A method and apparatus for screening mutation sites based on free energy change.

Abstract

This invention relates to a method and apparatus for screening mutation sites based on free energy change. The method includes: constructing a codebook retrieval model, a pre-training framework, a first prediction model, and a second prediction model; preparing first and second datasets; training the codebook retrieval model based on the first dataset using the pre-training framework; after training, setting the internal component parameters of the first and second prediction models based on the model parameters of the codebook retrieval model, and then training the first and second prediction models based on the second dataset; after training, combining native and mutant protein information based on user-input native protein sequences and mutant residue sets to obtain multiple original groups; using the first prediction model to perform a preliminary screening of all original groups to obtain N preferred groups; using the second prediction model to sort the N preferred groups to generate a final selected group sequence and provide feedback to the user. This invention can improve overall research efficiency.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a processing method and apparatus for screening mutation sites based on free energy variation. Background Technology

[0002] Protein-protein interaction (PPI) refers to the process by which two or more proteins bind through physical contact to achieve specific biological functions. Mutating amino acids at the interaction interface is a commonly used strategy for regulating protein-protein interactions. In this study, the free energy change ΔG before and after each mutation strategy is calculated; ΔG = post-mutation free energy ΔG. post - Free energy ΔG before mutation pre Then, mutation strategy screening is performed using △△G sorting.

[0003] Calculating the ΔG of all mutation strategies for a protein complex using a wet laboratory method is infeasible; however, it can be achieved through simulation. Simulation involves using molecular dynamics simulation tools to simulate the three-dimensional conformation of the protein complex before and after each mutation strategy, and then calculating the free energy (ΔG) based on the simulated conformation. pre ,△G post The free energy change ΔΔG for each mutation strategy can be obtained by performing calculations and then performing differential calculations on the free energy before and after the mutation. However, we have found in practice that even using simulation calculations results in long simulation calculation times and high simulation calculation costs. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a processing method, apparatus, electronic device, and computer-readable storage medium for screening mutation sites based on free energy change. To improve the screening efficiency of mutation strategies, this invention constructs three types of tool models: 1) a codebook retrieval model for mapping residue features of microenvironmental characteristics of specified residue sites in protein complexes using a codebook retrieval method, and a pre-training framework for this codebook retrieval model; 2) a coarse calculation model (i.e., the first prediction model) based on the codebook retrieval model, which first learns two microenvironmental mapping features of a single mutation site before and after mutation, and then predicts the free energy change ΔΔG before and after a single-point mutation based on the difference features of the two microenvironmental mapping features; 3) a fine calculation model (i.e., the second prediction model) based on the codebook retrieval model, which first learns two microenvironmental mapping features of a single mutation site before and after mutation, then reconstructs two local heteromorphic maps before and after mutation based on the two microenvironmental mapping features, and then predicts the free energy change ΔΔG before and after a single-point mutation based on the difference features of the two reconstructed local heteromorphic maps. Two training datasets (dataset 1 and dataset 2) were constructed for the three types of tool models. A two-stage training strategy was designed for the three types of tool models: 1) First stage: the codebook retrieval model is trained based on the pre-training framework using dataset 1; 2) Second stage: the internal component parameters of the first and second prediction models are set based on the codebook retrieval model trained in the first stage, and then the first and second prediction models are trained separately based on dataset 2. After training, the two prediction models are used for mutation strategy screening: first, the native protein sequence and mutant residue set input by the user are combined to obtain multiple original groups; then, the first prediction model is used to perform preliminary screening of all original groups according to the screening and sorting rules to obtain N preferred groups; then, the second prediction model is used to re-sort the N preferred groups according to the screening and sorting rules to obtain the final selected group sequence and provide feedback to the user. Through the two end-to-end prediction models of this invention, not only can the overall calculation cycle be shortened, the overall calculation cost reduced, and the overall research efficiency improved; the first prediction model, with higher prediction efficiency, can further improve the efficiency of massive pre-screening; and the second prediction model, with higher prediction accuracy, can further improve the prediction accuracy of the final selection results.

[0005] To achieve the above objectives, a first aspect of the present invention provides a method for screening mutation sites based on free energy variation, the method comprising:

[0006] A feature mapping model, denoted as the codebook retrieval model, is constructed to perform residue feature mapping on the microenvironmental features of a specified residue site in a protein complex using a codebook retrieval method. A pre-training framework is built for the codebook retrieval model. Based on the codebook retrieval model, two prediction models are constructed to predict the free energy change ΔΔG before and after a single-point mutation in the protein complex, denoted as the first prediction model and the second prediction model. The codebook retrieval model is used to perform residue feature mapping on the microenvironmental features of the current residue site based on the protein sequence S and residue site b input to the model, as well as the built-in residue feature codebook C, to obtain the corresponding residue mapping vector E. The pre-training framework is used to process the mask image G input to the framework. mask The three types of embedding vectors of all residue nodes are reconstructed to obtain the corresponding type reconstruction set X. 1 Dihedral Reconstruction Set X 2 Local orientation reconstruction set X 3 The first prediction model is used to predict the original protein sequence S1, the mutant protein sequence S2, and the mutant residue site b based on the input to the model. 1-2 Microenvironment embedding encoding is performed, and the free energy change is predicted based on the differential embedding of the two embedding encodings to obtain the corresponding first free energy change; the second prediction model is used to predict the free energy change based on the native protein sequence S3, the mutant protein sequence S4, and the mutant residue site b as input to the model. 3-4 Microenvironment embedding encoding is performed, and corresponding microenvironment reconstruction encoding is performed based on the two embedding encodings. Then, the free energy change is predicted based on the difference features of the two reconstruction features to obtain the corresponding second free energy change.

[0007] The training dataset is prepared by data collection to obtain the corresponding first and second datasets;

[0008] The codebook retrieval model is trained based on the pre-training framework using the first dataset. After the retrieval model is trained, the internal component parameters of the first and second prediction models are set based on the model parameters of the codebook retrieval model, and then the first and second prediction models are trained based on the second dataset.

[0009] After the two prediction models are trained, the native-mutant protein information is combined based on the first native protein sequence and the first set of mutant residues input by the user to obtain multiple original groups; the first prediction model is used to perform a preliminary screening of all the original groups according to a preset screening and sorting rule to obtain N preferred groups; the second prediction model is used to continue to sort the N preferred groups according to the screening and sorting rule to generate the corresponding final selected group sequence and feed it back to the current user; the first set of mutant residues includes multiple first mutant residues, the total number of the first mutant residues is greater than the preset total number N, where N is a positive integer; the first mutant residues include a first residue type and a first residue site; the original groups include the first native protein sequence, the first mutant protein sequence, and the first residue site; the screening and sorting rule is to sort in descending order of free energy change or in ascending order of free energy change.

[0010] Preferably, the codebook retrieval model includes a heterogeneous graph initialization module, a microenvironment subgraph extraction module, a microenvironment encoder, a codebook retrieval module, and the residue feature codebook C;

[0011] The heterogeneous graph initialization module is implemented based on a trained Uni-Fold model. This module is used to predict the tertiary structure of the protein sequence S using the Uni-Fold model and construct a corresponding heterogeneous graph G based on the predicted structure, which is then sent to the microenvironment subgraph extraction module. The protein sequence S is a one-dimensional residue sequence of a protein complex, composed of multiple residues ordered. The residue site b is a residue sequence index in the protein sequence S. The heterogeneous graph G consists of a node set and an edge set. The node set includes multiple residue nodes, each containing a type embedding vector, a dihedral embedding vector, and a local orientation embedding vector. The edge set includes multiple node edges, each containing a sequence distance embedding vector, a spatial distance embedding vector, and a residue pair orientation embedding vector.

[0012] The microenvironment subgraph extraction module is used to extract the near-neighbor subgraph corresponding to the residue site b in the heterogeneous graph G and send it to the microenvironment encoder as the corresponding microenvironment subgraph M.

[0013] The microenvironment encoder is implemented based on a type of SE3 equivariant graph neural network, which includes the SE(3)-Transformer model, the Equiformer model, and the EGNN model. The microenvironment encoder is used to perform feature encoding on the microenvironment subgraph M to obtain the corresponding microenvironment encoding vector H and send it to the codebook retrieval module.

[0014] The residue feature codebook C includes three sub-codebooks, namely sub-codebook C1, sub-codebook C2, and sub-codebook C3; each of sub-codebooks C1, C2, and C3 contains multiple corresponding token embedding vectors. Token embedding vector Token embedding vector Composition; 1 ≤ index i ≤ N C N C A preset positive integer; token embedding vector The vectors have the same shape and are consistent with the residue nodes; all embedding codes of the residue feature codebook C are initially composed of a set of preset initial codebook codes;

[0015] The codebook retrieval module is used to, according to a preset vector similarity algorithm, identify the N values ​​of the subcodebooks C1, C2, or C3 of the residue feature codebook C. C token embedding vector or The token embedding vector with the highest similarity to the microenvironment encoding vector H is used as the corresponding retrieval codebook vector e. 1 e 2 or e 3 ; and the three retrieved codebook vectors e 1 e 2 e 3 The corresponding residue mapping vector E is formed and output; wherein, the vector similarity algorithm includes cosine vector similarity algorithm and Euclidean distance similarity algorithm.

[0016] Preferably, the step of using the Uni-Fold model to predict the tertiary structure of the protein sequence S and constructing a corresponding isomorphism map G based on the predicted structure specifically includes:

[0017] The tertiary structure of the protein sequence S was predicted using the Uni-Fold model to obtain the atomic-level tertiary structure of the complex.

[0018] The tertiary structure of the complex is then subjected to atomic-level residue clustering to obtain multiple corresponding residue clusters. Each residue cluster is then labeled with four types of main chain atoms on the main chain backbone to obtain corresponding main chain atom groups. Each residue cluster corresponds to one residue. The four types of main chain atoms include nitrogen (N), alpha carbon (C), and carbon atoms. α Carbon atom C, oxygen atom O; the main chain atom group is composed of the four types of main chain atoms;

[0019] And with the alpha carbon atom C of each group of main chain atoms. α Origin, alpha carbon atom C α C to carbon atom C αA corresponding local coordinate system is constructed with the -C bond direction as the X-axis, a straight line passing through the origin and perpendicular to the X-axis in the main chain plane as the Y-axis, and a straight line passing through the origin and perpendicular to the main chain plane as the Z-axis; and the alpha carbon atom C of the first main chain atom group is used as the coordinate system. α Using the first local coordinate system as the global origin, and taking the first local coordinate system as the global coordinate system, establish corresponding local-global coordinate mapping relationships between the remaining local coordinate systems and the global coordinate system;

[0020] One-hot encoding is performed on the corresponding residue type of each residue cluster to obtain the corresponding type embedding vector; and preset quantum chemical tools are used to determine the three types of main chain dihedral angles φ, ψ, ω and the four types of side chain dihedral angles χ of the corresponding residues of each residue cluster. 1 , χ 2 , χ 3 , χ 4 Measurements were performed, and the seven dihedral angles obtained were encoded using dihedral angle embedding to obtain the corresponding dihedral angle embedding vectors. The quantum chemical tools were then used to correlate the nitrogen (N), carbon (C), and oxygen (O) atoms of each residue cluster with the alpha carbon atom (C) in the corresponding local coordinate system. α The relative directions are confirmed, and the three relative directions are encoded using local orientation embedding to obtain the corresponding local orientation embedding vectors; and each residue node is composed of the type embedding vector, the dihedral angle embedding vector, and the local orientation embedding vector; wherein, the quantum chemistry tools include MDAnalysis, BioPython, PyMOL, and VMD.

[0021] The residues in the protein sequence S are arranged in pairs to obtain multiple residue pairs; the relative sequence distance and spatial relative distance in three-dimensional space of the two residues before and after each residue pair in the protein sequence S are confirmed, and the two relative distances are encoded by distance embedding to obtain the corresponding sequence distance embedding vector and the spatial distance embedding vector; the four types of main chain atoms of the second residue of each residue pair are encoded by direction embedding in the four relative directions of the first residue in the local coordinate system to obtain the corresponding residue pair direction embedding vector; and the sequence distance embedding vector, the spatial distance embedding vector and the residue pair direction embedding vector of each residue pair form a corresponding node edge;

[0022] All the obtained residue nodes form a corresponding node set, all the obtained node edges form a corresponding edge set, and the node set and the edge set form the heterogeneous graph G.

[0023] Preferably, the step of extracting the near-neighbor subgraph corresponding to the residue site b in the isomorphic graph G as the corresponding microenvironment subgraph M specifically includes:

[0024] The residue node corresponding to residue site b in the heterogeneous graph G is taken as the current node; and the residue nodes in the heterogeneous graph G that have a total number of nodes and edges with the current node not exceeding a preset maximum hop count, a relative sequence distance from the current node not exceeding a preset sequence distance threshold, and a relative spatial distance from the current node not exceeding a preset spatial distance threshold are extracted to form a corresponding current node set; and the node and edge subsets corresponding to the current node set in the heterogeneous graph G are extracted to form a corresponding current edge set; and the current node set and the current edge set form the corresponding microenvironment subgraph M; wherein, the maximum hop count is a positive integer.

[0025] Preferably, the mask pattern G mask It consists of the node set and the edge set; the mask graph G mask Each residue node corresponds to a mask state information; the mask state information includes a mask marker and a mask type; the mask marker includes unmasked and masked; the mask mode consists of type mask mode, dihedral mask mode, and local orientation mask mode. All three modes include two states: on and off. The on state of the three modes is not uniquely restricted.

[0026] When the mask marker is unmasked, all three mask mode states are off; when the mask marker is masked and the type mask mode is on, the type embedding vector of the residue node is replaced by a randomly generated residue type one-hot encoded vector; when the mask marker is masked and the dihedral mask mode is on, the seven dihedral codes of the dihedral embedding vector of the residue node are set to zero; when the mask marker is masked and the local orientation mask mode is on, Gaussian noise is added to the three noisy orientation codes of the local orientation embedding vector of the residue node.

[0027] The mask image G mask The total number of residue nodes is denoted as the total number of nodes N. all The total number of residue nodes marked as masked and with the masking mode of the specified type enabled is denoted as the total number of first-class nodes N1; the total number of residue nodes marked as masked and with the dihedral angle masking mode enabled is denoted as the total number of second-class nodes N2; and the total number of residue nodes marked as masked and with the local orientation masking mode enabled is denoted as the total number of third-class nodes N3; N1, N2, or N3 are related to N... all The proportions do not exceed the preset maximum mask proportion threshold and are not lower than the preset minimum mask proportion threshold;

[0028] The type reconstruction set X 1 By N all The dihedral reconstruction set X is composed of several reconstruction type embedding vectors. 2 By N all The local orientation reconstruction set X is composed of a reconstructed dihedral embedding vector. 3 By N all It consists of a reconstructed local orientation embedding vector; the data format of the reconstructed type embedding vector, the reconstructed dihedral angle embedding vector, and the reconstructed local orientation embedding vector is consistent with the corresponding type embedding vector, dihedral angle embedding vector, and local orientation embedding vector;

[0029] The pre-training framework includes a full-structure microenvironment extraction module, a microenvironment encoder, a codebook forward lookup module, a codebook reverse lookup module, a residue feature codebook C, a first decoder, a second decoder, and a third decoder;

[0030] The full-structure microenvironment extraction module is used to extract the mask image G. mask Each of the aforementioned residue nodes is taken as the corresponding current node; and the mask image G is used as the current node. mask Residue nodes whose total number of edges between them and the current node does not exceed a preset maximum hop count, whose relative sequence distance to the current node does not exceed a preset sequence distance threshold, and whose relative spatial distance to the current node does not exceed a preset spatial distance threshold are extracted to form a corresponding current node set; and the mask image G is then used to form the current node set. mask The node edge subset corresponding to the current node set is extracted to form the corresponding current edge set; and the current node set and the current edge set form a corresponding micro-environment subgraph M. j 1 ≤ index j ≤ N all ; and from the obtained N all The microenvironment subgraph M j Form the corresponding sub-graph set {M j Send to the microenvironment encoder;

[0031] The microenvironment encoder is used to process the sub-map {M} j The various microenvironment subgraphs M of} j Feature encoding is performed to obtain the corresponding microenvironment encoding vector H. j ; and from the obtained N all The microenvironment encoding vector H j Form the corresponding vector set {H j Send to the codebook lookup module;

[0032] The codebook lookup module is used to find the vector set {H} jThe microenvironment encoding vectors H of each of the aforementioned microenvironment j As the corresponding current vector; and according to the vector similarity algorithm, embed the token in the subcodebook C1 that has the highest similarity to the current vector into the embedding vector. The index i is used as the corresponding codebook marker. Embed the token in the subcodebook C2 that has the highest similarity to the current vector into the vector. The index i is used as the corresponding codebook marker. Embed the token in the subcodebook C3 that has the highest similarity to the current vector into the vector. The index i is used as the corresponding codebook marker. And the three codebook markers obtained Form a corresponding codebook tag vector Z j ; and from the obtained N all The codebook tag vector Z j Form the corresponding vector set {Z j Send to the codebook reverse lookup module;

[0033] The codebook reverse lookup module is used to retrieve the vector set {Z} j The codebook tag vectors Z in} j As the corresponding current vector; and the index i in the subcodebooks C1, C2, and C3 is respectively compared with the codebook tag of the current tag vector. Matching token embedding vector As the corresponding retrieval codebook vector And from the obtained N all A retrieval codebook vector Form the corresponding vector set Send to the first decoder, by the obtained N all A retrieval codebook vector Form the corresponding vector set Send to the second decoder, by the obtained N all A retrieval codebook vector Form the corresponding vector set Send to the third decoder;

[0034] The first decoder is implemented based on a class of SE3 equivariant graph neural networks; the first decoder is used to process the vector set Each of the aforementioned retrieval codebook vectors As a corresponding first transformation node, and derived from N all Each of the first conversion nodes forms a corresponding first conversion node set; and the first conversion node set and the mask image G are combined. maskThe edge sets form a corresponding first transformation graph; and residue type decoding is performed on all the first transformation nodes of the first transformation graph to obtain the corresponding type reconstruction set X. 1 And output;

[0035] The second decoder is implemented based on a class of SE3 equivariant graph neural networks; the second decoder is used to process the vector set Each of the aforementioned retrieval codebook vectors As a corresponding second transformation node, and derived from N all Each of the second conversion nodes forms a corresponding second conversion node set; and the second conversion node set and the mask image G are combined. mask The edge sets form a corresponding second transformation graph; and dihedral decoding is performed on all the second transformation nodes of the second transformation graph to obtain the corresponding dihedral reconstruction set X. 2 And output;

[0036] The third decoder is implemented based on a class of SE3 equivariant graph neural networks; the third decoder is used to process the vector set Each of the aforementioned retrieval codebook vectors As a corresponding third transformation node, and derived from N all Each of the aforementioned third conversion nodes forms a corresponding third conversion node set; and the third conversion node set and the mask image G are combined. mask The edge set forms a corresponding third transformation graph; and local orientation decoding is performed on all the third transformation nodes of the third transformation graph to obtain the corresponding local orientation reconstruction set X. 3 And output it.

[0037] Preferably, the first prediction model includes a first retrieval model, a second retrieval model, a first self-attention module, a second self-attention module, a first difference module, and a first MLP model;

[0038] The model structures of the first retrieval model and the second retrieval model are consistent with the model structure of the codebook retrieval model;

[0039] The first retrieval model is used to search based on the native protein sequence S1 and the mutant residue site b, which are input to the model. 1-2 The built-in residue feature codebook performs residue feature mapping on the microenvironment features of the current residue site to obtain the corresponding residue mapping vector E1, which is then sent to the first self-attention module.

[0040] The second retrieval model is used to retrieve the mutant protein sequence S2 and the mutant residue site b from the model input. 1-2The built-in residue feature codebook performs residue feature mapping on the microenvironment features of the current residue site to obtain the corresponding residue mapping vector E2, which is then sent to the second self-attention module.

[0041] The first self-attention module is used to perform self-attention operation on the residue mapping vector E1 and send the operation result as the corresponding mapping vector P1 to the first difference module.

[0042] The second self-attention module is used to perform self-attention operation on the residue mapping vector E2 and send the operation result as the corresponding mapping vector P2 to the first difference module;

[0043] The first difference module is used to send the difference vector ΔP = P2 - P1 of the mapping vectors P1 and P2 to the first MLP model;

[0044] The first MLP model is used to perform regression prediction on the free energy change based on the difference vector ΔP to obtain the corresponding first free energy change and output it.

[0045] Preferably, the second prediction model includes a third retrieval model, a fourth retrieval model, a third self-attention module, a fourth self-attention module, a first reconstruction module, a second reconstruction module, a first encoder, a second encoder, a second difference module, and a second MLP model;

[0046] The model structures of the third retrieval model and the fourth retrieval model are consistent with the model structure of the codebook retrieval model;

[0047] The third retrieval model is used to search based on the native protein sequence S3 and the mutation residue site b, which are input into the model. 3-4 The built-in residue feature codebook performs residue feature mapping on the microenvironment features of the current residue site to obtain the corresponding residue mapping vector E3, which is then sent to the third self-attention module.

[0048] The fourth retrieval model is used to search based on the mutant protein sequence S4 and the mutant residue site b, which are input into the model. 3-4 The built-in residue feature codebook performs residue feature mapping on the microenvironment features of the current residue site to obtain the corresponding residue mapping vector E4, which is then sent to the fourth self-attention module.

[0049] The third self-attention module is used to perform self-attention operation on the residue mapping vector E3 and send the operation result as the corresponding mapping vector P3 to the first reconstruction module.

[0050] The fourth self-attention module is used to perform self-attention operation on the residue mapping vector E4 and send the operation result as the corresponding mapping vector P4 to the second reconstruction module.

[0051] The first reconstruction module is used to take the heterogeneous graph G and the microenvironment subgraph M generated during the processing of the third retrieval model as the corresponding current global graph and current local graph; and to perform a traversal of all residue nodes in the current local graph; and during this traversal, to take the currently traversed residue node as the corresponding current local node; and to record the current local node as the corresponding pre-superposition vector. And based on the mapping vector P3, the vector before superposition... The corresponding superimposed vector is obtained by superimposing the micro-environment mapping features. Using the superimposed vector The corresponding residue nodes in the current global graph are reset; and at the end of this round of traversal, the mutated residue site b is determined. 3-4 The corresponding residue node and the node set of the current local graph form a new first child node set, and the subgraph corresponding to the first child node set in the current global graph is extracted and sent to the first encoder as the corresponding reconstructed subgraph M1;

[0052] The second reconstruction module is used to take the heterogeneous graph G and the microenvironment subgraph M generated during the processing of the fourth retrieval model as the corresponding current global graph and current local graph; and to perform a traversal of all residue nodes in the current local graph; and during this traversal, to take the currently traversed residue node as the corresponding current local node; and to record the current local node as the corresponding pre-superposition vector. And based on the mapping vector P4, the vector before superposition... The corresponding superimposed vector is obtained by superimposing the micro-environment mapping features. Using the superimposed vector The corresponding residue nodes in the current global graph are reset; and at the end of this round of traversal, the mutated residue site b is determined. 3-4 The corresponding residue node and the node set of the current local graph form a new second child node set, and the subgraph corresponding to the second child node set in the current global graph is extracted and sent to the second encoder as the corresponding reconstructed subgraph M2;

[0053] The model structures of the first encoder and the second encoder are consistent with the model structure of the microenvironment encoder; the model parameters of the first encoder and the second encoder are consistent.

[0054] The first encoder is used to perform feature encoding on the reconstructed subgraph M1 to obtain the corresponding microenvironment encoding vector H1, which is then sent to the second difference module.

[0055] The second encoder is used to perform feature encoding on the reconstructed subgraph M2 to obtain the corresponding microenvironment encoding vector H2, which is then sent to the second difference module.

[0056] The second difference module is used to send the difference vector ΔH = H2 - H1 of the microenvironment encoding vectors H1 and H2 to the second MLP model;

[0057] The second MLP model is used to perform regression prediction on the free energy change based on the difference vector ΔH to obtain the corresponding second free energy change and output it.

[0058] Preferably, the first dataset includes multiple first data records; the first data records include a first label image, a first mask image, a set of first-class mask nodes, a set of second-class mask nodes, and a set of third-class mask nodes; the data structure of the first mask image is the same as that of the mask image G. mask Consistent with each other, the data structure of the first tag image is consistent with that of the heterogeneous image G; the first tag image is a microenvironment heterogeneous image of a single residue; the first mask image is a mask image obtained by performing three types of masking on the first tag image; the first type of mask node set includes multiple first-type node identifiers, the second type of mask node set includes multiple second-type node identifiers, and the third type of mask node set includes multiple third-type node identifiers;

[0059] The second dataset includes multiple second data records; the second data record includes a first training native sequence, a first training mutant sequence, a first training site, and a first tag free energy change; the first training native sequence and the first training mutant sequence are the residue sequences of a protein complex before and after a mutation occurs at the residue site corresponding to the first training site; the first tag free energy change is the free energy change of the current protein complex before and after the mutation.

[0060] Furthermore, the step of training the codebook retrieval model based on the pre-trained framework using the first dataset specifically includes:

[0061] Step 91: Take the first data record of the first dataset as the corresponding current training record;

[0062] Step 92: Extract the corresponding first label image, first mask image, first type mask node set, second type mask node set, and third type mask node set from the current training record as the corresponding current label image, current mask image, current first type node set, current second type node set, and current third type node set;

[0063] Step 93: Use the current mask image as the corresponding mask image G. mask The pre-trained framework is input into the data and processed to obtain the corresponding type reconstruction set X. 1 The dihedral angle reconstruction set X 2 The local orientation reconstruction set X 3 The microenvironment encoding vector set generated by the microenvironment encoder, the codebook tag vector set generated by the codebook forward lookup module, and the three types of retrieval codebook vector sets generated by the codebook reverse lookup module are saved during this processing.

[0064] Step 94: Denote each microenvironment coding vector in the microenvironment coding vector set stored this time as the corresponding H. a 1 ≤ index a ≤ N A N A Let Z be the total number of residue nodes in the current tag graph; and let Z be the corresponding codebook tag vector of the codebook tag vector set stored this time. a ; and mark the vector Z of each codebook a The three types of codebook markers in the text are denoted as the corresponding The retrieval codebook vector sets corresponding to subcodebooks C1, C2, and C3 in the three types of retrieval codebook vector sets stored this time are denoted as the corresponding vector sets. Vector set Vector set Retrieval codebook vector With codebook tags One-to-one correspondence; and the subcodebook C1 and the vector set Each retrieval codebook vector The matching token embedding vector is denoted as The subcodebook C2 and the vector set Each retrieval codebook vector The matching token embedding vector is denoted as The subcodebook C3 and the vector set Each retrieval codebook vector The matching token embedding vector is denoted as And by each codebook's tag vector Corresponding token embedding vector The corresponding token embedding vector set is formed; and each microenvironment encoding vector H is composed of the corresponding token embedding vector set. a The token embedding vector set and its corresponding set together form a first verification pair.

[0065] Step 95, combine the current label graph and the type reconstruction set X 1The type embedding vector and the reconstructed type embedding vector corresponding to each of the node identifiers of the current type of node set are combined to form a corresponding second check pair; and the current label graph and the dihedral reconstruction set X are combined. 2 The dihedral embedding vector and the reconstructed dihedral embedding vector corresponding to each of the two types of node identifiers in the current set of two types of nodes form a corresponding third check pair;

[0066] Wherein, the second check pair is denoted as The third verification pair is denoted as

[0067] 1≤index b≤N B 1 ≤ index c ≤ N C N B N C These are the total number of node identifiers corresponding to the current first-type node set and the current second-type node set, respectively.

[0068] These are the type embedding vector and the reconstructed type embedding vector of the b-th second check pair, respectively; These are the dihedral embedding vector and the reconstructed dihedral embedding vector of the c-th third verification pair, respectively.

[0069] Step 96, reconstruct the current label map and the local orientation set X. 3 In the process, the difference vector between the pair of local orientation embedding vectors and the reconstructed local orientation embedding vectors corresponding to each of the three types of node identifiers in the current three types of node sets is calculated, and the calculation result is used as the corresponding first prediction error vector; and the current mask image and the local orientation reconstruction set X are used to calculate the difference vector between the pair of local orientation embedding vectors and the reconstructed local orientation embedding vectors. 3 In the process, the difference vector between the pair of local orientation embedding vectors and the reconstructed local orientation embedding vectors corresponding to each of the three types of node identifiers in the current three types of node sets is calculated, and the calculation result is used as the corresponding first label error vector; and each of the first label error vectors and its corresponding first prediction error vectors forms a corresponding fourth check pair;

[0070] The fourth check pair is denoted as

[0071] 1≤index d≤N D N D The total number of node identifiers corresponding to the current three types of node sets;

[0072] These are the first label error vector and the first prediction error vector of the dth fourth check pair, respectively.

[0073] Step 97: Construct a first loss function L1 for the microenvironment encoder and the residue feature codebook C based on the constraint loss function; construct a second loss function L2 for the sub-codebook C1 and the first decoder based on the cross-entropy loss function; construct a third loss function L3 for the sub-codebook C2 and the second decoder based on the MSE loss function; construct a fourth loss function L4 for the sub-codebook C3 and the third decoder based on the Huber loss function; and construct an overall loss function L based on the first, second, fourth, and fourth loss functions. total ; and the obtained N A The first verification pair, N B The second check pair, N C The third check pair, N D The fourth check pair is substituted into the overall loss function L. total The corresponding first loss value is obtained through calculation;

[0074] in,

[0075] The first loss function L1 is:

[0076]

[0077] 1≤index s≤3, sg() is the stop gradient operation, and η is the preset tradeoff hyperparameter;

[0078] The second loss function L2 is:

[0079]

[0080] L CE () represents the cross-entropy loss function;

[0081] The third loss function L3 is:

[0082]

[0083] 1 ≤ index q ≤ 7, The dihedral embedding vector The q-th dihedral angle in the middle, For the reconstructed dihedral embedding vector The q-th dihedral angle in;

[0084] The fourth loss function L4 is:

[0085]

[0086] L Huber () represents the Huber loss function;

[0087] The overall loss function L total for:

[0088] L total =λL1+L2+L3+L4;

[0089] λ is a preset tradeoff hyperparameter;

[0090] Step 98: Identify whether the first loss value meets a preset first loss value range; if the first loss value meets the first loss value range, identify whether the current training record is the last first data record in the first dataset; if so, proceed to step 99; otherwise, take the next first data record in the first dataset as the new current training record and return to step 92; if the first loss value does not meet the first loss value range, first optimize the microenvironment encoder and the residue features based on a preset first model optimizer in the direction of minimizing the first loss function L1. Codebook C undergoes one round of parameter optimization. Then, based on a preset second model optimizer, a round of parameter optimization is performed on the sub-codebook C1 and the first decoder, aiming to minimize the second loss function L2. Next, based on a preset third model optimizer, a round of parameter optimization is performed on the sub-codebook C2 and the second decoder, aiming to minimize the third loss function L3. Finally, based on a preset fourth model optimizer, a round of parameter optimization is performed on the sub-codebook C3 and the third decoder, aiming to minimize the fourth loss function L4. Finally, based on a preset fifth model optimizer, a round of parameter optimization is performed on the overall loss function L... total The direction that reaches the minimum value is used to perform a round of parameter optimization on the microenvironment encoder and the residue feature codebook C, and the process returns to step 93 when the parameter optimization ends.

[0091] Wherein, the first model optimizer includes the Adam optimizer and the SGD optimizer; the second model optimizer includes the Adam optimizer and the SGD optimizer; the third model optimizer includes the Adam optimizer and the SGD optimizer; the fourth model optimizer includes the Adam optimizer and the SGD optimizer; and the fifth model optimizer includes the Adam optimizer and the SGD optimizer.

[0092] Step 99: Stop the current training and confirm that the retrieval model training has ended.

[0093] Furthermore, the step of first setting the internal component parameters of the first and second prediction models based on the model parameters of the codebook retrieval model, and then training the first and second prediction models based on the second dataset, specifically includes:

[0094] Based on the model parameters of the codebook retrieval model and the built-in residue feature codebook C, the model parameters of the first and second retrieval models inside the first prediction model and the built-in residue feature codebook are set, and the parameters of the first and second retrieval models are frozen after the setting is completed.

[0095] Based on the model parameters of the codebook retrieval model and the built-in residue feature codebook C, the model parameters of the third and fourth retrieval models inside the second prediction model and the built-in residue feature codebook are set, and the parameters of the third and fourth retrieval models are frozen after the settings are completed.

[0096] And based on the model parameters of the microenvironment encoder of the codebook retrieval model, the model parameters of the first and second encoders inside the second prediction model are initialized;

[0097] After the parameters of the first and second prediction models are set, the first prediction model is trained based on the second dataset, and the second prediction model is trained based on the second dataset.

[0098] After the training of both the first and second prediction models is completed, confirm that the training of both prediction models is finished.

[0099] Preferably, the step of combining native-mutant protein information based on the user-input first native protein sequence and first set of mutant residues to obtain multiple corresponding original groups specifically includes:

[0100] Each of the first mutated residues in the first set of mutated residues is taken as the corresponding current mutation; the first residue type and the first residue site of the current mutation are taken as the corresponding current mutated residue and current mutated site; a new residue sequence is obtained by replacing the original residues in the first native protein sequence corresponding to the current mutated site with the current mutated residues, and this new residue sequence is denoted as the corresponding first mutated protein sequence; the current mutated site is taken as a corresponding first residue site; and the first native protein sequence corresponding to the current mutation, the first mutated protein sequence, and the first residue site form a corresponding original set.

[0101] Preferably, the step of using the first prediction model to perform a preliminary screening of all the original groups according to a preset screening and sorting rule to obtain N preferred groups specifically includes:

[0102] The first native protein sequence, the first mutant protein sequence, and the first residue site of each of the original groups are taken as a corresponding set of native protein sequence S1, mutant protein sequence S2, and mutant residue site b. 1-2The first prediction model is input and processed to obtain the corresponding first free energy variable; all the obtained first free energy variables are sorted according to the screening and sorting rules; and the N original groups corresponding to the top N first free energy variables are extracted as the corresponding N preferred groups.

[0103] Preferably, the step of using the second prediction model to continue sorting the N preferred groups according to the screening and sorting rules to generate the corresponding final selection group sequence and feeding it back to the current user specifically includes:

[0104] The first native protein sequence, the first mutant protein sequence, and the first residue site of each of the preferred groups are taken as a corresponding set of native protein sequence S3, mutant protein sequence S4, and mutant residue site b. 3-4 The second prediction model is input and processed to obtain the corresponding second free energy change; and the first native protein sequence, the first mutant protein sequence, the first residue site and the second free energy change corresponding to each of the preferred groups are combined to form a corresponding final selection group; and the N final selection groups are sorted according to the screening and sorting rules to obtain the corresponding final selection group sequence and fed back to the current user.

[0105] A second aspect of the present invention provides an apparatus for implementing the processing method for screening mutation sites based on free energy variation as described in the first aspect above. The apparatus includes: a model building module, a data acquisition module, a model training module, and a model application module.

[0106] The model building module is used to construct a feature mapping model for mapping residue features of microenvironmental features of specified residue sites in a protein complex using a codebook retrieval method, denoted as the corresponding codebook retrieval model; and to construct a pre-training framework for the codebook retrieval model; and to construct two prediction models based on the codebook retrieval model for predicting the free energy change ΔΔG before and after a single-point mutation in the protein complex, denoted as the first prediction model and the second prediction model; the codebook retrieval model is used to perform residue feature mapping on the microenvironmental features of the current residue site according to the protein sequence S and residue site b input to the model, as well as the built-in residue feature codebook C, to obtain the corresponding residue mapping vector E; the pre-training framework is used to process the mask image G input to the framework. mask The three types of embedding vectors of all residue nodes are reconstructed to obtain the corresponding type reconstruction set X. 1 Dihedral Reconstruction Set X 2 Local orientation reconstruction set X 3 The first prediction model is used to predict the original protein sequence S1, the mutant protein sequence S2, and the mutant residue site b based on the input to the model. 1-2Microenvironment embedding encoding is performed, and the free energy change is predicted based on the differential embedding of the two embedding encodings to obtain the corresponding first free energy change; the second prediction model is used to predict the free energy change based on the native protein sequence S3, the mutant protein sequence S4, and the mutant residue site b as input to the model. 3-4 Microenvironment embedding encoding is performed, and corresponding microenvironment reconstruction encoding is performed based on the two embedding encodings. Then, the free energy change is predicted based on the difference features of the two reconstruction features to obtain the corresponding second free energy change.

[0107] The data acquisition module is used to prepare the training dataset through data acquisition to obtain the corresponding first dataset and second dataset;

[0108] The model training module is used to train the codebook retrieval model based on the pre-training framework and the first dataset. After the retrieval model training is completed, the internal component parameters of the first and second prediction models are set based on the model parameters of the codebook retrieval model, and then the first and second prediction models are trained based on the second dataset.

[0109] The model application module is used to, after the training of the two types of prediction models, combine native-mutant protein information based on the first native protein sequence and the first set of mutant residues input by the user to obtain multiple corresponding original groups; and use the first prediction model to perform a preliminary screening of all the original groups according to a preset screening and sorting rule to obtain N preferred groups; and use the second prediction model to continue to sort the N preferred groups according to the screening and sorting rule to generate the corresponding final selected group sequence and feed it back to the current user; the first set of mutant residues includes multiple first mutant residues, the total number of the first mutant residues is greater than the preset total number N, where N is a positive integer; the first mutant residues include a first residue type and a first residue site; the original groups include the first native protein sequence, the first mutant protein sequence, and the first residue site; the screening and sorting rule is to sort in descending order of free energy change or in ascending order of free energy change.

[0110] A third aspect of the present invention provides an electronic device, including: a memory, a processor, and a transceiver;

[0111] The processor is used to couple with the memory, read and execute instructions in the memory to implement the steps of the method described in the first aspect above;

[0112] The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

[0113] A fourth aspect of the present invention provides a computer-readable storage medium storing computer instructions that, when executed by a computer, cause the computer to perform the instructions described in the first aspect.

[0114] This invention provides a processing method, apparatus, electronic device, and computer-readable storage medium for screening mutation sites based on free energy change. As described above, this invention constructs three types of tool models to improve the screening efficiency of mutation strategies: 1) a codebook retrieval model for mapping residue features of microenvironmental characteristics of specified residue sites in protein complexes using codebook retrieval, and a pre-training framework is constructed for this codebook retrieval model; 2) a coarse calculation model (i.e., the first prediction model) based on the codebook retrieval model, which first learns the two microenvironmental mapping features of a single mutation site before and after mutation, and then predicts the free energy change ΔΔG before and after a single-point mutation based on the difference features of the two microenvironmental mapping features; 3) a fine calculation model (i.e., the second prediction model) based on the codebook retrieval model, which first learns the two microenvironmental mapping features of a single mutation site before and after mutation, then reconstructs two local heteromorphic maps before and after mutation based on the two microenvironmental mapping features, and then predicts the free energy change ΔΔG before and after a single-point mutation based on the difference features of the two reconstructed local heteromorphic maps. Two training datasets (dataset 1 and dataset 2) were constructed for the three types of tool models. A two-stage training strategy was designed for the three types of tool models: 1) First stage: the codebook retrieval model is trained based on the pre-trained framework using the first dataset; 2) Second stage: the internal component parameters of the first and second prediction models are set based on the codebook retrieval model trained in the first stage, and then the first and second prediction models are trained separately based on the second dataset. After training, the two prediction models are used for mutation strategy screening: first, the native protein sequence and mutant residue set input by the user are combined to obtain multiple original groups; then, the first prediction model is used to perform preliminary screening of all original groups according to the screening and sorting rules to obtain N preferred groups; then, the second prediction model is used to re-sort the N preferred groups according to the screening and sorting rules to obtain the final selected group sequence and provide feedback to the user. Through the two end-to-end prediction models in this embodiment of the invention, not only is the overall calculation cycle shortened, the overall calculation cost reduced, and the overall research efficiency improved; the first prediction model, with its higher prediction efficiency, further improves the efficiency of massive pre-screening; and the second prediction model, with its higher prediction accuracy, further improves the prediction accuracy of the final selection results. Attached Figure Description

[0115] Figure 1 This is a schematic diagram of a method for screening mutation sites based on free energy variation provided in Embodiment 1 of the present invention;

[0116] Figure 2 This is a schematic diagram of the codebook retrieval model provided in Embodiment 1 of the present invention;

[0117] Figure 3 A schematic diagram of the modules of the pre-training framework provided in Embodiment 1 of the present invention;

[0118] Figure 4 A schematic diagram of the modules of the first prediction model provided in Embodiment 1 of the present invention;

[0119] Figure 5 This is a schematic diagram of the modules of the second prediction model provided in Embodiment 1 of the present invention;

[0120] Figure 6 This is a module structure diagram of a processing device for screening mutation sites based on free energy change, provided in Embodiment 2 of the present invention;

[0121] Figure 7 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. Detailed Implementation

[0122] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0123] Embodiment 1 of the present invention provides a method for screening mutation sites based on free energy variation, such as... Figure 1 The schematic diagram shows a method for screening mutation sites based on free energy variation provided in Embodiment 1 of the present invention. The method mainly includes the following steps:

[0124] Step 1: Construct a feature mapping model for mapping the microenvironmental features of specified residue sites of a protein complex using codebook retrieval, denoted as the corresponding codebook retrieval model; construct a pre-training framework for the codebook retrieval model; and construct two prediction models based on the codebook retrieval model to predict the free energy change ΔΔG before and after a single-point mutation of the protein complex, denoted as the first prediction model and the second prediction model.

[0125] Here, as Figure 2The schematic diagram of the codebook retrieval model provided in Embodiment 1 of the present invention is shown. The codebook retrieval model of the present invention is used to perform residue feature mapping on the microenvironment features of the current residue site based on the protein sequence S and residue site b input to the model and the built-in residue feature codebook C to obtain the corresponding residue mapping vector E.

[0126] like Figure 2 As shown, the model components of the codebook retrieval model include: a heterogeneous graph initialization module, a microenvironment subgraph extraction module, a microenvironment encoder, a codebook retrieval module, and a residue feature codebook C.

[0127] The codebook retrieval model has a first input terminal for receiving the protein sequence S, a second input terminal for receiving residue sites b, and an output terminal for outputting the corresponding residue mapping vector E. The model components of the codebook retrieval model are connected as follows: the input terminal of the heterogeneous graph initialization module is connected to the first model input terminal, and its output terminal is connected to the first input terminal of the microenvironment subgraph extraction module; the second input terminal of the microenvironment subgraph extraction module is connected to the second model input terminal, and its output terminal is connected to the input terminal of the microenvironment encoder; the output terminal of the microenvironment encoder is connected to the input terminal of the codebook retrieval module; the output terminal of the codebook retrieval module is connected to the model output terminal, and the codebook retrieval module is also connected to the residue feature codebook C.

[0128] The functionality of the model components in the codebook retrieval model is shown below.

[0129] 1) Heterogeneous graph initialization module:

[0130] The heterogeneity initialization module is implemented based on the trained Uni-Fold model. Here, the Uni-Fold model is a general model for constructing three-dimensional conformations, i.e., tertiary structures, from one-dimensional residue sequences. Detailed information about the model can be found in the publicly available technical document "Uni-Fold: An Open-Source Platform for Developing ProteinFolding Models beyond AlphaFold", and will not be elaborated further here.

[0131] The heterogeneous graph initialization module is used to predict the tertiary structure of protein sequence S using the Uni-Fold model and construct the corresponding heterogeneous graph G based on the predicted structure, which is then sent to the microenvironment subgraph extraction module.

[0132] Wherein, protein sequence S is a one-dimensional residue sequence of a protein complex, which is composed of multiple residues arranged in order; residue site b is a residue sequence index in protein sequence S; heterogeneous graph G consists of node set and edge set; node set includes multiple residue nodes, and residue nodes include type embedding vector, dihedral embedding vector, and local orientation embedding vector; edge set includes multiple node edges, and node edges include sequence distance embedding vector, spatial distance embedding vector, and residue pair orientation embedding vector.

[0133] The specific processing steps for using the Uni-Fold model to predict the tertiary structure of protein sequence S and constructing the corresponding isomorphism map G based on the predicted structure are as follows:

[0134] Step A1: Use the Uni-Fold model to predict the tertiary structure of protein sequence S to obtain the atomic-level tertiary structure of the complex;

[0135] Step A2 involves atomic-level residue clustering of the tertiary structure of the complex to obtain multiple corresponding residue clusters; and labeling the four types of main chain atoms on the main chain backbone of each residue cluster to obtain the corresponding main chain atom group.

[0136] Each residue cluster corresponds to one residue; the four types of main chain atoms include nitrogen (N), alpha carbon (C) atoms, and so on. α Carbon atom C, oxygen atom O; the main chain atom group is composed of four types of main chain atoms;

[0137] Step A3, and with the alpha carbon atoms C of each main chain atom group α Origin, alpha carbon atom C α C to carbon atom C α A corresponding local coordinate system is constructed with the -C bond direction as the X-axis, a line passing through the origin and perpendicular to the X-axis in the main chain plane as the Y-axis, and a line passing through the origin and perpendicular to the main chain plane as the Z-axis; and the alpha carbon atom C of the first main chain atom group is used as the coordinate system. α Using the global origin and the first local coordinate system as the global coordinate system, establish corresponding local-global coordinate mapping relationships between the remaining local coordinate systems and the global coordinate system;

[0138] Step A4 involves one-hot encoding of the corresponding residue types for each residue cluster to obtain the corresponding type embedding vector; and using pre-defined quantum chemical tools, the three main chain dihedral angles φ, ψ, and ω and the four side chain dihedral angles χ for each residue cluster are determined. 1 , χ 2 , χ 3 , χ 4Measurements were performed, and the seven dihedral angles obtained were encoded using dihedral angle embedding to obtain the corresponding dihedral angle embedding vectors. Quantum chemical tools were then used to correlate the nitrogen (N), carbon (C), and oxygen (O) atoms of each residue cluster with the alpha carbon atom (C) in their respective local coordinate systems. α The relative directions are confirmed, and the three relative directions are encoded by local orientation embedding to obtain the corresponding local orientation embedding vectors; and each residue node is composed of the type embedding vector, dihedral angle embedding vector, and local orientation embedding vector.

[0139] Among them, quantum chemistry tools include MDAna ys is tool, BioPython tool, PyMOL tool, and VMD tool;

[0140] Step A5 involves arranging residues in protein sequence S in pairs to obtain multiple residue pairs; confirming the relative sequence distance and spatial distance in three-dimensional space between the two residues preceding and following each residue pair in protein sequence S; performing distance embedding encoding on the obtained two relative distances to obtain corresponding sequence distance embedding vectors and spatial distance embedding vectors; performing direction embedding encoding on the four relative directions of the four types of main chain atoms of the following residue in the local coordinate system of the preceding residue in each residue pair to obtain the corresponding residue pair direction embedding vector; and forming a corresponding node edge by the sequence distance embedding vector, spatial distance embedding vector, and residue pair direction embedding vector of each residue pair.

[0141] Step A6, and form a corresponding node set from all the obtained residue nodes, form a corresponding edge set from all the obtained node edges, and form a heterogeneous graph G from the node set and the edge set.

[0142] 2) Micro-environment subgraph extraction module:

[0143] The microenvironment subgraph extraction module is used to extract the near-neighbor subgraphs corresponding to residue sites b in the heterogeneous graph G and send them to the microenvironment encoder as the corresponding microenvironment subgraph M.

[0144] The specific processing steps for extracting the near-neighbor subgraph corresponding to residue site b in the isomorphic graph G as the corresponding microenvironment subgraph M are as follows:

[0145] The residue node corresponding to residue site b in the heterogeneous graph G is taken as the current node; the residue nodes in the heterogeneous graph G that have a total number of nodes and edges with the current node not exceeding the preset maximum number of hops, a relative sequence distance from the current node not exceeding the preset sequence distance threshold, and a relative spatial distance from the current node not exceeding the preset spatial distance threshold are extracted to form the corresponding current node set; the node edge subset corresponding to the current node set in the heterogeneous graph G is extracted to form the corresponding current edge set; and the current node set and the current edge set form the corresponding microenvironment subgraph M;

[0146] Among them, the maximum number of hops is a positive integer, and the sequence distance threshold and spatial distance threshold are two pre-set distance threshold parameters.

[0147] 3) Microenvironment encoder:

[0148] The microenvironment encoder is implemented based on a class of SE3 equivariant graph neural networks. The SE3 equivariant networks in this embodiment include the SE(3)-Transformer model, the Equiform model, and the EGNN model.

[0149] The microenvironment encoder is used to perform feature encoding on the microenvironment subgraph M to obtain the corresponding microenvironment encoding vector H, which is then sent to the codebook retrieval module.

[0150] 4) Residue signature codebook C:

[0151] The residue feature codebook C of this embodiment includes three sub-codebooks, namely sub-codebook C1, sub-codebook C2 and sub-codebook C3; each of sub-codebooks C1, C2 and C3 contains multiple corresponding token embedding vectors. Token embedding vector Token embedding vector Composition; 1 ≤ index i ≤ N C N C A preset positive integer; token embedding vector Token embedding vector Token embedding vector The vector shape is the same as and consistent with the vector shape of the residue node; all embedding codes of the residue feature codebook C are initially composed of a set of preset initial codebook codes.

[0152] The function of the residue feature codebook C is to map the microenvironment features to the residue feature space and generate a virtual residue feature with the same shape as the residue feature, namely the residue mapping vector E in the following text. The residue mapping vector E carries the overall perception information of the microenvironment. Superimposing the residue mapping vector E with the residue feature can increase the information richness of the residue node.

[0153] 5) Codebook retrieval module:

[0154] The codebook retrieval module is used to retrieve the N values ​​of subcodebooks C1, C2, or C3 of the residue feature codebook C according to a preset vector similarity algorithm. C token embedding vector or The token embedding vector with the highest similarity to the microenvironment encoding vector H is used as the corresponding retrieval codebook vector e. 1 e 2 or e 3 ; and the three retrieved codebook vectors e 1 e 2 e 3 The corresponding residue mapping vector E is constructed and output.

[0155] Here, the vector similarity algorithm in this embodiment of the invention includes the cosine vector similarity algorithm and the Euclidean distance similarity algorithm.

[0156] like Figure 3 As shown in the schematic diagram of the pre-training framework provided in Embodiment 1 of the present invention, the pre-training framework of this embodiment is used for processing the mask image G of the framework input. mask The three types of embedding vectors of all residue nodes are reconstructed to obtain the corresponding type reconstruction set X. 1 Dihedral Reconstruction Set X 2 Local orientation reconstruction set X 3 .

[0157] Mask diagram G of this invention embodiment mask It consists of a set of nodes and a set of edges; the mask graph G mask Each residue node corresponds to a mask state information; the mask state information includes mask marker and mask type; the mask marker includes unmasked and masked; the mask mode consists of type mask mode, dihedral mask mode, and local orientation mask mode. All three modes include two states: on and off. The on state of the three modes is not uniquely restricted.

[0158] Specifically, when the mask is marked as unmasked, all three mask modes are off; when the mask is marked as masked and the type mask mode is on, the type embedding vector of the residue node is replaced by a randomly generated one-hot encoding vector of the residue type; when the mask is marked as masked and the dihedral mask mode is on, the seven dihedral codes of the dihedral embedding vector of the residue node are set to zero; when the mask is marked as masked and the local orientation mask mode is on, the three noisy orientation codes of the local orientation embedding vector of the residue node are added with Gaussian noise.

[0159] Mask diagram G of the embodiment of the present invention mask The total number of residue nodes is denoted as the total number of nodes N. allThe total number of residue nodes marked as masked and with the type masking mode enabled is denoted as the total number of Class 1 nodes N1; the total number of residue nodes marked as masked and with the dihedral angle masking mode enabled is denoted as the total number of Class 2 nodes N2; and the total number of residue nodes marked as masked and with the local orientation masking mode enabled is denoted as the total number of Class 3 nodes N3; N1, N2, or N3 are related to N... all The proportions are all not greater than the preset maximum mask proportion threshold and are not lower than the preset minimum mask proportion threshold. Here, 0 < minimum mask proportion threshold < maximum mask proportion threshold < 1, and the maximum and minimum mask proportion thresholds are two preset threshold parameters.

[0160] Type Reconstruction Set X of this Invention Embodiment 1 By N all The dihedral reconstruction set X is composed of several embedded vectors of different reconstruction types. 2 By N all The local orientation reconstruction set X is composed of several reconstructed dihedral embedding vectors. 3 By N all It consists of a reconstructed local orientation embedding vector; the data format of the reconstructed type embedding vector, the reconstructed dihedral angle embedding vector, and the reconstructed local orientation embedding vector is consistent with the corresponding type embedding vector, dihedral angle embedding vector, and local orientation embedding vector.

[0161] like Figure 3 As shown, the input of the pre-trained framework is used to receive the mask image G. mask The first frame output is used to output the type reconstruction set X. 1 The second frame output is used to output the dihedral angle reconstruction set X. 2 The third frame output is used to output the local orientation reconstruction set X. 3 The pre-trained framework consists of the following components: a full-structure microenvironment extraction module, a microenvironment encoder, a codebook forward lookup module, a codebook reverse lookup module, a residue feature codebook C, a first decoder, a second decoder, and a third decoder.

[0162] The connection relationships of the pre-trained framework components are as follows: the input end of the full-structure microenvironment extraction module is connected to the framework input end, and the output end is connected to the input end of the microenvironment encoder; the output end of the microenvironment encoder is connected to the input end of the codebook forward lookup module; the output end of the codebook forward lookup module is connected to the input end of the codebook reverse lookup module, and the codebook forward lookup module is also connected to the residue feature codebook C; the first, second, and third output ends of the codebook reverse lookup module are connected to the input ends of the corresponding first, second, and third decoders, respectively, and the codebook reverse lookup module is also connected to the residue feature codebook C; the output ends of the first, second, and third decoders are connected to the corresponding first, second, and third framework output ends, respectively.

[0163] The framework components of the pre-trained framework are shown below.

[0164] 1) Full-structure microenvironment extraction module:

[0165] The full-structure microenvironment extraction module is used to extract the mask image G mask Each residue node is taken as the corresponding current node; and the mask image G is used as the current node. mask Residue nodes whose total number of edges between them and the current node does not exceed a preset maximum hop count, whose relative sequence distance to the current node does not exceed a preset sequence distance threshold, and whose relative spatial distance to the current node does not exceed a preset spatial distance threshold are extracted to form the corresponding current node set; and the mask image G is then used to form the current node set. mask The node edges corresponding to the current node set are extracted to form the corresponding current edge set; and the current node set and the current edge set form a corresponding micro-environment subgraph M. j 1 ≤ index j ≤ N all ; and from the obtained N all Microenvironment subgraph M j Form the corresponding sub-graph set {M j Send to the microenvironment encoder.

[0166] 2) Microenvironment encoder:

[0167] Microenvironment encoders are used for sub-atlases {M} j The various microenvironment subgraphs M of} j Feature encoding is performed to obtain the corresponding microenvironment encoding vector H. j ; and from the obtained N all A microenvironment encoding vector H j Form the corresponding vector set {H j Send to the codebook lookup module.

[0168] 3) Codebook lookup module:

[0169] The codebook lookup module is used to look up the vector set {H} j The encoding vectors H of each microenvironment of} j As the corresponding current vector; and according to the vector similarity algorithm, embed the token with the highest similarity to the current vector in subcodebook C1 into the embedding vector. The index i is used as the corresponding codebook marker. Embed the token in subcodebook C2 that has the highest similarity to the current vector into the vector. The index i is used as the corresponding codebook marker. Embed the token in subcodebook C3 that has the highest similarity to the current vector into the vector. The index i is used as the corresponding codebook marker. And the three codebook markers obtained Form a corresponding codebook tag vector Z j; and from the obtained N al l A codebook tag vector Z j Form the corresponding vector set {Z j Send to the codebook reverse lookup module.

[0170] 4) Codebook reverse lookup module:

[0171] The codebook reverse lookup module is used to transform the vector set {Z} j The codebook tag vectors Z in} j As the corresponding current vector; and respectively, the index i in the subcodebooks C1, C2, and C3 is compared with the codebook label of the current label vector. Matching token embedding vector As the corresponding retrieval codebook vector And from the obtained N al l A retrieval codebook vector Form the corresponding vector set Send to the first decoder, and obtain N al l A retrieval codebook vector Form the corresponding vector set Send to the second decoder, and obtain N al l A retrieval codebook vector Form the corresponding vector set Send to the third decoder.

[0172] 5) First decoder:

[0173] The first decoder is implemented based on a type of SE3 equivariant graph neural network.

[0174] The first decoder is used to process the vector set Each retrieval codebook vector As a corresponding first transformation node, and derived from N all Each first transformation node forms a corresponding first transformation node set; and the first transformation node set and the mask image G are combined. mask The edge sets form a corresponding first transformation graph; and residue type decoding is performed on all first transformation nodes of the first transformation graph to obtain the corresponding type reconstruction set X. 1 And output it.

[0175] 6) Second decoder:

[0176] The second decoder is implemented based on a type of SE3 equivariant graph neural network.

[0177] The second decoder is used to convert the vector set Each retrieval codebook vector As a corresponding second transformation node, and derived from N all Each second transformation node forms a corresponding second transformation node set; and the second transformation node set and the mask image G are combined. mask The edge sets form a corresponding second transformation graph; and dihedral decoding is performed on all second transformation nodes of the second transformation graph to obtain the corresponding dihedral reconstruction set X. 2 And output it.

[0178] 7) Third decoder:

[0179] The third decoder is implemented based on a type of SE3 equivariant graph neural network.

[0180] The third decoder is used to convert the vector set Each retrieval codebook vector As a corresponding third transformation node, and derived from N all Each third transformation node forms a corresponding third transformation node set; and the third transformation node set and the mask image G are combined. mask The edge sets form a corresponding third transformation graph; and local orientation decoding is performed on all third transformation nodes of the third transformation graph to obtain the corresponding local orientation reconstruction set X. 3 And output it.

[0181] like Figure 4 As shown in the schematic diagram of the first prediction model provided in Embodiment 1 of the present invention, the first prediction model of the present invention is used to predict the original protein sequence S1, the mutant protein sequence S2, and the mutant residue site b based on the input of the model. 1-2 Microenvironment embedding coding is performed, and the corresponding first free energy change is obtained by predicting the free energy change based on the differential embedding of the two embedding codes.

[0182] like Figure 4 As shown, the first model input of the first prediction model is used to receive the native protein sequence S1, the second model input is used to receive the mutated protein sequence S2, and the third model input is used to receive the mutated residue site b. 1-2 The model output is used to output the corresponding first free energy change. The model components of the first prediction model include: a first retrieval model, a second retrieval model, a first self-attention module, a second self-attention module, a first difference module, and a first MLP model.

[0183] The model component connections of the first prediction model are as follows: the first and second inputs of the first retrieval model are connected to the first model input and the third model input, respectively, and the output is connected to the input of the first self-attention module; the first and second inputs of the second retrieval model are connected to the second model input and the third model input, respectively, and the output is connected to the input of the second self-attention module; the outputs of the first and second self-attention modules are connected to the first and second inputs of the first difference module, respectively; the output of the first difference module is connected to the input of the first MLP model; and the output of the first MLP model is connected to the model output.

[0184] The model component functions of the first prediction model are shown below.

[0185] 1) First retrieval model, second retrieval model:

[0186] The model structure and function of the first retrieval model and the second retrieval model are consistent with the codebook retrieval model, and each of them also has a built-in residue feature codebook with the same format and size as the residue feature codebook C.

[0187] The first search model is used to search based on the native protein sequence S1 and the mutation residue site b input to the model. 1-2 The built-in residue feature codebook performs residue feature mapping on the microenvironment features of the current residue site to obtain the corresponding residue mapping vector E1, which is then sent to the first self-attention module.

[0188] The second search model is used to search based on the mutant protein sequence S2 and the mutant residue site b input to the model. 1-2 The built-in residue feature codebook performs residue feature mapping on the microenvironment features of the current residue site to obtain the corresponding residue mapping vector E2, which is then sent to the second self-attention module.

[0189] 2) First self-attention module, second self-attention module:

[0190] The first self-attention module is used to perform self-attention operation on the residue mapping vector E1 and send the operation result as the corresponding mapping vector P1 to the first difference module.

[0191] The second self-attention module is used to perform self-attention operation on the residue mapping vector E2 and send the operation result as the corresponding mapping vector P2 to the first difference module.

[0192] 3) First difference module:

[0193] The first difference module is used to send the difference vector ΔP = P2 - P1 of the mapping vectors P1 and P2 to the first MLP model.

[0194] 4) First MLP model:

[0195] The first MLP model is used to perform regression prediction on the free energy change based on the difference vector ΔP to obtain and output the corresponding first free energy change.

[0196] like Figure 5 As shown in the schematic diagram of the second prediction model provided in Embodiment 1 of the present invention, the second prediction model of the present invention is used to predict the original protein sequence S3, the mutant protein sequence S4, and the mutant residue site b based on the input of the model. 3-4 Microenvironment embedding encoding is performed, and corresponding microenvironment reconstruction encoding is performed based on the two embedding encodings. Then, the free energy change is predicted based on the difference features of the two reconstruction features to obtain the corresponding second free energy change.

[0197] like Figure 5 As shown, the first input terminal of the second prediction model is used to receive the native protein sequence S3, the second input terminal is used to receive the mutated protein sequence S4, and the third input terminal is used to receive the mutated residue site b. 3-4 The model output is used to output the corresponding second free energy change. The model components of the second prediction model include: a third retrieval model, a fourth retrieval model, a third self-attention module, a fourth self-attention module, a first reconstruction module, a second reconstruction module, a first encoder, a second encoder, a second difference module, and a second MLP model.

[0198] The connection relationships of the model components of the second prediction model are as follows: the first and second input terminals of the third retrieval model are connected to the input terminals of the first and third models, respectively, and the output terminal is connected to the input terminal of the third self-attention module; the first and second input terminals of the fourth retrieval model are connected to the input terminals of the second and third models, respectively, and the output terminal is connected to the input terminal of the fourth self-attention module; the output terminals of the third and fourth self-attention modules are connected to the input terminals of the corresponding first and second reconstruction modules, respectively; the output terminals of the first and second reconstruction modules are connected to the input terminals of the corresponding first and second encoders, respectively; the output terminals of the first and second encoders are connected to the first and second input terminals of the second difference module, respectively; the output terminal of the second difference module is connected to the input terminal of the second MLP model; and the output terminal of the second MLP model is connected to the model output terminal.

[0199] The functions of the model components in the second prediction model are shown below.

[0200] 1) Third retrieval model and fourth retrieval model:

[0201] The third and fourth retrieval models have the same model structure and model function as the codebook retrieval model, and each of them also has a built-in residue feature codebook with the same format and size as the residue feature codebook C.

[0202] The third search model is used to search based on the native protein sequence S3 and the mutation residue site b input to the model. 3-4 The built-in residue feature codebook performs residue feature mapping on the microenvironment features of the current residue site to obtain the corresponding residue mapping vector E3, which is then sent to the third self-attention module.

[0203] The fourth search model is used to search for the mutant protein sequence S4 and the mutant residue site b as input to the model. 3-4 The built-in residue feature codebook performs residue feature mapping on the microenvironment features of the current residue site to obtain the corresponding residue mapping vector E4, which is then sent to the fourth self-attention module.

[0204] 2) Third self-attention module and fourth self-attention module:

[0205] The third self-attention module is used to perform self-attention operation on the residue mapping vector E3 and send the operation result as the corresponding mapping vector P3 to the first reconstruction module.

[0206] The fourth self-attention module is used to perform self-attention operation on the residue mapping vector E4 and send the operation result as the corresponding mapping vector P4 to the second reconstruction module.

[0207] 3) First Restructuring Module and Second Restructuring Module:

[0208] The first reconstruction module is used to take the heterogeneous graph G and the microenvironment subgraph M generated during the processing of the third retrieval model as the corresponding current global graph and current local graph; and to perform a round of traversal on all residue nodes of the current local graph; and during this round of traversal, to take the currently traversed residue node as the corresponding current local node; and to record the current local node as the corresponding pre-superposition vector. And based on the mapping vector P3, the vectors before superposition are... The corresponding superimposed vector is obtained by superimposing the micro-environment mapping features. And use the superimposed vector Reset the corresponding residue node in the current global graph; and at the end of this round of traversal, use the mutated residue site b. 3-4 The corresponding residue node and the node set of the current local graph form a new first child node set, and the subgraph corresponding to the first child node set in the current global graph is extracted as the corresponding reconstructed subgraph M1 and sent to the first encoder.

[0209] The second reconstruction module is used to take the heterogeneous graph G and the microenvironment subgraph M generated during the processing of the fourth retrieval model as the corresponding current global graph and current local graph; and to perform a round of traversal on all residue nodes of the current local graph; and during this round of traversal, to take the currently traversed residue node as the corresponding current local node; and to record the current local node as the corresponding pre-superposition vector. And based on the mapping vector P4, the vectors before superposition are... The corresponding superimposed vector is obtained by superimposing the micro-environment mapping features. And use the superimposed vector Reset the corresponding residue node in the current global graph; and at the end of this round of traversal, use the mutated residue site b. 3-4 The corresponding residue node and the node set of the current local graph form a new second child node set, and the subgraph corresponding to the second child node set in the current global graph is extracted and sent to the second encoder as the corresponding reconstructed subgraph M2.

[0210] 4) First encoder, second encoder:

[0211] The model structure and function of the first encoder and the second encoder are consistent with those of the microenvironment encoder, and the model parameters of the first encoder and the second encoder are always consistent.

[0212] The first encoder is used to perform feature encoding on the reconstructed subgraph M1 to obtain the corresponding microenvironment encoding vector H1, which is then sent to the second differential module.

[0213] The second encoder is used to perform feature encoding on the reconstructed subgraph M2 to obtain the corresponding microenvironment encoding vector H2, which is then sent to the second difference module.

[0214] 5) Second difference module:

[0215] The second difference module is used to send the difference vector ΔH = H2 - H1 between the microenvironment encoding vectors H1 and H2 to the second MLP model.

[0216] 6) Second MLP model:

[0217] The second MLP model is used to perform regression prediction on the free energy change based on the difference vector ΔH to obtain and output the corresponding second free energy change.

[0218] Step 2: Prepare the training dataset by collecting data to obtain the corresponding first dataset and second dataset.

[0219] Here, the first dataset in this embodiment of the invention includes multiple first data records; the first data records include a first label image, a first mask image, a first type of mask node set, a second type of mask node set, and a third type of mask node set; the data structure of the first mask image is the same as that of the mask image G. mask Consistent, the data structure of the first label graph is consistent with that of the heterogeneous graph G; the first label graph is a microenvironment heterogeneous graph of a single residue; the first mask graph is a mask graph obtained by performing three types of masking on the first label graph; the first type of mask node set includes multiple first type node identifiers, the second type of mask node set includes multiple second type node identifiers, and the third type of mask node set includes multiple third type node identifiers.

[0220] The second dataset in this embodiment of the invention includes multiple second data records; the second data records include a first training native sequence, a first training mutant sequence, a first training site, and a first tag free energy change; the first training native sequence and the first training mutant sequence are the residue sequences of a protein complex before and after a mutation occurs at the residue site corresponding to the first training site; the first tag free energy change is the free energy change of the current protein complex before and after the mutation.

[0221] The current step 2 specifically includes:

[0222] Step 21: Collect one-dimensional residue sequences of multiple protein complexes through publicly available data acquisition methods to form a complex sequence set;

[0223] The protein complex is composed of any two types of proteins; the complex sequence set includes multiple complex sequences; the publicly available data acquisition methods include multiple publicly available structural databases of protein complexes, multiple publicly available experimental databases of protein complexes, and multiple publicly available databases of biochemical scientific and technological literature / technical papers.

[0224] Step 22: Take each complex sequence as the corresponding current sequence; and input the current sequence as the corresponding protein sequence S into the isomorphism initialization module of the codebook retrieval model for processing to obtain the corresponding complex tertiary structure and isomorphism G; and mutate any residue in the interface region between the two types of proteins on the current sequence to obtain a new residue sequence, which is recorded as the corresponding mutant sequence; and take the sequence site of the mutant residue of the mutant sequence as the corresponding mutation site; and input each mutant sequence as the corresponding protein sequence S into the isomorphism initialization module for processing to obtain the corresponding complex tertiary structure and isomorphism G.

[0225] Step 23: Take each residue of each complex sequence as the corresponding current residue, take the site of the current residue in the current complex sequence as the current residue site b, and input the isomorphic graph G corresponding to the current complex sequence and the current residue site b into the microenvironment subgraph extraction module of the codebook retrieval model for processing to obtain the corresponding microenvironment subgraph M.

[0226] Step 24: Each residue of each mutation sequence is taken as the corresponding current residue, the site of the current residue in the current mutation sequence is taken as the current residue site b, and the isomorphic graph G corresponding to the current mutation sequence and the current residue site b are input into the microenvironment subgraph extraction module for processing to obtain the corresponding microenvironment subgraph M.

[0227] Step 25, and take each obtained micro-environment subgraph M as the corresponding current subgraph; take the current subgraph as a corresponding first label graph; and calculate the total number of nodes N corresponding to the current subgraph.all Perform statistics, based on the total number of nodes N. all The maximum and minimum mask node counts are calculated based on preset maximum and minimum mask ratio thresholds. The maximum and minimum mask node counts then form the corresponding range for the total number of mask nodes. Within this range, three positive integers are randomly generated as the corresponding total number of type 1 nodes (N1), type 2 nodes (N2), and type 3 nodes (N3). N1 residue nodes are randomly selected from the current subgraph for type masking, N2 residue nodes for dihedral angle masking, and N3 residue nodes for local orientation masking. After masking, a corresponding mask is generated for each residue node based on its actual masking status in the current subgraph. The status information is used, and the current sub-graph that has completed masking is taken as the corresponding first mask graph; the node identifiers of each residue node that has undergone type masking are taken as the corresponding first-class node identifiers, the node identifiers of each residue node that has undergone dihedral corner masking are taken as the corresponding second-class node identifiers, and the node identifiers of each residue node that has undergone local direction masking are taken as the corresponding third-class node identifiers; and the corresponding first-class mask node set is composed of N1 first-class node identifiers, the corresponding second-class mask node set is composed of N2 second-class node identifiers, and the corresponding third-class mask node set is composed of N3 third-class node identifiers; and the first label graph, the first mask graph, the first-class mask node set, the second-class mask node set, and the third-class mask node set corresponding to the current sub-graph are combined to form a corresponding first data record;

[0228] Step 26, and record the tertiary structure of the complex corresponding to each complex sequence as the corresponding pre-mutation tertiary structure, and record the tertiary structure of the complex of each mutant sequence corresponding to the current complex sequence as the corresponding post-mutation tertiary structure; and form a corresponding original-mutation data set by each mutant sequence and its corresponding complex sequence, pre-mutation tertiary structure, mutation site, and post-mutation tertiary structure.

[0229] Step 27: Each original-mutated data set is taken as the corresponding current data set; the pre-mutation tertiary structure of the current data set is simulated using a preset molecular dynamics simulation tool, and the free energy ΔG of the pre-mutation tertiary structure is calculated to obtain the corresponding pre-mutation free energy; the post-mutation tertiary structure of the current data set is simulated using a molecular dynamics simulation tool, and the post-mutation free energy ΔG is calculated to obtain the corresponding post-mutation free energy; the corresponding mutation free energy change is calculated based on the post-mutation free energy and the pre-mutation free energy, where mutation free energy change = post-mutation free energy - pre-mutation free energy; the complex sequence, mutation sequence, mutation site, and mutation free energy change corresponding to the current data set are combined into a corresponding first training original sequence, first training mutation sequence, first training site, and first tag free energy change to form a corresponding second data record;

[0230] Among them, molecular dynamics simulation tools include GROMACS, AmberTools, and Desmond;

[0231] Step 28, and the first dataset is composed of all the first data records obtained; the second dataset is composed of all the second data records obtained.

[0232] Step 3: Train the codebook retrieval model based on the first dataset using the pre-trained framework; after the retrieval model is trained, first set the internal component parameters of the first and second prediction models based on the model parameters of the codebook retrieval model, and then train the first and second prediction models based on the second dataset.

[0233] Specifically, this includes: Step 31, training the codebook retrieval model based on the first dataset using a pre-trained framework;

[0234] Specifically, this includes: step 311, taking the first data record of the first dataset as the corresponding current training record;

[0235] Step 312: Extract the corresponding first label image, first mask image, first type mask node set, second type mask node set, and third type mask node set from the current training record as the corresponding current label image, current mask image, current first type node set, current second type node set, and current third type node set;

[0236] Step 313: Use the current mask image as the corresponding mask image G. mask The input to the pre-trained framework is processed to obtain the corresponding type reconstruction set X. 1 Dihedral Reconstruction Set X 2 Local orientation reconstruction set X 3Furthermore, the microenvironment encoding vector set generated by the microenvironment encoder, the codebook tag vector set generated by the codebook forward lookup module, and the three types of retrieval codebook vector sets generated by the codebook reverse lookup module are saved during this processing.

[0237] Step 314: Denote each microenvironment coding vector in the stored microenvironment coding vector set as the corresponding H. a 1 ≤ index a ≤ N A N A Let Z be the total number of residue nodes in the current tag graph; and let Z be the corresponding codebook tag vector of the stored codebook tag vector set. a ; and mark the vector Z of each codebook a The three types of codebook markers in the text are denoted as the corresponding The search codebook vector sets corresponding to subcodebooks C1, C2, and C3 in the three types of search codebook vector sets stored this time are denoted as the corresponding vector sets. Vector set Vector set Retrieval codebook vector With codebook tags One-to-one correspondence; and in the subcodebook C1 and the vector set Each retrieval codebook vector The matching token embedding vector is denoted as Subcodebook C2 and vector set Each retrieval codebook vector The matching token embedding vector is denoted as Subcodebook C3 and vector set Each retrieval codebook vector The matching token embedding vector is denoted as And by each codebook's tag vector Corresponding token embedding vector This forms the corresponding token embedding vector set; and is encoded by each microenvironment vector H. a The token embedding vector set and its corresponding token form a corresponding first verification pair.

[0238] Here, if we further set an index s, whose value range is 1≤s≤3, then the first check pair It can be further expressed as

[0239] Step 315, reconstruct the current label graph and type set X. 1 The type embedding vector and the reconstructed type embedding vector corresponding to each of the first-class node identifiers in the current first-class node set are combined to form a corresponding second check pair; and the current label graph and the dihedral reconstruction set X are then combined. 2The dihedral embedding vector corresponding to each of the two types of node identifiers in the current set of two types of nodes and the reconstructed dihedral embedding vector form a corresponding third check pair;

[0240] The second check pair is denoted as The third check pair is denoted as

[0241] 1≤index b≤N B 1 ≤ index c ≤ N C N B N C These represent the total number of node identifiers corresponding to the current first-type node set and the current second-type node set, respectively.

[0242] These are the type embedding vector and the reconstructed type embedding vector of the b-th second check pair, respectively; These are the dihedral embedding vector and the reconstructed dihedral embedding vector of the c-th third check pair, respectively;

[0243] Here, if we further set an index q, whose value range is 1≤q≤7, then the third check pair... It can be further expressed as Dihedral embedding vector The q-th dihedral angle in the middle, To reconstruct the dihedral embedding vector The q-th dihedral angle in;

[0244] Step 316, reconstruct the current label map and the local orientation set X. 3 In the process, the difference vector between a pair of local orientation embedding vectors corresponding to each of the three types of node identifiers in the current three types of node sets and the reconstructed local orientation embedding vector is calculated, and the calculation result is used as the corresponding first prediction error vector; and the current mask image and the local orientation reconstruction set X are then processed. 3 In the process, a pair of local orientation embedding vectors corresponding to each of the three types of node identifiers in the current three types of node sets and the difference vector of the reconstructed local orientation embedding vector are calculated, and the calculation result is used as the corresponding first label error vector; and each first label error vector and its corresponding first prediction error vector form a corresponding fourth check pair;

[0245] The fourth check pair is denoted as

[0246] 1≤index d≤N D N D This represents the total number of node identifiers corresponding to the current three types of node sets;

[0247] These are the first label error vector and the first prediction error vector of the d-th fourth check pair, respectively;

[0248] Step 317: Construct a first loss function L1 for the microenvironment encoder and residue feature codebook C based on the constraint loss function; construct a second loss function L2 for the sub-codebook C1 and the first decoder based on the cross-entropy loss function; construct a third loss function L3 for the sub-codebook C2 and the second decoder based on the MSE loss function; construct a fourth loss function L4 for the sub-codebook C3 and the third decoder based on the Huber loss function; and construct an overall loss function L based on the first, second, fourth, and fourth loss functions. total ; and the obtained N A The first check pair, N B The second check pair, N C A third check pair, N D Substituting the fourth check pair into the overall loss function L total The corresponding first loss value is obtained through calculation;

[0249] Here, the first loss function L1, the second loss function L2, the third loss function L3, the fourth loss function L4, and the overall loss function L in this embodiment of the invention are... total As shown below:

[0250]

[0251]

[0252] L total =λL1+L2+L3+L4;

[0253] Where sg() is the gradient stopping operation, η is a preset tradeoff hyperparameter, and λ is a preset tradeoff hyperparameter; L CE () is the cross-entropy loss function, L Huber () represents the Huber loss function; 1 ≤ index s ≤ 3; 1 ≤ index q ≤ 7. Dihedral embedding vector The q-th dihedral angle in the middle, To reconstruct the dihedral embedding vector The q-th dihedral angle in;

[0254] Step 318: Identify whether the first loss value meets the preset first loss value range; if the first loss value meets the first loss value range, identify whether the current training record is the last first data record of the first dataset; if so, proceed to step 319; otherwise, take the next first data record of the first dataset as the new current training record and return to step 312; if the first loss value does not meet the first loss value range, firstly, based on the preset first model optimizer, perform a round of parameter optimization on the microenvironment encoder and residue feature codebook C in the direction of minimizing the first loss function L1; then, based on the preset second model optimizer, perform a round of parameter optimization on the sub-codebook C1 and the first decoder in the direction of minimizing the second loss function L2; then, based on the preset third model optimizer, perform a round of parameter optimization on the sub-codebook C2 and the second decoder in the direction of minimizing the third loss function L3; then, based on the preset fourth model optimizer, perform a round of parameter optimization on the sub-codebook C3 and the third decoder in the direction of minimizing the fourth loss function L4; finally, based on the preset fifth model optimizer, perform a round of parameter optimization on the direction of minimizing the overall loss function L... total The direction that reaches the minimum value is used to perform a round of parameter optimization on the microenvironment encoder and residue feature codebook C, and the process returns to step 313 when the parameter optimization ends.

[0255] Wherein, the first loss value range is a pre-set numerical range; the first model optimizer includes the Adam optimizer and the SGD optimizer; the second model optimizer includes the Adam optimizer and the SGD optimizer; the third model optimizer includes the Adam optimizer and the SGD optimizer; the fourth model optimizer includes the Adam optimizer and the SGD optimizer; and the fifth model optimizer includes the Adam optimizer and the SGD optimizer.

[0256] Step 319: Stop the current training and confirm that the retrieval model training has ended;

[0257] Step 32: After the retrieval model training is completed, the internal component parameters of the first and second prediction models are set based on the model parameters of the codebook retrieval model, and then the first and second prediction models are trained based on the second dataset.

[0258] Specifically, it includes: Step 321, setting the model parameters and codebook of the first and second retrieval models inside the first prediction model based on the model parameters of the codebook retrieval model and the built-in residue feature codebook C, and freezing the parameters of the first and second retrieval models after the setting is completed;

[0259] Step 322: Based on the model parameters of the codebook retrieval model and the built-in residue feature codebook C, the model parameters of the third and fourth retrieval models inside the second prediction model and the built-in residue feature codebook are set, and the parameters of the third and fourth retrieval models are frozen after the settings are completed.

[0260] Step 323, and initialize the model parameters of the first and second encoders inside the second prediction model based on the model parameters of the microenvironment encoder of the codebook retrieval model;

[0261] Step 324: After setting the parameters of the first and second prediction models, train the first prediction model based on the second dataset, and train the second prediction model based on the second dataset.

[0262] Specifically, this includes: Step 3241, training the first prediction model based on the second dataset;

[0263] Specifically, it includes: step 32411, dividing the second dataset into two sub-datasets based on a preset first segmentation ratio, denoted as the corresponding first training set and first evaluation set;

[0264] Wherein, the first segmentation ratio is a pre-set ratio parameter, such as 8:2; both the first training set and the first evaluation set consist of multiple second data records; the ratio of the total number of records in the first training set and the first evaluation set satisfies the first segmentation ratio;

[0265] Step 32412: Take each second data record of the first training set as the corresponding current training record; and take the first training native sequence, the first training mutant sequence, and the first training site of the current training record as a set of corresponding native protein sequence S1, mutant protein sequence S2, and mutant residue site b. 1-2 The first prediction model is input and processed to obtain the corresponding first free energy change; and the first free energy change and the first label free energy change of the current training record are combined to form a corresponding first prediction-label pair;

[0266] Step 32413: Input all the obtained first prediction-label pairs into the preset first model loss function to calculate the corresponding second loss value;

[0267] The first model loss function is implemented based on either the L1 loss function or the L2 loss function.

[0268] Step 32414: Identify whether the second loss value meets the preset second loss value range; if the second loss value meets the second loss value range, proceed to step 32415; if the second loss value does not meet the second loss value range, perform a round of modulation on all non-frozen parameters of the first prediction model based on the preset sixth model optimizer in the direction of minimizing the first model loss function, and return to step 32412 when the current round of modulation ends.

[0269] The second loss value range is a pre-set numerical range; the sixth model optimizer includes the Adam optimizer and the SGD optimizer.

[0270] It should be noted that all non-frozen parameters of the first prediction model refer to the set of model parameters of the first self-attention module, the second self-attention module, the first difference module, and the first MLP model;

[0271] Step 32415: Take each second data record of the first evaluation set as the corresponding current evaluation record; and take the first training native sequence, the first training mutant sequence, and the first training site of the current evaluation record as a set of corresponding native protein sequence S1, mutant protein sequence S2, and mutant residue site b. 1-2 The first prediction model is input and processed to obtain the corresponding first free energy change; and the first free energy change and the first label free energy change of the current evaluation record are combined to form a corresponding second prediction-label pair; and all the obtained second prediction-label pairs are input into the preset first model evaluation function to calculate the corresponding first evaluation value.

[0272] The first model evaluation function is implemented based on the MAE function, MSE function, or RMSE function.

[0273] Step 32416: Identify whether the first evaluation value meets the preset first evaluation value range; if not, return to step 32411; if it meets, confirm that the training of the first prediction model has ended.

[0274] Here, the first evaluation value range is a pre-set evaluation range;

[0275] Step 3242: Train the second prediction model based on the second dataset;

[0276] Specifically, it includes: step 32421, dividing the second dataset into two subsets based on a preset second segmentation ratio, denoted as the corresponding second training set and second evaluation set;

[0277] The second segmentation ratio is a pre-set ratio parameter, such as 8:2; both the second training set and the second evaluation set consist of multiple second data records; the ratio of the total number of records in the second training set and the second evaluation set satisfies the second segmentation ratio.

[0278] Step 32422: Take each second data record of the second training set as the corresponding current training record; and take the first training native sequence, the first training mutant sequence, and the first training site of the current training record as a set of corresponding native protein sequence S3, mutant protein sequence S4, and mutant residue site b. 3-4 The second prediction model is input and processed to obtain the corresponding second free energy change; and the second free energy change and the first label free energy change of the current training record are combined to form a corresponding third prediction-label pair;

[0279] Step 32423: Input all the obtained third prediction-label pairs into the preset second model loss function to calculate the corresponding third loss value;

[0280] The second model loss function is implemented based on either the L1 loss function or the L2 loss function.

[0281] Step 32424: Identify whether the third loss value meets the preset range of the third loss value; if the third loss value meets the range of the third loss value, proceed to step 32425; if the third loss value does not meet the range of the third loss value, perform a round of modulation on all non-frozen parameters of the second prediction model based on the preset seventh model optimizer in the direction of minimizing the second model loss function, and return to step 32422 when the modulation ends.

[0282] Here, the third loss value range in this embodiment of the invention is a pre-set numerical range; the seventh model optimizer includes the Adam optimizer and the SGD optimizer; the maximum loss value of the third loss value range is less than the maximum loss value of the second loss value range mentioned above; the prediction accuracy constrained by the third loss value range is greater than the prediction accuracy constrained by the second loss value range.

[0283] It should be noted that all the unfrozen parameters of the second prediction model refer to the set of model parameters of the third self-attention module, the fourth self-attention module, the first encoder, the second encoder, the second difference module, and the second MLP model; and when modulating all the unfrozen parameters of the second prediction model, the model parameters of the first and second encoders must be kept consistent, which is actually the sharing of parameters between the first and second encoders.

[0284] Step 32425: Take each second data record of the second evaluation set as the corresponding current evaluation record; and take the first training native sequence, the first training mutant sequence, and the first training site of the current evaluation record as a set of corresponding native protein sequence S3, mutant protein sequence S4, and mutant residue site b. 3-4 The second prediction model is input to obtain the corresponding second free energy change; and the second free energy change and the first label free energy change of the current evaluation record are combined to form a corresponding fourth prediction-label pair; and all the obtained fourth prediction-label pairs are input into the preset second model evaluation function to calculate the corresponding second evaluation value.

[0285] The second model evaluation function is implemented based on the MAE function, MSE function, or RMSE function.

[0286] Step 32426: Identify whether the second evaluation value meets the preset range of the second evaluation value; if not, return to step 32421; if it does, confirm that the training of the second prediction model has ended.

[0287] Here, the second evaluation value range in this embodiment of the invention is a pre-set numerical range; the first and second model evaluation functions in this embodiment of the invention are actually an overall error statistics function, and the first and second evaluation values ​​are actually two error statistics; the maximum evaluation value of the second evaluation value range in this embodiment of the invention is less than the maximum evaluation value of the first evaluation value range mentioned above; the prediction accuracy constrained by the second evaluation value range is greater than the prediction accuracy constrained by the first evaluation value range;

[0288] Step 325: After the training of the first and second prediction models is completed, confirm that the training of the two prediction models is finished.

[0289] Step 4: After the training of the two prediction models is completed, the original-mutant protein information is combined based on the first original protein sequence and the first set of mutant residues input by the user to obtain multiple original groups; the first prediction model is used to perform a preliminary screening of all original groups according to the preset screening and sorting rules to obtain N preferred groups; the second prediction model is used to continue to sort the N preferred groups according to the screening and sorting rules to generate the corresponding final selection group sequence and feed it back to the current user.

[0290] The first set of mutant residues includes multiple first mutant residues, and the total number of first mutant residues is greater than the preset total number N, where N is a positive integer. The first mutant residues include the first residue type and the first residue site. The original set includes the first native protein sequence, the first mutant protein sequence, and the first residue site. The screening and sorting rules are to sort in descending order of free energy change or in ascending order of free energy change.

[0291] Specifically, this includes: Step 41, combining native-mutant protein information based on the first native protein sequence and the first set of mutant residues input by the user to obtain multiple corresponding original groups;

[0292] Specifically, this includes: taking each first mutation residue in the first mutation residue set as the corresponding current mutation; taking the first residue type and first residue site of the current mutation as the corresponding current mutation residue and current mutation site; obtaining a new residue sequence by replacing the original residues in the first native protein sequence corresponding to the current mutation site with the current mutation residues, and recording it as the corresponding first mutant protein sequence; taking the current mutation site as a corresponding first residue site; and forming a corresponding original set by the first native protein sequence corresponding to the current mutation, the first mutant protein sequence, and the first residue site.

[0293] Step 42: Use the first prediction model to perform a preliminary screening of all the original groups according to the preset screening and sorting rules to obtain N preferred groups;

[0294] Specifically, this includes: taking the first native protein sequence, the first mutant protein sequence, and the first residue site of each original group as a corresponding native protein sequence S1, mutant protein sequence S2, and mutant residue site b. 1-2 The first prediction model is input and processed to obtain the corresponding first free energy variable; all the obtained first free energy variables are sorted according to the screening and sorting rules; and the N original groups corresponding to the top N first free energy variables are extracted as the corresponding N preferred groups.

[0295] Step 43, and use the second prediction model to continue sorting the N preferred groups according to the screening and sorting rules to generate the corresponding final selection group sequence and feed it back to the current user;

[0296] Specifically, this includes: taking the first native protein sequence, the first mutant protein sequence, and the first residue site of each preferred group as a corresponding native protein sequence S3, mutant protein sequence S4, and mutant residue site b. 3-4 The second prediction model is input and processed to obtain the corresponding second free energy change; and the first native protein sequence, the first mutant protein sequence, the first residue site and the second free energy change corresponding to each preferred group are combined to form a corresponding final selection group; and the N final selection groups are sorted according to the screening and sorting rules to obtain the corresponding final selection group sequence and fed back to the current user.

[0297] Figure 6This is a module structure diagram of a processing device for screening mutation sites based on free energy variation, provided in Embodiment 2 of the present invention. This device can be a terminal device or server implementing the aforementioned method embodiment, or it can be a device that enables the aforementioned terminal device or server to implement the aforementioned method embodiment. For example, the device can be a device or chip system of the aforementioned terminal device or server. Figure 6 As shown, the device includes: a model building module 201, a data acquisition module 202, a model training module 203, and a model application module 204.

[0298] Model building module 201 is used to construct a feature mapping model for mapping residue features of microenvironmental features of specified residue sites in a protein complex using codebook retrieval, denoted as the corresponding codebook retrieval model; and to construct a pre-training framework for the codebook retrieval model; and to construct two prediction models based on the codebook retrieval model for predicting the free energy change ΔΔG before and after a single-point mutation in a protein complex, denoted as the first prediction model and the second prediction model; the codebook retrieval model is used to perform residue feature mapping on the microenvironmental features of the current residue site based on the protein sequence S and residue site b input to the model, as well as the built-in residue feature codebook C, to obtain the corresponding residue mapping vector E; the pre-training framework is used to process the mask image G input to the framework. mask The three types of embedding vectors of all residue nodes are reconstructed to obtain the corresponding type reconstruction set X. 1 Dihedral Reconstruction Set X 2 Local orientation reconstruction set X 3 The first prediction model is used to predict the original protein sequence S1, the mutant protein sequence S2, and the mutant residue site b as input to the model. 1-2 Microenvironment embedding encoding is performed, and the free energy change is predicted based on the differential embedding of the two embedding encodings to obtain the corresponding first free energy change; the second prediction model is used to predict the free energy change based on the native protein sequence S3, the mutant protein sequence S4, and the mutant residue site b as input to the model. 3-4 Microenvironment embedding encoding is performed, and corresponding microenvironment reconstruction encoding is performed based on the two embedding encodings. Then, the free energy change is predicted based on the difference features of the two reconstruction features to obtain the corresponding second free energy change.

[0299] The data acquisition module 202 is used to prepare the training dataset through data acquisition to obtain the corresponding first dataset and second dataset.

[0300] The model training module 203 is used to train the codebook retrieval model based on the first dataset using the pre-training framework. After the retrieval model is trained, the internal component parameters of the first and second prediction models are set based on the model parameters of the codebook retrieval model, and then the first and second prediction models are trained based on the second dataset.

[0301] The model application module 204 is used to combine native-mutant protein information based on the first native protein sequence and the first mutant residue set input by the user after the training of the two types of prediction models to obtain multiple original groups; and to use the first prediction model to perform a preliminary screening of all original groups according to the preset screening and sorting rules to obtain N preferred groups; and to use the second prediction model to continue to sort the N preferred groups according to the screening and sorting rules to generate the corresponding final selected group sequence and feed it back to the current user; the first mutant residue set includes multiple first mutant residues, the total number of first mutant residues is greater than the preset total number N, where N is a positive integer; the first mutant residues include the first residue type and the first residue site; the original groups include the first native protein sequence, the first mutant protein sequence, and the first residue site; the screening and sorting rules are to sort in descending order of free energy change or in ascending order of free energy change.

[0302] The present invention provides a processing device for screening mutation sites based on free energy variation, which can execute the method steps in the above method embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0303] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing elements; they can be fully implemented in hardware; or some modules can be implemented by processing elements calling software, while others are implemented in hardware. For example, the model building module can be a separate processing element, or it can be integrated into a chip in the above device. Alternatively, it can be stored as program code in the memory of the above device, and its functions can be called and executed by a processing element. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0304] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a System-on-a-Chip (SOC).

[0305] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the foregoing method embodiments are generated. The computer described above can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The aforementioned computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the aforementioned computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, Bluetooth, microwave, etc.) means. The aforementioned computer-readable storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The aforementioned available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0306] Figure 7 This is a schematic diagram of an electronic device provided in Embodiment 3 of the present invention. This electronic device can be a terminal device or server implementing the methods of the aforementioned embodiments, or it can be a terminal device or server connected to the aforementioned terminal device or server implementing the methods of the aforementioned embodiments. Figure 7As shown, the electronic device may include: a processor 301 (e.g., CPU), a memory 302, and a transceiver 303; the transceiver 303 is coupled to the processor 301, and the processor 301 controls the transmission and reception operations of the transceiver 303. The memory 302 may store various instructions for performing various processing functions and implementing the processing steps described in the foregoing embodiments. Preferably, the electronic device involved in the embodiments of the present invention further includes: a power supply 304, a system bus 305, and a communication port 306. The system bus 305 is used to realize communication connections between components. The communication port 306 is used for communication between the electronic device and other peripherals.

[0307] exist Figure 7 The system bus 305 mentioned can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, it is represented by only one thick line in the figure, but this does not indicate that there is only one bus or one type of bus. The communication interface is used to enable communication between the database access device and other devices (e.g., clients, read-write libraries, and read-only libraries). Memory may include Random Access Memory (RAM) and may also include Non-Volatile Memory, such as at least one disk storage device.

[0308] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), graphics processing units (GPUs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0309] It should be noted that the embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when run on a computer, cause the computer to perform the methods and processes provided in the above embodiments.

[0310] This invention provides a processing method, apparatus, electronic device, and computer-readable storage medium for screening mutation sites based on free energy change. As described above, this invention constructs three types of tool models to improve the screening efficiency of mutation strategies: 1) a codebook retrieval model for mapping residue features of microenvironmental characteristics of specified residue sites in protein complexes using codebook retrieval, and a pre-training framework is constructed for this codebook retrieval model; 2) a coarse calculation model (i.e., the first prediction model) based on the codebook retrieval model, which first learns the two microenvironmental mapping features of a single mutation site before and after mutation, and then predicts the free energy change ΔΔG before and after a single-point mutation based on the difference features of the two microenvironmental mapping features; 3) a fine calculation model (i.e., the second prediction model) based on the codebook retrieval model, which first learns the two microenvironmental mapping features of a single mutation site before and after mutation, then reconstructs two local heteromorphic maps before and after mutation based on the two microenvironmental mapping features, and then predicts the free energy change ΔΔG before and after a single-point mutation based on the difference features of the two reconstructed local heteromorphic maps. Two training datasets (dataset 1 and dataset 2) were constructed for the three types of tool models. A two-stage training strategy was designed for the three types of tool models: 1) First stage: the codebook retrieval model is trained based on the pre-trained framework using the first dataset; 2) Second stage: the internal component parameters of the first and second prediction models are set based on the codebook retrieval model trained in the first stage, and then the first and second prediction models are trained separately based on the second dataset. After training, the two prediction models are used for mutation strategy screening: first, the native protein sequence and mutant residue set input by the user are combined to obtain multiple original groups; then, the first prediction model is used to perform preliminary screening of all original groups according to the screening and sorting rules to obtain N preferred groups; then, the second prediction model is used to re-sort the N preferred groups according to the screening and sorting rules to obtain the final selected group sequence and provide feedback to the user. Through the two end-to-end prediction models in this embodiment of the invention, not only is the overall calculation cycle shortened, the overall calculation cost reduced, and the overall research efficiency improved; the first prediction model, with its higher prediction efficiency, further improves the efficiency of massive pre-screening; and the second prediction model, with its higher prediction accuracy, further improves the prediction accuracy of the final selection results.

[0311] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0312] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for screening mutation sites based on free energy change, characterized in that, The method includes: A feature mapping model, denoted as the codebook retrieval model, is constructed to perform residue feature mapping on the microenvironmental features of a specified residue site in a protein complex using a codebook retrieval method. A pre-training framework is built for the codebook retrieval model. Based on the codebook retrieval model, two prediction models are constructed to predict the free energy change ΔΔG before and after a single-point mutation in the protein complex, denoted as the first prediction model and the second prediction model. The codebook retrieval model is used to perform residue feature mapping on the microenvironmental features of the current residue site based on the protein sequence S and residue site b input to the model, as well as the built-in residue feature codebook C, to obtain the corresponding residue mapping vector E. The pre-training framework is used to process the mask image G input to the framework. mask The three types of embedding vectors of all residue nodes are reconstructed to obtain the corresponding type reconstruction set X. 1 Dihedral Reconstruction Set X 2 Local orientation reconstruction set X 3 The first prediction model is used to predict the original protein sequence S1, the mutant protein sequence S2, and the mutant residue site b based on the input of the model. 1-2 Microenvironment embedding encoding is performed, and the free energy change is predicted based on the differential embedding of the two embedding encodings to obtain the corresponding first free energy change; the second prediction model is used to predict the free energy change based on the native protein sequence S3, the mutant protein sequence S4, and the mutant residue site b as input to the model. 3-4 Microenvironment embedding encoding is performed, and corresponding microenvironment reconstruction encoding is performed based on the two embedding encodings. Then, the free energy change is predicted based on the difference features of the two reconstruction features to obtain the corresponding second free energy change. The training dataset is prepared by data collection, resulting in the corresponding first and second datasets. The codebook retrieval model is trained based on the pre-training framework using the first dataset. After the retrieval model is trained, the internal component parameters of the first and second prediction models are set based on the model parameters of the codebook retrieval model, and then the first and second prediction models are trained based on the second dataset. After the two prediction models are trained, the native-mutant protein information is combined based on the first native protein sequence and the first set of mutant residues input by the user to obtain multiple original groups; the first prediction model is used to perform a preliminary screening of all the original groups according to a preset screening and sorting rule to obtain N preferred groups; the second prediction model is used to continue to sort the N preferred groups according to the screening and sorting rule to generate the corresponding final selected group sequence and feed it back to the current user; the first set of mutant residues includes multiple first mutant residues, the total number of the first mutant residues is greater than the preset total number N, where N is a positive integer; the first mutant residues include a first residue type and a first residue site; the original groups include the first native protein sequence, the first mutant protein sequence, and the first residue site; the screening and sorting rule is to sort in descending order of free energy change or in ascending order of free energy change.

2. The processing method for screening mutation sites based on free energy change according to claim 1, characterized in that, The codebook retrieval model includes a heterogeneous graph initialization module, a microenvironment subgraph extraction module, a microenvironment encoder, a codebook retrieval module, and the residue feature codebook C; The heterogeneous graph initialization module is implemented based on a trained Uni-Fold model. This module is used to predict the tertiary structure of the protein sequence S using the Uni-Fold model and construct a corresponding heterogeneous graph G based on the predicted structure, which is then sent to the microenvironment subgraph extraction module. The protein sequence S is a one-dimensional residue sequence of a protein complex, composed of multiple residues ordered. The residue site b is a residue sequence index in the protein sequence S. The heterogeneous graph G consists of a node set and an edge set. The node set includes multiple residue nodes, each containing a type embedding vector, a dihedral embedding vector, and a local orientation embedding vector. The edge set includes multiple node edges, each containing a sequence distance embedding vector, a spatial distance embedding vector, and a residue pair orientation embedding vector. The microenvironment subgraph extraction module is used to extract the near-neighbor subgraph corresponding to the residue site b in the heterogeneous graph G and send it to the microenvironment encoder as the corresponding microenvironment subgraph M. The microenvironment encoder is implemented based on a type of SE3 equivariant graph neural network, which includes the SE(3)-Transformer model, the Equiformer model, and the EGNN model. The microenvironment encoder is used to perform feature encoding on the microenvironment subgraph M to obtain the corresponding microenvironment encoding vector H and send it to the codebook retrieval module. The residue feature codebook C includes three sub-codebooks, namely sub-codebook C1, sub-codebook C2, and sub-codebook C3; each of sub-codebooks C1, C2, and C3 contains multiple corresponding token embedding vectors. Token embedding vector Token embedding vector Composition; 1 ≤ index i ≤ N C N C A preset positive integer; token embedding vector , , The vectors have the same shape and are consistent with the residue nodes; all embedding codes of the residue feature codebook C are initially composed of a set of preset initial codebook codes; The codebook retrieval module is used to, according to a preset vector similarity algorithm, identify the N values ​​of the subcodebooks C1, C2, or C3 of the residue feature codebook C. C token embedding vector , or The token embedding vector with the highest similarity to the microenvironment encoding vector H is used as the corresponding retrieval codebook vector e. 1 e 2 or e 3 ; and the three retrieved codebook vectors e 1 e 2 e 3 The corresponding residue mapping vector E is formed and output; wherein, the vector similarity algorithm includes cosine vector similarity algorithm and Euclidean distance similarity algorithm.

3. The processing method for screening mutation sites based on free energy change according to claim 2, characterized in that, The step of using the Uni-Fold model to predict the tertiary structure of the protein sequence S and constructing a heteromorphic graph G based on the predicted structure specifically includes: The tertiary structure of the protein sequence S was predicted using the Uni-Fold model to obtain the atomic-level tertiary structure of the complex. The tertiary structure of the complex is then subjected to atomic-level residue clustering to obtain multiple corresponding residue clusters. Each residue cluster is then labeled with four types of main chain atoms on the main chain backbone to obtain corresponding main chain atom groups. Each residue cluster corresponds to one residue. The four types of main chain atoms include nitrogen (N), alpha carbon (C), and carbon atoms. α Carbon atoms (C) and oxygen atoms (O); the main chain atom group is composed of the four types of main chain atoms; And with the alpha carbon atom C of each group of main chain atoms. α Origin, alpha carbon atom C α C to carbon atom C α A corresponding local coordinate system is constructed with the -C bond direction as the X-axis, a straight line passing through the origin and perpendicular to the X-axis in the main chain plane as the Y-axis, and a straight line passing through the origin and perpendicular to the main chain plane as the Z-axis; and the alpha carbon atom C of the first main chain atom group is used as the coordinate system. α Using the first local coordinate system as the global origin, and taking the first local coordinate system as the global coordinate system, establish corresponding local-global coordinate mapping relationships between the remaining local coordinate systems and the global coordinate system; The corresponding residue types of each residue cluster are one-hot encoded to obtain the corresponding type embedding vectors; and pre-set quantum chemical tools are used to determine the three types of main chain dihedral angles φ, ψ, ω and the four types of side chain dihedral angles χ of the corresponding residues of each residue cluster. 1 , χ 2 , χ 3 , χ 4 Measurements were performed, and the seven dihedral angles obtained were encoded using dihedral angle embedding to obtain the corresponding dihedral angle embedding vectors. The quantum chemical tools were then used to correlate the nitrogen (N), carbon (C), and oxygen (O) atoms of each residue cluster with the alpha carbon atom (C) in the corresponding local coordinate system. α The relative directions are confirmed, and the three relative directions are encoded using local orientation embedding to obtain the corresponding local orientation embedding vectors; and each residue node is composed of the type embedding vector, the dihedral angle embedding vector, and the local orientation embedding vector; wherein, the quantum chemistry tools include MDAnalysis, Biopython, PyMOL, and VMD. The residues in the protein sequence S are arranged in pairs to obtain multiple residue pairs; the relative sequence distance and spatial relative distance in three-dimensional space of the two residues before and after each residue pair in the protein sequence S are confirmed, and the two relative distances are encoded by distance embedding to obtain the corresponding sequence distance embedding vector and the spatial distance embedding vector; the four types of main chain atoms of the second residue of each residue pair are encoded by direction embedding in the four relative directions of the first residue in the local coordinate system to obtain the corresponding residue pair direction embedding vector; and the sequence distance embedding vector, the spatial distance embedding vector and the residue pair direction embedding vector of each residue pair form a corresponding node edge; All the obtained residue nodes form a corresponding node set, all the obtained node edges form a corresponding edge set, and the node set and the edge set form the heterogeneous graph G.

4. The processing method for screening mutation sites based on free energy change according to claim 2, characterized in that, The step of extracting the neighborhood subgraph corresponding to the residue site b in the heterogeneous graph G as the corresponding microenvironment subgraph M specifically includes: The residue node corresponding to residue site b in the heterogeneous graph G is taken as the current node; and the residue nodes in the heterogeneous graph G that have a total number of nodes and edges with the current node not exceeding a preset maximum hop count, a relative sequence distance from the current node not exceeding a preset sequence distance threshold, and a relative spatial distance from the current node not exceeding a preset spatial distance threshold are extracted to form a corresponding current node set; and the node and edge subsets corresponding to the current node set in the heterogeneous graph G are extracted to form a corresponding current edge set; and the current node set and the current edge set form the corresponding microenvironment subgraph M; wherein, the maximum hop count is a positive integer.

5. The processing method for screening mutation sites based on free energy change according to claim 2, characterized in that, The mask image G mask It consists of the node set and the edge set; the mask graph G mask Each of the aforementioned residue nodes corresponds to a mask state information; The mask status information includes mask markers and mask modes; the mask markers include unmasked and masked. The mask mode consists of type mask mode, dihedral mask mode, and local orientation mask mode. All three types of modes include two states: on and off. The on state of the three types of modes is not unique. When the mask marker is unmasked, all three mask mode states are off; when the mask marker is masked and the type mask mode is on, the type embedding vector of the residue node is replaced by a randomly generated residue type one-hot encoded vector; when the mask marker is masked and the dihedral mask mode is on, the seven dihedral codes of the dihedral embedding vector of the residue node are set to zero; when the mask marker is masked and the local orientation mask mode is on, Gaussian noise is added to the three noisy orientation codes of the local orientation embedding vector of the residue node. The mask image G mask The total number of residue nodes is denoted as the total number of nodes N. all The total number of residue nodes marked as masked and with the masking mode of the specified type enabled is denoted as the total number of first-class nodes N1; the total number of residue nodes marked as masked and with the dihedral angle masking mode enabled is denoted as the total number of second-class nodes N2; and the total number of residue nodes marked as masked and with the local orientation masking mode enabled is denoted as the total number of third-class nodes N3; N1, N2, or N3 are related to N... all The proportions do not exceed the preset maximum mask proportion threshold and are not lower than the preset minimum mask proportion threshold; The type reconstruction set X 1 By N all The dihedral reconstruction set X is composed of several reconstruction type embedding vectors. 2 By N all The local orientation reconstruction set X is composed of a reconstructed dihedral embedding vector. 3 By N all It consists of a reconstructed local orientation embedding vector; the data format of the reconstructed type embedding vector, the reconstructed dihedral angle embedding vector, and the reconstructed local orientation embedding vector is consistent with the corresponding type embedding vector, dihedral angle embedding vector, and local orientation embedding vector; The pre-training framework includes a full-structure microenvironment extraction module, a microenvironment encoder, a codebook forward lookup module, a codebook reverse lookup module, a residue feature codebook C, a first decoder, a second decoder, and a third decoder; The full-structure microenvironment extraction module is used to extract the mask image G. mask Each of the aforementioned residue nodes is taken as the corresponding current node; and the mask image G is used as the current node. mask Residue nodes whose total number of edges between them and the current node does not exceed a preset maximum hop count, whose relative sequence distance to the current node does not exceed a preset sequence distance threshold, and whose relative spatial distance to the current node does not exceed a preset spatial distance threshold are extracted to form a corresponding current node set; and the mask image G is then used to form the current node set. mask The node edge subset corresponding to the current node set is extracted to form the corresponding current edge set; and the current node set and the current edge set form a corresponding micro-environment subgraph M. j 1 ≤ index j ≤ N all ; and from the obtained N all The microenvironment subgraph M j Form the corresponding sub-graph set {M j Send to the microenvironment encoder; The microenvironment encoder is used to process the sub-map {M} j The various microenvironment subgraphs M of} j Feature encoding is performed to obtain the corresponding microenvironment encoding vector H. j ; and from the obtained N all The microenvironment encoding vector H j Form the corresponding vector set {H j Send to the codebook lookup module; The codebook lookup module is used to find the vector set {H} j The microenvironment encoding vectors H of each of the aforementioned microenvironment j As the corresponding current vector; and according to the vector similarity algorithm, embed the token in the subcodebook C1 that has the highest similarity to the current vector into the embedding vector. The index i is used as the corresponding codebook marker. The token embedding vector in the subcodebook C2 that has the highest similarity to the current vector is embedded. The index i is used as the corresponding codebook marker. The token embedding vector in the subcodebook C3 that has the highest similarity to the current vector is embedded. The index i is used as the corresponding codebook marker. ; and the three codebook markers obtained , , Form a corresponding codebook tag vector Z j ; and from the obtained N all The codebook tag vector Z j Form the corresponding vector set {Z j Send to the codebook reverse lookup module; The codebook reverse lookup module is used to retrieve the vector set {Z} j The codebook tag vectors Z in} j As the corresponding current vector; and respectively, the index i in the subcodebooks C1, C2, and C3 is compared with the codebook marker of the current vector. , , Matching token embedding vector , , As the corresponding retrieval codebook vector , , ; and from the obtained N all A retrieval codebook vector Form the corresponding vector set { } is sent to the first decoder, and the obtained N all A retrieval codebook vector Form the corresponding vector set { } is sent to the second decoder, and the obtained N all A retrieval codebook vector Form the corresponding vector set { Send to the third decoder; The first decoder is implemented based on a type of SE3 equivariant graph neural network; The first decoder is used to convert the vector set { The respective retrieval codebook vectors of} As a corresponding first transformation node, and derived from N all Each of the first conversion nodes forms a corresponding first conversion node set; and the first conversion node set and the mask image G are combined. mask The edge sets form a corresponding first transformation graph; and residue type decoding is performed on all the first transformation nodes of the first transformation graph to obtain the corresponding type reconstruction set X. 1 And output; The second decoder is implemented based on a class of SE3 equivariant graph neural networks; the second decoder is used to process the vector set { The respective retrieval codebook vectors of} As a corresponding second transformation node, and derived from N all Each of the second conversion nodes forms a corresponding second conversion node set; and the second conversion node set and the mask image G are combined. mask The edge sets form a corresponding second transformation graph; and dihedral decoding is performed on all the second transformation nodes of the second transformation graph to obtain the corresponding dihedral reconstruction set X. 2 And output; The third decoder is implemented based on a type of SE3 equivariant graph neural network; The third decoder is used to process the vector set { The respective retrieval codebook vectors of} As a corresponding third transformation node, and derived from N all Each of the aforementioned third conversion nodes forms a corresponding third conversion node set; and the third conversion node set and the mask image G are combined. mask The edge set forms a corresponding third transformation graph; and local orientation decoding is performed on all the third transformation nodes of the third transformation graph to obtain the corresponding local orientation reconstruction set X. 3 And output it.

6. The processing method for screening mutation sites based on free energy change according to claim 2, characterized in that, The first prediction model includes a first retrieval model, a second retrieval model, a first self-attention module, a second self-attention module, a first difference module, and a first MLP model; The model structures of the first retrieval model and the second retrieval model are consistent with the model structure of the codebook retrieval model; The first retrieval model is used to search based on the native protein sequence S1 and the mutant residue site b, which are input to the model. 1-2 The built-in residue feature codebook performs residue feature mapping on the microenvironment features of the current residue site to obtain the corresponding residue mapping vector E1, which is then sent to the first self-attention module. The second retrieval model is used to retrieve the mutant protein sequence S2 and the mutant residue site b from the model input. 1-2 The built-in residue feature codebook performs residue feature mapping on the microenvironment features of the current residue site to obtain the corresponding residue mapping vector E2, which is then sent to the second self-attention module. The first self-attention module is used to perform self-attention operation on the residue mapping vector E1 and send the operation result as the corresponding mapping vector P1 to the first difference module. The second self-attention module is used to perform self-attention operation on the residue mapping vector E2 and send the operation result as the corresponding mapping vector P2 to the first difference module; The first difference module is used to send the difference vector ΔP = P2 - P1 of the mapping vectors P1 and P2 to the first MLP model; The first MLP model is used to perform regression prediction on the free energy change based on the difference vector ΔP to obtain the corresponding first free energy change and output it.

7. The processing method for screening mutation sites based on free energy change according to claim 2, characterized in that, The second prediction model includes a third retrieval model, a fourth retrieval model, a third self-attention module, a fourth self-attention module, a first reconstruction module, a second reconstruction module, a first encoder, a second encoder, a second difference module, and a second MLP model; The model structures of the third retrieval model and the fourth retrieval model are consistent with the model structure of the codebook retrieval model; The third retrieval model is used to search based on the native protein sequence S3 and the mutation residue site b, which are input into the model. 3-4 The built-in residue feature codebook performs residue feature mapping on the microenvironment features of the current residue site to obtain the corresponding residue mapping vector E3, which is then sent to the third self-attention module. The fourth retrieval model is used to search based on the mutant protein sequence S4 and the mutant residue site b, which are input into the model. 3-4 The built-in residue feature codebook performs residue feature mapping on the microenvironment features of the current residue site to obtain the corresponding residue mapping vector E4, which is then sent to the fourth self-attention module. The third self-attention module is used to perform self-attention operation on the residue mapping vector E3 and send the operation result as the corresponding mapping vector P3 to the first reconstruction module. The fourth self-attention module is used to perform self-attention operation on the residue mapping vector E4 and send the operation result as the corresponding mapping vector P4 to the second reconstruction module. The first reconstruction module is used to take the heterogeneous graph G and the microenvironment subgraph M generated during the processing of the third retrieval model as the corresponding current global graph and current local graph; and to perform a traversal of all residue nodes in the current local graph; and during this traversal, to take the currently traversed residue node as the corresponding current local node; and to record the current local node as the corresponding pre-superposition vector. And based on the mapping vector P3, the vector before superposition is... The corresponding superimposed vector is obtained by superimposing the microenvironment mapping features. ; and use the superimposed vector The corresponding residue nodes in the current global graph are reset; and at the end of this round of traversal, the mutated residue site b is determined. 3-4 The corresponding residue node and the node set of the current local graph form a new first child node set, and the subgraph corresponding to the first child node set in the current global graph is extracted and sent to the first encoder as the corresponding reconstructed subgraph M1; The second reconstruction module is used to take the heterogeneous graph G and the microenvironment subgraph M generated during the processing of the fourth retrieval model as the corresponding current global graph and current local graph; and to perform a traversal of all residue nodes in the current local graph; and during this traversal, to take the currently traversed residue node as the corresponding current local node; and to record the current local node as the corresponding pre-superposition vector. And based on the mapping vector P4, the vector before superposition is... The corresponding superimposed vector is obtained by superimposing the microenvironment mapping features. ; and use the superimposed vector The corresponding residue nodes in the current global graph are reset; and at the end of this round of traversal, the mutated residue site b is determined. 3-4 The corresponding residue node and the node set of the current local graph form a new second child node set, and the subgraph corresponding to the second child node set in the current global graph is extracted and sent to the second encoder as the corresponding reconstructed subgraph M2; The model structures of the first encoder and the second encoder are consistent with the model structure of the microenvironment encoder; the model parameters of the first encoder and the second encoder are consistent. The first encoder is used to perform feature encoding on the reconstructed subgraph M1 to obtain the corresponding microenvironment encoding vector H1 and send it to the second difference module; The second encoder is used to perform feature encoding on the reconstructed subgraph M2 to obtain the corresponding microenvironment encoding vector H2, which is then sent to the second difference module. The second difference module is used to send the difference vector ΔH = H2 - H1 of the microenvironment encoding vectors H1 and H2 to the second MLP model; The second MLP model is used to perform regression prediction on the free energy change based on the difference vector ΔH to obtain the corresponding second free energy change and output it.

8. The processing method for screening mutation sites based on free energy change according to claim 2, characterized in that, The first dataset includes multiple first data records; each first data record includes a first label image, a first mask image, a set of first-class mask nodes, a set of second-class mask nodes, and a set of third-class mask nodes; the data structure of the first mask image is the same as that of the mask image G. mask Consistent, the data structure of the first label graph is consistent with that of the heterogeneous graph G; The first tag map is a microenvironment isomer map of a single residue; The first mask image is a mask image obtained by performing three types of masking on the first label image; the first type of mask node set includes multiple first type node identifiers, the second type of mask node set includes multiple second type node identifiers, and the third type of mask node set includes multiple third type node identifiers; The second dataset includes multiple second data records; the second data record includes a first training native sequence, a first training mutant sequence, a first training site, and a first tag free energy change; the first training native sequence and the first training mutant sequence are the residue sequences of a protein complex before and after a mutation occurs at the residue site corresponding to the first training site; the first tag free energy change is the free energy change of the current protein complex before and after the mutation.

9. The processing method for screening mutation sites based on free energy change according to claim 8, characterized in that, The step of training the codebook retrieval model based on the pre-trained framework using the first dataset specifically includes: Step 91: Take the first data record of the first dataset as the corresponding current training record; Step 92: Extract the corresponding first label image, first mask image, first type mask node set, second type mask node set, and third type mask node set from the current training record as the corresponding current label image, current mask image, current first type node set, current second type node set, and current third type node set; Step 93: Use the current mask image as the corresponding mask image G. mask The pre-trained framework is input into the data and processed to obtain the corresponding type reconstruction set X. 1 The dihedral angle reconstruction set X 2 The local orientation reconstruction set X 3 The microenvironment encoding vector set generated by the microenvironment encoder, the codebook tag vector set generated by the codebook forward lookup module, and the three types of retrieval codebook vector sets generated by the codebook reverse lookup module are saved during this processing. Step 94: Denote each microenvironment coding vector in the microenvironment coding vector set stored this time as the corresponding H. a 1 ≤ index a ≤ N A N A Let Z be the total number of residue nodes in the current tag graph; and let Z be the corresponding codebook tag vector of the codebook tag vector set stored this time. a ; and mark the vector Z of each codebook a The three types of codebook markers in the text are denoted as the corresponding , , ; and denote the corresponding retrieval codebook vector sets in the three types of retrieval codebook vector sets stored this time, respectively, and the sub-codebooks C1, C2, and C3, as the corresponding vector sets { }, vector set { }, vector set { }, Retrieval codebook vector , , With codebook tags , , One-to-one correspondence; and the subcodebook C1 and the vector set { The various retrieval codebook vectors of} The matching token embedding vector is denoted as The subcodebook C2 and the vector set { The various retrieval codebook vectors of} The matching token embedding vector is denoted as The subcodebook C3 and the vector set { The various retrieval codebook vectors of} The matching token embedding vector is denoted as ; and is determined by the tag vector Z of each codebook. a { , , } corresponding token embedding vector , , The corresponding token embedding vector set is formed; and each microenvironment encoding vector H is composed of the corresponding token embedding vector set. a The token embedding vector set and its corresponding set form a first verification pair [H] a ,{ , , }]; Step 95, combine the current label graph and the type reconstruction set X 1 The type embedding vector and the reconstructed type embedding vector corresponding to each of the node identifiers of the current type of node set are combined to form a corresponding second check pair; and the current label graph and the dihedral reconstruction set X are combined. 2 The dihedral embedding vector and the reconstructed dihedral embedding vector corresponding to each of the two types of node identifiers in the current set of two types of nodes form a corresponding third check pair; Wherein, the second check pair is denoted as ( , The third check pair is denoted as ( , ); 1≤index b≤N B 1 ≤ index c ≤ N C N B N C These are the total number of node identifiers corresponding to the current first-type node set and the current second-type node set, respectively. , These are the type embedding vector and the reconstructed type embedding vector of the b-th second check pair, respectively; , These are the dihedral embedding vector and the reconstructed dihedral embedding vector of the c-th third verification pair, respectively. Step 96, reconstruct the current label map and the local orientation set X. 3 In the process, the difference vector between the pair of local orientation embedding vectors and the reconstructed local orientation embedding vectors corresponding to each of the three types of node identifiers in the current three types of node sets is calculated, and the calculation result is used as the corresponding first prediction error vector; and the current mask image and the local orientation reconstruction set X are used to calculate the difference vector between the pair of local orientation embedding vectors and the reconstructed local orientation embedding vectors. 3 In the process, the difference vector between the pair of local orientation embedding vectors and the reconstructed local orientation embedding vectors corresponding to each of the three types of node identifiers in the current three types of node sets is calculated, and the calculation result is used as the corresponding first label error vector; and each of the first label error vectors and its corresponding first prediction error vectors forms a corresponding fourth check pair; The fourth check pair is denoted as ( , ); 1≤index d≤N D N D The total number of node identifiers corresponding to the current three types of node sets; , These are the first label error vector and the first prediction error vector of the dth fourth check pair, respectively. Step 97: Construct a first loss function L1 for the microenvironment encoder and the residue feature codebook C based on the constraint loss function; construct a second loss function L2 for the sub-codebook C1 and the first decoder based on the cross-entropy loss function; construct a third loss function L3 for the sub-codebook C2 and the second decoder based on the MSE loss function; construct a fourth loss function L4 for the sub-codebook C3 and the third decoder based on the Huber loss function; and construct an overall loss function L based on the first, second, fourth, and fourth loss functions. total ; and the obtained N A The first verification pair, N B The second check pair, N C The third check pair, N D The fourth check pair is substituted into the overall loss function L. total The corresponding first loss value is obtained through calculation; in, The first loss function L1 is: ； 1≤index s≤3, sg() is the stop gradient operation, and η is the preset tradeoff hyperparameter; The second loss function L2 is: ； L CE () represents the cross-entropy loss function; The third loss function L3 is: ； 1 ≤ index q ≤ 7, The dihedral embedding vector The q-th dihedral angle in the middle, For the reconstructed dihedral embedding vector The q-th dihedral angle in; The fourth loss function L4 is: ； L Huber () represents the Huber loss function; The overall loss function L total for: ； λ is a preset tradeoff hyperparameter; Step 98: Identify whether the first loss value meets a preset first loss value range; if the first loss value meets the first loss value range, identify whether the current training record is the last first data record in the first dataset; if so, proceed to step 99; otherwise, take the next first data record in the first dataset as the new current training record and return to step 92; if the first loss value does not meet the first loss value range, first optimize the microenvironment encoder and the residue features based on a preset first model optimizer in the direction of minimizing the first loss function L1. Codebook C undergoes one round of parameter optimization. Then, based on a preset second model optimizer, a round of parameter optimization is performed on the sub-codebook C1 and the first decoder, aiming to minimize the second loss function L2. Next, based on a preset third model optimizer, a round of parameter optimization is performed on the sub-codebook C2 and the second decoder, aiming to minimize the third loss function L3. Finally, based on a preset fourth model optimizer, a round of parameter optimization is performed on the sub-codebook C3 and the third decoder, aiming to minimize the fourth loss function L4. Finally, based on a preset fifth model optimizer, a round of parameter optimization is performed on the overall loss function L... total The direction that reaches the minimum value is used to perform a round of parameter optimization on the microenvironment encoder and the residue feature codebook C, and the process returns to step 93 when the parameter optimization ends. Wherein, the first model optimizer includes the Adam optimizer and the SGD optimizer; the second model optimizer includes the Adam optimizer and the SGD optimizer; the third model optimizer includes the Adam optimizer and the SGD optimizer; the fourth model optimizer includes the Adam optimizer and the SGD optimizer; and the fifth model optimizer includes the Adam optimizer and the SGD optimizer. Step 99: Stop the current training and confirm that the retrieval model training has ended.

10. The processing method for screening mutation sites based on free energy change according to claim 8, characterized in that, The step of first setting the internal component parameters of the first and second prediction models based on the model parameters of the codebook retrieval model, and then training the first and second prediction models based on the second dataset, specifically includes: Based on the model parameters of the codebook retrieval model and the built-in residue feature codebook C, the model parameters of the first and second retrieval models inside the first prediction model and the built-in residue feature codebook are set, and the parameters of the first and second retrieval models are frozen after the setting is completed. Based on the model parameters of the codebook retrieval model and the built-in residue feature codebook C, the model parameters of the third and fourth retrieval models inside the second prediction model and the built-in residue feature codebook are set, and the parameters of the third and fourth retrieval models are frozen after the settings are completed. And based on the model parameters of the microenvironment encoder of the codebook retrieval model, the model parameters of the first and second encoders inside the second prediction model are initialized; After the parameters of the first and second prediction models are set, the first prediction model is trained based on the second dataset, and the second prediction model is trained based on the second dataset. After the training of both the first and second prediction models is completed, confirm that the training of both prediction models is finished.

11. The processing method for screening mutation sites based on free energy variation according to claim 1, characterized in that, The process of combining native and mutant protein information based on the user-input first native protein sequence and first mutant residue set to obtain multiple corresponding original groups specifically includes: Each of the first mutated residues in the first set of mutated residues is taken as the corresponding current mutation; the first residue type and the first residue site of the current mutation are taken as the corresponding current mutated residue and current mutated site; a new residue sequence is obtained by replacing the original residues in the first native protein sequence corresponding to the current mutated site with the current mutated residues, and this new residue sequence is taken as the corresponding first mutated protein sequence; the current mutated site is taken as a corresponding first residue site; and the first native protein sequence corresponding to the current mutation, the first mutated protein sequence, and the first residue site form a corresponding original set.

12. The processing method for screening mutation sites based on free energy variation according to claim 1, characterized in that, The step of using the first prediction model to perform a preliminary screening of all the original groups according to a preset screening and sorting rule to obtain N preferred groups specifically includes: The first native protein sequence, the first mutant protein sequence, and the first residue site of each of the original groups are taken as a corresponding set of native protein sequence S1, mutant protein sequence S2, and mutant residue site b. 1-2 The first prediction model is input and processed to obtain the corresponding first free energy variable; all the obtained first free energy variables are sorted according to the screening and sorting rules; and the N original groups corresponding to the top N first free energy variables are extracted as the corresponding N preferred groups.

13. The processing method for screening mutation sites based on free energy variation according to claim 1, characterized in that, The step of using the second prediction model to continue sorting the N preferred groups according to the screening and sorting rules to generate the corresponding final selection group sequence and feeding it back to the current user specifically includes: The first native protein sequence, the first mutant protein sequence, and the first residue site of each of the preferred groups are taken as a corresponding set of native protein sequence S3, mutant protein sequence S4, and mutant residue site b. 3-4 The second prediction model is input and processed to obtain the corresponding second free energy change; and the first native protein sequence, the first mutant protein sequence, the first residue site and the second free energy change corresponding to each of the preferred groups are combined to form a corresponding final selection group; and the N final selection groups are sorted according to the screening and sorting rules to obtain the corresponding final selection group sequence and fed back to the current user.

14. An apparatus for performing the processing method for screening mutation sites based on free energy change as described in any one of claims 1-13, characterized in that, The device includes: a model building module, a data acquisition module, a model training module, and a model application module; The model building module is used to construct a feature mapping model for mapping residue features of microenvironmental features of specified residue sites in a protein complex using a codebook retrieval method, denoted as the corresponding codebook retrieval model; and to construct a pre-training framework for the codebook retrieval model; and to construct two prediction models based on the codebook retrieval model for predicting the free energy change ΔΔG before and after a single-point mutation in the protein complex, denoted as the first prediction model and the second prediction model; the codebook retrieval model is used to perform residue feature mapping on the microenvironmental features of the current residue site according to the protein sequence S and residue site b input to the model, as well as the built-in residue feature codebook C, to obtain the corresponding residue mapping vector E; the pre-training framework is used to process the mask image G input to the framework. mask The three types of embedding vectors of all residue nodes are reconstructed to obtain the corresponding type reconstruction set X. 1 Dihedral Reconstruction Set X 2 Local orientation reconstruction set X 3 The first prediction model is used to predict the original protein sequence S1, the mutant protein sequence S2, and the mutant residue site b based on the input of the model. 1-2 Microenvironment embedding encoding is performed, and the free energy change is predicted based on the differential embedding of the two embedding encodings to obtain the corresponding first free energy change; the second prediction model is used to predict the free energy change based on the native protein sequence S3, the mutant protein sequence S4, and the mutant residue site b as input to the model. 3-4 Microenvironment embedding encoding is performed, and corresponding microenvironment reconstruction encoding is performed based on the two embedding encodings. Then, the free energy change is predicted based on the difference features of the two reconstruction features to obtain the corresponding second free energy change. The data acquisition module is used to prepare the training dataset through data acquisition to obtain the corresponding first dataset and second dataset; The model training module is used to train the codebook retrieval model based on the pre-training framework and the first dataset. After the retrieval model training is completed, the internal component parameters of the first and second prediction models are set based on the model parameters of the codebook retrieval model, and then the first and second prediction models are trained based on the second dataset. The model application module is used to, after the training of the two types of prediction models, combine native-mutant protein information based on the first native protein sequence and the first set of mutant residues input by the user to obtain multiple corresponding original groups; and use the first prediction model to perform a preliminary screening of all the original groups according to a preset screening and sorting rule to obtain N preferred groups; and use the second prediction model to continue to sort the N preferred groups according to the screening and sorting rule to generate the corresponding final selected group sequence and feed it back to the current user; the first set of mutant residues includes multiple first mutant residues, the total number of the first mutant residues is greater than the preset total number N, where N is a positive integer; the first mutant residues include a first residue type and a first residue site; the original groups include the first native protein sequence, the first mutant protein sequence, and the first residue site; the screening and sorting rule is to sort in descending order of free energy change or in ascending order of free energy change.

15. An electronic device, characterized in that, include: Memory, processor, and transceiver; The processor is configured to be coupled to the memory, read and execute instructions in the memory to implement the method according to any one of claims 1-3; The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a computer, cause the computer to perform the method according to any one of claims 1-13.

Images