Training methods, design methods, devices and equipment for protein reverse folding models

Abstract

This disclosure provides a training method, design method, apparatus, and device for a protein reverse folding model, relating to the field of computer science, and particularly to artificial intelligence and biological computing technologies. The training method includes: acquiring a training sample set, which includes structural information of the target protein and the binding protein; inputting the training samples into the protein reverse folding model to obtain multiple candidate amino acid sequences of the binding protein; calculating the binding performance indices of the target protein and each candidate amino acid sequence of the binding protein, including complex conformation quantification indices and / or binding affinity quantification indices; constructing positive and negative sample pairs based on the binding performance indices; and training the protein reverse folding model based on a reinforcement learning algorithm and the positive and negative sample pairs to obtain the trained protein reverse folding model. This disclosure can improve the ability of protein reverse folding models to generate binding protein sequences with superior binding performance.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to artificial intelligence and biological computing technology, specifically to a training method for a protein reverse folding model, a design method for binding protein, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product. Background Technology

[0002] Protein reverse folding is one of the core technologies in the fields of artificial intelligence and biological computing. Its core objective is to generate a binding protein amino acid sequence that can specifically bind to the target protein based on its three-dimensional structure, providing support for drug development, protein function regulation and other scenarios. Summary of the Invention

[0003] This disclosure provides a method for training a protein reverse folding model, a method for designing binding proteins, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product, to address the problem of low success rate of binding of binding protein sequences to target proteins generated by existing methods.

[0004] According to one aspect of this disclosure, a training method for a protein reverse folding model is provided, comprising: acquiring a training sample set, wherein the training samples in the training sample set include structural information of a target protein and a binding agent protein; inputting the training samples into the protein reverse folding model to obtain multiple candidate amino acid sequences of the binding agent protein; calculating the binding performance index between the target protein and the binding agent protein of each candidate amino acid sequence, wherein the binding performance index includes a complex conformation quantification index and / or a binding affinity quantification index, and the complex is a complex formed by the target protein and a binding agent protein of any candidate amino acid sequence; constructing positive and negative sample pairs based on the binding performance index; and training the protein reverse folding model based on a reinforcement learning algorithm and the positive and negative sample pairs to obtain a trained protein reverse folding model.

[0005] According to another aspect of this disclosure, a method for designing a binding agent protein is provided, comprising: inputting the structural information of a target protein and a target binding agent protein into a protein reverse folding model to obtain multiple candidate amino acid sequences of the target binding agent protein; calculating the binding performance index between the target protein and the target binding agent protein of each candidate amino acid sequence, the binding performance index including a complex conformation quantification index and / or a binding affinity quantification index, the complex being a complex formed by the target protein and a target binding agent protein of any candidate amino acid sequence; counting the number of candidate amino acid sequences whose binding performance index meets preset conditions, and outputting the candidate amino acid sequences that meet the preset conditions when the number is greater than or equal to a preset threshold.

[0006] According to another aspect of this disclosure, a training apparatus for a protein reverse folding model is provided, comprising: an acquisition module for acquiring a training sample set, wherein the training samples in the training sample set include structural information of a target protein and a binding agent protein; an input module for inputting the training samples into the protein reverse folding model to obtain multiple candidate amino acid sequences of the binding agent protein; a calculation module for calculating binding performance indices between the target protein and the binding agent protein of each candidate amino acid sequence, wherein the binding performance indices include a complex conformation quantification index and / or a binding affinity quantification index, and the complex is a complex formed by the target protein and a binding agent protein of any candidate amino acid sequence; a construction module for constructing positive and negative sample pairs based on the binding performance indices; and a training module for training the protein reverse folding model based on a reinforcement learning algorithm and the positive and negative sample pairs to obtain a trained protein reverse folding model.

[0007] According to another aspect of this disclosure, a design apparatus for a binding agent protein is provided, comprising: an input module for inputting structural information of a target protein and a target binding agent protein into a protein reverse folding model to obtain multiple candidate amino acid sequences of the target binding agent protein; a calculation module for calculating binding performance indices between the target protein and the target binding agent protein of each candidate amino acid sequence, the binding performance indices including complex conformation quantification indices and / or binding affinity quantification indices, the complex being a complex formed by the target protein and a target binding agent protein of any candidate amino acid sequence; and an output module for counting the number of candidate amino acid sequences whose binding performance indices meet preset conditions, and outputting the candidate amino acid sequences that meet the preset conditions when the number is greater than or equal to a preset threshold.

[0008] According to another aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method provided in any of the preceding aspects.

[0009] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause a computer to perform the method according to any of the preceding aspects.

[0010] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the method according to any of the foregoing aspects.

[0011] This disclosure uses the binding performance index between the target protein and the binding agent protein as the optimization basis, and uses a reinforcement learning algorithm to construct positive and negative sample pairs to train or fine-tune the protein reverse folding model. This allows the trained or fine-tuned protein reverse folding model to not only consider structural consistency when generating the amino acid sequence of the binding agent protein, but also to tend to generate amino acid sequences that perform better in terms of binding performance index, thereby improving the binding success rate of the generated amino acid sequence of the binding agent protein to the target protein.

[0012] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0013] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 A schematic flowchart illustrating a training method for a protein reverse folding model provided in this embodiment of the disclosure; Figure 2 A schematic flowchart of step S103 in the training method of the protein reverse folding model provided in the embodiments of this disclosure; Figure 3 Another flowchart illustrating step S103 of the training method for the protein reverse folding model provided in this embodiment of the disclosure; Figure 4 This is another flowchart illustrating step S103 of the training method for the protein reverse folding model provided in this embodiment of the present disclosure. Figure 5 This is another flowchart illustrating step S103 of the training method for the protein reverse folding model provided in this embodiment of the present disclosure. Figure 6 This is another flowchart illustrating step S103 of the training method for the protein reverse folding model provided in this embodiment of the present disclosure. Figure 7 This is another flowchart illustrating step S103 of the training method for the protein reverse folding model provided in this embodiment of the present disclosure. Figure 8 A schematic flowchart of step S104 in the training method of the protein reverse folding model provided in the embodiments of this disclosure; Figure 9 Another flowchart illustrating step S104 of the training method for the protein reverse folding model provided in this embodiment of the disclosure; Figure 10 A schematic flowchart of step S105 in the training method of the protein reverse folding model provided in the embodiments of this disclosure; Figure 11A schematic flowchart illustrating a method for designing binding protein according to an embodiment of this disclosure; Figure 12 Another flowchart illustrating the design method of the binding protein provided in this disclosure embodiment; Figure 13 A structural block diagram of a training device for a protein reverse folding model provided in an embodiment of this disclosure; Figure 14 A structural block diagram of a design device for binding protein provided in an embodiment of this disclosure; Figure 15 A schematic block diagram of an example electronic device that can be used to implement embodiments of the present disclosure is shown. Detailed Implementation

[0014] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0015] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0016] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0017] Various modifications and variations can be made to this disclosure without departing from its spirit or scope, as will be apparent to those skilled in the art. Therefore, this disclosure is intended to cover modifications and variations falling within the scope of the corresponding claims (the claimed technical solutions) and their equivalents. It should be noted that the embodiments provided in this disclosure can be combined with each other without contradiction.

[0018] Before describing the technical solutions provided by the embodiments of this disclosure, in order to facilitate understanding of the embodiments of this disclosure, this disclosure first specifically explains the problems existing in the related technologies: Protein reverse folding is one of the core technologies in the fields of artificial intelligence and biological computing. Its core objective is to generate the amino acid sequence of a binding protein (Binder) that can specifically bind to the target protein based on the three-dimensional structure of the target protein, providing support for drug development, protein function regulation and other scenarios.

[0019] However, existing methods for designing binding protein sequences suffer from problems such as insufficient success rate in binding the designed binding protein sequences to the target protein.

[0020] Specifically, on the one hand, current protein reverse folding models focus on sequence recovery rate or structural consistency, resulting in binding protein sequences that, while possessing reasonable structures, have weak binding abilities to target proteins. This not only leads to a low success rate in binding protein design but also increases the time and cost of subsequent sequence screening. On the other hand, in existing technologies, the sequence generation and structure evaluation stages are usually separated, exhibiting decoupling issues and lacking an effective feedback loop mechanism. During the sequence generation stage, protein reverse folding models struggle to identify the core characteristics of high-quality binding sequences through evaluation results, leading to a highly inefficient and blind sequence generation process. Furthermore, the models are easily limited by training data, tending to generate sequences similar to training samples, making it difficult to explore non-natural sequences with superior binding function, further restricting the overall performance of binding protein design.

[0021] In view of the above, embodiments of this disclosure provide a method for training a protein reverse folding model, a method for designing binding proteins, an apparatus, an electronic device, a computer-readable storage medium, and a computer program product.

[0022] The training method for the protein reverse folding model provided in the embodiments of this disclosure will be introduced first below.

[0023] Figure 1 This is a schematic flowchart illustrating a training method for a protein reverse folding model provided in an embodiment of this disclosure. Figure 1 As shown, the training method for the protein reverse folding model provided in this embodiment may include the following steps S101 to S105.

[0024] S101: Obtain the training sample set, which includes structural information of the target protein and the binding protein.

[0025] The training sample set may include multiple training samples, and the number of training samples can be flexibly adjusted according to the actual situation; this disclosure does not limit this. Each training sample may include structural information of the target protein and the binding protein.

[0026] For example, in some cases, the structural information of the target protein and the binding protein may include the skeletal structure information of the complex (combination) formed by the target protein and the binding protein. This skeletal structure information may include the skeletal structure information of both the target protein and the binding protein, which may exist in the form of a complex. Thus, based on the skeletal structure information of the complex, the relative spatial positions of the binding protein and the target protein can be determined, enabling the generated binding protein sequence to precisely bind to the target site of the target protein after folding.

[0027] The target protein's skeletal structure information can include the three-dimensional spatial coordinates of all atoms of the target protein. In some other examples, optionally, the structural information of the target protein and the binding protein can include not only the skeletal structure information of the target protein and the binding protein, but also the amino acid sequence and functional site information of the target protein. That is, the training sample can include the complete structural information of the target protein and the skeletal structure information of the binding protein.

[0028] For example, the backbone structure information of the binding protein may include the three-dimensional spatial coordinates of the main chain atoms (such as nitrogen, α-carbon, carbon, and oxygen) of the binding protein. The three-dimensional spatial coordinates of the main chain atoms of the binding protein can be used to provide basic spatial constraints for protein reverse folding models, so that the generated amino acid sequence can fold into a shape corresponding to the three-dimensional spatial coordinates of the main chain atoms.

[0029] In other examples, the backbone structure information of the binding protein may also include its secondary structure (such as α-helices, β-sheets, and random coils) and interfacial contact sites with the target protein. The secondary structure of the binding protein can be used to guide protein reverse folding models to generate amino acid sequences that meet structural stability requirements. For example, hydrophobic amino acids such as alanine and leucine are preferentially selected for the α-helical region, and suitable amino acid combinations are selected for the β-sheet region, thereby reducing the risk that the generated sequence cannot fold into the predetermined backbone shape.

[0030] S102: Input the training samples into the protein reverse folding model to obtain multiple candidate amino acid sequences of the binding protein.

[0031] Protein refolding models can be used to output candidate amino acid sequences that can fold into the backbone structure of the binding protein based on the structural information of the target protein and the binding protein. In some embodiments, protein refolding models include, but are not limited to, Protein Multiple Probability Neural Network (PMPNN) models or Evolutionary Scale Modeling Refolding (ESM-IF1) models. The core optimization objective of these models focuses on the matching degree between the sequence and the backbone structure, without considering the binding performance indicators between the target protein and the binding protein. Therefore, the generated candidate amino acid sequences may have reasonable structures but fail to bind efficiently to the target protein, making it difficult to meet the needs of practical applications. Therefore, further training of protein refolding models is required.

[0032] In S102, training samples can be input into the protein reverse folding model. The process by which the protein reverse folding model generates multiple candidate amino acid sequences for the binding protein is as follows: First, the structural features of the binding protein's backbone structure are encoded, converting the three-dimensional spatial coordinates of the backbone atoms into spatial feature vectors that the model can recognize, such as inter-atomic distances, angles, and dihedral angles. A structural feature matrix is ​​then constructed based on these spatial feature vectors. In some embodiments, secondary structure annotation and interface site annotation can be combined to add feature labels such as "interface site" and "α-helix region" to the backbone structure of different regions, ultimately outputting a structural feature matrix containing all spatial constraint information of the backbone.

[0033] Next, the protein reverse folding model predicts the probability of amino acid sequences based on the structural feature matrix. Starting from the amino terminus (N-terminus) of the backbone structure of the binding protein, it predicts the probability of the occurrence of 20 natural amino acids site by site, and outputs a 20-dimensional probability vector corresponding to the 20 amino acids for each site.

[0034] Next, based on the probability distribution, multiple candidate amino acid sequences can be generated through maximum probability sampling and / or random sampling. Maximum probability sampling directly selects the amino acid with the highest probability at each site, generating a candidate amino acid sequence with a high degree of matching to the backbone structure. Random sampling, on the other hand, randomly samples from each site based on the probability distribution, generating multiple different candidate amino acid sequences.

[0035] S103: Calculate the binding performance indicators between the target protein and the binding agent protein of each candidate amino acid sequence.

[0036] After obtaining multiple candidate amino acid sequences of the binding agent protein, for any given candidate amino acid sequence, a binding performance index can be calculated between the target protein and the binding agent protein of that candidate amino acid sequence. This binding performance index can be used to characterize the strength of the binding affinity between the binding agent protein and the target protein, and / or the structural stability of the complex formed by the binding agent protein and the target protein.

[0037] The binding performance metrics can include quantification metrics of complex conformation and / or quantification metrics of binding affinity. Quantification metrics of complex conformation can be used to characterize the structural stability of the complex formed between the binding agent protein and the target protein. Quantification metrics of binding affinity can be used to characterize the strength of the binding affinity between the binding agent protein and the target protein.

[0038] S104: Construct positive and negative sample pairs based on the combined performance index.

[0039] Based on the binding performance indicators corresponding to each candidate amino acid sequence, a predetermined number of candidate amino acid sequences with good binding performance can be selected as positive samples, and a predetermined number of candidate amino acid sequences with poor binding performance can be selected as negative samples. Positive and negative samples are then paired to obtain a predetermined number of positive and negative sample pairs. The predetermined number can be flexibly adjusted according to actual circumstances, and this disclosure does not limit it.

[0040] S105: Based on reinforcement learning algorithms and positive and negative sample pairs, the protein reverse folding model is trained to obtain the trained protein reverse folding model.

[0041] The principle of reinforcement learning algorithms is to guide the model to learn the sequence features of positive samples that are superior to those of negative samples, based on the difference in binding performance between positive and negative sample pairs. Through a feedback mechanism, the parameters of the protein refolding model are iteratively optimized, thereby improving the binding performance of the amino acid sequence of the binding agent protein generated by the protein refolding model with the target protein. In some embodiments, for example, a reinforcement learning algorithm can be used to calculate the loss value based on the positive and negative sample pairs, and the parameters of the protein refolding model can be updated in reverse based on the loss value. The training steps are iteratively executed until the convergence condition is met, resulting in the trained protein refolding model.

[0042] The protein reverse folding model training method provided in this disclosure uses the binding performance index between the target protein and the binding agent protein as the optimization basis, and uses a reinforcement learning algorithm to construct positive and negative sample pairs to train or fine-tune the protein reverse folding model. This allows the trained or fine-tuned protein reverse folding model to not only consider structural self-consistency when generating the amino acid sequence of the binding agent protein, but also to tend to generate amino acid sequences that perform better in terms of binding performance index, thereby improving the binding success rate of the generated amino acid sequence of the binding agent protein to the target protein.

[0043] To facilitate understanding, the training method of the protein reverse folding model provided in this disclosure will be illustrated below with some specific embodiments.

[0044] According to some embodiments of this disclosure, optionally, the conformation quantification index of the complex may include at least one of the following: the predicted local distance difference test (pLDDT) score of each residue in the complex, the predicted alignment error (PAE) of any residue pair in the complex, the confidence level of the residue-level interface structure of each residue in the complex, and the overall binding quality score of the complex (interfacepredictedTM-score, ipTM).

[0045] The Predicted Local Distance Difference Test (pLDDT) score characterizes the reliability of the three-dimensional structure and local conformational stability of individual residues in the complex; a higher score indicates greater conformational stability. Predicted Alignment Error (PAE) characterizes the accuracy of predicting the relative positions of any two residue pairs in three-dimensional space; a lower value indicates more accurate spatial arrangement of residue pairs. Residue-level interface structure confidence score characterizes the conformational reliability of individual residues at the complex binding interface and their contact stability with the target protein, directly reflecting the binding effectiveness of interface residues. Overall binding mass score (ipTM) characterizes the overall structural integrity and conformational rationality of the complex formed by the target protein and binding agent protein, comprehensively reflecting the overall stability of the complex.

[0046] Thus, by using residue-level complex conformation quantification indicators, the conformation quality of the complex can be accurately assessed from at least one level, such as local residues, residue pairs, interface sites, and overall structure. This provides a reliable quantitative basis for subsequent candidate sequence screening and positive and negative sample construction, and improves the structural conformation stability of the complex formed by the binding protein candidate sequence generated by the protein reverse folding model and the target protein.

[0047] Figure 2 This is a schematic flowchart of step S103 in the training method of the protein reverse folding model provided in this embodiment of the disclosure. Figure 2 As shown, according to some embodiments of this disclosure, optionally, S103: calculating the conformational quantification index of the complex between the target protein and the binding agent protein of each candidate amino acid sequence may include the following steps S201 to S203.

[0048] S201: Input the amino acid sequence of the target protein and the candidate amino acid sequence of the binding protein into the complex conformation prediction model.

[0049] Complex conformation prediction models can be used to predict the three-dimensional spatial conformation of complexes formed by target proteins and binding proteins, and the accuracy of conformation prediction can be improved through multiple rounds of iterative optimization. For example, complex conformation prediction models include, but are not limited to, the HelixFold3 model. In S201, the skeletal structure information of the complex, the amino acid sequence of the target protein, and the candidate amino acid sequences of the binding protein can be input into the complex conformation prediction model.

[0050] S202: Based on the complex conformation prediction model, calculate the degree of spatial position fluctuation of each residue in the complex during the conformation iteration optimization process.

[0051] In S202, the complex conformation prediction model can initiate multiple rounds of conformation iteration optimization, outputting the three-dimensional spatial coordinates (X, Y, Z) of each residue after each round. By extracting the coordinate information of the same residue in each round, the coordinate deviation of the residue between different iteration rounds is calculated, thereby statistically obtaining the degree of spatial position fluctuation of each residue. The degree of spatial position fluctuation of each residue can be characterized by the mean or maximum value of the coordinate deviation; the smaller the deviation, the more stable the residue conformation.

[0052] S203: Map the degree of spatial location fluctuation to a preset score range to obtain the predicted local distance difference test score of each residue in the complex.

[0053] In S203, the complex conformation prediction model maps the spatial positional variation of each residue to a preset score range. The smaller the spatial positional variation, the higher the mapped score, and the stronger the reliability of the corresponding residue conformation prediction. Finally, it outputs the predicted local distance difference test (pLDDT) score for each residue, forming a full residue score matrix. The preset score range can be flexibly adjusted according to actual conditions, and this disclosure does not limit it. For example, in some examples, the preset score range can be 0-100 points.

[0054] Thus, by using the conformation prediction model of the complex conformation prediction system and iterative fluctuation analysis, the local conformational stability of individual residues can be accurately quantified, providing accurate residue-level quantitative basis for subsequent screening of candidate amino acid sequences with reliable local structures. Furthermore, based on the complex conformation prediction model, both evaluation accuracy and computational efficiency can be balanced, eliminating the need for an additional complex evaluation system and reducing implementation costs.

[0055] Figure 3 This is another schematic diagram of step S103 in the training method of the protein reverse folding model provided in the embodiments of this disclosure. For example... Figure 3As shown, according to some embodiments of this disclosure, optionally, S103: calculating the conformational quantification index of the complex between the target protein and the binding agent protein of each candidate amino acid sequence may include the following steps S301 to S303.

[0056] S301: Based on the complex conformation prediction model, calculate the spatial distance between α carbon atoms of any residue pair in the complex.

[0057] For example, a complex conformation prediction model can extract the three-dimensional spatial coordinates (X, Y, Z) of the α carbon atoms (Cα) of all residues from the three-dimensional conformation of the complex. Then, iterates through all residue combinations in the complex to form residue pairs, and calculates the actual spatial distance between the α carbon atoms of each residue pair to obtain a distance dataset of all residue pairs.

[0058] S302: Calculate the deviation between the spatial distance and the predetermined reference spatial distance.

[0059] In some embodiments, the predetermined reference spatial distance can be determined based on the mean of the α-carbon atom spatial distances of the same residue pair output by the complex conformation prediction model during multiple rounds of iterative optimization. Specifically, for each residue pair, multiple sets of spatial distance data output during the iteration of the complex conformation prediction model are statistically analyzed, and their arithmetic mean is calculated as the reference spatial distance for that residue pair. This reference spatial distance can characterize the theoretical spatial distance of the residue pair in the stable conformation. Then, the actual spatial distance of the α-carbon atoms of the residue pair calculated in S301 is compared with the corresponding reference spatial distance to obtain the deviation value. The magnitude of the deviation value reflects the degree of deviation between the current spatial position of the residue pair and the stable conformation position.

[0060] S303: Map the deviation value to a preset numerical range to obtain the predicted alignment error of any residue pair in the complex.

[0061] The preset numerical range can be flexibly adjusted according to actual needs, and this disclosure does not limit it. For example, the preset numerical range can be 0-30 Å. In S303, the deviation value of each residue pair can be mapped to the preset numerical range. The smaller the deviation value, the lower the predicted alignment error (PAE) value obtained by mapping, indicating that the prediction accuracy of the relative spatial position of the residue pair is higher. Finally, the predicted alignment error of all residue pairs in the complex is output, forming a predicted alignment error matrix at the residue pair level.

[0062] Thus, by determining the reference spatial distance based on the iterative distance mean of the complex conformation prediction model, and combining bias calculation and numerical mapping, a quantitative evaluation index can be provided for the relative spatial position accuracy of each residue pair in the complex, forming a residue pair-level prediction alignment error matrix. This prediction alignment error (PAE) can serve as the basis for candidate amino acid sequence screening and positive / negative sample construction, thereby improving the training effect of the protein reverse folding model.

[0063] Figure 4 This is another schematic flowchart of step S103 in the training method of the protein reverse folding model provided in the embodiments of this disclosure. For example... Figure 4 As shown, according to some embodiments of this disclosure, optionally, S103: calculating the conformational quantification index of the complex between the target protein and the binding agent protein of each candidate amino acid sequence may include the following steps S401 to S403.

[0064] S401: Based on the complex conformation prediction model, obtain the prediction alignment error of any residue pair in the complex and the total number of residues in the complex.

[0065] In S401, the prediction alignment error (PAE) of any residue pair in the complex and the total number of residues in the complex can be obtained based on the complex conformation prediction model. The method for obtaining the PAE has been described in detail above and will not be repeated here. The total number of residues in the complex is the sum of the number of target protein residues and the number of binding protein residues.

[0066] S402: For each residue of the binding protein, calculate the spatial matching degree between the residue and multiple residues of the target protein based on the predicted alignment error between the residue and multiple residues of the target protein and the total number of residues.

[0067] In some embodiments, optionally, for each residue in the binding protein, the spatial matching degree between that residue and multiple residues of the target protein can be calculated according to the following expression: (1) (2) Where i represents the i-th residue in the binding protein, i∈B, and B is the set of all residues in the binding protein; j represents the j-th residue in the target protein, j∈T, and T is the set of all residues in the target protein; p ij d0(N) represents the predicted alignment error value in the i-th row and j-th column of the predicted alignment error matrix, d0(N) represents the normalization factor, N represents the total number of residues in the complex, and Score(i) represents the spatial matching degree between the i-th residue in the binding protein and multiple residues in the target protein.

[0068] S403: Normalize the spatial matching degree to obtain the confidence level of the residue-level interface structure.

[0069] In step S403, the spatial matching degree of each residue of the binding agent protein obtained in step S402 is mapped to a preset score range, such as 0-100, to obtain the confidence degree of the residue-level interface structure of each residue of the binding agent protein. The higher the confidence degree score of the residue-level interface structure corresponding to the i-th residue, the more stable the spatial matching between the i-th residue and the target protein residues, and the stronger the reliability of the interface conformation.

[0070] Thus, by employing the aforementioned method for calculating the confidence level of the interface structure at the residue level, the spatial matching stability of the interface between the binding agent protein and the target protein can be quantified at the individual residue level. This provides an assessment basis for the conformational stability of the interface at the residue level, supporting subsequent candidate sequence screening and the construction of positive and negative samples. The confidence level of the interface structure directly reflects the conformational reliability of the interface residues, helping protein reverse folding models to more accurately learn the interface residue features required for high-affinity binding, thereby improving the conformational stability of the interface between the binding agent protein sequence and the target protein generated by the protein reverse folding model.

[0071] Figure 5 This is another schematic flowchart of step S103 in the training method of the protein reverse folding model provided in the embodiments of this disclosure. For example... Figure 5 As shown, according to some embodiments of this disclosure, optionally, S103: calculating the conformational quantification index of the complex between the target protein and the binding agent protein of each candidate amino acid sequence may include the following steps S501 and S502.

[0072] S501: Based on the complex conformation prediction model, obtain the predicted local distance difference test score of each residue in the complex and the predicted alignment error of any residue pair in the complex.

[0073] In S501, based on the complex conformation prediction model, the predicted local distance difference test score (pLDDT) of each residue in the complex, the predicted alignment error (PAE) of any residue pair in the complex, and the total number of residues in the complex can be obtained. The methods for obtaining pLDDT and PAE have been described in detail above and will not be repeated here.

[0074] S502: The overall binding mass score of the complex is obtained by weighted summation based on the predicted local distance difference test score of each residue, the predicted alignment error of any residue pair, and the preset weight.

[0075] In this context, the weight of residues at the binding interface of the complex is greater than the weight of residues at the non-binding interface of the complex, and the weight of residue pairs at the binding interface is greater than the weight of residue pairs at the non-binding interface.

[0076] For example, different weights can be assigned to different types of pLDDTs and PAEs based on whether the residues and residue pairs are located at the binding interface. For instance, the weight of a pLDDT containing residues at the binding interface is set to w1, and the weight of a pLDDT containing residues not at the binding interface is set to w2, where w2 > w1. Similarly, the weight of a PAE containing residue pairs at the binding interface is set to w3, and the weight of a PAE containing residue pairs not at the binding interface is set to w4, where w4 > w3. The specific values ​​of each weight can be flexibly adjusted according to actual circumstances, and this disclosure does not impose any limitations on this.

[0077] Next, the pLDDT values ​​of all residues in the complex are summed according to their respective weights to obtain the pLDDT-weighted total score. Similarly, the PAE values ​​of all residue pairs in the complex are summed according to their respective weights to obtain the PAE-weighted total score. Since the Predicted Alignment Error (PAE) is a negative indicator that a smaller PAE value represents a more accurate spatial position of the residue pair, while the pLDDT is a positive indicator that a larger pLDDT value represents a more stable residue conformation, the PAE-weighted total score can be reversed, such as by taking the reciprocal, taking the reciprocal of the absolute value, or adding a negative sign.

[0078] Next, the pLDDT-weighted total score and the PAE-weighted total score after reverse processing are summed according to a preset weight ratio to obtain the overall binding mass score (ipTM) of the complex. The preset weight ratio can be flexibly adjusted according to the actual situation, and this disclosure does not limit it. For example, in some examples, the pLDDT-weighted total score accounts for 0.6 and the PAE-weighted total score accounts for 0.4.

[0079] In this embodiment, a corresponding table can be output for each candidate amino acid sequence of the binding protein. The table may include the pLDDT value of each residue, the PAE value of each residue pair, and the overall ipTM score of the complex. In other embodiments, the table may also include the residue-level interface structure confidence of each residue, which is not limited in this disclosure.

[0080] Thus, by assigning differentiated weights to residues and residue pairs at the binding and non-binding interfaces, the contribution of interface conformation can be highlighted in the overall binding quality score, more accurately reflecting the actual binding quality of the complex. Integrating two core types of information—local residue stability and the spatial accuracy of residue pairs—the resulting overall binding quality score can serve as a comprehensive quantitative basis for candidate sequence screening and positive / negative sample construction. This helps protein reverse folding models learn the global conformational features of high-affinity binding more efficiently, thereby improving the binding performance and overall conformational stability of the binding agent protein sequences generated by the protein reverse folding model.

[0081] According to some embodiments of this disclosure, optionally, binding affinity quantification indicators may include at least one of the binding energy between the target protein and the binding agent protein, and affinity parameters.

[0082] The binding energy between the target protein and the binding agent protein can be calculated using molecular mechanics simulation tools (such as FoldX). This energy can characterize the energy changes during the formation of the complex. A lower binding energy value (i.e., a larger absolute negative value) indicates a stronger binding interaction and a more stable complex structure. The affinity parameter between the target protein and the binding agent protein can be detected using wet methods such as surface plasmon resonance (SPR) and isothermal titration calorimetry (ITC). This parameter can characterize the tightness of the binding between the target protein and the binding agent protein.

[0083] Thus, by combining affinity quantification indicators such as energy and / or affinity parameters, the actual binding ability of the target protein and the binding agent protein can be quantified. This can provide a reliable affinity basis for the screening of candidate amino acid sequences and the construction of positive and negative samples, helping protein reverse folding models to learn more accurately the sequence and structural features required for high-affinity binding, thereby improving the model's ability to generate binding agent protein sequences with strong binding ability.

[0084] Figure 6 This is another schematic flowchart of step S103 in the training method of the protein reverse folding model provided in the embodiments of this disclosure. For example... Figure 6 As shown, according to some embodiments of this disclosure, optionally, S103: calculating the binding affinity quantitative index between the target protein and the binding agent protein of each candidate amino acid sequence may include the following steps S601 to S603.

[0085] S601: Input the atomic coordinates of the complex into the molecular mechanics simulation tool.

[0086] For example, atomic coordinate information of the complex, such as atom types, X / Y / Z axis coordinates, bond lengths, and bond angles, can be extracted from the three-dimensional conformation of the target protein and binding protein complex output by the complex conformation prediction model. Then, the atomic coordinate information of the complex is imported into a molecular mechanics simulation tool in a preset format (such as PDB format) to provide structured input data for subsequent energy calculations.

[0087] S602: Based on molecular mechanics simulation tools, calculate the total energy of the complex, the first stable state energy of the target protein, and the second stable state energy of the binding protein.

[0088] In molecular mechanics simulation tools, preset energy field parameters, such as van der Waals forces, electrostatic forces, and hydrogen bonding forces of amino acid residues, can be used to calculate the total energy maintaining the three-dimensional structure of the complex, i.e., the total energy of the complex. Additionally, by extracting the target protein monomer from the complex separately and performing energy minimization calculations, the first stable-state energy of the target protein monomer in a stable folded state can be obtained. Similarly, by extracting the binding protein monomer from the complex separately and performing energy minimization calculations, the second stable-state energy of the binding protein monomer in a stable folded state can be obtained.

[0089] S603: Calculate the binding energy between the target protein and the binding agent protein based on the total energy of the complex, the energy of the first stable state, and the energy of the second stable state.

[0090] The binding energy ΔE between the target protein and the binding agent protein can be obtained by calculating the difference between the complex energy and the first and second stable state energies. If ΔE < 0, the binding process is spontaneous; the more negative the value, the stronger the affinity. If ΔE > 0, the binding process is not spontaneous, and the candidate amino acid sequence of the binding agent protein is difficult to bind stably to the target protein.

[0091] Thus, calculating binding energy using molecular mechanics simulation tools such as FoldX offers significantly improved computational efficiency compared to wet experimental verification methods like SPR and ITC. It can simultaneously handle affinity assessments of large batches of candidate binding agent sequences at a lower cost. Furthermore, the calculated binding energy provides reliable affinity data for screening candidate amino acid sequences and constructing positive and negative samples. This helps protein reverse folding models more accurately learn the sequence and structural features required for high-affinity binding, thereby enhancing their ability to generate binding agent protein sequences with strong binding capacity.

[0092] According to some embodiments of this disclosure, optionally, S603: calculating the binding energy between the target protein and the binding agent protein based on the total energy of the complex, the energy of the first stable state, and the energy of the second stable state may include the following steps: Calculate the sum of the energies of the first and second stable states; The binding energy is obtained by calculating the difference between the total energy of the complex and the sum of the energy values.

[0093] In some embodiments, the binding energy between the target protein and the binding protein can be calculated, for example, according to the following expression: ΔE=E complex −(E target +E binder (3) Where ΔE represents the binding energy between the target protein and the binding agent protein, E complex E represents the total energy of the complex. target E represents the energy of the first stable state of the target protein.binder ΔE represents the energy of the second stable state of the binding protein; when ΔE is negative and the larger the absolute value, the stronger the binding effect between the target protein and the binding protein.

[0094] Thus, based on the above expression or the calculation logic of binding energy, the binding energy can be directly obtained from the energy output results of the molecular mechanics simulation tool, without the need for complex energy decomposition or conversion steps. This improves the calculation efficiency of binding energy and can be quickly adapted to high-throughput evaluation scenarios with large numbers of candidate amino acid sequences. Compared to wet experimental verification, which involves processing fewer sequences and is more costly, this calculation scheme can be considered a preferred option, efficiently providing affinity quantification for protein reverse folding models.

[0095] Figure 7 This is another schematic flowchart of step S103 in the training method of the protein reverse folding model provided in the embodiments of this disclosure. For example... Figure 7 As shown, according to some embodiments of this disclosure, optionally, S103: calculating the binding affinity quantitative index between the target protein and the binding agent protein of each candidate amino acid sequence may include the following steps S701 and S702.

[0096] S701: For any candidate amino acid sequence of binding protein, the binding process between the target protein and the binding protein is detected by wet experimental method, and the binding rate constant and dissociation rate constant of the target protein and the binding protein are obtained.

[0097] In S701, for any candidate amino acid sequence of binding agent protein, the binding process between the target protein and the binding agent protein is detected by wet experimental methods such as SPR or ITC, and the binding rate constant and dissociation rate constant of the target protein and the binding agent protein are obtained.

[0098] Taking SPR technology as an example, for a binding protein with any candidate amino acid sequence, the target protein can first be immobilized on the surface of a sensing chip. Then, binding protein solutions with different concentration gradients (such as 0.1 μM, 0.5 μM, 1 μM, 5 μM, etc.) are sequentially flowed through the chip surface. By detecting changes in the refractive index of the chip surface, the dynamic process of binding and dissociation is recorded in real time. The SPR detector can automatically output the binding rate constant ka and the dissociation rate constant kd based on dynamic curve fitting. The binding rate constant ka characterizes the speed at which the binding protein binds to the target protein; a larger ka value indicates faster binding. The dissociation rate constant kd characterizes the speed at which the binding protein dissociates from the target protein; a smaller kd value indicates slower dissociation.

[0099] S702: Calculate the affinity coefficient between the target protein and the binding agent protein based on the binding rate constant and the dissociation rate constant.

[0100] For example, the ratio of the dissociation rate constant to the binding rate constant can be used as the affinity coefficient, and the formula for calculating the affinity coefficient is KD = kd / ka. A smaller KD value indicates stronger binding specificity and higher affinity between the target protein and the binding agent protein. In some examples, the affinity coefficient can be calculated by averaging multiple experimental repetitions to reduce experimental error and improve the reliability of the calculation results.

[0101] Thus, obtaining affinity coefficients through wet experiments can directly reflect the actual interaction between the target protein and the binding agent protein under physiological conditions. The obtained affinity coefficients have high accuracy and can provide experimental support for the affinity performance of candidate amino acid sequences, thereby improving the sample quality for subsequent protein reverse folding model training.

[0102] It should be noted that either the binding energy obtained through molecular mechanics simulation tools or the affinity coefficient obtained through wet experimental methods can be used, or the two can be combined; this disclosure does not limit this.

[0103] According to some embodiments of this disclosure, optionally, S104: constructing positive and negative sample pairs based on the combined performance metrics may include the following steps: Positive and negative sample pairs are constructed based on binding affinity quantification indices; and / or, positive and negative sample pairs are constructed based on complex conformation quantification indices.

[0104] Specifically, the two sample pair construction methods mentioned above can be implemented individually or in combination to adapt to different model training needs, and this disclosure does not limit them.

[0105] In some embodiments, when constructing positive and negative sample pairs based on binding affinity quantification indicators, for example, binding energy ΔE and / or affinity coefficient KD can be used as the criteria, with a preset quantification threshold. Candidate amino acid sequences with an absolute negative binding energy value greater than the preset binding energy threshold and / or an affinity coefficient less than the preset threshold are designated as positive samples, indicating high affinity with the target protein. Candidate amino acid sequences with an absolute negative binding energy value less than the preset binding energy threshold and / or an affinity coefficient greater than the preset threshold are designated as negative samples, indicating low affinity with the target protein.

[0106] In some embodiments, when constructing positive and negative sample pairs based on complex conformation quantification metrics, indicators such as predicted local distance difference test score (pLDDT), predicted alignment error (PAE), residue-level interface structure confidence score, and / or overall binding quality score (ipTM) can be used as the determination criteria. Based on at least one of the above complex conformation quantification metrics, candidate amino acid sequences with relatively stable complex conformations with the target protein are selected as positive samples, and candidate amino acid sequences with unstable complex conformations with the target protein are selected as negative samples.

[0107] Thus, by constructing positive and negative sample pairs by combining affinity quantification indicators and / or complex conformation quantification indicators, high-quality training data can be provided for protein reverse folding models from the perspectives of binding affinity and / or structural stability. Positive samples can accurately convey the characteristics of candidate amino acid sequences with high affinity and / or structural stability, while negative samples can clearly identify low-performance characteristics that need to be avoided. The two complement each other, helping to improve the ability of protein reverse folding models to generate amino acid sequences of high-performance binding proteins.

[0108] Figure 8 This is a schematic flowchart of step S104 in the training method of the protein reverse folding model provided in this embodiment of the disclosure. Figure 8 As shown, according to some embodiments of this disclosure, optionally, S104: constructing positive and negative sample pairs based on the binding affinity quantification index may include the following steps S801 to S804.

[0109] S801: Multiple candidate amino acid sequences were sorted according to binding affinity quantification index.

[0110] In some embodiments, binding energy and / or affinity coefficient can be used as the ranking criteria for multiple candidate amino acid sequences. If binding energy is used, the candidate amino acid sequences can be ranked in descending order of the absolute value of their negative binding energy. A larger absolute value of the negative binding energy indicates stronger affinity. If affinity coefficient is used, the candidate amino acid sequences can be ranked in ascending order of their affinity coefficient. A smaller affinity coefficient indicates stronger affinity. If both indicators are used simultaneously, a weighted average affinity value can be calculated, and the sequences can then be ranked according to this average affinity value.

[0111] S802: Assign the candidate amino acid sequences that are ranked higher to the first affinity group, and assign the candidate amino acid sequences that are ranked lower to the second affinity group.

[0112] The division ratio intervals can be preset, such as the first 20% and the last 20%, and can be flexibly adjusted according to the actual situation. This disclosure does not limit this. In S802, the candidate amino acid sequences ranked in the first preset division ratio interval of the sorting results can be assigned to the first affinity group, which represents high affinity candidate amino acid sequences. The candidate amino acid sequences ranked in the last preset division ratio interval of the sorting results can be assigned to the second affinity group, which represents low affinity candidate amino acid sequences.

[0113] S803: Select a preset number of candidate amino acid sequences from the first affinity group as positive samples, and select a corresponding number of candidate amino acid sequences from the second affinity group as negative samples.

[0114] A predetermined number of candidate amino acid sequences are selected as positive samples from the first affinity group, and the same number of sequences are selected as negative samples from the second affinity group to ensure a balance between the number of positive and negative samples. The predetermined number can be flexibly adjusted according to actual conditions, and this disclosure does not limit it.

[0115] S804: Pair positive samples with negative samples and label the attributes of the paired positive and negative samples to obtain positive and negative sample pairs.

[0116] A one-to-one pairing method is used to sequentially combine selected positive and negative samples to form sample pairs. For each sample pair, the sequences can be labeled with attribute tags for positive samples (high affinity) and negative samples (low affinity), and the corresponding affinity quantitative index values ​​are associated to form a labeled positive and negative sample pair dataset.

[0117] Thus, by constructing positive and negative sample pairs through sorting, grouping, and pairing processes based on affinity quantification indicators, these positive and negative sample pairs can transmit the sequence feature differences of high affinity and low affinity to the protein reverse folding model. This helps the protein reverse folding model quickly learn the core sequence rules of high affinity binding to the target protein, and helps improve the ability of the protein reverse folding model to generate amino acid sequences of binding agent proteins with high affinity to the target protein.

[0118] As described above, according to some embodiments of this disclosure, optionally, S104: constructing positive and negative sample pairs based on the conformational quantification index of the complex may include the following steps: The determination criteria include the predicted local distance difference test score (pLDDT), the predicted alignment error (PAE), the confidence score of the residue-level interface structure, and / or the overall binding quality score (ipTM). Based on at least one of the above-mentioned complex conformation quantification indicators, candidate amino acid sequences with relatively stable complex conformations with the target protein are selected as positive samples, and candidate amino acid sequences with unstable complex conformations with the target protein are selected as negative samples.

[0119] According to other embodiments of this disclosure, alternatively, in addition to the basic screening strategy described above, a residue site-guided construction method may also be used when constructing positive and negative sample pairs based on complex conformation quantification indicators. Figure 9 This is another flowchart illustrating step S104 of the training method for the protein reverse folding model provided in this embodiment of the disclosure. Figure 9 As shown, S104: Constructing positive and negative sample pairs based on the conformation quantification index of the complex may include the following steps S901 to S904.

[0120] S901: Based on the conformation quantification index of the complex, obtain the contribution of residue-level conformation accuracy and / or residue-level affinity.

[0121] In some embodiments, the confidence score of the residue-level interface structure in the complex conformation quantification index can be used as the contribution score of residue-level conformation accuracy. The contribution score of residue-level conformation accuracy directly reflects the spatial matching stability and conformation reliability of a single residue with the target protein residue. The higher the score, the greater the contribution of that residue to the overall conformation accuracy of the complex. Residue-level conformation contribution data can be obtained quickly without additional complex calculations.

[0122] In some embodiments, energy calculation tools (such as Rosetta tools) can be used to calculate the contribution of each residue to the docking energy (i.e., binding energy) of the complex, thereby obtaining the residue-level affinity contribution of each residue.

[0123] In S901, one type of contribution can be obtained alone, or both types of contribution can be obtained simultaneously, providing a precise basis for subsequent target residue screening.

[0124] S902: Select target residue sites whose contribution to residue-level conformation accuracy and / or residue-level affinity is greater than or equal to a preset threshold.

[0125] For each type of contribution—residue-level conformational accuracy contribution and residue-level affinity contribution—a threshold can be set. If the residue-level conformational accuracy contribution of a residue is greater than or equal to a first preset threshold, or the residue-level affinity contribution is greater than or equal to a second preset threshold, or both types of contributions meet their respective threshold requirements, then that residue site is identified as a target residue site. Target residue sites are core key sites that affect the conformational stability and / or binding affinity of the complex, and can provide target objects for subsequent site-directed mutagenesis.

[0126] S903: Select candidate amino acid sequences that meet the preset conditions for quantification of complex conformation as reference sequences, and perform site-directed mutagenesis on the target residue sites in the reference sequences to obtain positive and negative mutagenesis sequences.

[0127] In S903, candidate amino acid sequences that meet preset conditions for quantification of complex conformation can be selected as reference sequences to ensure that the reference sequences themselves have better conformational stability. The preset conditions can be set according to the predicted local distance difference test score (pLDDT), predicted alignment error (PAE), residue-level interface structure confidence and / or overall binding mass score (ipTM), and this disclosure does not limit them.

[0128] Then, site-directed mutagenesis is performed on the target residue sites in the baseline sequence. Positive mutations can employ conserved mutations or mutation types known to improve conformation and / or affinity, such as replacing the target residue site in the baseline sequence with an amino acid that forms a hydrogen bond with the target protein residue, resulting in better conformational stability and / or affinity of the mutated amino acid sequence. Negative mutations can employ mutation types known to weaken conformation and / or affinity, such as replacing the target residue site in the baseline sequence with a sterically hindered amino acid, resulting in decreased conformational stability and / or affinity of the mutated amino acid sequence, thus obtaining positive and negative mutated sequences, respectively.

[0129] S904: Use the positive mutation sequence as the positive sample and the negative mutation sequence as the negative sample to obtain positive and negative sample pairs.

[0130] Positive mutation sequences are used as positive samples, and negative mutation sequences are used as negative samples. Positive and negative mutation sequences are paired up, and each pair of positive and negative mutation sequences forms a positive-negative sample pair. The sample attributes (such as positive or negative), mutation site information, and corresponding residue-level contribution change data are labeled to form a positive-negative sample pair dataset for use in training protein reverse folding models.

[0131] Thus, by identifying target residue sites through residue-level conformational accuracy contribution and / or residue-level affinity contribution, and constructing positive and negative sample pairs through site-directed mutagenesis at these target residue sites, it is possible to focus on sites influencing conformational stability and / or affinity, providing valuable structure-function correlation data for protein reverse folding model training. The differences between positive and negative mutated sequences are mainly concentrated at the target residue sites, which can guide the protein reverse folding model to learn the influence of key residues on complex stability and / or affinity, reduce interference from irrelevant site differences in model training, and help improve the training efficiency and accuracy of the protein reverse folding model.

[0132] According to some embodiments of this disclosure, optionally, S901: obtaining the residue-level conformational accuracy contribution and / or residue-level affinity contribution based on the complex conformational quantification index may include step one and / or step two.

[0133] Step 1: Use the confidence level of the residue-level interface structure in the complex conformation quantification index as the contribution to the accuracy of residue-level conformation.

[0134] The confidence level of the residue-level interface structure of each residue is directly used as the contribution to the accuracy of the residue-level conformation of each residue.

[0135] Step 2: Based on the three-dimensional structure of the complex verified by the complex conformation quantification index, the residue-level affinity contribution is obtained using energy calculation tools.

[0136] In some embodiments, complex conformation quantification metrics such as pLDDT, PAE, ipTM, and / or residue-level interface structure confidence can be used, such as pLDDT > 80 and PAE < 5 Å, to verify the three-dimensional structure of the complex. Then, the verified three-dimensional structure file of the complex is imported into an energy calculation tool (such as Rosetta), and the contribution of each residue to the complex binding energy is calculated using the energy calculation tool. This contribution is the residue-level affinity contribution.

[0137] Thus, through steps one and two described above, the residue-level conformational accuracy contribution and residue-level affinity contribution can be obtained. These contributions can be used to identify target residue sites, thereby focusing on sites influencing conformational stability and / or affinity. This provides valuable structure-function correlation data for protein reverse folding model training, improving the training efficiency and accuracy of the model. Furthermore, validating the complex's three-dimensional structure using conformational quantification metrics can eliminate 3D structural data with low conformational prediction reliability and poor interface matching, ensuring the 3D structure used to calculate the residue-level affinity contribution has high reliability.

[0138] According to some embodiments of this disclosure, optionally, the protein reverse folding model may include interface site constraint parameters, which can be used to constrain the protein reverse folding model to preferentially select amino acids that can form hydrogen bonds and / or hydrophobic interactions with the target protein when generating amino acids in the binding protein that come into contact with the target protein.

[0139] Specifically, to improve the interface matching between the binding protein and the target protein and reduce the probability of the model generating invalid or low-affinity sequences, the interface sites where the binding protein contacts the target protein can be set to a semi-fixed state. An amino acid selection preference rule can be set for these semi-fixed interface sites, prioritizing amino acids that can form hydrogen bonds and / or hydrophobic interactions with the target protein. For example, for charged residues at the target protein interface, amino acids with opposite charges are preferred to enhance electrostatic interactions and form stable hydrogen bonds. For nonpolar residues at the target protein interface, hydrophobic amino acids are preferred to enhance intermolecular hydrophobic interactions. Furthermore, the protein backfolding model can be configured with interface site constraint parameters. These parameters can be used as one of the terms in the constraint loss function. When the amino acids at the generated interface sites do not conform to the amino acid selection preference rule, the constraint loss increases, pushing the model to adjust its parameters until a constrained amino acid sequence is generated.

[0140] Thus, by setting interface site constraint parameters, protein reverse folding models can be guided to generate sequences with better interface matching, increasing the probability of generating binding protein sequences with high affinity for target proteins and stable complex conformation. Furthermore, setting the interface site between the binding protein and the target protein to a semi-fixed state can preserve appropriate sequence diversity and address the structural rigidity and insufficient adaptability issues that easily occur in fully fixed modes.

[0141] Figure 10 This is a schematic flowchart of step S105 in the training method for the protein reverse folding model provided in this embodiment of the disclosure. Figure 10 As shown, according to some embodiments of this disclosure, optionally, S105: training the protein reverse folding model based on reinforcement learning algorithm and positive and negative sample pairs may include the following steps S1001 to S1005.

[0142] S1001: Converts the structural information of positive and negative sample pairs, target proteins, and binding proteins into input vectors.

[0143] In S1001, the amino acid sequence, target protein and binding agent protein structural information in positive and negative sample pairs can be encoded and converted into input vectors that can be recognized by the protein reverse folding model.

[0144] S1002: Input the input vector into the protein defolding model and calculate the first probability of the protein defolding model generating a positive sample amino acid sequence and the second probability of the protein defolding model generating a negative sample amino acid sequence.

[0145] The input vector is fed into the protein defolding model. The model learns and infers from the input sample features through a sequence generation module, outputting the generation probability for each sample sequence. The probability of the protein defolding model generating a positive sample amino acid sequence is defined as the first probability, and the probability of it generating a negative sample amino acid sequence is defined as the second probability. A higher probability value indicates a stronger tendency for the protein defolding model to generate the corresponding sequence.

[0146] S1003: Based on the reinforcement learning algorithm, the total loss value is calculated according to the first probability, the second probability, and the predetermined reference probability.

[0147] The reference probabilities can include a first reference probability of the original protein defolding model (or reference protein defolding model) generating positive sample amino acid sequences and a second reference probability of generating negative sample amino acid sequences. Specifically, the input vector can be input into the original protein defolding model to obtain the first reference probability of the original protein defolding model generating positive sample amino acid sequences and the second reference probability of generating negative sample amino acid sequences.

[0148] Reinforcement learning algorithms include, but are not limited to, Direct Preference Optimization (DPO), Proximal Policy Optimization (PPO), and Reinforcement algorithms. Taking DPO as an example, its core advantage is that it eliminates the need to train an additional reward model (RM), directly utilizing the preference information of positive and negative sample pairs to optimize the protein backfolding model. The core assumption of DPO is that, given the same structural constraints, the probability of generating positive sample sequences (high affinity) should be much higher than the probability of generating negative sample sequences (low affinity). The optimization objective of DPO is to minimize the probability of the model generating negative samples and maximize the probability of the model generating positive samples, which can be achieved through a loss function.

[0149] In S1003, the reference probability is set as the baseline generation probability of the original protein reverse folding model before training, used to measure the performance change during the training process of the protein reverse folding model. Based on the reinforcement learning algorithm, the difference between the first probability and the first reference probability, and the difference between the second probability and the second reference probability are calculated. At the same time, a reward coefficient is introduced to assign a positive reward to the behavior of increasing the probability of positive samples and a negative penalty to the behavior of increasing the probability of negative samples. Finally, the total loss value of the protein reverse folding model training is obtained through weighted calculation.

[0150] S1004: Update the parameters of the protein reverse folding model based on the total loss value.

[0151] The total loss value is backpropagated to the protein reverse folding model network to adjust the model's parameters. During the adjustment process, parameters related to interface site constraints, structural conformation, and affinity can be prioritized for optimization, causing the protein reverse folding model to iterate towards generating amino acid sequences with stable structural conformation and high affinity.

[0152] S1005: Repeat the steps of input vector calculation, probability calculation, loss value calculation and parameter update until the convergence condition is met to obtain the trained protein reverse folding model.

[0153] Repeat steps S1001-S1004 until the convergence condition is met. Stop iterating and save the final model parameters to obtain the trained protein reverse folding model. The convergence condition can be flexibly set according to the actual situation, and this disclosure does not limit it. For example, in some examples, the convergence condition may include the total loss value remaining stable for multiple consecutive iterations with a fluctuation amplitude less than a preset threshold, or the positive sample generation probability reaching a preset target value, or the number of iterations reaching a preset number, etc.

[0154] Thus, training the model using reinforcement learning algorithms and positive / negative sample pairs can guide the model to learn the relevant features of high-affinity, structurally stable amino acid sequences by rewarding positive samples and penalizing negative samples. This improves the ability of protein refolding models to generate amino acid sequences of high-affinity, structurally stable binding proteins. Furthermore, introducing a reference probability as a benchmark allows for the measurement of the model's iterative optimization progress.

[0155] According to some embodiments of this disclosure, optionally, S1003: Calculating the total loss value based on a reinforcement learning algorithm and according to a first probability, a second probability, and a predetermined reference probability may include steps three and four.

[0156] Step 3: Based on the loss function in the reinforcement learning algorithm, and according to the first probability, the second probability, and the reference probability, calculate the preference loss value and the reference contrast loss value.

[0157] The preference loss value characterizes the degree of preference between the protein refolding model's probability of generating positive sample amino acid sequences and its probability of generating negative sample amino acid sequences. The reference contrast loss value characterizes the degree of prediction deviation between the protein refolding model and the original protein refolding model when generating positive sample amino acid sequences, preventing excessive model updates from deviating from the initial capabilities.

[0158] In some embodiments, optionally, taking the DPO algorithm as an example, the preference loss value can be calculated according to the following expression: (4) in, For the preference loss value, The first probability of generating a positive sample amino acid sequence for the current protein reverse folding model. The second probability for generating negative sample amino acid sequences for the current protein reverse folding model. The first reference probability for generating positive sample amino acid sequences for the original protein reverse folding model. The second reference probability for generating negative sample amino acid sequences for the original protein reverse folding model. For temperature coefficient, This is the Sigmoid function.

[0159] In some embodiments, the reference contrast loss value can be calculated according to the following expression: (5) in, The reference loss value is J; J is the set of key residue sites that satisfy Score(i) > βrcl, where βrcl is a preset threshold, such as 0.2. This is a minimum value used to ensure numerical stability. The smaller the reference contrast loss value, the more consistent the current protein refolding model is with the original protein refolding model in predicting key residues, and the more stable the training process is.

[0160] Step 4: The preference loss value and the reference comparison loss value are weighted and summed using preset weights to obtain the total loss value.

[0161] In some embodiments, the total loss value can be calculated according to the following expression: (6) in, This represents the total loss value. The hyperparameters used to control the loss weights, which balance the contribution ratios of preference optimization and reference constraints, can be dynamically adjusted based on training.

[0162] Thus, by fusing preference loss and reference contrast loss, the dual objectives of preference guidance and stable iterative constraints can be achieved in reinforcement learning training. The preference loss directly utilizes the preference information of positive and negative sample pairs, guiding the protein refolding model towards optimizing for high-affinity positive sample sequences without requiring additional training of the reward model. The reference contrast loss, by constraining the prediction deviation between the current protein refolding model and the original model, can reduce overfitting during training. The weighted summation of the total loss function can be flexibly adjusted using hyperparameters to adjust the weights of preference loss and reference contrast loss, adapting to the training needs at different stages. Ultimately, this allows the protein refolding model to efficiently learn the characteristics of high-affinity amino acid sequences while maintaining the stability of the training process.

[0163] Based on the training method of the protein reverse folding model provided in any of the above embodiments, this disclosure also provides a method for designing binding protein.

[0164] Figure 11 This is a schematic flowchart illustrating a method for designing binding proteins according to embodiments of this disclosure. Figure 11 As shown, the design method for binding proteins provided in this disclosure may include the following steps: S1101: Input the structural information of the target protein and the target binding protein into the protein reverse folding model to obtain multiple candidate amino acid sequences of the target binding protein; S1102: Calculate the binding performance index between the target protein and the target binding agent protein of each candidate amino acid sequence. The binding performance index includes the complex conformation quantification index and / or binding affinity quantification index. The complex is the complex formed by the target protein and the target binding agent protein of any candidate amino acid sequence. S1103: Count the number of candidate amino acid sequences that meet the preset conditions for performance indicators. If the number is greater than or equal to the preset threshold, output the candidate amino acid sequences that meet the preset conditions.

[0165] The target protein can be any target protein, and the target binding protein can be any binding protein used to bind to the target protein. This disclosure does not limit this. The specific processes of S1101 and S1102 are similar to those of S102 and S103 described above, and will not be repeated here. The preset conditions and preset thresholds can be flexibly adjusted according to the actual situation, and this disclosure does not limit this.

[0166] The binding protein design method provided in this disclosure uses the binding performance index between the target protein and the target binding protein as the optimization basis. It screens and evaluates multiple candidate amino acid sequences of the target binding protein generated by the protein reverse folding model. It can screen out candidate amino acid sequences with high binding affinity and stable complex conformation from multiple candidate amino acid sequences, so that the amino acid sequence of the final output target binding protein has a high binding success rate with the target protein.

[0167] According to some embodiments of this disclosure, optionally, the protein reverse folding model used in S1101 can be trained using a training method based on the protein reverse folding model provided in any of the above embodiments. Of course, in other embodiments, the protein reverse folding model used in S1101 can also be an existing protein reverse folding model. When the number of candidate amino acid sequences that meet the preset conditions is less than a preset threshold, the protein reverse folding model can be fine-tuned in the following manner until the number of candidate amino acid sequences that meet the preset conditions is greater than or equal to the preset threshold or the number of iterations reaches a preset number.

[0168] Figure 12 This is another schematic flowchart illustrating the design method of the binding protein provided in this disclosure. Figure 12 As shown, according to some embodiments of the present disclosure, optionally, the design method of the binding protein provided in the embodiments of the present disclosure may further include the following steps S1201 to S1203.

[0169] S1201: When the number is less than the preset threshold, construct positive and negative sample pairs based on the combined performance index.

[0170] When the number of candidate amino acid sequences that meet the preset conditions is less than a preset threshold, positive and negative sample pairs can be constructed based on the binding performance index. The specific process for constructing positive and negative sample pairs is the same as or similar to the specific process of S104 above, and will not be repeated here.

[0171] S1202: Based on reinforcement learning algorithms and positive and negative sample pairs, the protein reverse folding model is fine-tuned to obtain the fine-tuned protein reverse folding model.

[0172] After obtaining the positive and negative sample pairs, the protein reverse model can be fine-tuned based on the reinforcement learning algorithm and the positive and negative sample pairs to obtain the fine-tuned protein reverse model. The process of fine-tuning the protein reverse model is the same as or similar to the process of training the protein reverse model in S105 above, and will not be repeated here.

[0173] S1203: Input the structural information into the fine-tuned protein reverse folding model, regenerate the candidate amino acid sequence of the target binding agent protein, and repeat the steps of calculating the binding performance index and fine-tuning the protein reverse folding model until the number is greater than or equal to the preset threshold or the number of iterations reaches the preset number.

[0174] The structural information is input into the fine-tuned protein reverse folding model to regenerate the candidate amino acid sequence of the target binding protein. S1102, S1201 and S1202 are repeated until the number of candidate amino acid sequences that meet the preset conditions is greater than or equal to the preset threshold or the number of iterations reaches the preset number.

[0175] Thus, when the number of candidate amino acid sequences that meet the preset conditions is insufficient, the fine-tuning process of the protein reverse folding model is automatically triggered. Positive and negative sample pairs are constructed using the binding performance data of the current batch, and the model is optimized through reinforcement learning. This enables the protein reverse folding model to generate amino acid sequences of binding agent proteins with high affinity and strong conformational stability to the target protein.

[0176] According to some embodiments of this disclosure, optionally, the preset conditions may include at least one of the following: the prediction alignment error (PAE) is less than a preset error threshold, the overall binding mass score (ipTM) is greater than a preset score threshold, the confidence score of the residue-level interface structure is greater than a preset confidence threshold, the predicted local distance difference test score (pLDDT) is greater than a preset difference threshold, the absolute value of the binding energy is greater than a preset threshold, and the affinity coefficient is less than a preset affinity threshold. The magnitude of each threshold can be flexibly adjusted according to the actual situation, and this disclosure does not limit it.

[0177] According to other embodiments of this disclosure, optionally, the preset conditions may also include a number of interaction pairs greater than a preset number and / or hotspot residue coverage greater than or equal to a preset coverage threshold.

[0178] The number of interaction pairs can be obtained by combining the three-dimensional structure of the complex output by the complex conformation prediction model with the statistical count of interfacial atomic contact pairs using molecular mechanics simulation tools. The number of interaction pairs can be used to characterize the sufficiency of interfacial contact between the binding agent protein and the target protein; the more interaction pairs, the tighter the interfacial binding between the binding agent protein and the target protein. Hotspot residue coverage can be obtained by decomposing the residue-level affinity contribution obtained by energy calculation tools, identifying the affinity hotspot residues of the target protein, and then comparing the proportion of hotspot residues covered by the binding site of the binding agent protein to the total number of hotspot residues. Hotspot residue coverage is used to characterize the targeting of the binding agent protein to the core binding site of the target protein; the higher the coverage, the stronger the binding specificity and affinity between the binding agent protein and the target protein.

[0179] According to some embodiments of this disclosure, optionally, the conformation quantification index of the complex includes at least one of the following: the predicted local distance difference test score of each residue in the complex, the predicted alignment error of any residue pair in the complex, the confidence of the residue-level interface structure of each residue in the complex, and the overall binding mass score of the complex.

[0180] The methods for obtaining the conformational quantification indicators of the above-mentioned complexes and their technical effects have been described in detail above, and will not be repeated here.

[0181] According to some embodiments of this disclosure, optionally, the binding affinity quantification index includes at least one of the binding energy and affinity parameters between the target protein and the target binding agent protein.

[0182] The methods for obtaining the aforementioned affinity-based indicators and their technical effects have been detailed above and will not be repeated here.

[0183] According to some embodiments of this disclosure, optionally, S1201: Constructing positive and negative sample pairs based on the combined performance metrics may include the following steps: Positive and negative sample pairs are constructed based on binding affinity quantification indices; and / or, positive and negative sample pairs are constructed based on complex conformation quantification indices.

[0184] Optionally, according to some embodiments of this disclosure, constructing positive and negative sample pairs based on a binding affinity quantification index may include the following steps: Multiple candidate binding agent protein sequences were sorted according to binding affinity quantification indicators; The candidate amino acid sequences that are ranked higher are assigned to the first affinity group, and the candidate amino acid sequences that are ranked lower are assigned to the second affinity group. A predetermined number of candidate amino acid sequences are selected from the first affinity group as positive samples, and a corresponding number of candidate amino acid sequences are selected from the second affinity group as negative samples. Positive samples are paired with negative samples, and the attributes of the paired positive and negative samples are labeled to obtain positive and negative sample pairs.

[0185] Optionally, according to some embodiments of this disclosure, constructing positive and negative sample pairs based on a complex conformation quantification index may include the following steps: Based on the conformation quantification index of the complex, the contribution of residue-level conformation accuracy and / or residue-level affinity is obtained; Target residue sites that are selected based on their contribution to residue-level conformation accuracy and / or residue-level affinity being greater than or equal to a preset threshold. Candidate amino acid sequences that meet the preset conditions for complex conformation quantification are selected as reference sequences, and site-directed mutagenesis is performed on target residue sites in the reference sequences to obtain positive and negative mutagenesis sequences. Positive mutation sequences are used as positive samples, and negative mutation sequences are used as negative samples to obtain positive-negative sample pairs.

[0186] According to some embodiments of this disclosure, optionally, the contribution of residue-level conformational accuracy and / or residue-level affinity is obtained based on the conformational quantification index of the complex, including: The confidence level of the residue-level interface structure in the quantification index of complex conformation is used as the contribution of residue-level conformation accuracy. And / or, based on the three-dimensional structure of the complex verified by the complex conformation quantification index, the residue-level affinity contribution is obtained using energy calculation tools.

[0187] The specific process of constructing positive and negative sample pairs described above is the same as or similar to the specific process of S104 mentioned above. Both can achieve the same or similar technical effects, and will not be repeated here.

[0188] According to some embodiments of this disclosure, optionally, S1202: fine-tuning the protein reverse folding model based on reinforcement learning algorithms and positive and negative sample pairs may include the following steps: Transform positive and negative sample pairs and structural information into input vectors; Input the input vector into the protein defolding model and calculate the first probability of the protein defolding model generating a positive sample amino acid sequence and the second probability of the protein defolding model generating a negative sample amino acid sequence; Based on the reinforcement learning algorithm, the total loss value is calculated according to the first probability, the second probability and the predetermined reference probability; wherein, the reference probability includes the first reference probability of the original protein reverse folding model generating positive sample amino acid sequences and the second reference probability of negative sample amino acid sequences. The parameters of the protein reverse folding model are updated based on the total loss value; Repeat the steps of input vector calculation, probability calculation, loss value calculation and parameter update until the convergence condition is met to obtain the fine-tuned protein reverse folding model.

[0189] According to some embodiments of this disclosure, optionally, the total loss value is calculated based on a reinforcement learning algorithm and according to a first probability, a second probability, and a predetermined reference probability, including: Based on the loss function in the reinforcement learning algorithm, and according to the first probability, the second probability and the reference probability, the preference loss value and the reference contrast loss value are calculated; The total loss value is obtained by weighting and summing the preference loss value and the reference comparison loss value using preset weights. Among them, the preference loss value is used to characterize the degree of preference of the protein reverse folding model in generating positive sample amino acid sequences relative to the probability of generating negative sample amino acid sequences, and the reference contrast loss value is used to characterize the degree of prediction deviation between the protein reverse folding model and the original protein reverse folding model when generating positive sample amino acid sequences.

[0190] The process of fine-tuning the protein reverse folding model described above is the same as or similar to the process of training the protein reverse folding model in S105 above. Both can achieve the same or similar technical effects, so they will not be described again here.

[0191] According to some embodiments of this disclosure, optionally, the technical solutions of this disclosure can be widely applied in the fields of biomedicine and bioengineering, including but not limited to affinity maturation and de novo design of protein drugs, catalytic activity and substrate-specific modification of industrial enzymes, development of functional protein elements in synthetic biology, and design and optimization of novel vaccine antigens. The technical solutions of this disclosure are applicable to application scenarios where, given the three-dimensional structure of the target protein-binder complex, targeted optimization design is carried out on the binder protein sequence, specifically including but not limited to the following two core scenarios: Binder protein affinity maturation: This disclosure can target wild-type binder protein sequences and improve the binding affinity of binder proteins to target proteins by optimizing interfacial binding characteristics and strengthening intermolecular forces.

[0192] Synergistic optimization of multiple properties of binding agents: This disclosure can integrate multiple property prediction models and quantitative evaluation methods to simultaneously carry out synergistic optimization of multiple target properties such as binding affinity, conformational stability, and binding specificity for wild-type binding agents, so as to meet the diversified performance requirements of different application scenarios.

[0193] Based on the protein reverse folding model training method provided in any of the above embodiments, this disclosure also provides a protein reverse folding model training device.

[0194] Figure 13 This is a structural block diagram of a training device for a protein reverse folding model provided in an embodiment of this disclosure. Figure 13 As shown, the protein reverse folding model training device 130 provided in this embodiment may include the following modules: The acquisition module 131 is used to acquire the training sample set, which includes the structural information of the target protein and the binding protein. Input module 132 is used to input training samples into the protein reverse folding model to obtain multiple candidate amino acid sequences of the binding protein; The calculation module 133 is used to calculate the binding performance index between the target protein and the binding agent protein of each candidate amino acid sequence. The binding performance index includes the complex conformation quantification index and / or binding affinity quantification index. The complex is a complex formed by the target protein and the binding agent protein of any candidate amino acid sequence. Module 134 is used to construct positive and negative sample pairs based on the combination performance metrics; Training module 135 is used to train the protein reverse folding model based on reinforcement learning algorithm and positive and negative sample pairs to obtain the trained protein reverse folding model.

[0195] The protein reverse folding model training device provided in this embodiment uses the binding performance index between the target protein and the binding agent protein as the optimization basis, and uses a reinforcement learning algorithm to construct positive and negative sample pairs to train or fine-tune the protein reverse folding model. This allows the trained or fine-tuned protein reverse folding model to not only consider structural self-consistency when generating the amino acid sequence of the binding agent protein, but also to tend to generate amino acid sequences that perform better in terms of binding performance index, thereby improving the binding success rate of the generated amino acid sequence of the binding agent protein to the target protein.

[0196] According to some embodiments of this disclosure, optionally, the conformation quantification index of the complex includes at least one of the following: the predicted local distance difference test score of each residue in the complex, the predicted alignment error of any residue pair in the complex, the confidence of the residue-level interface structure of each residue in the complex, and the overall binding mass score of the complex.

[0197] According to some embodiments of this disclosure, optionally, the calculation module 133 can be used to input the amino acid sequence of the target protein and the candidate amino acid sequence of the binding agent protein into the complex conformation prediction model; based on the complex conformation prediction model, calculate the degree of spatial position fluctuation of each residue in the complex during the conformation iteration optimization process; map the degree of spatial position fluctuation to a preset score range to obtain the predicted local distance difference test score of each residue in the complex.

[0198] According to some embodiments of this disclosure, optionally, the calculation module 133 can be used to calculate the spatial distance between α carbon atoms of any residue pair in the complex based on the complex conformation prediction model; calculate the deviation value between the spatial distance and the predetermined reference spatial distance; map the deviation value to a preset numerical range to obtain the prediction alignment error of any residue pair in the complex.

[0199] According to some embodiments of this disclosure, optionally, the calculation module 133 may be used to obtain the prediction alignment error of any residue pair in the complex and the total number of residues in the complex based on the complex conformation prediction model; for each residue of the binding protein, calculate the spatial matching degree between the residue and multiple residues of the target protein based on the prediction alignment error between the residue and multiple residues of the target protein and the total number of residues; and normalize the spatial matching degree to obtain the residue-level interface structure confidence degree of the residue.

[0200] According to some embodiments of this disclosure, optionally, the calculation module 133 can be specifically used to obtain the predicted local distance difference test score of each residue in the complex and the predicted alignment error of any residue pair in the complex based on the complex conformation prediction model; and to perform weighted summation calculation based on the predicted local distance difference test score of each residue, the predicted alignment error of any residue pair and the preset weight to obtain the overall binding quality score of the complex; wherein, the weight of residues on the binding interface of the complex is greater than the weight of residues on the non-binding interface of the complex, and the weight of residue pairs on the binding interface is greater than the weight of residue pairs on the non-binding interface.

[0201] According to some embodiments of this disclosure, optionally, the binding affinity quantification index includes at least one of the binding energy and affinity parameters between the target protein and the binding agent protein.

[0202] According to some embodiments of this disclosure, optionally, the calculation module 133 may be used to input the atomic coordinate information of the complex into a molecular mechanics simulation tool; based on the molecular mechanics simulation tool, calculate the total energy of the complex, the first stable state energy of the target protein, and the second stable state energy of the binding agent protein respectively; and calculate the binding energy between the target protein and the binding agent protein based on the total energy of the complex, the first stable state energy, and the second stable state energy.

[0203] According to some embodiments of this disclosure, optionally, the calculation module 133 may be used to calculate the sum of the first stable state energy and the second stable state energy; calculate the difference between the total energy of the complex and the sum, and obtain the binding energy.

[0204] According to some embodiments of this disclosure, optionally, for a binding protein with any candidate amino acid sequence, the binding process between the target protein and the binding protein is detected by a wet experimental method to obtain the binding rate constant and dissociation rate constant between the target protein and the binding protein; and the affinity coefficient between the target protein and the binding protein is calculated based on the binding rate constant and the dissociation rate constant.

[0205] According to some embodiments of this disclosure, optionally, the construction module 134 may be used to construct positive and negative sample pairs based on binding affinity quantification indices; and / or, to construct positive and negative sample pairs based on complex conformation quantification indices.

[0206] According to some embodiments of this disclosure, optionally, the construction module 134 can be specifically used to sort multiple candidate amino acid sequences according to binding affinity quantitative indicators; classify the candidate amino acid sequences with higher rankings into a first affinity group and classify the candidate amino acid sequences with lower rankings into a second affinity group; select a preset number of candidate amino acid sequences from the first affinity group as positive samples and select a corresponding number of candidate amino acid sequences from the second affinity group as negative samples; pair the positive samples and negative samples and label the attributes of the paired positive and negative samples to obtain positive and negative sample pairs.

[0207] According to some embodiments of this disclosure, optionally, the construction module 134 can be specifically used to obtain the residue-level conformational accuracy contribution and / or residue-level affinity contribution based on the complex conformational quantification index; screen out target residue sites whose residue-level conformational accuracy contribution and / or residue-level affinity contribution is greater than or equal to a preset threshold; select candidate amino acid sequences whose complex conformational quantification index meets preset conditions as reference sequences, and perform site-directed mutagenesis on the target residue sites in the reference sequences to obtain positive mutagenesis sequences and negative mutagenesis sequences; use the positive mutagenesis sequences as positive samples and the negative mutagenesis sequences as negative samples to obtain positive and negative sample pairs.

[0208] According to some embodiments of this disclosure, optionally, the construction module 134 may be used to use the confidence level of the residue-level interface structure in the complex conformation quantification index as the contribution of residue-level conformation accuracy; and / or, based on the three-dimensional structure of the complex verified by the complex conformation quantification index, to obtain the residue-level affinity contribution through an energy calculation tool.

[0209] According to some embodiments of this disclosure, optionally, the protein reverse folding model includes interface site constraint parameters, which are used to constrain the protein reverse folding model to preferentially select amino acids that can form hydrogen bonds and / or hydrophobic interactions with the target protein when generating amino acids in the binding protein that come into contact with the target protein.

[0210] According to some embodiments of this disclosure, optionally, the training module 135 can be used to convert positive and negative sample pairs and structural information into input vectors; input the input vectors into a protein defolding model and calculate a first probability and a second probability of the protein defolding model generating positive sample amino acid sequences and negative sample amino acid sequences; calculate a total loss value based on a reinforcement learning algorithm and according to the first probability, the second probability, and a predetermined reference probability; wherein the reference probability includes a first reference probability and a second reference probability of the original protein defolding model generating positive sample amino acid sequences and negative sample amino acid sequences before training; update the parameters of the protein defolding model based on the total loss value; and repeat the steps of input vector calculation, probability calculation, loss value calculation, and parameter update until the convergence condition is met to obtain the trained protein defolding model.

[0211] According to some embodiments of this disclosure, optionally, the training module 135 can be specifically used to calculate a preference loss value and a reference contrast loss value based on the loss function in the reinforcement learning algorithm, according to the first probability, the second probability, and the reference probability; and to obtain a total loss value by weighting the preference loss value and the reference contrast loss value with preset weights; wherein, the preference loss value is used to characterize the degree of preference of the protein refolding model in generating positive sample amino acid sequences relative to the probability of generating negative sample amino acid sequences, and the reference contrast loss value is used to characterize the degree of prediction deviation between the protein refolding model and the original protein refolding model when generating positive sample amino acid sequences.

[0212] It should be noted that the protein reverse folding model training device 130 has the same or corresponding technical features as the protein reverse folding model training method provided in any of the above embodiments, and achieves the same technical effect. For the sake of brevity, it will not be described in detail here.

[0213] Based on the design method for binding proteins provided in any of the above embodiments, this disclosure also provides a design apparatus for binding proteins.

[0214] Figure 14 A structural block diagram of a design device for binding protein provided in an embodiment of this disclosure. (See diagram below.) Figure 14 As shown, the binding protein design apparatus 140 provided in this embodiment may include the following modules: Input module 141 is used to input the structural information of the target protein and the target binding protein into the protein reverse folding model to obtain multiple candidate amino acid sequences of the target binding protein; Calculation module 142 is used to calculate the binding performance index between the target protein and the target binding agent protein of each candidate amino acid sequence. The binding performance index includes the complex conformation quantification index and / or binding affinity quantification index. The complex is a complex formed by the target protein and the target binding agent protein of any candidate amino acid sequence. Output module 143 is used to count the number of candidate amino acid sequences that meet the preset conditions for binding performance indicators, and output the candidate amino acid sequences that meet the preset conditions when the number is greater than or equal to the preset threshold.

[0215] The binding protein design apparatus provided in this embodiment uses the binding performance index between the target protein and the target binding protein as the optimization basis. It screens and evaluates multiple candidate amino acid sequences of the target binding protein generated by the protein reverse folding model. It can screen out candidate amino acid sequences with high binding affinity and stable complex conformation from multiple candidate amino acid sequences, so that the amino acid sequence of the final output target binding protein has a high binding success rate with the target protein.

[0216] According to some embodiments of this disclosure, optionally, the protein reverse folding model is trained by the training method of the protein reverse folding model as described in any of the above embodiments.

[0217] Optionally, according to some embodiments of this disclosure, the design apparatus 140 for binding protein provided in the embodiments of this disclosure may further include a fine-tuning module, used to construct positive and negative sample pairs according to binding performance indicators when the number is less than a preset threshold; fine-tune the protein reverse folding model based on reinforcement learning algorithm and positive and negative sample pairs to obtain a fine-tuned protein reverse folding model; input structural information into the fine-tuned protein reverse folding model, regenerate the candidate amino acid sequence of the target binding protein, and repeat the steps of calculating binding performance indicators and fine-tuning the protein reverse folding model until the number is greater than or equal to the preset threshold or the number of iterations reaches the preset number.

[0218] According to some embodiments of this disclosure, optionally, the conformation quantification index of the complex includes at least one of the following: the predicted local distance difference test score of each residue in the complex, the predicted alignment error of any residue pair in the complex, the confidence of the residue-level interface structure of each residue in the complex, and the overall binding mass score of the complex.

[0219] According to some embodiments of this disclosure, optionally, the binding affinity quantification index includes at least one of the binding energy and affinity parameters between the target protein and the target binding agent protein.

[0220] According to some embodiments of this disclosure, optionally, the fine-tuning module can be used to convert positive and negative sample pairs and structural information into input vectors; input the input vectors into a protein reverse folding model and calculate a first probability and a second probability of the protein reverse folding model generating a positive sample amino acid sequence and a negative sample amino acid sequence; calculate a total loss value based on a reinforcement learning algorithm and according to the first probability, the second probability, and a predetermined reference probability; wherein, the reference probability includes a first reference probability and a second reference probability of the original protein reverse folding model generating a positive sample amino acid sequence and a negative sample amino acid sequence before fine-tuning; update the parameters of the protein reverse folding model based on the total loss value; and repeat the steps of input vector calculation, probability calculation, loss value calculation, and parameter update until the convergence condition is met to obtain the fine-tuned protein reverse folding model.

[0221] According to some embodiments of this disclosure, optionally, the fine-tuning module can be specifically used to calculate a preference loss value and a reference contrast loss value based on the loss function in the reinforcement learning algorithm, and according to the first probability, the second probability, and the reference probability; the preference loss value and the reference contrast loss value are weighted and summed by preset weights to obtain the total loss value; wherein, the preference loss value is used to characterize the degree of preference of the protein refolding model in generating positive sample amino acid sequences relative to the probability of generating negative sample amino acid sequences, and the reference contrast loss value is used to characterize the degree of prediction deviation between the protein refolding model and the original protein refolding model when generating positive sample amino acid sequences.

[0222] It should be noted that the binding protein design device 140 has the same or corresponding technical features as the binding protein design method provided in any of the above embodiments, and achieves the same technical effect. For the sake of brevity, it will not be described in detail here.

[0223] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0224] Figure 15 A schematic block diagram of an example electronic device 1500 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0225] like Figure 15As shown, device 1500 includes a computing unit 1501, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1502 or a computer program loaded from storage unit 1508 into random access memory (RAM) 1503. The RAM 1503 may also store various programs and data required for the operation of device 1500. The computing unit 1501, ROM 1502, and RAM 1503 are interconnected via bus 1504. Input / output (I / O) interface 1505 is also connected to bus 1504.

[0226] Multiple components in device 1500 are connected to I / O interface 1505, including: input unit 1506, such as keyboard, mouse, etc.; output unit 1507, such as various types of monitors, speakers, etc.; storage unit 1508, such as disk, optical disk, etc.; and communication unit 1509, such as network card, modem, wireless transceiver, etc. Communication unit 1509 allows device 1500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0227] The computing unit 1501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1501 performs the various methods and processes described above, such as methods for training protein reverse-folding models or methods for designing binding proteins. For example, in some embodiments, methods for training protein reverse-folding models or methods for designing binding proteins can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 1508. In some embodiments, part or all of the computer program can be loaded and / or installed on device 1500 via ROM 1502 and / or communication unit 1509. When the computer program is loaded into RAM 1503 and executed by the computing unit 1501, one or more steps of the methods for training protein reverse-folding models or methods for designing binding proteins described above can be performed. Alternatively, in other embodiments, computing unit 1501 may be configured by any other suitable means (e.g., by means of firmware) to perform a method for training a protein reverse folding model or a method for designing a binding protein.

[0228] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0229] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0230] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0231] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0232] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0233] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0234] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0235] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A training method for a protein reverse folding model, comprising: Obtain a training sample set, wherein the training samples in the training sample set include structural information of the target protein and the binding protein; The training samples are input into a protein reverse folding model to obtain multiple candidate amino acid sequences of the binding protein; Calculate the binding performance indicators between the target protein and the binding agent protein of each candidate amino acid sequence. The binding performance indicators include complex conformation quantification indicators and / or binding affinity quantification indicators. The complex is a complex formed by the target protein and the binding agent protein of any candidate amino acid sequence. Based on the aforementioned performance metrics, positive and negative sample pairs are constructed; The protein reverse folding model is trained based on the reinforcement learning algorithm and the positive and negative sample pairs to obtain the trained protein reverse folding model.

2. The method according to claim 1, wherein, The conformation quantification index of the complex includes at least one of the following: the predicted local distance difference test score of each residue in the complex, the predicted alignment error of any residue pair in the complex, the confidence of the residue-level interface structure of each residue in the complex, and the overall binding mass score of the complex.

3. The method according to claim 1 or 2, wherein, Calculate conformational quantification metrics for the complex between the target protein and binding proteins of each candidate amino acid sequence, including: The amino acid sequence of the target protein and the candidate amino acid sequence of the binding protein are input into the complex conformation prediction model. Based on the complex conformation prediction model, the degree of spatial position fluctuation of each residue in the complex during the conformation iteration optimization process is calculated. The degree of spatial location fluctuation is mapped to a preset score range to obtain the predicted local distance difference test score of each residue in the complex.

4. The method according to claim 1 or 2, wherein, Calculate conformational quantification metrics for the complex between the target protein and binding proteins of each candidate amino acid sequence, including: Based on the complex conformation prediction model, the spatial distance between α carbon atoms of any residue pair in the complex is calculated. Calculate the deviation between the spatial distance and the predetermined reference spatial distance; The deviation value is mapped to a preset numerical range to obtain the predicted alignment error of any residue pair in the complex.

5. The method according to claim 1 or 2, wherein, Calculate conformational quantification metrics for the complex between the target protein and binding proteins of each candidate amino acid sequence, including: Based on the complex conformation prediction model, the prediction alignment error of any residue pair in the complex and the total number of residues in the complex are obtained. For each residue of the binding protein, the spatial matching degree between the residue and multiple residues of the target protein is calculated based on the predicted alignment error between the residue and multiple residues of the target protein and the total number of residues. The spatial matching degree is normalized to obtain the residue-level interface structure confidence degree of the residue.

6. The method according to claim 1 or 2, wherein, Calculate conformational quantification metrics for the complex between the target protein and binding proteins of each candidate amino acid sequence, including: Based on the complex conformation prediction model, the predicted local distance difference test score of each residue in the complex and the predicted alignment error of any residue pair in the complex are obtained. The overall binding mass score of the complex is obtained by weighted summation based on the predicted local distance difference test score of each residue, the predicted alignment error of any residue pair, and the preset weight. In this context, the weights of residues at the binding interface of the complex are greater than the weights of residues at the non-binding interface of the complex, and the weights of residue pairs at the binding interface are greater than the weights of residue pairs at the non-binding interface.

7. The method according to claim 1, wherein, The binding affinity quantification parameters include at least one of the binding energy and affinity parameters between the target protein and the binding agent protein.

8. The method according to claim 1 or 7, wherein, Calculate quantitative indices of binding affinity between the target protein and binding agent proteins for each candidate amino acid sequence, including: The atomic coordinates of the complex are input into a molecular mechanics simulation tool; Based on molecular mechanics simulation tools, the total energy of the complex, the first stable state energy of the target protein, and the second stable state energy of the binding protein were calculated respectively. The binding energy between the target protein and the binding agent protein is calculated based on the total energy of the complex, the first stable state energy, and the second stable state energy.

9. The method according to claim 8, wherein, The binding energy between the target protein and the binding agent protein is calculated based on the total energy of the complex, the first stable state energy, and the second stable state energy, including: Calculate the sum of the energy of the first stable state and the energy of the second stable state; The binding energy is obtained by calculating the difference between the total energy of the complex and the sum of the given energy.

10. The method according to claim 1 or 7, wherein, Calculate quantitative indices of binding affinity between the target protein and binding agent proteins for each candidate amino acid sequence, including: For any candidate amino acid sequence of the binding agent protein, the binding process between the target protein and the binding agent protein is detected by wet experimental method to obtain the binding rate constant and dissociation rate constant of the target protein and the binding agent protein. The affinity coefficient between the target protein and the binding agent protein is calculated based on the binding rate constant and the dissociation rate constant.

11. The method according to claim 1, wherein, Based on the aforementioned performance metrics, positive and negative sample pairs are constructed, including: Positive and negative sample pairs are constructed based on the binding affinity quantification index; and / or, positive and negative sample pairs are constructed based on the complex conformation quantification index.

12. The method according to claim 11, wherein, Based on the combined affinity strength quantification index, positive and negative sample pairs are constructed, including: Multiple candidate amino acid sequences were sorted according to the binding affinity quantification index. The candidate amino acid sequences that are ranked higher are assigned to the first affinity group, and the candidate amino acid sequences that are ranked lower are assigned to the second affinity group. A predetermined number of candidate amino acid sequences are selected from the first affinity group as positive samples, and a corresponding number of candidate amino acid sequences are selected from the second affinity group as negative samples. The positive sample is paired with the negative sample, and the attributes of the paired positive sample and the negative sample are labeled to obtain the positive and negative sample pair.

13. The method according to claim 11, wherein, Based on the conformational quantification index of the complex, positive and negative sample pairs are constructed, including: Based on the conformation quantification index of the complex, the contribution of residue-level conformation accuracy and / or residue-level affinity is obtained; Target residue sites that are selected based on the contribution of the residue-level conformation accuracy and / or the contribution of the residue-level affinity to a preset threshold. Candidate amino acid sequences that meet the preset conditions for conformation quantification of the complex are selected as reference sequences, and site-directed mutagenesis is performed on the target residue sites in the reference sequences to obtain positive and negative mutagenesis sequences. The positive mutation sequence is used as a positive sample, and the negative mutation sequence is used as a negative sample to obtain the positive-negative sample pair.

14. The method according to claim 13, wherein, Based on the conformation quantification metrics of the complex, the contribution of residue-level conformation accuracy and / or residue-level affinity is obtained, including: The confidence level of the residue-level interface structure in the conformation quantification index of the complex is used as the contribution of the residue-level conformation accuracy. And / or, based on the three-dimensional structure of the complex verified by the conformation quantification index of the complex, the residue-level affinity contribution is obtained by energy calculation tools.

15. The method according to claim 1, wherein, The protein reverse folding model includes interface site constraint parameters, which are used to constrain the protein reverse folding model to preferentially select amino acids that can form hydrogen bonds and / or hydrophobic interactions with the target protein when generating amino acids in the binding protein that come into contact with the target protein.

16. The method according to claim 1, wherein, The protein reverse folding model is trained based on a reinforcement learning algorithm and the positive and negative sample pairs, including: The positive and negative sample pairs and the structural information are converted into an input vector; The input vector is input into the protein defolding model, and the first probability of the protein defolding model generating a positive sample amino acid sequence and the second probability of generating a negative sample amino acid sequence are calculated. Based on the reinforcement learning algorithm, and according to the first probability, the second probability and the predetermined reference probability, the total loss value is calculated; wherein, the reference probability includes the first reference probability of the original protein reverse folding model generating positive sample amino acid sequences and the second reference probability of negative sample amino acid sequences. The parameters of the protein reverse folding model are updated based on the total loss value; Repeat the steps of input vector calculation, probability calculation, loss value calculation and parameter update until the convergence condition is met to obtain the trained protein reverse folding model.

17. The method according to claim 16, wherein, Based on a reinforcement learning algorithm, and according to the first probability, the second probability, and a predetermined reference probability, the total loss value is calculated, including: Based on the loss function in the reinforcement learning algorithm, and according to the first probability, the second probability and the reference probability, the preference loss value and the reference contrast loss value are calculated; The total loss value is obtained by weighting and summing the preference loss value and the reference comparison loss value using preset weights. The preference loss value is used to characterize the degree of preference between the probability of the protein defolding model generating a positive sample amino acid sequence and the probability of generating a negative sample amino acid sequence, and the reference contrast loss value is used to characterize the degree of prediction deviation between the protein defolding model and the original protein defolding model when generating a positive sample amino acid sequence.

18. A method for designing a binding protein, comprising: The structural information of the target protein and the target binding protein is input into the protein reverse folding model to obtain multiple candidate amino acid sequences of the target binding protein; Calculate the binding performance index between the target protein and the target binding agent protein of each candidate amino acid sequence. The binding performance index includes the complex conformation quantification index and / or binding affinity quantification index. The complex is a complex formed by the target protein and the target binding agent protein of any candidate amino acid sequence. The number of candidate amino acid sequences that meet the preset conditions for binding performance is counted. If the number is greater than or equal to a preset threshold, the candidate amino acid sequences that meet the preset conditions are output.

19. The method according to claim 18, wherein, The protein reverse folding model is obtained by training the protein reverse folding model according to any one of claims 1 to 17.

20. The method according to claim 18, wherein, The method further includes: If the number is less than the preset threshold, positive and negative sample pairs are constructed according to the combination performance index. Based on the reinforcement learning algorithm and the positive and negative sample pairs, the protein reverse model is fine-tuned to obtain the fine-tuned protein reverse model. The structural information is input into the fine-tuned protein reverse folding model to regenerate the candidate amino acid sequence of the target binding agent protein. The steps of calculating the binding performance index and fine-tuning the protein reverse folding model are repeated until the number is greater than or equal to a preset threshold or the number of iterations reaches a preset number.

21. The method according to claim 18, wherein, The conformation quantification index of the complex includes at least one of the following: the predicted local distance difference test score of each residue in the complex, the predicted alignment error of any residue pair in the complex, the confidence of the residue-level interface structure of each residue in the complex, and the overall binding mass score of the complex.

22. The method according to claim 18, wherein, The binding affinity quantification parameters include at least one of the binding energy and affinity parameters between the target protein and the target binding agent protein.

23. The method of claim 20, wherein, Based on the reinforcement learning algorithm and the positive and negative sample pairs, the protein reverse folding model is fine-tuned, including: The positive and negative sample pairs and the structural information are converted into an input vector; The input vector is input into the protein defolding model, and the first probability of the protein defolding model generating a positive sample amino acid sequence and the second probability of generating a negative sample amino acid sequence are calculated. Based on the reinforcement learning algorithm, and according to the first probability, the second probability and the predetermined reference probability, the total loss value is calculated; wherein, the reference probability includes the first reference probability of the original protein reverse folding model generating positive sample amino acid sequences and the second reference probability of negative sample amino acid sequences. The parameters of the protein reverse folding model are updated based on the total loss value; Repeat the steps of input vector calculation, probability calculation, loss value calculation and parameter update until the convergence condition is met to obtain the fine-tuned protein reverse folding model.

24. The method according to claim 23, wherein, Based on a reinforcement learning algorithm, and according to the first probability, the second probability, and a predetermined reference probability, the total loss value is calculated, including: Based on the loss function in the reinforcement learning algorithm, and according to the first probability, the second probability and the reference probability, the preference loss value and the reference contrast loss value are calculated; The total loss value is obtained by weighting and summing the preference loss value and the reference comparison loss value using preset weights. The preference loss value is used to characterize the degree of preference between the probability of the protein defolding model generating a positive sample amino acid sequence and the probability of generating a negative sample amino acid sequence, and the reference contrast loss value is used to characterize the degree of prediction deviation between the protein defolding model and the original protein defolding model when generating a positive sample amino acid sequence.

25. A training device for a protein reverse folding model, comprising: The acquisition module is used to acquire a training sample set, wherein the training samples in the training sample set include structural information of the target protein and the binding protein; The input module is used to input the training samples into the protein reverse folding model to obtain multiple candidate amino acid sequences of the binding protein; The calculation module is used to calculate the binding performance index between the target protein and the binding agent protein of each candidate amino acid sequence. The binding performance index includes the complex conformation quantification index and / or binding affinity quantification index. The complex is a complex formed by the target protein and the binding agent protein of any candidate amino acid sequence. A construction module is used to construct positive and negative sample pairs based on the combined performance metrics; The training module is used to train the protein reverse folding model based on the reinforcement learning algorithm and the positive and negative sample pairs to obtain the trained protein reverse folding model.

26. A design device for a binding protein, comprising: The input module is used to input the structural information of the target protein and the target binding protein into the protein reverse folding model to obtain multiple candidate amino acid sequences of the target binding protein; The calculation module is used to calculate the binding performance index between the target protein and the target binding agent protein of each candidate amino acid sequence. The binding performance index includes the complex conformation quantification index and / or binding affinity quantification index. The complex is a complex formed by the target protein and the target binding agent protein of any candidate amino acid sequence. The output module is used to count the number of candidate amino acid sequences that meet the preset conditions for binding performance indicators, and output the candidate amino acid sequences that meet the preset conditions when the number is greater than or equal to a preset threshold.

27. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-24.

28. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-24.

29. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-24.

Images