Protein disulfide bond mutation site prediction method and device, equipment, storage medium

Abstract

The application belongs to the technical field of bioinformatics, and discloses a protein disulfide bond mutation site prediction method and device, equipment and a storage medium. The predicted structure of a to-be-transformed protein is acquired to extract a plurality of candidate residue pairs, and the structural feature information of each candidate residue pair is extracted, including the inter-residue distance matrix of the two residues, the sequence distance, and the relative solvent accessible surface area, depth index and predicted local distance difference test value of each residue. The predicted model is input to obtain the probability value of each candidate residue pair forming a disulfide bond. The candidate residue pair with a probability greater than a specified probability is determined as a disulfide bond mutation site. Therefore, more comprehensive features can be used to comprehensively predict the bonding probability of the disulfide bond, the model is less dependent on individual traditional features, the high sensitivity of the traditional method to atomic coordinates and the uncertainty of the prediction result are overcome to a certain extent, the robustness is higher, and the prediction accuracy of the protein disulfide bond mutation site is improved.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a method, apparatus, device, and storage medium for predicting protein disulfide bond mutation sites. Background Technology

[0002] Protein structural stability is crucial for the proper functioning of proteins in biological processes, drug design, and disease treatment. Rational artificial design or modification methods can effectively improve protein structural stability, with the introduction of disulfide bonds being a common approach. A disulfide bond, also known as a disulfide bridge, is formed by the oxidative folding of the thiol groups of two cysteine ​​residues and is the most common covalent link in proteins besides peptide bonds. Many proteins, such as antibodies and membrane receptors, are rich in disulfide bonds in their native conformations, and their formation is vital for the structural stability and functional integrity of these proteins. Introducing disulfide bonds at appropriate sites in wild-type proteins is a common and effective strategy to enhance protein stability. Disulfide bonds are believed to reduce conformational entropy, imposing geometric constraints on the protein backbone to limit the separation of peptide chains and help maintain a specific conformation. In the biopharmaceutical field, the formation and stability mechanisms of protein disulfide bonds can help design more stable protein drugs, improving efficacy and safety.

[0003] Determining suitable mutation sites is crucial for the introduction of disulfide bonds. Traditional experience-based disulfide bond mutation methods rely heavily on the designer's subjective judgment, which can be time-consuming and costly. Structure-based disulfide bond mutation site prediction algorithms extract spatial information features (distances, angles, dihedral angles, etc.) of residues in the protein structure, starting from the structural characteristics of disulfide bonds. They then predict the score, energy, or probability of disulfide bond formation between residues using empirical rule-based methods or machine learning. These disulfide bond mutation site prediction programs include SSBONDPredict, MODIP, and Disulfide by Design (DbD and DbD2). However, these traditional prediction methods primarily score residues based on spatial distances, angles, and dihedral angles, making them highly sensitive to the spatial coordinates of related atoms. Since the atomic coordinates of specific protein residue pairs differ in different experimental structures, the calculated spatial distances vary, leading to different prediction results. In other words, different experimental structures of the same protein may yield different prediction results, resulting in significant uncertainty in the prediction outcomes. Summary of the Invention

[0004] The purpose of this invention is to provide a method, apparatus, device, and storage medium for predicting protein disulfide bond mutation sites, which can have higher robustness and improve the accuracy of predicting protein disulfide bond mutation sites.

[0005] The first aspect of this invention discloses a method for predicting protein disulfide bond mutation sites, comprising:

[0006] Obtain the target predicted structure of the protein to be modified;

[0007] Multiple candidate residue pairs are extracted based on the target predicted structure of the protein to be modified, and candidate residue pairs are determined based on the candidate residue pairs.

[0008] Structural feature information is extracted for each candidate residue pair; the structural feature information includes the residue distance matrix between the two residues in each candidate residue pair, the sequence distance, and the relative solvent-accessible surface area, depth index, and predicted local distance difference test value of each residue in each candidate residue pair;

[0009] The structural feature information of each candidate residue pair is input into a pre-trained prediction model to obtain the probability value of each candidate residue pair forming a disulfide bond;

[0010] Candidate residue pairs with a probability value greater than a specified probability are identified as disulfide bond mutation sites in the protein to be modified.

[0011] In some embodiments, the training process of the prediction model includes:

[0012] In each protein, the cysteine ​​residues that have formed disulfide bonds in the three-dimensional structure are reverse-mutated into alanine pairs to obtain the template structure and corresponding mutation sequence of each protein. Based on the template structure of each protein, the corresponding mutation sequence is used to predict the structure and obtain multiple predicted structures of each protein. The alanine pairs resulting from the mutation of the original disulfide bond residue pairs are extracted from each predicted structure of each protein as positive samples.

[0013] Extract four adjacent amino acids from each positive sample, combine them in any pair to obtain four pairs of cross amino acids, calculate the CA-CA distance between the CA atoms of the four pairs of cross amino acids, and select the pair of amino acids with the shortest CA-CA distance as the negative sample.

[0014] The structural feature information of the positive and negative samples is extracted respectively, and the fully connected neural network is trained to obtain a prediction model.

[0015] In some embodiments, obtaining the target predicted structure of the protein to be modified includes:

[0016] Determine if the protein to be modified has a corresponding experimental structure;

[0017] If so, the experimental structure of the protein to be modified is used as a template, and the predicted structure obtained by optimizing the structure of the protein to be modified using ColabFold is determined as the target predicted structure.

[0018] If not, the predicted structure obtained by de novo structure prediction of the protein to be modified using ColabFold will be determined as the target predicted structure.

[0019] In some embodiments, multiple candidate residue pairs are extracted based on the target predicted structure of the protein to be modified, including:

[0020] When focusing solely on the target conformation and requiring its stability, the predicted target structure of the protein to be modified is determined as the target conformation.

[0021] Based on the sequence distance between each pair of amino acid residues in the target conformation, residue pairs with a sequence distance greater than a specified threshold are selected as candidate residue pairs.

[0022] In some embodiments, multiple candidate residue pairs are extracted based on the target predicted structure of the protein to be modified, including:

[0023] When simultaneously focusing on the target conformation and the competing conformation, and requiring a stable target conformation while keeping the competing conformation unstable, the predicted target structure of the protein to be modified is determined as the target conformation, and another specific conformation of the protein to be modified is determined as the competing conformation; wherein, the target conformation and the competing conformation are different conformations;

[0024] Calculate the CA interatomic distance matrix of amino acid residues in the target conformation and the competing conformation respectively. Subtract the CA interatomic distance matrix of the target conformation from the CA interatomic distance matrix of the competing conformation to obtain the CA interatomic distance difference matrix.

[0025] From the CA interatomic distance difference matrix, select residue pairs whose CA interatomic distance difference is greater than a specified threshold as candidate residue pairs.

[0026] In some embodiments, determining candidate residue pairs based on alternative residue pairs includes:

[0027] Based on the CA interatomic distance matrix of amino acid residues in the target conformation, candidate residue pairs with CA interatomic distances within a specified range are selected as candidate residue pairs.

[0028] In some embodiments, structural feature information of each candidate residue pair is extracted, including:

[0029] Calculate the residue distance matrix and sequence distance between the two residues in each candidate residue pair, as well as the relative solvent-accessible surface area, depth index, and predicted local distance difference test value for each residue in each candidate residue pair;

[0030] The inter-residue distance matrix, the sequence distance, the relative solvent-accessible surface area of ​​each residue, the depth index, and the predicted local distance difference test value are used to determine the structural feature information of each candidate residue pair.

[0031] A second aspect of this invention discloses a device for predicting protein disulfide bond mutation sites, comprising:

[0032] The first acquisition unit is used to acquire the target predicted structure of the protein to be modified.

[0033] The second acquisition unit is used to extract multiple candidate residue pairs based on the target predicted structure of the protein to be modified, and to determine candidate residue pairs based on the candidate residue pairs.

[0034] The feature extraction unit is used to extract the structural feature information of each candidate residue pair; the structural feature information includes the residue distance matrix between the two residues of each candidate residue pair, the sequence distance, and the relative solvent-accessible surface area, depth index, and predicted local distance difference test value of each residue in each candidate residue pair;

[0035] The prediction unit is used to input the structural feature information of each candidate residue pair into the pre-trained prediction model to obtain the probability value of each candidate residue pair forming a disulfide bond.

[0036] A determining unit is used to identify candidate residue pairs with a probability value greater than a specified probability as disulfide bond mutation sites in the protein to be modified.

[0037] In some embodiments, it also includes:

[0038] The first acquisition unit is used to reverse the virtual mutation of cysteine ​​residues that have formed disulfide bonds in the three-dimensional structure of each protein into alanine pairs, obtain the template structure and corresponding mutation sequence of each protein, perform structural prediction on the corresponding mutation sequence based on the template structure of each protein, and obtain multiple predicted structures for each protein; extract the alanine pairs after mutation of the original disulfide bond residue pairs from each predicted structure of each protein as positive samples.

[0039] The second acquisition unit is used to extract four adjacent amino acids from each positive sample, combine them in any pair to obtain four pairs of cross amino acids, calculate the CA-CA distance between the CA atoms of the four pairs of cross amino acids, and select the pair of amino acids with the shortest CA-CA distance as the negative sample.

[0040] The training unit is used to extract the structural feature sample information of the positive samples and the negative samples respectively, and to train the fully connected neural network to obtain a prediction model.

[0041] A third aspect of the present invention discloses an electronic device, including a memory storing executable program code and a processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the protein disulfide bond mutation site prediction method disclosed in the first aspect.

[0042] The fourth aspect of the present invention discloses a computer-readable storage medium storing a computer program, wherein the computer program causes a computer to execute the protein disulfide bond mutation site prediction method disclosed in the first aspect.

[0043] The beneficial effects of this invention are as follows: by obtaining the target predicted structure of the protein to be modified, multiple candidate residue pairs are extracted, and the structural feature information of each candidate residue pair is extracted, including the residue distance matrix and sequence distance between the two residues in each candidate residue pair, as well as the relative solvent accessible surface area, depth index, and predicted local distance difference test value of each residue in each candidate residue pair. The results are input into a pre-trained prediction model to obtain the probability value of each candidate residue pair forming a disulfide bond. Candidate residue pairs with a probability value greater than a specified probability are identified as disulfide bond mutation sites of the protein to be modified. This allows for the use of more comprehensive features to comprehensively predict the bonding probability of disulfide bonds, and more comprehensively considers the factors affecting the probability of disulfide bond formation. It also reduces the model's high dependence on individual traditional features. Compared with traditional methods, it overcomes to some extent the high sensitivity of traditional methods to atomic coordinates and the uncertainty of prediction results, and has higher robustness, thus improving the prediction accuracy of protein disulfide bond mutation sites. Attached Figure Description

[0044] The accompanying drawings illustrate specific examples of the technical solutions described in this invention and, together with the detailed embodiments, form part of the specification, serving to explain the technical solutions, principles, and effects of this invention.

[0045] Unless otherwise specified or defined, the same reference numerals in different figures represent the same or similar technical features, and different reference numerals may be used to represent the same or similar technical features.

[0046] Figure 1 This is the reverse virtual mutation and structure prediction process for PDB protein structures disclosed in the embodiments of the present invention;

[0047] Figure 2 This is a schematic diagram of the structure before and after reverse virtual mutation and structure prediction of disulfide bonds disclosed in an embodiment of the present invention;

[0048] Figure 3 This is a detailed implementation flowchart of a protein disulfide bond mutation site prediction method disclosed in an embodiment of the present invention;

[0049] Figure 4 This is a general flowchart of the prediction of disulfide bond mutation sites in the protein to be modified, as disclosed in the embodiments of the present invention.

[0050] Figure 5 This is a schematic diagram of the pre-fusion conformation of the RSV F protein disclosed in an embodiment of the present invention;

[0051] Figure 6 This is a schematic diagram of the conformation of the RSV F protein after fusion, as disclosed in an embodiment of the present invention;

[0052] Figure 7 This is a flowchart of the disulfide bond mutation prediction based on two predicted structures of the RSV F protein disclosed in this embodiment of the invention;

[0053] Figure 8 This is a schematic diagram of the A-chain structure of the RSV F protein before fusion as disclosed in an embodiment of the present invention;

[0054] Figure 9 This is a schematic diagram of the A-chain structure of the RSV F protein after fusion, as disclosed in an embodiment of the present invention.

[0055] Figure 10 This is a schematic diagram of the structure of a protein disulfide bond mutation site prediction device disclosed in an embodiment of the present invention;

[0056] Figure 11 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of the present invention.

[0057] Explanation of reference numerals in the attached figures:

[0058] 1001, First acquisition unit; 1002, Second acquisition unit; 1003, Feature extraction unit; 1004, Prediction unit; 1005, Determination unit; 1101, Memory; 1102, Processor. Detailed Implementation

[0059] Unless otherwise specified or defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. When combined with the technical solutions of the invention in a real-world scenario, all technical and scientific terms used herein may also have meanings corresponding to the purpose of achieving the technical solutions of the invention. The terms "first," "second," etc., used herein are merely for distinguishing names and do not represent a specific number or order. The term "and / or," as used herein, includes any and all combinations of one or more of the associated listed items.

[0060] It should be noted that when a component is considered "fixed" to another component, it can be directly fixed to the other component or there can be an intervening component; when a component is considered "connected" to another component, it can be directly connected to the other component or there can be an intervening component; when a component is considered "mounted" on another component, it can be directly mounted on the other component or there can be an intervening component; when a component is considered "placed" on another component, it can be directly placed on the other component or there can be an intervening component.

[0061] Unless otherwise specified or defined, the terms "described" or "the" as used herein refer to the technical features or technical content mentioned or described prior to the relevant section, which may be the same as or similar to the technical features or technical content mentioned herein. Furthermore, the terms "comprising" and "having," and any variations thereof, as used herein, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0062] This invention discloses a method for predicting protein disulfide bond mutation sites, which can be implemented through computer programming. The method can be executed by electronic devices such as computers, laptops, and tablets, or by a protein disulfide bond mutation site prediction device embedded in an electronic device; this invention does not limit this. To facilitate understanding of this invention, specific embodiments will be described in more detail below with reference to the accompanying drawings.

[0063] To facilitate understanding of the present invention, the embodiments of the present invention first provide a training method for a protein disulfide bond mutation site prediction model, including the following steps S01~S02 (not shown):

[0064] S01. In the three-dimensional structure of each protein, the cysteine ​​residues that have formed disulfide bonds are reverse-mutated into alanine pairs to obtain the template structure and corresponding mutant sequence of each protein. Based on the template structure of each protein, the structure of the corresponding mutant sequence is predicted to obtain multiple predicted structures of each protein. The alanine pairs after mutation of the original disulfide bond residue pairs are extracted from each predicted structure of each protein as positive samples. In addition, the four adjacent amino acids of each positive sample are extracted, and they are randomly paired to obtain four pairs of cross amino acids. The CA-CA distance between the CA atoms of the four pairs of cross amino acids is calculated, and the pair of amino acids with the shortest CA-CA distance is selected as the negative sample.

[0065] Database preparation for the machine learning model: Specifically, the dataset preparation process for the prediction model of this invention is as follows: First, referring to the approach of SSBondPredict, 14,647 protein three-dimensional structures from the Protein Data Bank (PDB) database were selected to form a non-redundant PDB dataset. The definitions of positive and negative samples are as follows: using the "SSBOND" marker in the PDB file, the alanine pairs (Ala-Ala) formed by the mutation of the original disulfide bond residue pair are extracted as positive samples, and the positive sample residue pairs are denoted as C. i Cj And, extracting C i C j The four adjacent amino acids in the sequence, namely X i-1 X i+1 and X j-1 X j+1 By combining any two of them, four pairs of cross amino acids (X) are obtained. i-1 X j-1 X i-1 X j+1 X i+1 X j-1 X i+1 X j+1 Then, calculate the CA-CA distance between the CA atoms of the four pairs of crossed amino acids, and select the pair of amino acids with the shortest CA-CA distance as the negative sample.

[0066] It's important to note that machine learning requires both positive and negative samples. The alanine pair resulting from the mutation of the disulfide bond residues is considered a positive sample, while its neighboring residue pairs are considered negative samples. These neighboring residue pairs do not form disulfide bonds in the original protein structure, hence they are negative samples. There are two reasons for choosing negative samples in this way: First, it's rare for two adjacent disulfide bonds to exist in a protein. For example, adjacent residues 10 and 11 forming disulfide bonds 10-50 and 11-51 with another pair of adjacent residues 50 and 51, respectively, is uncommon. Second, the CA-CA spatial distance of these negative samples is close to that of the positive samples. Compared to randomly selecting two distant residues from the protein as negative samples, these negative samples are harder to distinguish from the positive samples, thus helping to train a more accurate model.

[0067] In this embodiment of the invention, residue numbers of 12,496 pairs of positive samples (residue pairs that have formed disulfide bonds) and 12,496 pairs of negative samples (residue pairs that have not formed disulfide bonds) were extracted.

[0068] Then, as Figure 1 As shown, the original PDB structure of each protein in the PDB dataset was reverse-mutated using the pymol mutagenesis tool. That is, the two cysteine ​​residues Cys-Cys involved in the formation of each disulfide bond in the three-dimensional structure of each protein were reverse-mutated to alanine Ala-Ala to obtain the template structure and the corresponding mutation sequence query. The template structure is the mutated PDB structure. Except for the mutated residues, the coordinates of the other residues are the same as the original PDB structure.

[0069] Subsequently, ColabFold was used to predict the structure of the mutated sequence query after reverse virtual mutation based on the template structure, obtaining multiple predicted structures. Unlike the usual de novo structure prediction process, using the mutated structure as the template structure ensures the similarity between the predicted structure of the mutated sequence after reverse mutation and the original PDB structure. The number of predicted structures can be preset by the developers. For example, if the number is set to 5, each PDB protein will obtain 5 predicted structures through the above prediction process. Each predicted structure can extract the same number of residue pairs, so the actual number of samples extracted from the predicted structures is 5 times that extracted from the original experimental structure.

[0070] S02. Extract the structural feature information of positive and negative samples respectively, and train the fully connected neural network to obtain the prediction model.

[0071] For residues i and j of each sample, the following feature combinations are calculated to form structural feature sample information:

[0072] (1) The distance matrix between two residues:

[0073] The elements of this inter-residue distance matrix are composed of the pairwise distances between multiple atoms (N, CA, C, O, CB) of residue i and multiple atoms (N, CA, C, O, CB) of residue j. The inter-residue distance matrix has a dimension of 5×5=25. For example... Figure 2 As shown, Figure 2 The diagram illustrates the structures before and after reverse virtual mutation and structure prediction of the disulfide bond Cys1-Cys2. The inter-residue distance features were extracted from the N, CA, C, O, and CB atomic coordinates of the alanine Ala1-Ala2 in the predicted structure after the mutation. Figure 2 The dashed lines represent the distances between the CA atom of one residue Ala1 and the N, CA, C, O, and CB atoms of another residue Ala2, and so on, for a total of 5×5=25 distance features between residues.

[0074] Using the N-atom coordinates (x) of residue i i y i , z i ) and the N-atom coordinates (x) of residue j j y j , z j For example, the distance d between the N atom of residue i and the N atom of residue j. iN_jN Calculated using the following formula (1):

[0075] (1)

[0076] The distances between the other atoms of residues i and j are calculated in a similar manner.

[0077] (2) Sequence distance between two residues:

[0078] Sequence distance between residues i and j (1-dimensional), determined by the position of the residue in the primary sequence of the protein. Assuming residue i is the i-th residue in the protein sequence and residue j is the j-th residue in the protein sequence, the sequence distance between them is: .

[0079] (3) Relative solvent-accessible surface area of ​​each residue:

[0080] It should be noted that the solvent accessible surface area (ASA) represents the atomic surface area that the solvent can access. RASA (Relative ASA) refers to the ratio of the ASA value of a specific residue to the standard ASA value of that residue. The specific calculation formula is as follows, taking residue i as an example:

[0081] (2)

[0082] Among them, ASA i This represents the calculated ASA value for residue i. stand,i The standard ASA value represents residue i. This invention calculates the RASA value of residues in the predicted structure using a Protein Structure and Interaction Analyzer (PSAIA). Similarly, the relative solvent accessible surface area (RASA) of residues i and j can be calculated. i RASA j (2-dimensional).

[0083] (4) Depth index of each residue: Depth index DPX of residue i and residue j i DPX j (2D), or can be calculated using PSAIA.

[0084] The depth index DPX of a residue is the average of the depth indices of all atoms in the residue, where the depth index of an atom is defined as the distance d between that atom and the nearest solvent-accessible atom. Taking residue i containing N atoms as an example:

[0085] (3)

[0086] in, It represents the depth index of the nth atom contained in residue i.

[0087] (5) Test values ​​of predicted local distance differences for each residue:

[0088] The Predicted Local-Distance Difference Test (pLDDT) reflects the confidence level of each residue in the predicted structure, with a value ranging from 0 to 100. A higher pLDDT value indicates a more reliable predicted structure. The pLDDT values ​​of residues i and j can be extracted from the predicted structure PDB file (columns 61-66 of the predicted structure PDB file).

[0089] After extracting the structural feature information of the samples, 80% and 20% of the total samples were used as the training and testing sets, respectively, to train and test the fully connected neural network. The basic model structure uses a four-layer fully connected neural network, consisting of an input layer, a first hidden layer, a second hidden layer, and an output layer connected sequentially. The middle two hidden layers use the ReLU activation function, and the output layer uses the Sigmoid activation function. Model hyperparameters with different gradients (number of neurons in the first hidden layer l1, number of neurons in the second hidden layer l2, learning rate lr, batch size, number of epochs) were set, and Ray Tune was used for automatic hyperparameter tuning. The final prediction model was obtained by determining the hyperparameters (neurons in each hidden layer, learning rate, etc.) that achieve the relatively optimal performance. The ASHA scheduler was used to automatically search for the best hyperparameter configuration to optimize model performance (i.e., maximize accuracy). 10,000 independent trials were run, and Ray Tune was used to search and determine the optimal hyperparameter combination among these trials. After optimization, the optimal hyperparameter configuration of the model is as follows: the number of neurons in the two hidden layers, l1 and l2, are 256 and 64 respectively; the learning rate, lr, is 0.008; the batch size is 16; and the number of epochs is 50.

[0090] like Figure 3 As shown, the method for predicting protein disulfide bond mutation sites provided in this embodiment of the invention includes the following steps 110-150:

[0091] 110. Obtain the target predicted structure of the protein to be modified.

[0092] As an optional implementation, step 110 includes the following steps 1101-1103 (not shown):

[0093] 1101. Determine if the protein to be modified has a corresponding experimental structure. If yes, proceed to step 1102; otherwise, proceed to step 1103.

[0094] 1102. Using the experimental structure of the protein to be modified as a template, the predicted structure obtained by optimizing the structure of the protein to be modified using ColabFold is determined as the target predicted structure.

[0095] 1103. The predicted structure obtained by de novo structure prediction of the protein to be modified using ColabFold is determined as the target predicted structure.

[0096] In the process of specific protein modification, if the experimental structure is available in the PDB file, it can be used as a template for structure optimization using ColabFold (see reference). Figure 1 The steps are as follows, but without the need for reverse virtual mutation; if no experimental structure is available, de novo structure prediction is performed using the ColabFold method. Then, as... Figure 4 As shown, starting from the target predicted structure of the protein to be modified, and combined with the design requirements of the protein to be modified, candidate residue pairs that may undergo Cys mutation are extracted from the target predicted structure; for each candidate residue pair, its structural feature information is extracted, and the probability value of each candidate residue pair forming a disulfide bond is predicted using a trained prediction model. That is, steps 120-140 are executed.

[0097] 120. Extract multiple candidate residue pairs based on the target predicted structure of the protein to be modified, and determine the candidate residue pairs based on the candidate residue pairs.

[0098] It should be noted that disulfide bond mutations stabilize specific protein conformations (i.e., target conformations) to achieve purposes such as structural characterization, enhanced immunogenicity, improved catalytic activity, and increased affinity. The specific modification requirements may differ for different proteins, including but not limited to ensuring the correct folding of the enzyme's catalytic pocket and antigenic epitopes. For example, if the goal is to improve enzyme catalytic activity, candidate residue pairs are identified from the target conformation of the enzyme to be modified. Further screening and filtering of these candidate residue pairs can be performed, for instance, considering only appropriate disulfide bond mutations around the enzyme's active site.

[0099] Generally, there are two ways to select candidate residue pairs. First, when only the target conformation is of interest and stability is required, residue pairs with a sequence distance greater than a specified threshold between residue i and residue j are selected as candidate residue pairs based on the sequence distance between amino acid residues in the target conformation. Second, when both the target and competing conformations are of interest and the target conformation needs to be stabilized while the competing conformation remains unstable (that is, when a specific conformation of the protein needs to be selectively stabilized while its competing conformation remains unstable, in which case both specific conformations of the protein are of interest), the CA atomic distance matrices of amino acid residues in the target and competing conformations are calculated separately. The CA atomic distance matrix of the target conformation is subtracted from the CA atomic distance matrix of the competing conformation to obtain the CA atomic distance difference matrix. From the CA atomic distance difference matrix, residue pairs with a CA atomic distance difference greater than a specified threshold are selected as candidate residue pairs.

[0100] It's important to clarify that the terms "target conformation" and "competitive conformation" are defined based on the specific modification requirements of the protein; that is, any conformation of the protein can be identified as the target conformation / competitive conformation. "If only the target conformation is considered and its stability is required" means that only the target conformation needs to be stabilized, regardless of the stability of other conformations of the protein to be modified. "If both the target and competitive conformations are considered and the target conformation needs to be stabilized while the competitive conformation remains unstable" means that both conformations need to be considered; and the target conformation needs to be stabilized while the competitive conformation remains unstable. For example, if the protein to be modified is a viral fusion protein (such as RSV F protein, spike protein, etc.), viral fusion proteins have two conformations: a pre-fusion conformation and a post-fusion conformation. In some modification scenarios, only the target conformation can be considered (the target conformation can be either the pre-fusion or post-fusion conformation). In another modification scenario, the target conformation needs to be stabilized while the competitive conformation remains unstable; the competitive conformation can be any other conformation different from the target conformation.

[0101] In the first embodiment above, when only the target conformation is of interest and a stable target conformation is required, step 120 may include the following steps S11-S13 (not shown):

[0102] S11. The predicted target structure of the protein to be modified is determined as the target conformation.

[0103] The target predicted structure of the protein to be modified is the conformation to be stabilized. Therefore, the target predicted structure of the protein to be modified can be determined as the target conformation.

[0104] S12. Based on the sequence distance between each pair of amino acid residues in the target conformation, filter residue pairs with a sequence distance greater than a specified threshold as candidate residue pairs.

[0105] S13. Determine candidate residue pairs based on alternative residue pairs.

[0106] In the second embodiment above, when both the target conformation and the competing conformation are of interest and the target conformation needs to be stable while the competing conformation needs to be unstable, step 120 may include the following steps S21 to S24 (not shown):

[0107] S21. The target predicted structure of the protein to be modified is determined as the target conformation, and another specific conformation of the protein to be modified is determined as the competing conformation.

[0108] The "other specific conformation of the protein to be modified" is determined based on research and development needs and can be any conformation different from the target conformation. Generally speaking, competing conformations exist naturally.

[0109] S22. Calculate the CA atom distance matrix of amino acid residues in the target conformation and the competing conformation respectively. Subtract the CA atom distance matrix of the target conformation from the CA atom distance matrix of the competing conformation to obtain the CA atom distance difference matrix.

[0110] S23. From the CA inter-atomic distance difference matrix, select residue pairs whose CA inter-atomic distance difference is greater than a specified threshold as candidate residue pairs.

[0111] S24. Determine candidate residue pairs based on alternative residue pairs.

[0112] In steps S13 and S24, all candidate residue pairs can be identified as candidate residue pairs, or the candidate residue pairs can be further screened and filtered. Optionally, candidate residue pairs with CA interatomic distances within a specified range can be selected as candidate residue pairs based on the CA interatomic distance matrix of amino acid residues in the target conformation.

[0113] 130. Extract the structural feature information of each candidate residue pair; the structural feature information includes the residue distance matrix between the two residues in each candidate residue pair, the sequence distance, and the relative solvent-accessible surface area, depth index, and predicted local distance difference test value of each residue in each candidate residue pair.

[0114] Specifically, the method for extracting the structural feature information of each candidate residue pair in step 130 is as follows: calculate the residue distance matrix and sequence distance of the two residues in each candidate residue pair, and calculate the relative solvent accessible surface area, depth index and predicted local distance difference test value of each residue in each candidate residue pair; and determine the residue distance matrix, sequence distance, and the relative solvent accessible surface area, depth index and predicted local distance difference test value of each residue as the structural feature information of each candidate residue pair.

[0115] In practical applications, the method for extracting the structural feature information of each candidate residue pair is the same as the method for extracting the structural feature sample information described above, and will not be elaborated upon here. It should be noted that the structural feature information of candidate residue pairs can be extracted from the target conformation. That is, if the protein to be modified is a viral fusion protein, it is necessary to selectively stabilize the pre-fusion or post-fusion conformation of the protein to be modified. When extracting the structural feature information of each candidate residue pair, it is also extracted from the predicted target conformation rather than the competing conformation.

[0116] 140. Input the structural feature information of each candidate residue pair into the pre-trained prediction model to obtain the probability value of each candidate residue pair forming a disulfide bond.

[0117] 150. Identify candidate residue pairs with a probability value greater than a specified probability as disulfide bond mutation sites in the protein to be modified.

[0118] The probability value refers to the probability of a candidate residue pair on a single-chain protein structure, or the average inter-chain probability of a candidate residue pair on a multi-chain protein structure. That is, if the protein to be modified is not a homopolymer protein, the probability value is specifically the probability of a candidate residue pair on a single-chain protein structure; if the protein to be modified is a homopolymer protein, the probability value is specifically the average inter-chain probability of a candidate residue pair on a multi-chain protein structure. The average inter-chain probability is obtained by averaging the probabilities of the candidate residue pair on multiple chains of the homopolymer protein. The probability of a candidate residue pair on a single-chain protein structure ranges from 0 to 1.

[0119] After obtaining the probability value of each candidate residue pair forming a disulfide bond, the candidate residue pairs can be sorted from highest to lowest probability. Then, candidate residue pairs with a probability value greater than a specified probability are identified as disulfide bond mutation sites in the protein to be modified. For the predicted disulfide bond mutation sites, Cys substitution can be performed subsequently, and the mutation effect can be verified experimentally.

[0120] In this embodiment of the invention, the RSV F protein is used as an example for illustration. Since the RSV F protein is a trimer, it has two conformations: pre-fusion and post-fusion, as shown below... Figure 5 and Figure 6 As shown, the colored structures represent ColabFold predicted structures, and the gray structures represent experimental structures. The predicted structures are gradient-colored based on pLDDT, with red indicating regions with high pLDDT and blue indicating regions with low pLDDT. Studies have found that the main target sites of RSV neutralizing antibodies induced by natural infection are located in the pre-fusion conformation of RSV F. To enhance the activation of potent RSV antibodies, it is necessary to stabilize the F protein in its pre-fusion conformation. Research has shown that introducing disulfide bonds (such as S155C-S290C) at specific positions can stabilize the pre-fusion conformation of RSV F. In this invention, the goal is to stabilize the pre-fusion conformation of the RSV F protein. The entire process of predicting disulfide bond mutations and conducting experimental verification based on two predicted structures of the RSV F protein is described in [reference needed]. Figure 7 As shown, but not limited to this protein, the specific steps include:

[0121] The first step is the preparation of the input structure: First, download the pre-fusion experimental structure of RSV F protein (PDB: 4JHW) from the protein database. Remove irrelevant components such as antibodies, water, and ions from the structure. Then, predict the structure using ColabFold. To ensure a high similarity between the predicted and experimental structures, the template used for prediction is the experimental structure itself. Additionally, perform the same ColabFold prediction process on the post-fusion conformation of RSV F protein (PDB: 3RRR). The pre-fusion and post-fusion conformations obtained by ColabFold prediction are shown below. Figure 5 and Figure 6As shown, the predicted structure is highly similar to the experimental structure used as a template, and it fills in the missing parts of the experimental structure.

[0122] The second step is to obtain candidate residue pairs: First, to design disulfide bond mutations that selectively stabilize the pre-fusion conformation, the CA atomic distance matrices of amino acid residues in the pre-fusion and post-fusion conformations are calculated. The CA atomic distance matrix of the pre-fusion conformation is subtracted from the CA atomic distance matrix of the post-fusion conformation to obtain the CA atomic distance difference matrix. Candidate residue pairs with CA atomic distance differences greater than a specified threshold are selected. This step filters out residue pairs that might simultaneously stabilize both the pre-fusion and post-fusion conformations. In this embodiment, the specified threshold for the distance difference is 20 angstroms (but not limited to this threshold; it can be determined based on the specific protein and experimental throughput). The number of candidate residue pairs obtained after filtering is 55775 pairs. Figure 8 As shown, Figure 8 This is a schematic diagram of the A chain structure of the RSV F protein before fusion, where red indicates the candidate residue pairs obtained through filtering.

[0123] The third step is the preliminary screening of candidate residue pairs: Preliminary filtering is performed based on the CA-CA distance between residues in the pre-fusion conformation, retaining only candidate residue pairs with CA-CA distances between 3 and 8 angstroms. This yields a final list of 204 candidate residue pairs, primarily involving multiple domains at the N-terminus and C-terminus of the F1 subunit of the RSV F protein, as well as the β1 and β2 domains of the F2 subunit. Specifically, the conformation of residues 137-216 and 460 to the C-terminus on the F1 subunit shows significant changes from pre-fusion to post-fusion. For example... Figure 9 As shown, Figure 9 This is a schematic diagram of the A-chain structure of the RSV F protein after fusion. The red areas represent candidate disulfide bond residue sites obtained by screening based on a specified threshold of distance difference and a specified range of CA-CA in the pre-fusion conformation. These are mainly residue pairs between residue numbers 27-62, 137-200, 286-302, 386-399, and 464-503.

[0124] The third step is feature extraction: for each pair of candidate residues, various structural feature information is calculated from the predicted pre-fusion conformation.

[0125] The fourth step is scoring and ranking: The trained prediction model is used to predict each candidate residue pair to assess the probability of forming a disulfide bond after mutating the residue pair to Cys. Since the RSV F protein is a trimeric structure, the scores of the candidate residue pairs are averaged across chains. This inter-chain averaging process considers the structural differences between different chains when predicting disulfide bonds in homologous multimeric proteins, and averages the prediction results of different chains as the final prediction result. Specifically, assuming the homotrimeric protein RSV F has three chains, A, B, and C, and RSV F has a candidate residue pair 27-62, the predicted probability of this residue pair forming a disulfide bond in chain A of RSV F is p1, in chain B it is p2, and in chain C it is p3. Then the average interchain probability of residue pair 27-62 is (p1+p2+p3) / 3. Then, the residues are sorted from high to low according to the average probability. The results of the top 20 are shown in Table 1. Some disulfide bond mutation sites have been reported in the literature (see the reference column in Table 1). Among them, 4 pairs of mutation sites have been experimentally verified to effectively stabilize the pre-fusion conformation of RSV F (see the effect column in Table 1), 1 pair has been verified to be ineffective, and 7 other pairs have been reported in the literature, but due to insufficient data, their individual stabilizing effects cannot be determined.

[0126] Step 5: Experimental Validation: Novel mutation sites that were ranked high and not previously reported in the literature were selected for experimental testing. Four novel mutation sites were chosen (as shown in Table 2), and the combined effects of these disulfide bond mutation sites with cavity-filling mutations (S190F, V207L) and proline mutations (I214P) were tested. The main experimental results showed that among the four pairs of disulfide bond mutations tested, the RSV-KW4 (D392C-S493C) mutant was the most effective, stabilizing the pre-fusion conformation and promoting trimer formation. Its pre-fusion trimer conformation content was superior to Moderna mRNA-1345, Ds-Cav1, and RSV-4 (S466C-S443C). Further addition of cavity-filling mutations (S190F, V207L) slightly increased trimer expression levels (RSV-KW8), while adding I214P had little effect (RSV-KW12). Secondly, the stabilizing effect of introducing the D392C-S491C disulfide bond alone was not significant. However, adding cavity-filling mutations (S190F, V207L) to this disulfide bond could maintain the pre-fusion conformation and stabilize trimerization (RSV-KW6). The introduction of I214P further improved the expression level of trimerization (RSV-KW10). Structurally, the introduction of the disulfide bond effectively prevented the transition from the pre-fusion conformation to the post-fusion conformation, and may also have stabilized the conformation of the α-helix involved in trimerization.

[0127] Table 1. Disulfide bond mutation sites predicted by this invention that may have a stabilizing effect on the pre-fusion conformation of RSV F protein.

[0128]

[0129]

[0130] Four previously unreported novel disulfide bond mutations (i.e., residue pairs 321-475, 392-491, 392-493, and 334-475 in Table 1 with empty references) were selected for testing. These four mutations constitute the first four mutation sites RSV-KW1 to RSV-KW4 in Table 2. The other mutation sites in Table 2 are combinations of these four novel disulfide bond mutations with other types of mutations (non-disulfide bond mutations).

[0131] Table 2. List of some novel disulfide bond mutation sites predicted in the embodiments of the present invention and their combinations with other mutations.

[0132]

[0133] In summary, by implementing the embodiments of the present invention, starting from the three-dimensional structure of proteins, candidate residue pairs that may introduce disulfide bonds are extracted. A machine learning model is used to predict the probability of disulfide bond formation after the candidate residue pairs mutate into two Cys residues. These probabilities are then sorted from high to low, providing disulfide bond design schemes to improve protein stability. Specifically, the structural databases of the training and testing sets for the machine learning model are obtained by reverse-mutating disulfide bonds in the PDB structure to alanine and then predicting them using the ColabFold method. Figure 1 Compared to the crystal structure, this reverse-mutated structure simulates the wild-type protein in actual protein engineering scenarios. Then, multi-faceted features of positive and negative samples are extracted from the predicted structure to train a fully connected neural network that uses residue pair features as input to predict the probability of disulfide bond formation. It is important to emphasize that the positive samples used in this invention are residue pairs that have already formed disulfide bonds in the original PDB structure, but in the predicted structure, they have all been reverse-mutated to alanine pairs (…). Figure 2 The negative samples are residue pairs that have not formed disulfide bonds in the original PDB structure, and the types of residues have not changed; all features of the positive and negative samples are extracted from the predicted structure.

[0134] The training and test sets used in this invention are not directly derived from the PDB structure. Instead, they are obtained by performing a reverse virtual mutation on all disulfide bonds (Cys-Cys) in the PDB structure to Ala-Ala, followed by ColabFold prediction. This approach offers several advantages:

[0135] (a) The structures predicted after these mutations no longer contain disulfide bonds, simulating the wild-type protein to be modified, and are more in line with actual application scenarios as training and testing sets;

[0136] (b) The ColabFold structure contains pLDDT (value range 0-100). pLDDT reflects the flexibility of residues to a certain extent, similar to the temperature factor, but is not affected by factors such as solvent environment and crystal quality, making different structures comparable. pLDDT can be used as an additional kinetic feature to guide prediction.

[0137] (c) ColabFold itself has a certain optimization effect on the experimental structure, reducing the dependence of the prediction method on the quality of the experimental structure;

[0138] (d) ColabFold can provide multiple prediction structures, which can play a role in data augmentation.

[0139] like Figure 10 As shown, this embodiment of the invention discloses a protein disulfide bond mutation site prediction device, including a first acquisition unit 1001, a second acquisition unit 1002, a feature extraction unit 1003, a prediction unit 1004, and a determination unit 1005, wherein,

[0140] The first acquisition unit 1001 is used to acquire the target predicted structure of the protein to be modified.

[0141] The second acquisition unit 1002 is used to extract multiple candidate residue pairs based on the target predicted structure of the protein to be modified, and to determine candidate residue pairs based on the candidate residue pairs.

[0142] The feature extraction unit 1003 is used to extract the structural feature information of each candidate residue pair; the structural feature information includes the residue distance matrix between the two residues of each candidate residue pair, the sequence distance, and the relative solvent accessible surface area, depth index, and predicted local distance difference test value of each residue in each candidate residue pair;

[0143] The prediction unit 1004 is used to input the structural feature information of each candidate residue pair into the pre-trained prediction model to obtain the probability value of each candidate residue pair forming a disulfide bond.

[0144] The determination unit 1005 is used to identify candidate residue pairs with a probability value greater than a specified probability as disulfide bond mutation sites in the protein to be modified.

[0145] As an optional implementation, the protein disulfide bond mutation site prediction device further includes the following units (not shown):

[0146] The first acquisition unit is used to reverse the virtual mutation of cysteine ​​residues that have formed disulfide bonds in the three-dimensional structure of each protein into alanine pairs, obtain the template structure and corresponding mutation sequence of each protein, perform structural prediction on the corresponding mutation sequence based on the template structure of each protein, and obtain multiple predicted structures for each protein; extract the alanine pairs after mutation of the original disulfide bond residue pairs from each predicted structure of each protein as positive samples.

[0147] The second acquisition unit is used to extract four adjacent amino acids from each positive sample, combine them in any pair to obtain four pairs of cross amino acids, calculate the CA-CA distance between the CA atoms of the four pairs of cross amino acids, and select the pair of amino acids with the shortest CA-CA distance as the negative sample.

[0148] The training unit is used to extract structural feature information from positive and negative samples respectively, and to train the fully connected neural network to obtain a prediction model.

[0149] As an optional implementation, the first acquisition unit 1001 includes the following sub-units (not shown):

[0150] The judgment subunit is used to determine whether the protein to be modified has a corresponding experimental structure.

[0151] The first prediction subunit is used to use the experimental structure of the protein to be modified as a template when the judgment subunit determines that there is an experimental structure, and to determine the predicted structure obtained by using ColabFold to optimize the structure of the protein to be modified as the target predicted structure.

[0152] The second prediction subunit is used to determine the predicted structure obtained by de novo structure prediction of the protein to be modified using ColabFold as the target predicted structure when the judgment subunit determines that there is no experimental structure.

[0153] As an optional implementation, the second acquisition unit 1002 includes the following sub-units (not shown):

[0154] The first determining subunit is used to determine the target predicted structure of the protein to be modified as the target conformation when only the target conformation is of concern and the target conformation needs to be stabilized.

[0155] The screening subunit is used to select residue pairs whose sequence distance is greater than a specified threshold as candidate residue pairs based on the sequence distance between each pair of amino acid residues in the target conformation.

[0156] The second determining subunit is used to determine candidate residue pairs based on alternative residue pairs.

[0157] Further optionally, the second determining subunit is specifically used to select candidate residue pairs whose CA interatomic distances are within a specified range, based on the CA interatomic distance matrix of amino acid residues in the target conformation.

[0158] Alternatively, as another optional implementation, the second acquisition unit 1002 includes the following sub-units (not shown):

[0159] The third determining subunit is used to determine the target predicted structure of the protein to be modified as the target conformation and another specific conformation of the protein to be modified as the competing conformation when simultaneously focusing on the target conformation and competing conformations and needing to stabilize the target conformation while stabilizing the competing conformation; wherein the target conformation and the competing conformation are different conformations.

[0160] The computational subunit is used to calculate the CA interatomic distance matrix of amino acid residues in the target conformation and the competing conformation respectively. The CA interatomic distance matrix of the competing conformation is subtracted from the CA interatomic distance matrix of the target conformation to obtain the CA interatomic distance difference matrix.

[0161] The selection subunit is used to select residue pairs from the CA inter-atomic distance difference matrix where the CA inter-atomic distance difference value is greater than a specified threshold as candidate residues;

[0162] The fourth determining subunit is used to determine candidate residue pairs based on alternative residue pairs.

[0163] Further optionally, the fourth determining subunit is specifically used to select candidate residue pairs whose CA interatomic distances are within a specified range based on the CA interatomic distance matrix of amino acid residues in the target conformation.

[0164] As an optional implementation, the feature extraction unit 1003 is specifically used to calculate the inter-residue distance matrix and sequence distance of the two residues in each candidate residue pair, and to calculate the relative solvent-accessible surface area, depth index and predicted local distance difference test value of each residue in each candidate residue pair; the inter-residue distance matrix, sequence distance, and the relative solvent-accessible surface area, depth index and predicted local distance difference test value of each residue are used to determine the structural feature information of each candidate residue pair.

[0165] like Figure 11 As shown, an embodiment of the present invention discloses an electronic device, including a memory 1101 storing executable program code and a processor 1102 coupled to the memory 1101;

[0166] The processor 1102 calls the executable program code stored in the memory 1101 to execute the protein disulfide bond mutation site prediction method described in the above embodiments.

[0167] This invention also discloses a computer-readable storage medium storing a computer program that causes a computer to execute the protein disulfide bond mutation site prediction method described in the above embodiments.

[0168] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

[0169] The above embodiments are not an exhaustive list based on the present invention, and there may be many other embodiments not listed. Any substitutions and improvements made without departing from the concept of the present invention are within the protection scope of the present invention.

Claims

1. A method for predicting protein disulfide bond mutation sites, characterized in that, include: Obtain the target predicted structure of the protein to be modified; Multiple candidate residue pairs are extracted based on the target predicted structure of the protein to be modified, and candidate residue pairs are determined based on the candidate residue pairs. Structural feature information is extracted for each candidate residue pair; the structural feature information includes the residue distance matrix between the two residues in each candidate residue pair, the sequence distance, and the relative solvent-accessible surface area, depth index, and predicted local distance difference test value of each residue in each candidate residue pair; The structural feature information of each candidate residue pair is input into a pre-trained prediction model to obtain the probability value of each candidate residue pair forming a disulfide bond; Candidate residue pairs with a probability value greater than a specified probability are identified as disulfide bond mutation sites in the protein to be modified. The training process of the prediction model includes: In each protein, the cysteine ​​residues that have formed disulfide bonds in the three-dimensional structure are reverse-mutated into alanine pairs to obtain the template structure and corresponding mutation sequence of each protein. Based on the template structure of each protein, the corresponding mutation sequence is used to predict the structure and obtain multiple predicted structures of each protein. The alanine pairs resulting from the mutation of the original disulfide bond residue pairs are extracted from each predicted structure of each protein as positive samples. Extract four adjacent amino acids from each positive sample, combine them in any pair to obtain four pairs of cross amino acids, calculate the CA-CA distance between the CA atoms of the four pairs of cross amino acids, and select the pair of amino acids with the shortest CA-CA distance as the negative sample. Structural feature information of the positive and negative samples is extracted respectively, and a prediction model is obtained by training a fully connected neural network. Based on the target predicted structure of the protein to be modified, multiple candidate residue pairs were extracted, including: When focusing solely on the target conformation and requiring its stability, the predicted target structure of the protein to be modified is determined as the target conformation; based on the sequence distance between each pair of amino acid residues in the target conformation, residue pairs with a sequence distance greater than a specified threshold are selected as candidate residue pairs. When simultaneously focusing on the target conformation and competing conformations, and requiring a stable target conformation while maintaining an unstable competing conformation, the predicted target structure of the protein to be modified is identified as the target conformation, and another specific conformation of the protein to be modified is identified as the competing conformation; wherein, the target conformation and the competing conformation are different conformations; the CA atomic distance matrices of amino acid residues in the target conformation and the competing conformation are calculated respectively, and the CA atomic distance matrix of the target conformation is subtracted from the CA atomic distance matrix of the competing conformation to obtain the CA atomic distance difference matrix; from the CA atomic distance difference matrix, residue pairs with CA atomic distance differences greater than a specified threshold are selected as candidate residue pairs.

2. The method for predicting protein disulfide bond mutation sites as described in claim 1, characterized in that, Obtain the target predicted structure of the protein to be modified, including: Determine if the protein to be modified has a corresponding experimental structure; If so, the experimental structure of the protein to be modified is used as a template, and the predicted structure obtained by optimizing the structure of the protein to be modified using ColabFold is determined as the target predicted structure. If not, the predicted structure obtained by de novo structure prediction of the protein to be modified using ColabFold will be determined as the target predicted structure.

3. The method for predicting protein disulfide bond mutation sites as described in claim 1, characterized in that, Candidate residue pairs are determined based on alternative residue pairs, including: Based on the CA interatomic distance matrix of amino acid residues in the target conformation, candidate residue pairs with CA interatomic distances within a specified range are selected as candidate residue pairs.

4. The method for predicting protein disulfide bond mutation sites as described in claim 1 or 2, characterized in that, Extract the structural feature information of each candidate residue pair, including: Calculate the residue distance matrix and sequence distance between the two residues in each candidate residue pair, as well as the relative solvent-accessible surface area, depth index, and predicted local distance difference test value for each residue in each candidate residue pair; The inter-residue distance matrix, the sequence distance, the relative solvent-accessible surface area of ​​each residue, the depth index, and the predicted local distance difference test value are used to determine the structural feature information of each candidate residue pair.

5. A device for predicting protein disulfide bond mutation sites, characterized in that, include: The first acquisition unit is used to acquire the target predicted structure of the protein to be modified. The second acquisition unit is used to extract multiple candidate residue pairs based on the target predicted structure of the protein to be modified, and to determine candidate residue pairs based on the candidate residue pairs. The feature extraction unit is used to extract the structural feature information of each candidate residue pair; the structural feature information includes the residue distance matrix between the two residues of each candidate residue pair, the sequence distance, and the relative solvent-accessible surface area, depth index, and predicted local distance difference test value of each residue in each candidate residue pair; The prediction unit is used to input the structural feature information of each candidate residue pair into the pre-trained prediction model to obtain the probability value of each candidate residue pair forming a disulfide bond. The determining unit is used to determine the candidate residue pairs with a probability value greater than a specified probability as disulfide bond mutation sites in the protein to be modified. Also includes: The first acquisition unit is used to reverse the virtual mutation of cysteine ​​residues that have formed disulfide bonds in the three-dimensional structure of each protein into alanine pairs, obtain the template structure and corresponding mutation sequence of each protein, perform structural prediction on the corresponding mutation sequence based on the template structure of each protein, and obtain multiple predicted structures for each protein; extract the alanine pairs after mutation of the original disulfide bond residue pairs from each predicted structure of each protein as positive samples. The second acquisition unit is used to extract four adjacent amino acids from each positive sample, combine them in any pair to obtain four pairs of cross amino acids, calculate the CA-CA distance between the CA atoms of the four pairs of cross amino acids, and select the pair of amino acids with the shortest CA-CA distance as the negative sample. The training unit is used to extract the structural feature sample information of the positive sample and the negative sample respectively, and to train the fully connected neural network to obtain a prediction model. Based on the target predicted structure of the protein to be modified, multiple candidate residue pairs were extracted, including: When focusing solely on the target conformation and requiring its stability, the predicted target structure of the protein to be modified is determined as the target conformation; based on the sequence distance between each pair of amino acid residues in the target conformation, residue pairs with a sequence distance greater than a specified threshold are selected as candidate residue pairs. When simultaneously focusing on the target conformation and competing conformations, and requiring a stable target conformation while maintaining an unstable competing conformation, the predicted target structure of the protein to be modified is identified as the target conformation, and another specific conformation of the protein to be modified is identified as the competing conformation; wherein, the target conformation and the competing conformation are different conformations; the CA atomic distance matrices of amino acid residues in the target conformation and the competing conformation are calculated respectively, and the CA atomic distance matrix of the target conformation is subtracted from the CA atomic distance matrix of the competing conformation to obtain the CA atomic distance difference matrix; from the CA atomic distance difference matrix, residue pairs with CA atomic distance differences greater than a specified threshold are selected as candidate residue pairs.

6. An electronic device, characterized in that, It includes a memory storing executable program code and a processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the protein disulfide bond mutation site prediction method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program causes a computer to perform the protein disulfide bond mutation site prediction method according to any one of claims 1 to 4.

Images