites, and mining co-evolutionary relationships of residues. It is the fundamental data for inferring the evolutionary energy matrix and extracting evolutionary constraint information. The specific determination of the evolutionary energy matrix is ​​detailed below.

[0027] Specifically, protein mutation prediction can include the following steps: First, the complete amino acid sequence and high-precision three-dimensional structural information of the protein to be predicted are accurately obtained, ensuring that there are no missing sequence features and no deviations in structural features, serving as the basic input for subsequent inference. Then, based on the sequence and structural information, a first multimodal feature map, i.e., the feature map used for inference, is constructed. During this process, zero vectors are filled into the input channels originally used to receive evolutionary energy features in the model, thereby achieving efficient feature integration without calculating Potts model features. Subsequently, the first multimodal feature map is input into a pre-trained graph neural network prediction model. Since this model has already undergone enhanced training using the second multimodal feature map during the training phase, its weight parameters have fully internalized the physical constraints of evolutionary energy. Therefore, even if explicit energy input is missing during the inference phase, the model can still utilize its learned robust representation capabilities to quickly complete feature extraction and logical inference. Finally, the model outputs the predicted mutational effect of the protein to be predicted, thereby achieving low-computational-cost, high-precision protein function prediction.

[0028] The following details the predicted mutation results of this disclosure, including: The predicted mutation results include: the mutation type of the amino acid and the mutation score of the amino acid.

[0029] In this disclosure, the mutation type of an amino acid can be understood as the variant form of a specific amino acid residue in the protein to be predicted, that is, the wild-type amino acid is replaced by one of the other 20 natural amino acids (e.g., if the wild-type residue is alanine (A) and the mutant is valine (V), then the mutation type is denoted as A→V), which can characterize the specific change of the mutation site.

[0030] Specifically, the 20 (natural) amino acids are as follows: glycine (G), alanine (A), valine (V), leucine (L), isoleucine (I), methionine (M), proline (P), tryptophan (W), serine (S), tyrosine (Y), cysteine ​​(C), phenylalanine (F), asparagine (N), glutamine (Q), threonine (T), aspartic acid (D), glutamic acid (E), lysine (K), arginine (R), and histidine (H).

[0031] In this disclosure, the mutation score of an amino acid can be understood as a numerical indicator that quantifies the degree of influence of a mutation on protein function, stability, or conformation. The positive or negative value or magnitude of the score corresponds to the harmfulness / benefit of the mutation and the intensity of its influence, and is the core output result for predicting mutation effects.

[0032] For example, this disclosure provides a specific formula for calculating the amino acid mutation score, which can satisfy the following:

[0033] in, This represents the set of mutations to be predicted, that is, the set of all mutation types in the target protein that need to be evaluated. This represents the probability distribution of amino acid residues output by the multimodal fusion prediction model in a specific sequence context, reflecting the likelihood of a certain type of amino acid appearing at that position. Indicates the first The position of an amino acid residue (i.e., the site where the mutation occurs). Indicates the first Wild-type amino acids at each position (i.e., the original amino acids without mutations). Represents the set excluding mutations Besides the corresponding site, the amino acid residues at other positions are replaced by masked markers (i.e., sequence context information). Indicates the first The mutant amino acid at each position (i.e., the target amino acid after replacement).

[0034] The following will explain in detail how to obtain amino acid sequence information, including: Obtain the label of the protein to be predicted; Using identifiers, raw amino acid sequence data can be retrieved from protein sequence databases; The raw amino acid sequence data is preprocessed to extract the amino acid sequence order, amino acid type, and amino acid sequence length to obtain amino acid sequence information. The preprocessing includes at least one of the following: removing redundant sequence fragments, removing unknown amino acid residues, or invalid whitespace characters.

[0035] In this disclosure, the identifier of the protein to be predicted can be understood as information used to uniquely identify the target protein, including but not limited to the Universal Protein Resource Number (UniProt), gene name, Protein Data Bank Number (PDB), National Center for Biotechnology Information Sequence Accession Number (NCBI), etc., which are the core indexes for accurately retrieving target protein sequences from public databases.

[0036] In this disclosure, protein sequence databases can be understood as public / private bioinformatics databases that store massive amounts of known protein amino acid sequences and related information, including but not limited to UniProt database, NCBI database, PDB sequence library, etc., which provide the source for retrieving raw amino acid sequence data for this disclosure.

[0037] In this disclosure, preprocessing can be understood as a standardization and cleaning process performed on the raw amino acid sequence data retrieved from the database, with the aim of eliminating data noise and ensuring the accuracy and consistency of sequence information. Specifically, removing redundant sequence fragments refers to removing repetitive sequence regions or additional fragments that are not part of the target protein from the raw data; removing unknown amino acid residues refers to deleting or marking unknown residues represented by X (or other placeholders) in the sequence; and removing invalid whitespace characters refers to removing non-amino acid characters such as spaces and newlines from the sequence.

[0038] Specifically, obtaining amino acid sequence information includes the following steps: First, a unique identifier (such as UniProt ID) corresponding to the protein to be predicted is obtained through at least one method, such as user input or association with experimental records. Next, the identifier is input into the retrieval interface of the protein sequence database to match and obtain the original amino acid sequence data corresponding to the protein. Then, the original amino acid sequence data is cleaned by sequentially removing redundant sequence fragments, replacing / deleting unknown amino acid residues, and filtering invalid whitespace characters to obtain a standardized sequence. Finally, the order of amino acids is read from the standardized sequence, the type of amino acid at each position is identified, and the total number of residues in the sequence is counted, ultimately obtaining amino acid sequence information containing the order, type, and length of the amino acid sequence.

[0039] For example, the complete process of obtaining amino acid sequence information is as follows: Obtain the unique identifier of the protein to be predicted (such as UniProt ID or PDB ID), and use this identifier to retrieve and download the corresponding raw amino acid sequence data from a protein sequence database (such as UniProt or PDB database).

[0040] The raw amino acid sequence data is preprocessed and cleaned, including removing redundant non-target sequence fragments, removing unknown non-standard amino acid residues (such as 'X' markers), and removing invalid whitespace characters, to obtain a standardized amino acid sequence.

[0041] The sequence of amino acid residues, the type of residues at each position, and the full length of the sequence are extracted from the standardized sequence and used as the sequence feature basis for subsequent model input.

[0042] The following will explain in detail how to obtain three-dimensional structural information, including: In this disclosure, the backbone atom can be understood as the core atom that constitutes the main chain structure of amino acid residues, specifically including nitrogen atom (N), alpha carbon atom (CA), carbon atom (C), and oxygen atom (O), which is a set of atoms that characterize the spatial position of amino acid residues and the folding state of peptide chains.

[0043] In this disclosure, the dihedral angle can be understood as the torsion angle formed by the backbone atoms of adjacent amino acid residues in a protein peptide chain. Specifically, it includes the Phi angle (φ, the angle between the C atom of the previous residue and the N, CA and C atoms of the current residue), the Psi angle (ψ, the angle between the N, CA and C atoms of the current residue and the N atom of the next residue), and the Omega angle (ω, the angle between the CA and C atoms of the current residue and the N and CA atoms of the next residue), which are indicators reflecting the local conformation of the peptide chain.

[0044] In this disclosure, Euclidean distance can be understood as the straight-line distance between any two amino acid spatial coordinates in a three-dimensional Cartesian coordinate system. It is a fundamental parameter for quantifying the spatial proximity between amino acid residues and determining the interaction relationship between residues.

[0045] Specifically, determining the three-dimensional structural information includes the following steps: First, the three-dimensional Cartesian coordinates of the N, CA, C, and O backbone atoms corresponding to each amino acid residue are extracted from the three-dimensional structural data of the protein to be predicted. Next, the average of the N, CA, C, and O atom coordinates of each amino acid residue is taken to obtain the single spatial coordinates of that amino acid, serving as a spatial position representation of the residue. Then, based on the spatial coordinates of adjacent amino acids, the Phi angle, Psi angle, and Omega angle corresponding to each residue are calculated sequentially to obtain the dihedral angle parameters of each residue. Furthermore, all pairs of amino acid residues are traversed, and the Euclidean distance between the spatial coordinates of each pair of residues is calculated to obtain the relative spatial distance matrix between residues. Finally, the spatial coordinates, dihedral angles, and relative spatial distance matrix of each amino acid are summarized to form the three-dimensional structural information of the protein to be predicted.

[0046] For example, the complete process of obtaining three-dimensional structural information is as follows: For each amino acid residue in the protein to be predicted, the Cartesian coordinates of its backbone atoms (including nitrogen atom N, alpha carbon atom CA, carbonyl carbon atom C and oxygen atom O) in three-dimensional space are extracted from the PDB structure file.

[0047] Based on the spatial coordinates of the backbone atoms of adjacent residues, the dihedral features of each residue are calculated, specifically including the Phi angle (φ), Psi angle (ψ), and Omega angle (ω) that reflect the torsion state of the backbone.

[0048] Calculate the Euclidean distance (e.g., CA for alpha carbon atom) between any two amino acid residues to construct a relative spatial distance matrix between residues. This matrix is ​​used to subsequently construct a geometric nearest neighbor graph (k-NN Graph) and define the connection relationships of the graph edges.

[0049] The extracted skeleton atomic coordinates, the calculated dihedral angles, and the relative spatial distance matrix are combined to determine the three-dimensional structural information.

[0050] The following section will elaborate on how to construct the first multimodal feature map for the inference stage, including: The amino acid sequence information and the three-dimensional structure information are both subjected to feature encoding and dimension mapping to obtain the initial node feature vector and the initial edge feature vector of the amino acid. Construct node features for inference; the node features for inference are obtained by concatenating the dimensions of the initial node feature vector and the first zero vector; wherein, the dimension of the first zero vector is consistent with the dimension of the single-point potential energy vector in the evolutionary energy feature, and is used to mask the single-point energy feature; Construct inference edge features; the inference edge features are obtained by determining the nearest neighbor connection relationship based on the relative spatial distance of the amino acids, and concatenating the geometric feature vector extracted based on the nearest neighbor connection relationship with the second zero vector; wherein, the dimension of the second zero vector is consistent with the dimension of the pairwise coupling potential energy matrix in the evolutionary energy feature, and is used to shield the coupling energy feature; The first multimodal feature map is generated based on the node features and edge features used for inference.

[0051] In this disclosure, the node features used for inference can be understood as multimodal fusion feature vectors with amino acid residues as graph nodes. It is a comprehensive representation obtained by encoding and dimensionally aligning the three-dimensional structural features (spatial coordinates, dihedral angles, etc.) of amino acids with evolutionary frequency features (PSSM), and then splicing them together. It is used to characterize the sequence and structural composite properties of a single residue.

[0052] In this disclosure, the nearest neighbor connection relationship used for inference can be understood as a graph edge connection rule constructed based on the relative spatial distance between amino acid residues. Typically, the spatial distance is less than a preset threshold or the k nearest neighbor residues of each residue are selected as connection objects to define the association relationship between residue nodes in the multimodal graph.

[0053] In this disclosure, the edge features used for inference can be understood as the feature vectors corresponding to the edges connecting adjacent residue nodes in a multimodal graph, which are used to characterize the spatial association strength and physical interaction properties between nodes.

[0054] Specifically, acquiring the first multimodal image includes the following steps: Step 1: Extract structural features by encoding the three-dimensional structural information of the protein to be predicted, calculating the dihedral angle features of the main chain, and obtaining structural node feature vectors; extract geometric features by processing the relative spatial distances of the amino acids using radial basis functions (RBF) and sequence position encoding to obtain geometric edge feature vectors; extract frequency features by encoding the amino acid sequence information, calculating the position-specific score matrix (PSSM), and obtaining evolutionary frequency feature vectors.

[0055] Step 2: Construct inference node features. For the initial node feature vector, the combination of the extracted structural node feature vector and the evolution frequency feature vector is called the initial node feature vector. For the node information, a first all-zero vector with the same dimension as the Potts single-point potential energy feature used in the training phase is constructed. The initial node vector and the first all-zero vector are logically concatenated. In specific implementations, they can be directly concatenated as input; alternatively, the structural features can be used as graph node attributes, while the evolution frequency and the first all-zero vector are used as auxiliary inputs, and physical concatenation is performed in the model's fusion layer. Both methods achieve the technical effect of masking single-point energy features.

[0056] Step 3: Construct inference edge features. Determine the nearest neighbor connections between residues based on relative spatial distance (k-NN graph). The geometric edge feature vector extracted based on the connection relationship is called the initial edge feature vector. Construct a second all-zero vector with the same dimension as the Potts pairwise coupling potential energy feature used in the training phase. Concatenate the initial edge feature vector with the second all-zero vector in terms of dimension to obtain the edge feature vector. By filling the zero vector, the Potts energy features not calculated in the inference phase are masked, making the input dimension fit the pre-trained structural encoder, while eliminating noise interference.

[0057] Step 4: Generate graph structure and auxiliary features. Using amino acid residues as graph nodes and nearest neighbor connections as graph edges, assign the features of the inference nodes to the nodes or associate them with the nodes, and assign the features of the inference edges to the edges, thereby generating the first multimodal feature graph for inference.

[0058] The following section will elaborate on the prediction model, including: The sequence encoding module is used to determine sequence dependencies based on multiple sequence alignment data from the training protein dataset, and to determine sequence semantic embedding vectors based on sequence dependencies. The structure encoding module is used to extract geometric features based on the three-dimensional structural information in the training protein dataset. Furthermore, it uses the evolutionary energy matrix in the training protein dataset for normalization constraints to learn the joint representation of geometric structure and evolutionary energy, thereby obtaining the structure embedding vector. The multimodal fusion module is used to concatenate and fuse sequence semantic embedding vectors, structural embedding vectors, and position-specific score matrices to determine the fused embedding vector. The predicted mutation decoding module is used to decode the fused embedding vector to obtain the predicted mutation result.

[0059] In this disclosure, the sequence encoding module can be understood as the core functional module in the multimodal fusion prediction model responsible for extracting deep sequence semantic features of proteins. This module, based on a protein language model (such as the MSA Transformer), processes the input multiple sequence alignment (MSA) data through an attention mechanism to mine sequence dependencies. These sequence dependencies can be the contextual relationships between residues in a protein's amino acid sequence, including short-range dependencies between adjacent residues and long-range dependencies between non-adjacent residues, reflecting the inherent patterns of sequence arrangement and the co-evolutionary effects between residues. The sequence semantic embedding vector can be understood as the core output of the sequence encoding module; it is a low-dimensional dense vector obtained by feature encoding long-range sequence dependencies, capable of accurately characterizing the sequence contextual semantic attributes of individual amino acid residues.

[0060] In this disclosure, the structure encoding module can be understood as the core functional module in the multimodal fusion prediction model responsible for extracting the three-dimensional geometric features of proteins and internalizing physical energy constraints. This module, based on a geometric graph neural network (such as GVP-GNN), integrates the three-dimensional structural information of the training dataset with the evolutionary energy matrix (output of the Potts model) to achieve simultaneous learning of spatial correlation rules and physical energy rules. Spatial correlation rules can be understood as the distribution and interaction rules of protein amino acid residues in three-dimensional space, including the spatial proximity relationships between residues and peptide chain folding patterns. Physical energy constraints can be understood as the energy information provided by the Potts model, characterizing the strength of co-evolutionary coupling between residues, guiding the model to focus on key sites that are geometrically distant but energy-closely coupled. The structure embedding vector can be understood as the core output of the structure encoding module, a low-dimensional dense vector obtained by deeply fusing geometric structural features and evolutionary energy features, capable of accurately characterizing the local spatial environment and physical energy constraints of individual amino acid residues.

[0061] In this disclosure, the multimodal fusion module can be understood as the core functional module in the multimodal fusion prediction model responsible for integrating sequence features, structural features, and evolutionary frequency features. This module concatenates the sequence semantic embedding vector, the structural embedding vector, and the position-specific score matrix (PSSM) along the feature channel dimension. The PSSM reflects the evolutionary frequency patterns of amino acids at specific positions. Through this fusion, the model can address the problem of insufficient information from a single modality, achieving deep complementarity between sequence semantics, geometric structure, and evolutionary frequency information, and outputting a fused embedding vector that comprehensively characterizes the multidimensional properties of proteins.

[0062] In this disclosure, the predicted mutation decoding module can be understood as the core functional module in the multimodal fusion prediction model responsible for transforming the fusion embedding vector into the final mutation prediction result. This module performs feature decoding and probability modeling on the fusion embedding vector using a graph neural network decoder, outputting the probability distribution (Logits) of 20 natural amino acids at each amino acid residue position. Furthermore, based on the log-likelihood ratio of wild-type and mutant amino acids, the predicted mutation result (i.e., the mutation effect score) is calculated, achieving accurate prediction of the protein mutation effect.

[0063] The following will explain in detail how to obtain the evolutionary energy matrix, including: Based on the MSA of the training protein dataset, the location-specific score matrix is ​​calculated; Based on the MSA of the training protein dataset, the single-point potential vector and the paired coupled potential matrix are calculated using a statistical physics model. Obtain the order of the first vocabulary output by the statistical model and the order of the second vocabulary of the multimodal fusion prediction model; Based on the mapping relationship between the order of the first and second vocabularies and the position-specific score matrix, the dimensions of the single-point potential vector and the paired coupling potential matrix are rearranged to determine the evolutionary energy matrix for semantic alignment.

[0064] In layman's terms, based on multiple sequence alignment (MSA) data from the training protein dataset, parameter inference is performed using a statistical energy mechanism to obtain the original single-point potential and paired coupling potential matrices. The first vocabulary order output by the statistical energy mechanism and the second vocabulary order of the graph neural network model are obtained. Based on the mapping relationship between the two, the original single-point potential vector and paired coupling potential matrix are subjected to feature semantic alignment processing. The aligned single-point potential vector and paired coupling potential matrix are then determined as the evolutionary energy matrix.

[0065] In this disclosure, the Position-Specific Scoring Matrix (PSSM) can be understood as an evolutionary feature matrix constructed based on the results of multiple sequence alignment (MSA) of the training protein dataset. The dimension can be the sequence length L×21, with each row corresponding to an amino acid site of the protein and each column corresponding to a native amino acid or vacancy. The numerical values ​​in the matrix represent the frequency of occurrence and evolutionary conservation score of various amino acids at the corresponding sites. The higher the score, the stronger the evolutionary conservation of that amino acid at that site.

[0066] In this disclosure, the statistical energy mechanism can be understood as an energy landscape inference mechanism based on statistical physics, specifically employing the Potts model. This mechanism takes multiple sequence alignment (MSA) data as direct input and learns the co-evolutionary coupling parameters between amino acid residues by maximizing the pseudo-likelihood function of the sequence data. It transforms biological co-evolutionary frequencies into a physical energy potential function, thereby explicitly capturing long-range interactions between residues.

[0067] In this disclosure, a single-point potential vector can be understood as a one-dimensional representation of the evolutionary energy matrix, also known as a field vector. The dimension can be the sequence length L×20. Each row corresponds to an amino acid site in the protein, and each column corresponds to a natural amino acid. The numerical value in the vector represents the inherent potential energy of the corresponding amino acid at a single amino acid site, reflecting the evolutionary preference of the amino acid at that site.

[0068] In this disclosure, the pairwise coupling potential matrix can be understood as the two-dimensional representation of the evolutionary energy matrix, also known as the coupling matrix. Its dimension can be the sequence length L×L×20×20, corresponding to the coupling effect between any two amino acid sites in the protein. The numerical values ​​in the matrix represent the coupling potential of the corresponding amino acid combination at the two sites, reflecting the co-evolutionary dependence strength between residues, and are the core indicator for characterizing long-range residue interactions.

[0069] Specifically, obtaining the evolutionary energy matrix may include the following steps: Step 1: Perform homology sequence search on the training protein dataset to obtain high-quality multiple sequence alignment (MSA) data, and then perform redundancy removal and quality filtering; Step 2: Using Potts model inference tools (such as CCMpred), with the MSA data as input, the pseudo-likelihood maximization algorithm is used for iterative optimization to infer the model parameters that can restore the statistical characteristics of MSA to the greatest extent, namely the original single-point potential energy vector and the paired coupled potential energy matrix. Step 3: Perform feature semantic alignment. To address the spatial misalignment issue caused by the alphabetical order (e.g., ACDEF...) in Potts tool outputs compared to the custom order (e.g., ARNDC...) used by deep learning models, a permutation index matrix is ​​constructed. This index is then used to rearrange the amino acid dimension of the original features to ensure that the energy features are semantically consistent with the model input channels. Step 4: Compress and normalize the aligned features. For example, extract the energy scalar map and mutation potential spectrum (profile map) corresponding to the current sequence from the four-dimensional coupling matrix, and use the Tanh function for numerical normalization to finally obtain the complete evolutionary energy matrix adapted to the deep learning model.

[0070] The following section will elaborate on how to construct a second multimodal feature map for training, specifically during the training phase: Obtain the amino acid sequence information, the three-dimensional structure information, and the evolutionary energy matrix of the training protein dataset; The amino acid sequence information and the three-dimensional structure information are encoded to obtain initial node feature vectors and initial edge feature vectors; Construct training node features; the training node features are obtained by concatenating the initial node feature vector with the single-point potential vector in the evolutionary energy matrix in terms of dimension. Construct training edge features; the training edge features are obtained by concatenating the initial edge feature vector with the pairwise coupling potential energy matrix in the evolutionary energy matrix. The second multimodal feature map is generated based on the training node features and the training edge features.

[0071] In this disclosure, the training node features can be understood as a multimodal graph node comprehensive feature vector used for model training. It is constructed by fusing structural features, evolutionary frequency features (PSSM), and evolutionary energy features (Field) with the amino acid residues of the training protein as graph nodes. It not only describes the sequence and structural composite properties of a single residue, but also injects single-point preference constraints at the evolutionary level, providing more comprehensive feature support for model training.

[0072] In this disclosure, the training nearest neighbor connection can be understood as a graph edge connection rule constructed based on the relative spatial distance between amino acid residues of the training protein. It adopts the same construction logic as the inference nearest neighbor connection (such as k nearest neighbor or distance threshold screening) to ensure the topological consistency of the graph structure during the training and inference stages.

[0073] In this disclosure, the training edge features can be understood as multimodal graph edge comprehensive feature vectors used for model training. Based on training the nearest neighbor connection relationship, it integrates the spatial geometric features and evolutionary energy features (Couplings) of residue pairs, which not only characterizes the spatial interaction strength between nodes, but also injects co-evolutionary coupling constraints between residues, helping the model internalize the intrinsic relationship between sequence, structure and evolution.

[0074] Specifically, the construction of the second multimodal feature map includes the following steps: Step 1: Data Acquisition and Cleaning. Acquire the amino acid sequence information, three-dimensional structural information, and original evolutionary features of the training protein dataset. Perform redundancy removal on the sequence information; extract the backbone atomic coordinates from the structural information; obtain the original Potts evolutionary energy matrix (including single-point potential and paired coupling potential) and PSSM matrix.

[0075] Step 2: To address the inconsistency between the original features output by the Potts tool (usually alphabetical) and the input vocabulary of the deep learning model (custom order), a feature rearrangement algorithm is executed. A permutation index matrix is ​​constructed to rearrange the amino acid dimension of the Potts energy features, ensuring strict semantic alignment between the physical energy values ​​and the model input channels. Subsequently, the aligned energy features are subjected to Tanh normalization, compressing the numerical range to the [-1, 1] interval to prevent gradient explosion during training.

[0076] Step 3: Constructing Training Node Features. First, extract the geometric structure features of each amino acid residue and obtain the corresponding PSSM frequency feature vector. The combination of the geometric structure features and the PSSM frequency feature vector is called the initial node feature vector. Subsequently, the initial node feature vector is concatenated with the aligned Potts single-point potential vector to generate training node features rich in physical evolutionary constraints.

[0077] Step 4: Construct training edge features. A geometric nearest neighbor graph (e.g., a k-NN graph) is constructed based on the relative spatial distances between residues. For each edge in the graph, geometric distance features are extracted. These geometric distance features are called the initial edge feature vector. The initial edge feature vector is then coupled with the corresponding aligned Potts pairs of potential energy features (containing energy scalars and mutation potential spectra) and their dimensions are concatenated to generate the training edge features.

[0078] Step 5: Generate the complete training graph structure. Using amino acid residues as graph nodes and nearest-neighbor connections as graph edges, assign the training node features to the corresponding nodes and the training edge features to the corresponding edges, outputting a complete second multimodal feature map. This feature map contains all sequence, structural, and evolutionary energy information, specifically designed to drive the augmentation training process of the graph neural network.

[0079] In some instances, the multimodal fusion prediction model disclosed herein also includes: An adaptive gating module is used to dynamically adjust the weight ratio of the evolutionary energy matrix features in the structure encoding module according to the effective sequence depth of the input MSA.

[0080] The adaptive gating module can be understood as the core functional module in the multimodal fusion prediction model used to dynamically adjust the weight ratio of the evolutionary energy matrix. Based on the quality indicators of multiple sequence alignment (MSA) of the training protein dataset (such as the number of effective sequences Neff), it adaptively outputs weight coefficients to precisely control the contribution of evolutionary energy features (single-point potential vector, paired coupled potential matrix) in the model training process, thereby achieving dynamic adaptation of evolutionary constraint information.

[0081] For example, Figure 2 This is a schematic flowchart illustrating a protein mutation prediction method provided in an embodiment of this disclosure. Figure 2 As can be seen, this process uses protein sequence and protein structure as dual inputs. First, an MSA (Multimodal Feature Analysis) is generated for the protein sequence. Then, evolutionary and sequence features are extracted from the MSA, and structural features are extracted from the protein structure. These three types of features are then fused and input into the decoding and prediction module, ultimately outputting the predicted probabilities of 20 amino acids. The mutation result is then calculated through mutation effects. This process clearly embodies the core idea of ​​multimodal feature fusion disclosed in this paper: simultaneously utilizing evolutionary / sequence information at the sequence level and spatial information at the structural level, avoiding the limitations of single-feature modeling, and ensuring that mutation effect prediction takes into account both sequence conservation and spatial conformational influences.

[0082] For example, Figure 3 This is a schematic flowchart illustrating another protein mutation prediction process provided in an embodiment of this disclosure. From Figure 3As can be seen, the process consists of three core stages: input and preprocessing, asymmetric training and inference, and fusion prediction and output. In the input stage, an MSA (Sequence Subsequence Analysis) is generated from the protein sequence. This is used to extract sequence features and to calculate Potts energy features and PSSM (Population-Symptom-Mode) frequency features. Simultaneously, geometric features are extracted from the protein structure. Furthermore, feature alignment is performed, and dimensionality rearrangement addresses the semantic misalignment between Potts features and the model vocabulary. In the training stage, the geometric features, PSSM features, and complete Potts energy features are concatenated and fused, and then input into the model for enhanced training. In the inference stage, the geometric features, PSSM features, and constructed zero vectors are concatenated, reflecting the asymmetric training and inference strategy of this method. Sequence embeddings are extracted using the MSA Transformer on the right, and energy-constrained structural embeddings are extracted using the Geometric Vector Perceptron (GVP) encoder. After fusion in the Fusion layer, the two are decoded by the GVP decoder, ultimately outputting the Logits distribution and mutation effect score. This process fully demonstrates the efficient prediction logic of introducing evolutionary constraints in the training phase and simplifying input in the inference phase, while also clarifying the importance of collaboration among the modules.

[0083] For example, this disclosure also provides a specific experimental data comparison process, including: Table 1

[0084] As shown in Table 1, on 10 different multiple variation effect detection datasets (covering various functional proteins such as NUD15, P53, and PTEN), the average Spearman correlation coefficient of the model disclosed in this publication (0.535) significantly outperforms the sequence-based Gemme model (0.483) and the structure-based Rosetta model (0.421). Except for the P53 and MAPK datasets, the model disclosed in this publication achieves state-of-the-art performance on the other 8 datasets, especially showing significant advantages on the CP2C9 and PABP datasets (correlation coefficients exceeding 0.6). This fully demonstrates that by fusing sequence, structural, and evolutionary synergistic information, this publication can more accurately capture the association between amino acid mutations and protein function compared to baseline models with a single modality, resulting in more comprehensive and robust predictive performance.

[0085] Table 2

[0086]

[0087] As shown in Table 2, on the Venusmuthub dataset (which covers a variety of protein functional properties and is based on real biochemical experimental measurements), the Spearman correlation coefficient (0.321) of the model disclosed in this publication is slightly better than that of the sequence and structure embedding model (0.315), and the normalized loss cumulative gain (0.894) is also slightly higher than that of the SSEmb model (0.890). Compared with single-modal models such as the VenusREM mutation effect prediction model, MSA Transformer, and ESM 2, the predictive performance of the model disclosed in this publication is significantly superior, with the Spearman correlation coefficient leading by more than 40%.

[0088] Table 3

[0089]

[0090] As shown in Table 3, on the VenusMutHub activity-related mutation dataset, the Spearman correlation coefficient of the model disclosed in this publication (0.398) is significantly higher than all baseline models, improving by 0.009 compared to the second-best SSEmb model (0.389) and by 0.107 compared to the VenusREM model (0.291), demonstrating a clear advantage. Although the NDCG value of the model disclosed in this publication (0.857) is slightly lower than that of the SSEmb model (0.862), it is still significantly higher than other single-modality models. This indicates that the model disclosed in this publication can more accurately quantify the impact of mutations on protein activity in the core task of predicting protein activity-related mutations, especially performing better in capturing key activity mutation sites, meeting the practical needs of activity optimization in protein engineering.

[0091] Table 4

[0092]

[0093] As shown in Table 4, on the VenusMutHub selectivity-related mutation dataset, the model disclosed in this publication achieved the best performance in both Spearman correlation coefficient (0.321) and NDCG value (0.894), improving by 0.006 and 0.004 respectively compared to the SSEmb model, and showing a greater improvement compared to other single-modality models. Selectivity prediction requires accurately capturing the impact of mutations on the specificity of protein-ligand and target binding, placing extremely high demands on the model's feature fusion capabilities. The model disclosed in this publication maintains its advantage on this task, fully demonstrating its design approach of fusing sequence, structural, and evolutionary information. This approach can effectively correct the prediction bias of single-modality models, accurately identify key mutations affecting protein selectivity, and provide reliable support for drug target optimization and reducing off-target effects.

[0094] This disclosure also provides a protein mutation prediction device. Figure 4This is a structural block diagram of a protein mutation prediction device provided in an embodiment of the present disclosure, as shown below. Figure 4 As shown, the protein mutation prediction device 400 includes: The first acquisition unit 401 is used to acquire the amino acid sequence information of the protein to be predicted; the amino acid sequence information includes the amino acid sequence order, the type of amino acid, and the sequence length of the amino acid. The second acquisition unit 402 is used to acquire the three-dimensional structural information of the protein to be predicted; the three-dimensional structural information includes the spatial coordinates of amino acids, the dihedral angles of amino acids, and the relative spatial distances between amino acids; Construction unit 403 is used to construct a first multimodal feature map based on the amino acid sequence information and the three-dimensional structure information, and using a zero vector filling strategy; the first multimodal feature map includes node features and edge features, but does not include explicit evolutionary energy features; Prediction unit 404 inputs the first multimodal feature map into the pre-trained prediction model to obtain the predicted mutation result; the prediction model is a multimodal fusion prediction model determined based on a graph neural network, and is obtained by training using the second multimodal feature map; the second multimodal feature map contains at least the amino acid sequence information, three-dimensional structure information, and evolutionary energy matrix calculated based on multiple sequence alignment (MSA) and processed by feature alignment of the training protein dataset; the evolutionary energy matrix participates in the training of the prediction model as a regularization constraint.

[0095] In one exemplary embodiment, the prediction unit 404 is specifically used to predict mutation results including: the mutation type of amino acid and the mutation score of amino acid.

[0096] In one exemplary embodiment, the first acquisition unit 401 is specifically used to: acquire the identifier of the protein to be predicted; use the identifier to retrieve the original amino acid sequence data from the protein sequence database; preprocess the original amino acid sequence data and extract the amino acid sequence order, amino acid type and amino acid sequence length to obtain amino acid sequence information; the preprocessing includes at least one of the following: removing redundant sequence fragments, removing unknown amino acid residues or invalid whitespace characters.

[0097] In one exemplary embodiment, the second acquisition unit 402 is specifically used to: extract the spatial coordinates of the backbone atoms of any amino acid of the protein to be predicted, and use the average spatial coordinates of the backbone atoms as the spatial coordinates of the amino acid; determine the torsion angle between adjacent peptide bonds of the amino acid based on the spatial coordinates of the amino acid to obtain the dihedral angle of the amino acid; calculate the Euclidean distance between the spatial coordinates of any two amino acids to obtain the relative spatial distance of the amino acid; and determine the spatial coordinates of the amino acid, the dihedral angle of the amino acid, and the relative spatial distance of the amino acid as three-dimensional structural information.

[0098] In one exemplary embodiment, the construction unit 403 is specifically used to: construct a first multimodal feature map based on amino acid sequence information and three-dimensional structure information, including: performing feature encoding and dimension mapping processing on both the amino acid sequence information and the three-dimensional structure information to obtain initial node feature vectors and initial edge feature vectors of amino acids; constructing node features for inference; the node features for inference are obtained by concatenating the initial node feature vectors with a first zero vector in terms of dimension; wherein, the dimension of the first zero vector is consistent with the dimension of the single-point potential energy vector in the evolutionary energy matrix, and is used to shield single-point energy features; constructing edge features for inference; the edge features for inference are obtained by determining the nearest neighbor connection relationship based on the relative spatial distance of amino acids, and concatenating the geometric feature vector extracted based on the nearest neighbor connection relationship with a second zero vector in terms of dimension; wherein, the dimension of the second zero vector is consistent with the dimension of the pairwise coupled potential energy matrix in the evolutionary energy matrix, and is used to shield coupled energy features; and generating the first multimodal feature map based on the node features for inference and the edge features for inference.

[0099] In one exemplary embodiment, the prediction unit 404 is specifically configured as follows: a sequence encoding module, used to learn long-range dependencies between residues based on the amino acid sequence information of the training protein dataset and through an attention mechanism, to obtain a sequence semantic embedding vector; a structure encoding module, used to extract geometric features based on the three-dimensional structural information in the training protein dataset, and to learn the joint representation of geometric structure and evolutionary energy using the evolutionary energy matrix in the training protein dataset for normalization constraints, to obtain a structure embedding vector; a multimodal fusion module, used to concatenate and fuse the sequence semantic embedding vector, the structure embedding vector, and the position-specific score matrix to obtain a fused embedding vector; and a predicted mutation decoding module, used to decode the fused embedding vector and output the amino acid probability distribution at each residue position, thereby obtaining the predicted mutation result.

[0100] In one exemplary embodiment, the prediction unit 404 is specifically configured to: calculate a position-specific score matrix based on the MSA of the training protein dataset; calculate a single-point potential vector and a pairwise coupling potential matrix using a statistical physical model based on the MSA of the training protein dataset; obtain the first vocabulary order output by the statistical model and the second vocabulary order of the multimodal fusion prediction model; and perform dimensional rearrangement of the single-point potential vector and the pairwise coupling potential matrix based on the mapping relationship between the first vocabulary order and the second vocabulary order and the position-specific score matrix to determine the semantically aligned evolutionary energy matrix.

[0101] In one exemplary embodiment, the prediction unit 404 is specifically configured to: acquire amino acid sequence information, three-dimensional structure information, and evolutionary energy matrix of the training protein dataset; encode the amino acid sequence information and three-dimensional structure information to obtain initial node feature vectors and initial edge feature vectors; construct training node features; the training node features are obtained by dimensionally concatenating the initial node feature vectors with the single-point potential vectors in the evolutionary energy matrix; construct training edge features; the training edge features are obtained by dimensionally concatenating the initial edge feature vectors with the paired coupling potential matrices in the evolutionary energy matrix; and determine a second multimodal feature map based on the training node features and training edge features.

[0102] In summary, this disclosure provides a method and apparatus for predicting protein mutations. This disclosure involves acquiring the amino acid sequence information of the protein to be predicted; the amino acid sequence information includes the amino acid order, amino acid type, and amino acid sequence length; acquiring the three-dimensional structural information of the protein to be predicted; the three-dimensional structural information includes the spatial coordinates of the amino acids, the dihedral angles of the amino acids, and the relative spatial distances between the amino acids; constructing a first multimodal feature map for inference based on the amino acid sequence information and the three-dimensional structural information; inputting the first multimodal feature map into a pre-trained multimodal fusion prediction model to obtain the predicted mutation result; the multimodal fusion prediction model is trained and determined based on a second multimodal feature map; the second multimodal feature map, in addition to containing the amino acid sequence information and three-dimensional structural information based on the training protein dataset, also contains an evolutionary energy matrix calculated based on multiple sequence alignment (MSA) and processed by feature semantic alignment; the evolutionary energy matrix participates in model training as a regularization constraint. In this way, compared with existing single sequence / structure models or simple combination models, this disclosure can construct a multimodal fusion architecture through graph neural networks. This not only achieves deep fusion of sequence and structure but also introduces an evolutionary energy matrix based on statistical physics, compensating for the lack of explicit physical / evolutionary constraints in deep learning models. This fundamentally improves the model's ability to capture the correlation between mutations and protein conformation and function. Simultaneously, by adopting an asymmetric design that integrates energy features during training and only requires sequence and structural features during inference, the computationally time-consuming evolutionary energy features are shielded during inference, solving the problems of high inference cost and slow speed in traditional statistical physics models. Furthermore, through energy constraint guidance during training and the complementarity of multimodal information, the model reduces its dependence on high-quality multi-sequence alignment data during inference. Even in scenarios where homologous sequences are scarce or data quality is low, the model can still maintain high prediction accuracy. In summary, the technical solution provided by this disclosure can improve the accuracy and efficiency of protein mutation prediction and can be adapted to various application scenarios.

[0103] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0104] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0105] The block diagrams of devices, apparatuses, devices, and systems disclosed herein are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0106] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0107] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.

[0108] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0109] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0110] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for predicting protein mutations, characterized in that, The method includes: Obtain the amino acid sequence information of the protein to be predicted; the amino acid sequence information includes the amino acid sequence order, amino acid type, and amino acid sequence length; Obtain the three-dimensional structural information of the protein to be predicted; the three-dimensional structural information includes the spatial coordinates of amino acids, the dihedral angles of amino acids, and the relative spatial distances between amino acids; Based on the amino acid sequence information and the three-dimensional structure information, a first multimodal feature map is constructed; the first multimodal feature map includes node features and edge features, but does not include explicit evolutionary energy features; The first multimodal feature map is input into the pre-trained prediction model to obtain the predicted mutation result; the prediction model is a multimodal fusion prediction model determined based on a graph neural network, and is trained using the second multimodal feature map; the second multimodal feature map at least includes the amino acid sequence information, the three-dimensional structure information, and the evolutionary energy matrix calculated based on multiple sequence alignment (MSA) and processed by feature alignment of the training protein dataset; the evolutionary energy matrix is ​​used as a regularization constraint in the training of the prediction model.

2. The method according to claim 1, characterized in that, The predicted mutation results include: the mutation type of the amino acid and the mutation score of the amino acid.

3. The method according to claim 1, characterized in that, The process of obtaining the amino acid sequence information of the protein to be predicted includes: Obtain the identifier of the protein to be predicted; Using the aforementioned identifier, the original amino acid sequence data can be retrieved from a protein sequence database; The original amino acid sequence data is preprocessed, and the amino acid sequence order, amino acid type, and amino acid sequence length are extracted to obtain the amino acid sequence information. The preprocessing includes at least one of the following: removing redundant sequence fragments, removing unknown amino acid residues, or invalid whitespace characters.

4. The method according to claim 1, characterized in that, The process of obtaining the three-dimensional structural information of the protein to be predicted includes: For any amino acid of the protein to be predicted, the spatial coordinates of its backbone atoms are extracted, and the average spatial coordinates of the backbone atoms are used as the spatial coordinates of the amino acid. Based on the spatial coordinates of the amino acid, the torsion angle between adjacent peptide bonds of the amino acid is determined, and the dihedral angle of the amino acid is obtained. Calculate the Euclidean distance between the spatial coordinates of any two amino acids to obtain the relative spatial distance between the amino acids; The spatial coordinates of the amino acid, the dihedral angle of the amino acid, and the relative spatial distance between the amino acids are determined as the three-dimensional structural information.

5. The method according to claim 1, characterized in that, The construction of the first multimodal feature map based on the amino acid sequence information and the three-dimensional structural information includes: The amino acid sequence information and the three-dimensional structure information are both subjected to feature encoding and dimension mapping to obtain the initial node feature vector and the initial edge feature vector of the amino acid. Construct node features for inference; the node features for inference are obtained by concatenating the dimensions of the initial node feature vector and the first zero vector; wherein, the dimension of the first zero vector is consistent with the dimension of the single-point potential energy vector in the evolutionary energy feature, and is used to mask the single-point energy feature; Construct inference edge features; the inference edge features are obtained by determining the nearest neighbor connection relationship based on the relative spatial distance of the amino acids, and concatenating the geometric feature vector extracted based on the nearest neighbor connection relationship with the second zero vector; wherein, the dimension of the second zero vector is consistent with the dimension of the pairwise coupling potential energy matrix in the evolutionary energy feature, and is used to shield the coupling energy feature; The first multimodal feature map is generated based on the node features and edge features used for inference.

6. The method according to claim 1, characterized in that, The prediction model includes: The sequence encoding module is used to obtain a sequence semantic embedding vector based on the amino acid sequence information of the training protein dataset and by learning the long-range dependencies between residues through an attention mechanism. The structure encoding module is used to extract geometric features based on the three-dimensional structure information in the training protein dataset, and to learn the joint representation of geometric structure and evolutionary energy by using the evolutionary energy matrix in the training protein dataset for normalization constraints, thereby obtaining a structure embedding vector. The multimodal fusion module is used to concatenate and fuse the sequence semantic embedding vector with the structural embedding vector and the position-specific score matrix to obtain a fused embedding vector; The predicted mutation decoding module is used to decode the fusion embedding vector and output the amino acid probability distribution at each residue position, thereby obtaining the predicted mutation result.

7. The method according to claim 6, characterized in that, The evolutionary energy matrix includes: Based on the MSA of the training protein dataset, the location-specific score matrix is ​​calculated. Based on the MSA of the training protein dataset, the single-point potential vector and the paired coupled potential matrix are calculated using a statistical physics model. Obtain the order of the first vocabulary output by the statistical model and the order of the second vocabulary of the multimodal fusion prediction model; Based on the mapping relationship between the order of the first vocabulary and the second vocabulary and the position-specific score matrix, the single-point potential vector and the paired coupling potential matrix are rearranged in dimension to obtain the semantically aligned evolutionary energy matrix.

8. The method according to claim 6, characterized in that, The second multimodal feature map includes: Obtain the amino acid sequence information, the three-dimensional structure information, and the evolutionary energy matrix of the training protein dataset; The amino acid sequence information and the three-dimensional structure information are encoded to obtain initial node feature vectors and initial edge feature vectors; Construct training node features; the training node features are obtained by concatenating the initial node feature vector with the single-point potential vector in the evolutionary energy matrix in terms of dimension. Construct training edge features; the training edge features are obtained by concatenating the initial edge feature vector with the pairwise coupling potential energy matrix in the evolutionary energy matrix. The second multimodal feature map is generated based on the training node features and the training edge features.

9. The method according to claim 6, characterized in that, The multimodal fusion prediction model also includes: An adaptive gating module is used to dynamically adjust the weight ratio of the evolutionary energy matrix features in the structure encoding module according to the effective sequence depth of the input MSA.

10. A protein mutation prediction device, characterized in that, The device includes: The first acquisition unit is used to acquire the amino acid sequence information of the protein to be predicted; the amino acid sequence information includes the amino acid sequence order, the type of amino acid, and the sequence length of the amino acid. The second acquisition unit is used to acquire the three-dimensional structural information of the protein to be predicted; the three-dimensional structural information includes the spatial coordinates of amino acids, the dihedral angles of amino acids, and the relative spatial distances between amino acids; A construction unit is used to construct a first multimodal feature map based on the amino acid sequence information and the three-dimensional structure information; the first multimodal feature map includes node features and edge features, but does not include explicit evolutionary energy features; The prediction unit is used to input the first multimodal feature map into a pre-trained prediction model to obtain the predicted mutation result; the prediction model is a multimodal fusion prediction model determined based on a graph neural network, and is trained using a second multimodal feature map; the second multimodal feature map at least includes the amino acid sequence information, the three-dimensional structure information, and the evolutionary energy matrix calculated based on multiple sequence alignment (MSA) and processed by feature alignment of the training protein dataset; the evolutionary energy matrix participates in the training of the prediction model as a regularization constraint.

Images