Method for training affinity prediction model of protein complex and application thereof

Abstract

The application relates to a training method of an affinity prediction model of a protein complex and application thereof. The training method comprises the following steps: obtaining training data, at least part of mutant protein complexes being obtained by mutation of corresponding wild-type protein complexes; obtaining multi-modal features of the mutant protein complex according to three-dimensional structure information of the mutant complex, wherein the multi-modal features comprise molecular physical property features of the mutant protein complex, and / or residue interaction features of the mutant protein complex and residue interaction features of the corresponding wild-type protein complex; and training a preset affinity prediction model according to the training data and the multi-modal features to obtain a trained target affinity prediction model. The scheme provided by the application can improve the accuracy and generalization of the affinity prediction model in predicting the relative affinity of the mutant protein complex.

Description

Technical Field

[0001] This application relates to the field of protein affinity prediction technology, and in particular to a training method for a protein complex affinity prediction model and its application. Background Technology

[0002] In the field of protein computation, affinity refers to the strength of binding between biomolecules (such as proteins). Changes in relative affinity are typically quantified as the difference in binding free energy (ΔΔG or ddG), used to quantitatively characterize the impact of amino acid mutations on intermolecular affinity. Affinity prediction is one of the central tasks of structural biology and protein engineering, while relative affinity prediction focuses even further on capturing subtle changes in binding free energy caused by mutations. This is crucial for elucidating pathogenic mutation mechanisms and optimizing protein binding properties.

[0003] In relative affinity prediction tasks, commonly used methods include traditional Computer-Aided Drug Design (CADD) methods based on physics and experience, and data-driven deep learning methods. Traditional CADD methods mainly rely on the knowledge of domain experts, directly calculating energy changes through manually designed energy functions and force fields. However, these methods typically suffer from drawbacks such as limited data, limited features, and low prediction accuracy in multi-site mutation scenarios. Deep learning methods can learn the relationship between protein sequences, structures, and energy from data, and their performance has the potential to surpass traditional methods as the amount of data increases. However, current deep learning methods generally struggle to generalize to real-world applications due to data scarcity or significant distribution bias, especially in protein engineering and drug development, where numerous mutations exist, making it difficult for current affinity prediction models to effectively predict binding energy changes from multiple mutations.

[0004] Therefore, improving the generalization and accuracy of affinity prediction models for predicting the relative affinity between mutated proteins, thereby further enhancing the precision of antibody screening for mutated proteins and their practical application capabilities, is a problem that urgently needs to be solved. Summary of the Invention

[0005] To address or partially address the problems existing in related technologies, this application provides a training method for a protein complex affinity prediction model and its application, which can improve the accuracy and generalization of the affinity prediction model in predicting the relative affinity of mutant protein complexes.

[0006] The first aspect of this application provides a method for training a protein complex affinity prediction model, comprising: Acquire training data, which includes the three-dimensional structural information of the mutant protein complex and training tags, wherein at least a portion of the mutant protein complex is obtained by mutation of the corresponding wild-type protein complex, and the training tags include the relative affinity of the mutant protein complex. Based on the three-dimensional structural information of the mutant complex, the multimodal features of the mutant protein complex are obtained, wherein the multimodal features include the molecular physical properties of the mutant protein complex, and / or the residue interaction features of the mutant protein complex and the corresponding residue interaction features of the wild-type protein complex. The preset affinity prediction model is trained based on the training data and the multimodal features to obtain a trained target affinity prediction model.

[0007] A second aspect of this application provides a method for predicting the affinity of mutant protein complexes, comprising: Obtain the information to be predicted, which includes the target mutant three-dimensional structure information of the target mutant protein complex, or the target wild-type three-dimensional structure information and target mutation information of the target wild-type protein complex. Based on the information to be predicted, corresponding multimodal features are generated; the multimodal features include the molecular physical properties of the target mutant protein complex, and / or the residue interaction features of the target mutant protein complex and the corresponding residue interaction features of the target wild-type protein complex. Based on the information to be predicted and the multimodal features, the target affinity prediction model trained by the training method described in the first aspect generates the relative affinity corresponding to the target mutant protein complex.

[0008] A third aspect of this application provides a training apparatus for a protein complex affinity prediction model, comprising: A data acquisition module is used to acquire training data, which includes the three-dimensional structural information of the mutant protein complex and training tags, wherein at least a portion of the mutant protein complex is obtained by mutation of the corresponding wild-type protein complex, and the training tags include the relative affinity of the mutant protein complex. The feature extraction module is used to obtain the multimodal features of the mutant protein complex based on the three-dimensional structural information of the mutant complex, wherein the multimodal features include the molecular physical properties of the mutant protein complex, and / or the residue interaction features of the mutant protein complex and the corresponding residue interaction features of the wild-type protein complex. The model training module is used to train the preset affinity prediction model based on the training data and the multimodal features to obtain the trained target affinity prediction model.

[0009] A fourth aspect of this application provides a training apparatus for a protein complex affinity prediction model, comprising: The information acquisition module is used to acquire the information to be predicted, which includes the target mutant three-dimensional structure information of the target mutant protein complex, or the target wild-type three-dimensional structure information and target mutation information of the target wild-type protein complex. The feature encoding module is used to generate corresponding multimodal features based on the information to be predicted; the multimodal features include the molecular physical properties of the target mutant protein complex, and / or the residue interaction features of the target mutant protein complex and the corresponding residue interaction features of the target wild-type protein complex; The affinity prediction module is used to generate the relative affinity corresponding to the target mutant protein complex based on the information to be predicted and the multimodal features, using the target affinity prediction model trained by the training method described in the first aspect.

[0010] The fifth aspect of this application provides an electronic device, comprising: Processor; and A memory having executable code stored thereon, which, when executed by the processor, causes the processor to perform the method described in the first or second aspect above.

[0011] A sixth aspect of this application provides a computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method described in the first or second aspect above.

[0012] The seventh aspect of this application provides a computer program product comprising computer instructions that, when executed by a processor, implement the method described in the first or second aspect above.

[0013] The technical solution provided in this application may include the following beneficial results: The training method of this application trains a preset affinity prediction model by combining the three-dimensional structural information and multimodal features of the mutant protein complex. During the training process, the model learns molecular physical properties such as the physical energy term, geometric complementarity, and hydrophobic core integrity of the mutant complex, and / or residue interaction features focusing on the conformational energy and interaction energy of residues. This deeply integrates the three-dimensional structural information of the mutant complex with the above-mentioned multi-level features, which can improve the generalization and accuracy of the affinity prediction model, thereby further improving the accuracy of antibody screening containing mutant proteins and its practical application capability.

[0014] The training method in this application overcomes the limitations of single data types by constructing a high-quality, multi-dimensional dataset. First, a rigorous cleaning and correction process is employed to screen all wild-type protein complex structural data by resolution and residue count, and site alignment is performed on mutant data to ensure the accuracy and consistency of the three-dimensional structural information of the mutant protein complexes. Second, structural data of mutant protein complexes with affinity tags (three-dimensional modalities) are added, and priority rules and strict deduplication are established to ensure the uniqueness and reliability of each group of interaction data. Finally, multi-clustering of protein sequences and complex structures contained in the mutant protein complexes is performed using protein clustering tools, with stringent thresholds set to ensure no data leakage between the training, validation, and test sets, thereby improving the model's prediction accuracy and generalization ability in scenarios with limited mutant data.

[0015] Furthermore, the training method of this application, through multimodal feature extraction and a multi-granularity feature extraction strategy, can grasp the overall energy distribution and stability characteristics of the mutant protein complex from a global perspective, and also capture the sensitivity differences of key residues to mutation effects from a local perspective, providing rich multi-level information for the affinity prediction model. By integrating the three-dimensional structural information of the mutant protein complex and local residue interaction features in a unified feature representation space, data from different modalities are mapped to a shared embedding space for fusion, achieving deep fusion and information complementarity between different modal features. Through the above feature fusion strategy, the affinity prediction model can fully explore the synergistic information between various feature data, overcome the limitations of incomplete single feature information, and thus significantly improve the generalization ability and prediction stability of the affinity prediction model in diverse application scenarios.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0017] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of exemplary embodiments thereof in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments thereof.

[0018] Figure 1 This is a flowchart illustrating the training method of the protein complex affinity prediction model shown in the embodiments of this application; Figure 2 This is a schematic diagram illustrating the fusion of multimodal features in the first prediction model shown in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the fusion of multimodal features in the second prediction model shown in an embodiment of this application; Figure 4 This is a schematic flowchart illustrating the affinity prediction method for mutant protein complexes according to embodiments of this application; Figure 5 This is a schematic diagram of the structure of the training device for the protein complex affinity prediction model shown in the embodiments of this application; Figure 6 This is another schematic diagram of the training device for the protein complex affinity prediction model shown in the embodiments of this application; Figure 7 This is a schematic diagram of the affinity prediction device for mutant protein complexes shown in the embodiments of this application; Figure 8 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application. Detailed Implementation

[0019] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0020] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0021] It should be understood that although the terms "first," "second," "third," etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, initial information may also be referred to as second information, and similarly, second information may also be referred to as initial information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0022] In related technologies, when predicting the mutant affinity of mutant protein-protein complexes, training data is scarce and data features are limited, using only the three-dimensional structural features of the complex. This makes it difficult for affinity prediction models to effectively predict changes in mutant binding energy, and consequently, to reliably assess the impact of mutations on protein complexes.

[0023] To address the aforementioned issues, this application provides a training method for a protein complex affinity prediction model and its application, which can improve the accuracy and generalization of the affinity prediction model for mutant protein complexes.

[0024] The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0025] See Figure 1 An embodiment of this application illustrates a method for training a protein complex affinity prediction model, which includes: S110, Acquire training data, which includes the three-dimensional structural information of the mutant protein complex and training labels, wherein at least a portion of the mutant protein complex is obtained by mutation of the corresponding wild-type protein complex, and the training labels include the relative affinity of the mutant protein complex.

[0026] The protein complex of this application can be a complex of protein 1 and protein 2. Protein 1 and protein 2 can be the same or different types of proteins, and can be wild-type proteins without mutation or mutant proteins after mutation. The mutation of a single protein can be a single-point mutation or a multi-point mutation, without limitation.

[0027] In this application, when at least one of protein 1 and protein 2 has mutated, the protein complex is a mutant protein complex. When neither protein 1 nor protein 2 has mutated, the protein complex is a wild-type protein complex. To facilitate the differentiation of the various proteins in this application, the two proteins in the mutant protein complex are named mutant protein 1 and mutant protein 2 (it should be noted that although both are named mutant proteins, one of them can be a wild-type protein without mutation; this naming is only to differentiate the two proteins), and the two proteins in the wild-type protein complex are named wild-type protein 1 and wild-type protein 2.

[0028] In the training data of this application, the three-dimensional structural information of the mutant protein complex may include data stored in PDB format, containing the three-dimensional coordinates of each heavy atom of the complex. Specifically, PDB is a structure file format (.pdb) that uses a fixed-width plain text format and describes atomic coordinates, residue information, and connection relationships line by line using record types such as ATOM and HETATM. It is a standard input / output format commonly used by molecular visualization and simulation software. For consistency, this application refers to data files that store the three-dimensional coordinate information of heavy atoms of amino acid residues in PDB format as PDB format files.

[0029] In addition to the mutant protein complexes already existing in protein databases, to obtain a richer collection of mutant protein complexes, this application can pre-mutate wild-type protein complexes to obtain corresponding mutant protein complexes. The wild-type protein complexes can be mutated using computer simulation based on mutation information, or mutated based on real experiments, thereby obtaining the corresponding mutant protein complexes. For each mutant protein complex, a specific value of its relative affinity can be obtained as a training label based on experimental determination, literature search, computer simulation calculation, etc. Preferably, the actual experimental value of the relative affinity can be used as the training label for the mutant protein complex.

[0030] The affinity prediction model of this application is used to predict the relative affinity of mutant protein complexes. In some embodiments, the relative affinity is quantified by the difference in binding free energy (ΔG) (ΔΔG or ddG), used to quantitatively characterize the effect of amino acid mutations on intermolecular affinity. When ΔΔG > 0 (positive value), it indicates that the binding of the mutated protein complex is weaker, that is, the mutation leads to dissociation of the complex; conversely, ΔΔG < 0 (negative value), it indicates that the binding of the mutated protein complex is stronger, that is, the mutation makes the structure of the complex more fitted. In other embodiments, the relative affinity can also be quantified using other parameters, such as the ratios of Ki, Kd, ​​IC50, EC50, etc., or the difference of their logarithmic values ​​(ΔpKi, ΔpIC50, etc.), as well as the differences in experimental physical quantities such as ΔTm, retention time ratio, and inhibition rate difference, without limitation.

[0031] To obtain higher quality training data, some embodiments employ a cleaning and correction process to screen the three-dimensional structural information of all wild-type protein complexes based on resolution and residue number, and perform site alignment on the PDB format files of mutant protein complexes obtained by mutation to ensure the accuracy and consistency of the three-dimensional structural information of the complexes before and after mutation.

[0032] S111, Cleans the wild-type protein complex to obtain the three-dimensional structural information of the wild-type complex.

[0033] In this application, a large amount of PDB format data corresponding to wild-type protein complexes can be obtained from publicly available protein databases (such as PDB databases, UniProt databases, etc.), that is, the three-dimensional structure information of wild-type complexes is stored in PDB format files. In this step, the cleaned wild-type protein complex and its corresponding three-dimensional structure information are obtained according to the following method. It should be noted that, based on the three-dimensional structure information of the wild-type complex, the three-dimensional structure information of wild-type protein 1 and wild-type protein 2 can be extracted.

[0034] In some implementations, only the information corresponding to the ATOM record line and the natural amino acid information in the PDB format file of the wild-type protein complex are retained. The ATOM record line indicates that the line records the three-dimensional coordinate information of the heavy atom (non-hydrogen atom) in a standard amino acid. That is, in this application, the three-dimensional structural information stored in the PDB format files of each wild-type protein 1, wild-type protein 2, and wild-type protein complex only retains the information of the ATOM record line. Specifically, all non-ATOM record lines (such as HETATM, ANISOU, CONECT, etc.) are removed from the PDB format file of the wild-type protein complex, and only the three-dimensional coordinate information of the heavy atom recorded in the standard ATOM line is retained; non-standard amino acids, including modified amino acids, non-natural amino acids, and residue types that cannot be clearly assigned in crystal analysis, are filtered out to ensure that subsequent processing is only performed on the 20 standard amino acids, thereby avoiding calculation errors or feature deviations caused by abnormal residue types.

[0035] In some implementations, wild-type protein complexes with a resolution greater than or equal to 3.5 Å and a residue count greater than or equal to 1000 are filtered out. This processing ensures both the quality of the input data and computational efficiency. After the above cleaning process, high-quality raw structural data of wild-type protein complexes are obtained, providing a reliable computational foundation for subsequent generation of mutant protein complex data, training data partitioning, and feature generation.

[0036] S112, based on the mutation information, the wild-type protein complex is mutated to obtain a mutant protein complex.

[0037] It is understood that, in order to increase the amount of training data, at least some of the mutant protein complexes in this application are obtained by mutagenesis of wild-type protein complexes. The mutation method for the wild-type protein complexes in this application can be substitution, insertion, or deletion of amino acid residues. To ensure the quality and stability of training data, in some embodiments, the same mutation method can be used to mutate each wild-type protein complex. Preferably, for example, the mutation method is unified as "the type of amino acid residue replaced".

[0038] In some specific implementations, the three-dimensional structural information and mutation information of the wild-type protein complex are obtained. The three-dimensional structural information of the wild-type complex includes the three-dimensional coordinates of heavy atoms corresponding to each residue, and the residue category is a natural amino acid. The mutation information includes at least one residue substitution information. At least one residue of the wild-type protein complex is substituted according to the mutation information to obtain the three-dimensional structural information of the mutant protein complex. That is, at least one natural amino acid at one site in wild-type protein 1 and / or wild-type protein 2 in the wild-type protein complex is replaced with another natural amino acid. The mutation information includes the position of the amino acid to be replaced on the wild-type protein complex and the type of the replaced amino acid residue. For example, when the mutation information is AA39G, the first A represents alanine, the second A represents chain A, 39 indicates a mutation occurring at the 39th amino acid position, and G indicates that the mutated amino acid is glycine. The mutation information indicates that "the alanine at position 39 of chain A in the wild-type protein complex is mutated to glycine."

[0039] S113, Correct the PDB format file of the wild-type protein complex with the site correspondence in the mutation information to obtain the corrected PDB format file of the mutant protein complex.

[0040] This step systematically calibrates the PDB format file of the mutant protein complex obtained using mutation information to ensure consistency between the mutation sites and the three-dimensional structure. Its accuracy directly affects the quality of subsequent feature extraction and model training. The specific correction method after adding mutation information to the PDB format file of the wild-type protein complex is as follows: If the mutation annotation occurs on an insertion code in the PDB file of the wild-type protein complex, the residue numbers in that wild-type PDB file undergo a renaming process. For example, when the mutation annotation is W35AA, "35AA" indicates an additional residue inserted after the standard number 35. Such insertion codes are typically represented as "35A" in the wild-type complex structure PDB file. The calibration process requires finding the residue corresponding to 35A in the wild-type complex structure PDB file and mapping it to the consecutive number 36. This eliminates the numbering discontinuity problem caused by the insertion code, ensuring that the mutation site can be accurately located to the correct residue in the wild-type complex structure. This guarantees accurate mutation completion and ultimately yields an accurate PDB file of the mutant protein complex.

[0041] Furthermore, significant numbering "skips" appear in the PDB format files for wild-type protein complexes, such as jumping directly from residue 139 to 141, omitting residue 140. This numbering discontinuity is common in protein databases and usually stems from various reasons: it could be due to naturally occurring insertions or deletions in the sequence; it could be numbering traces left after partial sequence removal during precursor protein processing; or it could be artificial renumbering simply to distinguish different isoforms, mutants, or crystallographic symmetry units. For such numbering anomalies, the calibration process must identify and accurately process each case. By comparing the correspondence between the wild-type amino acid sequence and the amino acid sequence in the PDB format file, the true mutation location of the mutation site in the continuous numbering system can be determined, thereby obtaining an accurate PDB format file for the mutant protein complex.

[0042] To accurately map the mutation site's number in the amino acid sequence to its number in the PDB file, the starting point for calculating the mutation site (i.e., the first amino acid residue of the protein complex) should correspond to the sequence number of the corresponding chain in the PDB file. This is especially important for cases involving amino acid deletions, additional artificially added residues, or disordered regions and terminal extensions in the crystal structure, which require separate, targeted corrections.

[0043] In some implementations, when the starting point for calculating the mutation site is the same as the sequence number in the corresponding strand's PDB format file, then... .

[0044] In some implementations, when the calculated starting point of the mutation site is the same as the sequence starting point in the corresponding strand's PDB format file, then... .

[0045] In some implementations, when the calculated starting point of the mutation site is the same as the given wild-type amino acid sequence starting point, but the starting point of the wild-type amino acid sequence is different from the sequence starting point in the PDB format file, then... .

[0046] in, This indicates the actual mutation location in the PDB file. This indicates the mutation points labeled in the training data. The value is read from the PDB format file.

[0047] Through the above systematic calibration process, it was ensured that each mutation site of the wild-type protein complex could be accurately located and mapped in the three-dimensional structural information of the corresponding mutant protein complex, providing a reliable data foundation for subsequent modeling of relative affinity after mutation.

[0048] S114, perform deduplication and screening on each mutant protein complex to ensure the uniqueness of the three-dimensional structural information of each mutant protein complex and its corresponding training label in the training data.

[0049] To ensure the uniqueness and reliability of the relative affinity data in each training dataset, this step, after obtaining the mutant protein complex from the wild-type protein complex through mutation information, requires ensuring that the three-dimensional structural information of each mutant complex retains only the unique relative affinity data.

[0050] First, for the same interaction pair (i.e., a combination of the same three-dimensional structural information of the mutant complex and the same mutation information), a priority ranking of relative affinity data was established based on the authority and accuracy of the experimental measurement methods. Specifically, considering technological maturity, throughput, sample consumption, and applicability, relative affinity data measured by surface plasmon resonance (SPR) technology (considered one of the standard methods for affinity determination due to its advantages such as real-time, label-free, and rich kinetic information) were prioritized. Next, relative affinity data measured by biolayer interferometry (BLI), isothermal titration calorimetry (ITC), and microthermophoresis (MST) were retained. Then, the data were filtered based on the publication date, retaining only the most recently published records to ensure that the relative affinity data used in the training data reflects the latest understanding and measurement level at the current research stage. Furthermore, for redundant entries caused by identical sequence composition but interchanged chain identifiers, such as duplicate records where the ligand and receptor sequences are completely identical but the chain identifiers (e.g., chain A and chain B) are interchanged, only one of them was retained. Such redundancy usually stems from the arbitrary naming of chain links in the PDB format files of mutant protein complexes. If left unaddressed, this will cause the three-dimensional structure of the same mutant protein complex to appear repeatedly in the training data, affecting the independence of subsequent model training and the objectivity of evaluation.

[0051] Through the above multi-level deduplication and screening process, only the record with the highest priority, the latest publication time and the unique chain identifier is retained for each interaction pair. All other duplicate entries are deleted to ensure that only one relative affinity data is retained in the PDB format file of each mutant protein complex. This avoids data redundancy and potential bias caused by multiple similar or duplicate records corresponding to the same structure, and provides a high-quality, low-redundancy data foundation for subsequent model training.

[0052] S115 divides the training data to obtain the training set, validation set, and test set.

[0053] After cleaning and filtering the data, this step involves partitioning the final high-quality training data. To ensure the independence of the model during training, validation, and testing, and to avoid performance evaluation bias due to data leakage, this application employs a strict data partitioning strategy. In some implementations, the training data is divided into multiple clusters, and the mutant protein complexes in each cluster are clustered based on structural similarity, interface region confidence, and / or sequence homology.

[0054] In some implementations, a structural similarity score is obtained for every two mutant protein complexes. When the structural similarity score is greater than a preset structural similarity threshold, the corresponding two mutant protein complexes are aggregated to form a corresponding cluster. The structural similarity score (TM-score) is an indicator that measures the similarity of three-dimensional protein structures, ranging from 0 to 1; a higher score indicates greater structural similarity. The preset structural similarity threshold can be set according to actual needs, for example, it can be 0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, or other values, without limitation. For example, when the TM-score is greater than 0.5, it indicates that the two protein structures have the same folding type; when the TM-score is greater than 0.7, it indicates that the two structures are highly similar. In this application, mutant protein complexes with a TM-score of 0.7 or higher are grouped into the same cluster during the clustering process, thereby ensuring that highly similar mutant protein complexes are not scattered across different data partitions and ensuring the correlation between data.

[0055] In some implementations, the confidence level of the interface region for each mutant protein complex is obtained. When the confidence level of the interface region is greater than a preset confidence level threshold, the mutant protein complexes are aggregated to form a corresponding cluster. For example, the interface local distance difference test (iLDDT) score is used to measure the confidence level of the interface region prediction. The preset confidence level threshold can be set according to actual needs, such as 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, or other values, without limitation. Alternatively, the pLDDT score can be used to measure the confidence level of the interface region prediction. The preset confidence level threshold can be set according to actual needs, such as 70, 75, 80, 85, 90, or other values, without limitation. This implementation aims to screen structures with high confidence levels in the interface region prediction, ensuring that the structural information on which clustering is based has sufficient reliability and stability.

[0056] In some implementations, the sequence homology score of every two mutant protein complexes is obtained. When the sequence homology score is greater than or equal to a preset minimum sequence homology threshold, the corresponding two mutant protein complexes are aggregated to form a corresponding cluster. The minimum sequence homology threshold means that mutant protein complexes with sequence homology reaching or exceeding this threshold are considered to have significant homology and will be assigned to the same cluster during data partitioning, thereby avoiding the risk of overestimation due to sequence similarity. The preset minimum sequence homology threshold can be set according to actual needs, for example, it can be 0.25, 0.28, 0.3, 0.35, 0.4, 0.45, 0.5, 0.55, 0.6, or other values, without limitation. For example, when the preset minimum sequence homology threshold is 0.3, mutant protein complexes with sequence homology of 30% or more can be grouped into the same cluster.

[0057] It is understood that one or more of the above dimensions can be selected to cluster the training data to obtain corresponding clusters. In some preferred embodiments, when multiple dimensions are selected for clustering simultaneously, mutant protein complex structures that simultaneously meet the structural similarity threshold, interface region confidence threshold, and sequence homology threshold are grouped into the same cluster. This ensures that the training data in the same cluster simultaneously meet the following criteria: high structural similarity (e.g., TM-score ≥ 0.7), high interface region confidence (e.g., iLDDT ≥ 0.65), and significant sequence homology (e.g., sequence identity ≥ 0.3). This application ensures the strictness of data partitioning through the above multi-dimensional clustering.

[0058] In some implementations, the training data includes a training set, a validation set, and a test set, wherein the training set, validation set, and test set have distinct clusters, and each cluster contains distinct mutant protein complexes. This design ensures that there is no structural or sequence overlap between the training set, validation set, and test set, and ensures that each piece of training data exists in only one of the datasets, providing a solid foundation for objective evaluation of model performance.

[0059] In some other embodiments, before training the preset affinity model using training data, this application can also pre-train the preset affinity model using pre-training data to obtain a pre-trained preset affinity prediction model. Accordingly, this application also includes a step of acquiring pre-training data. Specifically, the pre-training data includes the three-dimensional structural information of the mutant protein complex and pre-training labels, wherein the pre-training labels are exploitability data. In some embodiments, exploitability data includes at least one of the following: aggregation tendency, thermal stability, pharmacokinetic properties, solubility, and chemical stability of the mutant protein complex. The pre-training data is data independent of the aforementioned training data. The pre-training data can consist of mutant protein complexes with exploitability data; that is, compared to the already collected training data, additional exploitability data corresponding to the mutant protein complexes is collected to pre-train the preset affinity prediction model, thereby ensuring the fine-tuning efficiency of the affinity prediction model. By using one or more of the above-mentioned exploitability data as pre-training labels in the pre-training stage to pre-train the model, the network parameters of the preset affinity prediction model are fine-tuned.

[0060] In other words, exploitability data and relative affinity of mutant protein complexes serve as labels for different training stages, enabling phased training of the pre-defined affinity prediction model. Exploitability data includes specific indicators such as aggregation tendency (e.g., aggregation temperature Tagg, aggregation fraction), thermal stability (e.g., melting temperature Tm), pharmacokinetic properties (e.g., half-life, clearance rate), solubility (e.g., soluble expression level), and chemical stability (e.g., deamidation rate, oxidation rate). Exploitability data can include data corresponding to mutant protein complexes, and can also include data corresponding to mutant monomeric proteins (e.g., mutant protein 1 and / or mutant protein 2 corresponding to the mutant protein complex). For example, in the pre-training stage, pre-training is performed uniformly based on the exploitability data corresponding to at least one of the mutant protein complex, mutant protein 1, and mutant protein 2. The predicted exploitability data values ​​are obtained, and the loss value is calculated before backpropagation iterates the model parameters. For example, pre-training can be performed in parallel based on the exploitability data corresponding to mutant protein complexes, mutant protein 1, and mutant protein 2. After obtaining the predicted values ​​of their respective exploitability data, the joint loss value is calculated, and then backpropagation is performed to iterate the model parameters. It should be noted that since the pre-training data and the training data have no overlap, the mutant protein complexes in the pre-training data are different from those in the training data, and their corresponding mutant monomeric proteins are also different.

[0061] Developable data can be obtained through public databases, literature mining, and high-throughput experimental platforms (such as differential scanning fluorometry, static light scattering, and size exclusion chromatography). Although this type of data does not directly reflect the binding affinity between proteins in a complex, these indicators can indirectly reflect the binding tendency between molecules, the rationality of structural folding, and the stability of the overall conformation. This helps to enhance the model's global perception of the protein molecular space and improve its generalization performance in low-data scenarios. Based on the prediction strategy of the above-mentioned multi-scale pre-trained labels, the physicochemical differences of proteins in both independent and bound states can be captured simultaneously. By fine-tuning the network parameters of the model, a better starting point for training the preset affinity prediction model can be obtained.

[0062] In some specific implementations, when using multiple exploitability data as pre-training labels, a joint loss function is used to calculate the loss value of the preset affinity prediction model. It can be understood that multiple parallel and independent task heads can be set at the output of the preset affinity prediction model. Each task head is used to independently output the predicted value of one type of exploitability data, and each task head has its own corresponding loss function for calculating the loss value. The loss values ​​calculated based on the prediction values ​​of each task head are further calculated to obtain the corresponding joint loss value. Backpropagation is performed based on the joint loss value, and the model parameters are iterated through multiple rounds until the pre-training termination condition is met, thus obtaining the pre-trained preset affinity prediction model. The pre-trained preset affinity prediction model is then formally trained based on the aforementioned training data. It can be understood that when only one type of exploitability data is used as the corresponding pre-training label, only one task head needs to be set at the output of the preset affinity prediction model.

[0063] To unify exploitability data across different dimensions, the exploitability data can be pre-cleaned and processed. In some implementations, this involves standardizing the format and units of the exploitability data, removing training samples with missing or incomplete measurements, filtering out heterogeneous data from inconsistent experimental conditions (such as different pH values ​​or temperatures), and normalizing various exploitability indicators to comparable intervals. After these processes, various exploitability data from different dimensions can be unified, resulting in high-quality exploitability data that can be used within the same feature space.

[0064] As shown above, this application overcomes the limitations of single data types by constructing a high-quality, multi-dimensional dataset. First, a rigorous cleaning and correction process is employed to screen all wild-type protein complex structural data by resolution and residue count, and site alignment is performed on mutant data to ensure the accuracy and consistency of the three-dimensional structural information of the mutant protein complexes. Second, structural data of mutant protein complexes with affinity tags (three-dimensional modalities) are added, and priority rules and strict deduplication are established to ensure the uniqueness and reliability of each group of interaction data. Finally, multiple clustering of protein sequences and complex structures contained in the mutant protein complexes is performed using protein clustering tools, with stringent thresholds set to ensure no data leakage between the training, validation, and test sets, thereby improving the model's prediction accuracy and generalization ability in scenarios with limited mutant data.

[0065] S120. Based on the three-dimensional structural information of the mutant complex, obtain the multimodal characteristics of the mutant protein complex. The multimodal characteristics include the molecular physical properties of the mutant protein complex and / or the residue interaction characteristics of the mutant protein complex and the residue interaction characteristics of the corresponding wild-type protein complex.

[0066] In this application, the features used in the affinity prediction model are not limited to three-dimensional structural information, but also include multimodal features. Specifically, multimodal features may include molecular physical properties of the mutant protein complex, including at least one of the following: interface analysis features of the mutant complex, overall structure and hydrophobic core features, or energy features of coarse-grained conformation. Multimodal features may include residue interaction features of the mutant protein complex and the corresponding wild-type protein complex, including at least one of the following: van der Waals interactions, electrostatic interactions, hydrogen bond networks, solvation effects, backbone dihedral angle preference, side chain rotomer preference, hydrophobic interactions, or pairwise interactions.

[0067] By systematically integrating multi-dimensional information such as the three-dimensional structure and physicochemical properties of protein complexes, this method effectively overcomes the inherent limitations of single data modalities in terms of sample size and diversity. It can fully explore and leverage the complementary advantages between different data modalities, thereby significantly improving the prediction accuracy of affinity prediction models for changes in protein binding affinity in training scenarios with only a small amount of mutation data, and enhancing the model's generalization ability in different types of protein systems.

[0068] To obtain feature data that can be recognized by affinity prediction models, in some embodiments of this application, the three-dimensional structural information of the mutant complex is encoded as graph structure data that can be learned by a graph neural network. During the construction of the full residue graph, each amino acid residue in the mutant protein complex is defined as a node in the graph. Node features are used to encode the multidimensional attribute information of that residue, specifically including the residue's identity type (i.e., the category encoding of 20 standard amino acids), and three-dimensional spatial coordinates (taking C...). α The coordinates of atoms (representing the positions of residues) and at least one of the following: local structural geometric properties (such as main chain dihedral angles φ and ψ, and related parameters such as side chain conformation). Assuming the mutant protein complex has Q residues, then the diagram is defined as follows: , where the node set , node The corresponding feature vector can be represented as:

[0069] in, i Indicates the first in a protein complex i One residue, One-hot encoding for residue types, For residues i of The three-dimensional coordinates of atoms in a mutant protein complex. For residues i Local geometric properties, Represents residues i The key dihedral features of the side chains. These node features together constitute a comprehensive characterization of the properties of individual residues and their spatial location in the three-dimensional space of the mutant protein complex.

[0070] Meanwhile, edge features are used to characterize the spatial proximity relationships between residues, by analyzing the relationships between residues. Atomic distance is calculated, and residue pairs with a distance less than a preset threshold are linked. Geometric constraint information between the two residues is encoded in the edge features. This geometric constraint information includes at least one of the following: spatial distance, relative orientation, and possible interaction types (such as hydrogen bonding, hydrophobic contact, electrostatic interaction, etc.). Specifically, for any two residues... and Calculate the two Euclidean distance between atoms:

[0071] like (in (If the distance threshold is preset), then at the node and Establish edges between The geometric constraint information between two residues is further encoded in the edge features, and the edge feature vector can be represented as:

[0072] in, The spatial distance between two residues This represents the relative orientation characteristics between residues, specifically the spatial angle between the side chain direction vectors of two residues in a local coordinate system (the local coordinate system is based on the residues). With the origin, The key is constructed for the x-axis. The edges in the graph structure are encoded to indicate the presence of key interaction modes such as hydrogen bonding, hydrophobic contact, and electrostatic interactions. In this way, the edges can effectively capture the geometric constraints and interaction modes between residues, reflecting the spatial dependencies in protein folding and molecular recognition processes.

[0073] By constructing the full residue map described above, this application transforms the complex spatial information of the three-dimensional structure of mutant protein complexes into a structured graph representation, which not only preserves the fine-grained information at the residue level but also depicts the global topological relationships between residues, providing a high-quality structural semantic foundation for the subsequent deep fusion of multimodal features.

[0074] Based on the above method of constructing a full residue graph of the three-dimensional structure information of mutant protein complexes, the complex three-dimensional structure information is transformed into a vector representation of node features and edge features in a graph structure. Similarly, the three-dimensional structure information of wild-type protein complexes can be transformed into a vector representation of node features and edge features in a graph structure, and the three-dimensional structure information of individual proteins, mutant / wild-type protein 1 and mutant / wild-type protein 2, can be transformed into a vector representation of node features and edge features in a graph structure.

[0075] It should be noted that in the field of protein engineering, accurately characterizing the overall properties of protein structure and its intermolecular interactions is a core prerequisite and key challenge for predicting mutation effects and designing high-affinity binding agents. The functional performance of a protein depends not only on its static three-dimensional structure but also on the combined influence of multiple factors such as residue interactions, energy distribution, and the integrity of the hydrophobic core. Therefore, how to comprehensively extract and integrate feature information closely related to binding affinity from different dimensions has become a core issue in improving the predictive ability of models. In this application, the rich energy term calculation modules and structural analysis tools provided by Rosetta software are used to perform refined characterization of protein structures from multiple perspectives, including thermodynamics, geometry, and statistics. Based on the high-quality three-dimensional structural information of mutant complexes obtained in the above steps, Rosetta software's powerful capabilities in calculating one-dimensional (e.g., residue sequence features, energy distribution curves), two-dimensional (e.g., residue contact diagrams, interaction networks), and three-dimensional (e.g., spatial coordinates, interface geometry) features of protein complexes are fully utilized to systematically extract multi-dimensional features for each mutant protein complex.

[0076] In order to obtain the multimodal features of a standard mutant protein complex and obtain a more accurate feature description, in some implementations, the multimodal features can be extracted in the following manner.

[0077] S121, structural relaxation is performed on the mutant protein complex to obtain the lowest energy mutant complex conformation; the three-dimensional structural information of the mutant complex is the three-dimensional structural information of the lowest energy mutant complex conformation.

[0078] In this step, the structure of the mutant protein complex can be processed using the all-atom constrained relaxation protocol in Rosetta software. In some specific implementations, firstly, coordinate constraints are applied to the backbone heavy atoms and side-chain heavy atoms of the mutant protein complex to ensure that all heavy atoms always reference their initial coordinate positions during relaxation. Secondly, a multi-cycle FastRelax protocol is executed, searching for low-energy conformations by alternating between side-chain rearrangement and gradient-based torsion angle minimization. In each relaxation process, a strategy of gradually increasing the repulsion force weight is adopted, for example, gradually increasing the repulsion force weight from 2% to 25%, 55%, and finally 100%, using a pulsed increment to allow atoms to enter a low-energy state through alternating expansion and contraction. Each relaxation process includes two stages: side-chain rearrangement and gradient minimization. After multiple cycles of structural relaxation (e.g., after 5 cycles), the mutant complex conformation with the lowest Rosetta energy is selected from all generated conformations as the baseline structure for subsequent analysis. Through this processing, potential atomic conflicts and unreasonable conformations in the original complex structure can be effectively eliminated, while each... The RMSD of atoms is controlled to within 2. This standardized processing procedure lays a high-quality structural foundation for the accurate extraction of subsequent multimodal features.

[0079] S122, Based on the conformation of the mutant complex, extract the molecular physical characteristics of the mutant protein complex. The molecular physical characteristics include at least one of the following: interface analysis characteristics of the mutant complex, overall structure and hydrophobic core characteristics, or coarse-grained conformation energy characteristics.

[0080] In this step, based on the relaxed, high-quality mutant protein complex structure described above, the molecular physical characteristics of the mutant protein complex can be calculated using Rosetta software. These molecular physical characteristics include at least one of the following: interface analysis features of the mutant complex, overall structure and hydrophobic core features, or energy features of coarse-grained conformations. Specifically, the first type of molecular physical characteristic is the interface analysis feature of the mutant complex. By independently analyzing the binding and dissociation states of the mutant protein complex using Rosetta software, the system can calculate and output a series of energy terms and geometric features, such as binding free energy related terms (e.g., interfacial solvent-accessible surface area, hydrophobic contribution area, and polar contribution area), energy term decomposition features (e.g., van der Waals attraction and repulsion, electrostatic interactions, hydrogen bond networks, solvation effects), and geometric and statistical features (e.g., packing density, binding energy per unit area, number of interfacial residues), totaling K features. These features comprehensively characterize the strength and patterns of intermolecular interactions from both thermodynamic and geometric perspectives. The second category of molecular physical characteristics involves the overall structure and hydrophobic core features of the mutant protein complex. Rosetta software was used to extract refined features from the entire mutant protein complex and its structural regions (such as the core and boundary regions). The main feature categories include hydrophobic core features (such as atomic contact number, average node density, hydrophobic residue count, and hydrophobic energy terms), overall solvent accessibility features (such as the overall solvent-accessible surface area and buried area of ​​the protein), and secondary structure integrity features, totaling M features. These features effectively reflect the tightness of protein folding, the stability of the hydrophobic core, and the conservation of secondary domains, which is of great value for understanding the impact of mutations on structure and function. The third category of molecular physical characteristics is coarse-grained conformational energy features. Using the Rosetta coarse-grained scoring function, a rapid energy assessment of the overall conformation of the mutant protein complex was performed while reducing atomic resolution. This outputs a set of global scalar features reflecting structural compactness, secondary structure integrity, and chain continuity, resulting in a total of P features. K + M + P = N, where N is the dimension of the molecular physical characteristics of the mutant protein complex. By splicing together the various features, a 1*N dimensional molecular physical characteristic feature vector of the mutant protein complex can be obtained.

[0081] To further improve the feature quality of molecular physical properties, some implementations employ zero-filling for features with missing values ​​in the Rosetta software calculations. For example, the three standardized parameters in the Rosetta energy function used to describe the geometry of protein dimer interfaces—shape complementarity score (Sc), interface buried surface area (Area), and interface energy score (dG)—may have missing calculated values. These three features are only validly calculated in mutant complex structures with clearly defined contact interfaces; for monomeric proteins or free binding units, Rosetta calculations result in empty values. Analysis of existing structures shows that the calculated values ​​are mostly negative, reflecting the distribution characteristics of interface geometric features at effective interfaces. To maintain the integrity of the feature matrix and compatibility with subsequent calculations, a zero-filling strategy is uniformly adopted to ensure that the N-dimensional molecular physical property feature vectors of all samples have complete feature representations.

[0082] In some implementations, standardization adjusts the molecular physical properties of each mutant protein complex to a standard normal distribution with a mean of 0 and a standard deviation of 1. Since features from different mutant protein complex structural sources may have different dimensions and numerical ranges, this design makes features from different mutant protein complex structural sources comparable at the same scale, effectively eliminating the impact of feature scale differences on model training. This provides standardized input data for subsequent models, ensuring that the model can fairly learn the contribution of each feature.

[0083] To further improve the quality of the molecular physical properties of mutant protein complexes, this application may also optimize and integrate the initially obtained molecular physical properties according to any of the following feature combinations to obtain the optimized target molecular physical properties, which can then be used as the multimodal features of the mutant protein complex.

[0084] In some preferred embodiments, mutant molecular physical property features and wild-type molecular physical property features corresponding to the mutant protein complex and wild-type protein complex are obtained respectively. Based on the corresponding differences between the various features of the mutant and wild-type molecular physical property features, a target molecular physical property feature with a vector size of 1*N is generated. That is, after obtaining the 1*N-dimensional mutant molecular physical property features and the 1*N-dimensional wild-type molecular physical property features of the mutant protein complex using Rosetta software, the differences of the same type of features in the two sets of molecular physical property features are calculated, thereby obtaining a 1*N-dimensional target molecular physical property feature composed of the differences in each dimension of the feature.

[0085] In some preferred embodiments, target molecular physical property features with a vector size of 3*N are generated based on the mutant molecular physical property features of the mutant protein complex, the molecular physical property features of mutant protein 1, and the molecular physical property features of mutant protein 2. Similarly, 1*N dimensional mutant molecular physical property features corresponding to the mutant protein complex, the 1*N dimensional molecular physical property features corresponding to mutant protein 1, and the 1*N dimensional molecular physical property features corresponding to mutant protein 2 are obtained respectively using Rosetta software and then spliced ​​together to form the 3*N dimensional target molecular physical property features.

[0086] In some preferred embodiments, the corresponding differences between the molecular physical characteristics of the mutant and wild-type proteins, the corresponding differences between the molecular physical characteristics of mutant protein 1 and wild-type protein 1, and the corresponding differences between the molecular physical characteristics of mutant protein 2 and wild-type protein 2 are used to generate target molecular physical characteristics with a vector size of 3*N. Similarly, the feature differences of the three sets of 1*N dimensional molecular physical characteristics are concatenated, and multiple difference processing is used to maximize the extraction of mutational anomaly signals, thereby obtaining 3*N dimensional target molecular physical characteristics.

[0087] Step S122 and step S123 below are two independent extraction steps corresponding to two different types of features. The corresponding multimodal features can be extracted according to the affinity prediction model of different model architectures.

[0088] S123, Based on the three-dimensional structural information of the mutant complex and the three-dimensional structural information of the wild-type protein complex, the residue interaction features of the mutant protein complex and the residue interaction features of the wild-type protein complex are extracted respectively. The residue interaction features include at least one of the energy contribution values ​​of each residue in van der Waals interaction, electrostatic interaction, hydrogen bond network, solvation effect, backbone dihedral angle preference, side chain rotomer preference, hydrophobic interaction or pairwise interaction.

[0089] In this step, residue interaction features of the mutant protein complex can be extracted based on the three-dimensional structural information of the relaxed mutant complex. Residue interaction features of the wild-type protein complex can also be extracted based on the three-dimensional structural information of the original wild-type protein complex.

[0090] It is understandable that the binding affinity and conformational stability of proteins depend not only on the global molecular physics of the overall structure, but also on the finely regulated interactions of each residue itself and its surrounding residues. To fully capture this key information at the local level, this application systematically extracts standard energy terms from the Rosetta scoring function, and combines them with internal interaction energy terms to form a basic feature set describing the conformational stability and energy distribution of mutant protein complexes as residue interaction features. In related technologies, the Rosetta scoring function decomposes various physicochemical contributions in protein structure into multiple interpretable energy terms, each with a clear physical or statistical meaning. During the calculation process, this application outputs energy terms residue-by-residue for the mutant protein complex structure, systematically recording the energy contribution values ​​of each residue in multiple dimensions such as van der Waals interactions, electrostatic interactions, hydrogen bond networks, solvation effects, backbone dihedral angle preference, side chain rotomer preference, hydrophobic interactions, and pairwise interactions. Specifically, van der Waals interaction terms reflect the spatial packing density of residues, while electrostatic interaction terms characterize the attraction and repulsion between charged residues; hydrogen bond network terms record the contribution of hydrogen bonds formed between residues and between residues and the solvent to stability; solvation effect terms reflect the energy cost of residues transferring from the hydrophobic core to the solvent environment; and backbone dihedral preference terms assess the rationality of the current main chain conformation based on statistical distributions in known protein structure databases. These energy terms collectively characterize the physicochemical state of residues in the three-dimensional structure from different perspectives.

[0091] As can be seen from the multimodal feature extraction strategies in S122 and S123 above, the multi-granularity feature extraction strategy can not only grasp the overall energy distribution and stability characteristics of the mutant protein complex from a global perspective, but also capture the sensitivity differences of key residues to mutation effects from a local perspective, providing rich multi-level information for affinity prediction models.

[0092] S130, Train the preset affinity prediction model based on training data and multimodal features to obtain a well-trained target affinity prediction model.

[0093] Specifically, a preset affinity prediction model can be trained based on training data and at least one feature from multimodal features to obtain a target affinity prediction model.

[0094] In some implementations, the target affinity prediction model includes a first prediction model and / or a second prediction model. The first prediction model is trained based on the three-dimensional structural information of the mutant complex, the training tag, and the molecular physical characteristics of the mutant protein complex. The second prediction model is trained based on the three-dimensional structural information of the mutant complex, the training tag, the residue interaction characteristics of the mutant protein complex, and the residue interaction characteristics of the wild-type protein complex.

[0095] In this step, the first prediction model and the second prediction model can be two independent models with the same or different model architectures. Both the first and second prediction models can predict the same mutant protein complex based on the same training data and their respective corresponding multimodal features, obtaining the predicted values ​​of the corresponding relative affinity. It can be understood that when the first and second prediction models predict the same mutant protein complex, they can obtain different predicted values ​​of relative affinity. Furthermore, the target predicted value can be the average of the predicted values ​​from the first and second prediction models, or either the first or second prediction model can be selected as the target predicted value.

[0096] The following combination Figure 2 The training process of the first prediction model is introduced.

[0097] S131A: Input the three-dimensional structural information of the mutant complex into the protein message passing neural network, obtain the feature difference tensor output by the encoder of the protein message passing neural network, and fuse the feature difference tensor with the molecular physical property features of the mutant protein complex to generate the first residual correction feature.

[0098] The input data for the first prediction model of this application includes the three-dimensional structural information of the mutant protein complex. Based on the three-dimensional structural information of the mutant complex, the three-dimensional structural information of mutant protein 1 and mutant protein 2 can be extracted. The feature calculation module can call Rosetta software to obtain the molecular physical property features of the corresponding mutant protein complex based on the input data. Each three-dimensional structural information can be represented using a PDB format file. The molecular physical property features can be one set of target molecular physical property features optimized according to the above feature combination method.

[0099] The first prediction model of this application can employ the encoder of the ProteinMPNN model (Protein Message Passing Neural Network). The encoder of the ProteinMPNN model outputs an initial prediction value of relative affinity through a task head connection. The initial prediction value is then corrected by the residual deviation output of a fusion layer used to fuse molecular physical property features to obtain the final prediction value. In other words, the first prediction model of this application does not include a decoder for the ProteinMPNN model. It can be understood that the ProteinMPNN model automatically encodes the corresponding embedding features, node features, and edge features based on the PDB format file corresponding to the input mutant complex's three-dimensional structure information, and then outputs the corresponding feature difference tensors (the feature differences between wt_hv and mut_hv). Here, wt_hv and mut_hv are high-dimensional tensors generated by the ProteinMPNN encoder during computation; wt_hv is the encoding result corresponding to the wild-type protein complex, and mut_hv corresponds to the encoding result of the mutant protein complex. The ProteinMPNN model predicts the 3D structure of the wild-type complex based on the input 3D structure information of the mutant complex and its corresponding amino acid sequence, generating wt_hv and mut_hv respectively. wt_hv and mut_hv each encode a high-dimensional tensor representing the complex's embedding, node, and edge features. Differences are calculated between corresponding features in these two high-dimensional tensors to form a tensor composed of embedding, node, and edge differences. This feature difference tensor, along with the molecular physical properties of the mutant protein complex, is input into the fusion layer. Specifically, the embedding difference tensor represents the global representation difference of the protein before and after mutation; the node difference feature tensor represents changes in the local chemical environment of each amino acid residue; and the edge difference feature tensor represents changes in the spatial relationships between residue pairs. Furthermore, the feature difference tensor and the molecular physical properties of the mutant protein complex are fused through the fusion layer. In the fusion layer, the corresponding residual offset is output after deep fusion through mutation-aware position-weighted pooling, linear projection layer, gated / Cross-Attention layer, MLP layer, etc., and this residual offset is the first residual correction feature.Specifically, in the fusion layer, regarding mutation-aware location weighting, the minimum distance from each node location to the mutation site is calculated through block broadcasting, the distance weight is calculated using a Gaussian decay function, and then normalized after multiplying with the feature importance weight to obtain the final weight of each node location. Regarding cross-modal projection, for example, three independent linear projection layers are set up to map the three types of signals in the aforementioned feature difference tensor to a unified hidden dimension and expand them into unit-length sequences. Regarding multi-layer cross-attention, node projections and edge projections are concatenated into a key-value sequence, and two stacked multi-head cross-attention layers are applied with the embedded projection as the query. Each layer contains a multi-head attention sublayer and a feedforward sublayer, and each has residual connections and layer normalization. Regarding gated residual fusion, a gating network adaptively determines whether information in each dimension is transmitted through the main transform path or a skip path. Finally, the fusion result of the output projection is compressed into a scalar prediction value, i.e., the residual offset.

[0100] S132A, based on the three-dimensional structural information of mutant protein 1 and mutant protein 2, the protein message passing neural network outputs the corresponding first affinity prediction value of mutant protein 1 and the second affinity prediction value of mutant protein 2, respectively, and based on the three-dimensional structural information of the mutant complex, the protein message passing neural network outputs the corresponding first complex affinity prediction value.

[0101] The three-dimensional structural information of mutant protein 1 and mutant protein 2 can be directly extracted from the corresponding three-dimensional structural information of the mutant complex.

[0102] See also Figure 2 The first prediction model comprises three sets of independent and parallel ProteinMPNN model encoders. Each ProteinMPNN model has a corresponding input layer and a hidden layer for the encoder. The three encoders are connected to corresponding task heads to output corresponding initial prediction values. Based on the three-dimensional structure information of the input mutant protein 1, the task head connected to one ProteinMPNN model in S131A outputs the corresponding first affinity prediction value Binder1_ddG for mutant protein 1. Based on the three-dimensional structure information of the input mutant protein 2, the task head connected to another ProteinMPNN model outputs the corresponding second affinity prediction value Binder2_ddG for mutant protein 2. Based on the three-dimensional structure information of the input mutant complex, the task head connected to yet another ProteinMPNN model outputs the corresponding first complex affinity prediction value Complex ddG. The first affinity prediction value, the second affinity prediction value, and the first complex affinity prediction value are all initial prediction values ​​of relative affinity. Based on different input information, the task heads connected to each ProteinMPNN model independently output their corresponding initial prediction values.

[0103] S133A calculates the affinity change value based on the first affinity prediction value, the second affinity prediction value, and the first complex affinity prediction value, and uses the first residual correction feature to perform a fusion correction on the affinity change value, outputting the predicted value of the relative affinity of the mutant protein complex.

[0104] After each ProteinMPNN model outputs the initial predicted values ​​of the three relative affinities, this step combines them into a binding affinity change value based on the thermodynamic cycle: ddG = Binder1_ddG + Binder2_ddG Complex_ddG is the sum of the stability changes of the two individual proteins caused by mutations, minus the overall stability change of the mutant protein complex, thus obtaining the effect of mutation on the binding free energy of the protein complex.

[0105] Furthermore, based on the first residual correction feature, the binding affinity change value ddG is corrected through a fully connected layer, that is, the first residual correction feature is added to the binding affinity change value ddG to form fusion ddG as the predicted value of relative affinity of the mutant protein complex as the final output of the first prediction model.

[0106] S134A calculates the loss based on the predicted value of the relative affinity of the mutant protein complex and the training labels, and then backpropagates to obtain the first well-trained prediction model.

[0107] In this step, the predicted relative affinity (fusionddG) of the mutant protein complex generated by S133A and the experimental relative affinity in the training labels are used to calculate the loss. The obtained loss value is then backpropagated to iterate the model parameters multiple times to obtain the trained first prediction model.

[0108] The above describes the training process of the first prediction model after incorporating multimodal features, specifically molecular physical property features. The loss is calculated and backpropagated based on the predicted relative affinity value and the experimental (true) relative affinity value in the training labels. Simultaneously, according to relevant techniques, the weight parameters in the hidden layers and fusion layers of multiple ProteinMPNN model encoders in the first prediction model are tuned, resulting in a well-trained first prediction model. This allows the model to learn the residual offset between the predicted value (fusion ddG) and the true value. By using residual connections, the direct regression task is transformed into a residual fitting task, thereby reducing the learning difficulty of the model, enabling more stable convergence during training, and effectively improving the final prediction accuracy.

[0109] In this embodiment, by integrating the three-dimensional structural information of the mutant protein complex and global molecular physical property data, deep fusion and information complementarity between different modalities are achieved by mapping data from different modalities to a shared embedding space within a unified feature representation space. Through the above feature fusion strategy, the affinity prediction model can fully explore the synergistic information between various feature data, overcome the limitations of incomplete single feature information, and thus significantly improve the model's generalization ability and prediction stability in diverse application scenarios.

[0110] The following combination Figure 3 The training process of the second prediction model is introduced.

[0111] S131B, input the three-dimensional structural information of the mutant complex, the three-dimensional structural information of mutant protein 1 and mutant protein 2 into the protein message passing neural network respectively, and obtain the first embedding vector and the second complex affinity prediction value, the second embedding vector and the third affinity prediction value, and the third embedding vector and the fourth affinity prediction value output by the encoder of the protein message passing neural network.

[0112] The input data in this embodiment includes the three-dimensional structural information of the mutant complex, the three-dimensional structural information of mutant protein 1, and the three-dimensional structural information of mutant protein 2. Each three-dimensional structural information can be in PDB format. The feature calculation module can call Rosetta software to obtain the corresponding residue interaction features of the mutant protein complex, mutant protein 1, mutant protein 2, wild-type protein complex, wild-type protein 1, and wild-type protein 2 based on the input data.

[0113] The second prediction model in this embodiment can be considered an extension of the first prediction model architecture. Specifically, the second prediction model also includes three sets of parallel ProteinMPNN model encoders. Each encoder is independently connected to its corresponding task head and outputs an initial prediction value of the corresponding relative affinity based on different input data. Each ProteinMPNN model encoder is connected to a corresponding feature fusion layer to output corresponding residual correction features, and the initial relative affinity is corrected based on the residual correction features. It can be understood that the training objective of the second prediction model is to iteratively obtain the weight parameters of the encoder hidden layers and feature fusion layers of each set of ProteinMPNN models.

[0114] See Figure 3The encoder of the first set of ProteinMPNN models can extract the corresponding first embedding vector based on the three-dimensional structure information of the mutant complex and output the second complex affinity prediction value (Complex ddG based) through the first task head. Similarly, the encoder of the second set of ProteinMPNN models can extract the second embedding vector based on the three-dimensional structure information of mutant protein 1 and output the third affinity prediction value (Binder1 ddG based) through the second task head. The encoder of the third set of ProteinMPNN models can extract the third embedding vector based on the three-dimensional structure information of mutant protein 2 and output the fourth affinity prediction value (Binder2 ddG based) through the third task head.

[0115] In some implementations, the feature fusion layers connected to the second and third sets of ProteinMPNN models share the same network parameters. By making the feature fusion layers of the two sets of ProteinMPNN models share the same network parameters, not only is the number of model parameters effectively reduced and the risk of overfitting lowered, but the model also ensures consistent processing of the two sets of binding units (wild-type and mutant individual proteins) at the feature extraction level. This better captures the symmetric relationship between the two in their interaction, laying a balanced and unified feature foundation for subsequent interaction modeling.

[0116] S132B, the residue interaction features of the mutant protein complex and the wild-type protein complex are fused with the first embedding vector to generate the second residual correction feature; the residue interaction features of mutant protein 1 and wild-type protein 1 are fused with the second embedding vector to generate the third residual correction feature; the residue interaction features of mutant protein 2 and wild-type protein 2 are fused with the third embedding vector to generate the fourth residual correction feature.

[0117] Furthermore, through the first feature fusion layer, the residue interaction features of the mutant protein complex and the wild-type protein complex are fused with the first embedding vector to generate the corresponding residual correction value, namely the second residual correction feature Complex ddG residual. Through the second feature fusion layer, the residue interaction features of mutant protein 1 and wild-type protein 1 are fused with the second embedding vector to generate the third residual correction feature Binder1 ddGresidual. Through the third feature fusion layer, the residue interaction features of mutant protein 2 and wild-type protein 2 are fused with the third embedding vector to generate the fourth residual correction feature Binder2 ddGresidual.

[0118] The feature fusion layer in this embodiment is basically the same in principle as the fusion layer in the above embodiments, except that the feature fusion layer in this embodiment also includes a mask layer. In each feature fusion layer, the mask layer is set between the linear projection layer and the gated / Cross-Attention layer, and is used to explicitly mark mutation sites, so that the model focuses on the structural change region caused by the mutation. The mutation-aware location-weighted pooling is transformed into a distance-attenuation soft masking mechanism in mutation-aware location-weighting. A Gaussian decay function is used to apply higher weights to the positions within the window near the mutation site, while the weights outside the window approach zero.

[0119] S133B outputs the predicted value of the relative affinity of the mutant protein complex based on the predicted value of the second complex affinity corrected by the second residual correction feature, the predicted value of the third affinity corrected by the third residual correction feature, and the predicted value of the fourth affinity corrected by the fourth residual correction feature.

[0120] In this step, the second residual correction feature is added to the predicted affinity of the second complex to obtain the corrected predicted affinity of the second complex. Similarly, the third residual correction feature is added to the predicted affinity of the third complex to obtain the corrected predicted affinity of the third complex. Finally, the fourth residual correction feature is added to the predicted affinity of the fourth complex to obtain the corrected predicted affinity of the fourth complex.

[0121] Similarly, based on thermodynamic cycles, after obtaining the above three corrected affinity prediction values, these three are combined into the final predicted value of the relative affinity of the mutant protein complex output by the second prediction model: ddG = corrected Binder1 ddG based + corrected Binder2 ddG based The revised Complex ddG based.

[0122] S134B calculates the loss based on the predicted value of the relative affinity of the mutant protein complex and the training labels, and then backpropagates to obtain a trained second prediction model.

[0123] In this step, the loss can be calculated based on the experimental value of the relative affinity between ddG and the training label obtained in step S133B, and the model parameters can be backpropagated multiple times based on the loss to obtain the trained second prediction model.

[0124] The above describes the training process of the second prediction model after incorporating multimodal features, namely residue interaction features. The loss is calculated and backpropagated based on the predicted relative affinity values ​​from the final output and the experimental (true) relative affinity values ​​from the training labels. Using relevant techniques, the weight parameters of the encoder hidden layer and feature fusion layer of multiple ProteinMPNN models for the second prediction model are tuned, resulting in a well-trained second prediction model.

[0125] In this embodiment, by integrating the three-dimensional structural information and local residue interaction features of the mutant protein complex into a unified feature representation space, data from different modalities are mapped to a shared embedding space for fusion, achieving deep fusion and information complementarity between features of different modalities. Through the above feature fusion strategy, the affinity prediction model can fully explore the synergistic information between various feature data, overcome the limitations of incomplete information from a single feature, and thus significantly improve the model's generalization ability and prediction stability in diverse application scenarios.

[0126] As this example demonstrates, the training method of this application, by constructing a multimodal feature dataset, systematically integrates multi-dimensional information such as the three-dimensional structure and physicochemical properties of protein complexes, effectively overcoming the inherent limitations of single structural data modalities in terms of sample size and diversity. Based on this, this application further proposes a deep fusion strategy for multimodal features, which integrates the three-dimensional structural information of wild-type / mutant protein complexes with molecular physical characteristics calculated using Rosetta software, including physical energy terms, geometric complementarity, and hydrophobic core integrity, as well as residue interaction characteristics of interacting protein regions, such as the energy contribution of residues themselves and interactions between residues, performing corresponding deep feature fusion. Through the synergistic utilization and deep fusion of the aforementioned multimodal data, this strategy can fully explore and leverage the complementary advantages between different data modalities, thereby significantly improving the prediction accuracy of affinity prediction models for changes in protein binding affinity in training scenarios with only a small amount of mutation data, and enhancing the generalization ability of affinity prediction models in different types of protein systems. The trained model in this application provides computational support for designing mutant proteins by predicting the difference in free energy changes (i.e., ΔΔG or ddG) when mutant protein complexes bind, thereby further improving the accuracy of antibody screening containing mutant proteins and its ability to be applied in practical scenarios.

[0127] To verify the effectiveness of the multimodal features used in this application, the first prediction model (Ours_global) that integrates molecular physical property features was systematically compared with several state-of-the-art methods on the same test and validation sets. The selected comparison methods included the energy function-based FoldX model, the flexible backbone-based FlexddG model, and the deep learning-based StaddG model. The comparison results on the validation and test sets are shown in Tables 1 and 2 below.

[0128] Table 1 Results of the first prediction model on the validation set

[0129] Table 2 Results of the first prediction model on the test set

[0130] Similarly, to evaluate the impact of the fusion of residue interaction features proposed in this application on the performance of mutant affinity prediction, comparative experiments were conducted on the validation and test sets. Ours represents the basic model using only the three-dimensional structural features of the mutant protein complex; Ours_local introduces residue interaction features on top of the basic model. StabddG was used as the representative baseline for comparison. The experimental results are shown in Tables 3 and 4.

[0131] Table 3 Results of the second prediction model on the validation set

[0132] Table 4 Results of the second prediction model on the test set

[0133] The experimental results above demonstrate that the affinity prediction model obtained by the training method proposed in this application exhibits a significant improvement in prediction performance across the Spearman correlation coefficient, Pearson correlation coefficient, and RMSE. This performance advantage fully validates the effectiveness and advancement of the training data and multimodal feature fusion strategy constructed in this application for the mutation affinity prediction task, providing a high-precision computational prediction tool for the fields of protein engineering and antibody drug development.

[0134] See Figure 4 An embodiment of this application illustrates a method for predicting the affinity of a mutant protein complex, comprising: S210, Obtain the information to be predicted, which includes the target mutant three-dimensional structure information of the target mutant protein complex, or the target wild-type three-dimensional structure information and target mutation information of the target wild-type protein complex.

[0135] In this application, the information to be predicted can be the three-dimensional structural information of the target mutant protein complex that has already undergone mutation. Alternatively, the information to be predicted can be the three-dimensional structural information of the target wild-type protein complex that has not undergone mutation, along with the target mutation information, and then the corresponding target mutant protein complex can be obtained through mutation using relevant techniques. The target mutation information can include mutations at one or more amino acid sites.

[0136] It is understandable that both the target mutant 3D structural information and the target wild-type 3D structural information are stored in PDB format files, and all data in non-ATOM record rows are deleted from the PDB format files.

[0137] S220, Based on the information to be predicted, generate corresponding multimodal features; the multimodal features include the molecular physical properties of the target mutant protein complex, and / or the residue interaction features of the target mutant protein complex and the residue interaction features of the corresponding target wild-type protein complex.

[0138] Based on the Rosetta software in related technologies, the molecular physical characteristics of the target mutant protein complex are extracted according to the three-dimensional structural information of the target mutant; or, the residue interaction characteristics of the target mutant protein complex and the residue interaction characteristics of the target wild-type protein complex are extracted respectively according to the three-dimensional structural information of the target mutant and the target wild-type protein complex.

[0139] S230, based on the information to be predicted and multimodal features, generates the relative affinity of the target mutant protein complex through a trained target affinity prediction model.

[0140] The trained affinity prediction model includes a first prediction model and / or a second prediction model. The first prediction model generates a first predicted value for the relative affinity of the target mutant protein complex based on the three-dimensional structural information of the target mutant and the molecular physical characteristics of the target mutant protein complex. The second prediction model generates a second predicted value for the relative affinity of the target mutant protein complex based on the three-dimensional structural information of the target mutant, the residue interaction characteristics of the target mutant protein complex, and the residue interaction characteristics of the target wild-type protein complex. Users can choose either the first or second predicted value as the final predicted value for the target mutant protein complex, or use the average of the two as the final predicted value.

[0141] As this example demonstrates, the affinity prediction method of this application, through precise prediction of changes in binding affinity caused by mutations in single or multiple amino acids in protein complexes, provides robust computational support for key aspects of antibody drug development, such as affinity maturation and exploitability optimization. This enhanced predictive capability facilitates the rapid screening of candidate molecules with higher binding affinity and superior druggability in the early stages of development, thereby effectively reducing experimental screening costs, shortening the development cycle, and ultimately improving the overall success rate of protein antibody drug development.

[0142] Corresponding to the aforementioned application function implementation method embodiments, this application also provides a training device, electronic device, and corresponding embodiments for a protein complex affinity prediction model.

[0143] Figure 5 This is a schematic diagram of the structure of the training device for the protein complex affinity prediction model shown in the embodiments of this application.

[0144] See Figure 5 An embodiment of this application illustrates a training apparatus for a protein complex affinity prediction model, comprising a data acquisition module 310, a feature extraction module 320, and a model training module 330. Wherein: The data acquisition module 310 is used to acquire training data, which includes the three-dimensional structural information of the mutant protein complex and training labels. At least some of the mutant protein complexes are obtained by mutation of the corresponding wild-type protein complexes. The training labels include the relative affinity of the mutant protein complexes.

[0145] The feature extraction module 320 is used to obtain multimodal features of the mutant protein complex based on the three-dimensional structural information of the mutant complex. The multimodal features include the molecular physical properties of the mutant protein complex, and / or the residue interaction features of the mutant protein complex and the residue interaction features of the corresponding wild-type protein complex.

[0146] The model training module 330 is used to train the preset affinity prediction model based on training data and multimodal features to obtain a trained target affinity prediction model.

[0147] The target affinity prediction model includes a first prediction model and / or a second prediction model. The first prediction model is trained based on the three-dimensional structural information of the mutant complex, the training tag, and the molecular physical characteristics of the mutant protein complex. The second prediction model is trained based on the three-dimensional structural information of the mutant complex, the training tag, the residue interaction characteristics of the mutant protein complex, and the residue interaction characteristics of the wild-type protein complex.

[0148] See Figure 6 In some embodiments, the data acquisition module 310 includes a data clustering module 312, which is used to divide the training data into multiple clusters. The mutant protein complexes in each cluster are clustered based on structural similarity, interface region confidence, and / or sequence homology. The data acquisition module 310 also includes a data partitioning module 313, which is used to divide the training data into a training set, a validation set, and a test set. The training set, validation set, and test set have distinct clusters, and each cluster contains distinct mutant protein complexes.

[0149] In some embodiments, the data acquisition module 310 further includes a protein mutation module 311, which is used to acquire the three-dimensional structural information and mutation information of the wild-type protein complex; wherein, the three-dimensional structural information of the wild-type complex includes the three-dimensional coordinate information of the heavy atoms corresponding to each residue, the category of the residue is natural amino acid, and the mutation information includes at least one residue substitution information; at least one residue of the wild-type protein complex is substituted according to the mutation information to obtain the three-dimensional structural information of the mutant protein complex.

[0150] In some embodiments, the feature extraction module 320 includes a structure relaxation module 321, a first extraction module 322, and a second extraction module 323. The structure relaxation module 321 is used to perform structural relaxation on the mutant protein complex to obtain the mutant complex conformation with the lowest energy; the three-dimensional structural information of the mutant complex is the three-dimensional structural information of the mutant complex conformation with the lowest energy. The first extraction module 322 is used to extract the molecular physical property features of the mutant protein complex based on the mutant complex conformation. The molecular physical property features include at least one of the following: interface analysis features of the mutant complex, overall structure and hydrophobic core features, or energy features of the coarse-grained conformation. The second extraction module 323 is used to extract the residue interaction features of the mutant protein complex and the wild-type protein complex based on the three-dimensional structural information of the mutant complex and the three-dimensional structural information of the wild-type protein complex, respectively. The residue interaction features include at least one of the energy contribution values ​​of each residue in van der Waals interactions, electrostatic interactions, hydrogen bond networks, solvation effects, backbone dihedral angle preference, side chain rotomer preference, hydrophobic interactions, or pairwise interactions.

[0151] In some specific embodiments, the first extraction module 322 is used to obtain the mutant molecular physical property features and wild-type molecular physical property features corresponding to the mutant protein complex and the wild-type protein complex, respectively; and generate target molecular physical property features with a vector size of 1*N based on the corresponding differences of each feature of the mutant molecular physical property features and the wild-type molecular physical property features; where N is the feature dimension of the molecular physical property features; or the first extraction module 322 is used to generate target molecular physical property features with a vector size of 3*N based on the mutant molecular physical property features of the mutant protein complex, the molecular physical property features of mutant protein 1, and the molecular physical property features of mutant protein 2; or the first extraction module 322 is used to generate target molecular physical property features with a vector size of 3*N based on the corresponding differences of each feature of the mutant molecular physical property features and the wild-type molecular physical property features, the corresponding differences of each feature of the molecular physical property features of mutant protein 1 and wild-type protein 1, and the corresponding differences of each feature of the molecular physical property features of mutant protein 2 and wild-type protein 2.

[0152] In some specific implementations, the model training module 330 includes a first training module 331. The first training module 331 is used to input the three-dimensional structural information of the mutant complex into a protein message passing neural network, obtain the feature difference tensor output by the encoder of the protein message passing neural network, and fuse the feature difference tensor with the molecular physical characteristics of the mutant protein complex to generate a first residual correction feature. Based on the three-dimensional structural information of mutant protein 1 and mutant protein 2, the protein message passing neural network outputs the corresponding first affinity prediction value of mutant protein 1 and the second affinity prediction value of mutant protein 2, respectively, and based on the three-dimensional structural information of the mutant complex, the protein message passing neural network outputs the corresponding first complex affinity prediction value. Based on the first affinity prediction value, the second affinity prediction value, and the first complex affinity prediction value, the affinity change value is calculated, and the affinity change value is fused and corrected using the first residual correction feature to output the relative affinity prediction value of the mutant protein complex. Based on the relative affinity prediction value of the mutant protein complex and the training labels, the loss is calculated and backpropagated to obtain the trained first prediction model.

[0153] In some specific implementations, the model training module 330 includes a second training module 332. The second training module 332 is used to input the three-dimensional structural information of the mutant complex, the three-dimensional structural information of mutant protein 1, and the three-dimensional structural information of mutant protein 2 into a protein message passing neural network, respectively, to obtain the first embedding vector and the second complex affinity prediction value, the second embedding vector and the third affinity prediction value, and the third embedding vector and the fourth affinity prediction value output by the encoder of the protein message passing neural network; to fuse the residue interaction features of the mutant protein complex and the residue interaction features of the wild-type protein complex with the second embedding vector to generate a second residual correction feature; and to input the residue interaction features of mutant protein 1... The interaction features of the mutant protein 1 and the residue interaction features of wild-type protein 1 are fused with the third embedding vector to generate the third residual correction feature; the residue interaction features of the mutant protein 2 and the residue interaction features of wild-type protein 2 are fused with the fourth embedding vector to generate the fourth residual correction feature; based on the predicted affinity of the second complex corrected by the second residual correction feature, the predicted affinity of the third complex corrected by the third residual correction feature, and the predicted affinity of the fourth complex corrected by the fourth residual correction feature, the predicted affinity of the mutant protein complex is output; based on the predicted affinity of the mutant protein complex and the training labels, the loss is calculated and backpropagated to obtain the trained second prediction model.

[0154] In some embodiments, the model training module 330 of this application further includes a pre-training module 333, and the data acquisition module 310 is further used to acquire pre-training data. The pre-training data includes the three-dimensional structural information of the mutant protein complex and pre-training labels, where the pre-training labels are exploitability data. The pre-training module 333 is used to pre-train the preset affinity prediction model based on the pre-training data to obtain a pre-trained preset affinity prediction model.

[0155] In some embodiments, the exploitability data include at least one of the following: aggregation tendency of mutant protein complexes and / or mutant monomeric proteins, thermal stability, pharmacokinetic properties, solubility, and chemical stability.

[0156] As can be seen from this example, the training device of this application can train the affinity prediction model based on high-quality training data and multimodal features extracted from it through feature fusion, which can significantly improve the accuracy and generalization of affinity prediction.

[0157] See Figure 7 An embodiment of this application also provides an affinity prediction device for mutant protein complexes, comprising: an information acquisition module 410, a feature encoding module 420, and an affinity prediction module 430, wherein: The information acquisition module 410 is used to acquire the information to be predicted, which includes the target mutant three-dimensional structure information of the target mutant protein complex, or the target wild-type three-dimensional structure information and target mutation information of the target wild-type protein complex.

[0158] The feature encoding module 420 is used to generate corresponding multimodal features based on the information to be predicted; the multimodal features include the molecular physical properties of the target mutant protein complex, and / or the residue interaction features of the target mutant protein complex and the residue interaction features of the corresponding target wild-type protein complex.

[0159] The affinity prediction module 430 is used to generate the relative affinity of the target mutant protein complex based on the information to be predicted and multimodal features, through a trained target affinity prediction model.

[0160] In some implementations, the target affinity prediction model includes a first prediction model, and the affinity prediction module 430 is used to generate a first predicted value of the relative affinity of the target mutant protein complex based on the information to be predicted and the molecular physical characteristics of the target mutant protein complex through the first prediction model.

[0161] In some embodiments, the target affinity prediction model includes a second prediction model. The affinity prediction module 430 is used to generate a second predicted value of the relative affinity of the target mutant protein complex based on the information to be predicted and the residue interaction characteristics of the target mutant protein complex and the residue interaction characteristics of the target wild-type protein complex.

[0162] In some embodiments, the target affinity prediction model includes a first prediction model and a second prediction model. The affinity prediction module 430 is used to generate a first predicted value of the relative affinity corresponding to the target mutant protein complex through the first prediction model based on the information to be predicted and the molecular physical characteristics of the target mutant protein complex; generate a second predicted value of the relative affinity corresponding to the target mutant protein complex through the second prediction model based on the information to be predicted, the residue interaction characteristics of the target mutant protein complex and the residue interaction characteristics of the target wild-type protein complex; calculate the average value of the first prediction value and the second prediction value, and use the average value as the relative affinity corresponding to the target mutant protein complex.

[0163] The affinity prediction device of this application, based on a trained affinity prediction model, can improve the prediction accuracy of the relative affinity of mutant protein complexes and provide computational support for the design of mutant proteins.

[0164] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated further here.

[0165] Figure 8 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application.

[0166] See Figure 8 The electronic device 1000 includes a memory 1010 and a processor 1020.

[0167] The processor 1020 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. Memory 1010 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. ROM may store static data or instructions required by processor 1020 or other modules of the computer. Permanent storage devices may be read-write storage devices. Permanent storage devices may be non-volatile storage devices that retain stored instructions and data even when the computer is powered off. In some embodiments, permanent storage devices use mass storage devices (e.g., magnetic or optical disks, flash memory) as permanent storage devices. In other embodiments, permanent storage devices may be removable storage devices (e.g., floppy disks, optical drives). System memory may be a read-write storage device or a volatile read-write storage device, such as dynamic random access memory. System memory may store some or all of the instructions and data required by the processor during operation. Furthermore, memory 1010 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (e.g., DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and / or optical disks may also be used. In some embodiments, memory 1010 may include a removable storage device that is readable and / or writable, such as a laser disc (CD), a read-only digital multifunction optical disc (e.g., DVD-ROM, dual-layer DVD-ROM), a read-only Blu-ray disc, an ultra-high density optical disc, a flash memory card (e.g., SD card, mini SD card, Micro-SD card, etc.), a magnetic floppy disk, etc. Computer-readable storage media do not contain carrier waves or transient electronic signals transmitted wirelessly or via wired connections.

[0168] The memory 1010 stores executable code, which, when processed by the processor 1020, can cause the processor 1020 to execute part or all of the methods described above.

[0169] Furthermore, the method according to this application can also be implemented as a computer program or computer program product, which includes computer program code instructions for performing some or all of the steps in the method described above.

[0170] Alternatively, this application may be implemented as a computer-readable storage medium (or a non-transitory machine-readable storage medium or a machine-readable storage medium) storing executable code (or computer program or computer instruction code) that, when executed by a processor of an electronic device (or server, etc.), causes the processor to perform part or all of the steps of the methods described above according to this application.

[0171] This application also provides a computer program product, which includes computer instructions that, when executed by a processor, implement the method described above.

[0172] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for training a protein complex affinity prediction model, characterized in that, include: Acquire training data, which includes the three-dimensional structural information of the mutant protein complex and training tags, wherein at least a portion of the mutant protein complex is obtained by mutation of the corresponding wild-type protein complex, and the training tags include the relative affinity of the mutant protein complex. Based on the three-dimensional structural information of the mutant complex, the multimodal features of the mutant protein complex are obtained, wherein the multimodal features include the molecular physical properties of the mutant protein complex, and / or the residue interaction features of the mutant protein complex and the corresponding residue interaction features of the wild-type protein complex. The preset affinity prediction model is trained based on the training data and the multimodal features to obtain a trained target affinity prediction model.

2. The training method according to claim 1, characterized in that, The target affinity prediction model includes a first prediction model and / or a second prediction model. The first prediction model is trained based on the three-dimensional structure information of the mutant complex, the training tag, and the molecular physical characteristics of the mutant protein complex. The second prediction model is trained based on the three-dimensional structure information of the mutant complex, the training tag, the residue interaction characteristics of the mutant protein complex, and the residue interaction characteristics of the wild-type protein complex.

3. The training method according to claim 1, characterized in that, Methods for obtaining mutant protein complexes by mutation of wild-type protein complexes include: Obtain the three-dimensional structure information and mutation information of the wild-type protein complex; wherein, the three-dimensional structure information of the wild-type complex includes the three-dimensional coordinate information of the heavy atoms corresponding to each residue, the category of the residue is natural amino acid, and the mutation information includes at least one residue substitution information; Based on the mutation information, at least one residue of the wild-type protein complex is replaced to obtain the mutant protein complex and the corresponding three-dimensional structural information of the mutant complex.

4. The training method according to claim 1, characterized in that, The acquisition of the multimodal features of the mutant protein complex includes: The mutant protein complex is subjected to structural relaxation to obtain the mutant complex conformation with the lowest energy; the three-dimensional structural information of the mutant complex is the three-dimensional structural information of the mutant complex conformation with the lowest energy. Based on the conformation of the mutant complex, the molecular physical characteristics of the mutant protein complex are extracted, including at least one of the following: interface analysis characteristics of the mutant complex, overall structure and hydrophobic core characteristics, or energy characteristics of the coarse-grained conformation; and / or Based on the three-dimensional structural information of the mutant complex and the three-dimensional structural information of the wild-type protein complex, residue interaction features of the mutant protein complex and the wild-type protein complex are extracted respectively. The residue interaction features include at least one of the following: van der Waals interaction, electrostatic interaction, hydrogen bond network, solvation effect, backbone dihedral angle preference, side chain rotomer preference, hydrophobic interaction, or energy contribution value of pairwise interaction for each residue.

5. The training method according to any one of claims 1 to 4, characterized in that, When the multimodal features include the molecular physical property features, the step of obtaining the multimodal features of the mutant protein complex further includes: Obtain the mutant molecular physical property features and wild-type molecular physical property features corresponding to the mutant protein complex and the wild-type protein complex, respectively; based on the corresponding differences between the mutant molecular physical property features and the wild-type molecular physical property features, generate target molecular physical property features with a vector size of 1*N; where N is the feature dimension of the molecular physical property features; or Based on the mutant molecular physical characteristics of the mutant protein complex, the molecular physical characteristics of mutant protein 1, and the molecular physical characteristics of mutant protein 2, a target molecular physical characteristic vector with a size of 3*N is generated; or The corresponding differences between the molecular physical characteristics of mutant and wild-type proteins, the corresponding differences between the molecular physical characteristics of mutant protein 1 and wild-type protein 1, and the corresponding differences between the molecular physical characteristics of mutant protein 2 and wild-type protein 2 are used to generate target molecular physical characteristics with a vector size of 3*N.

6. The training method according to claim 2, characterized in that, The training process of the first prediction model includes: The three-dimensional structural information of the mutant complex is input into a protein message passing neural network to obtain the feature difference tensor output by the encoder of the protein message passing neural network. The feature difference tensor is then fused with the molecular physical property features of the mutant protein complex to generate a first residual correction feature. Based on the three-dimensional structural information of mutant protein 1 and mutant protein 2, the protein message passing neural network outputs the corresponding first affinity prediction value of mutant protein 1 and the second affinity prediction value of mutant protein 2, respectively; and based on the three-dimensional structural information of the mutant complex, the protein message passing neural network outputs the corresponding first complex affinity prediction value. Based on the first affinity prediction value, the second affinity prediction value, and the first complex affinity prediction value, the affinity change value is calculated, and the first residual correction feature is used to perform a fusion correction on the affinity change value, and the predicted value of the relative affinity of the mutant protein complex is output. The loss is calculated and backpropagated based on the predicted relative affinity of the mutant protein complex and the training labels to obtain the trained first prediction model.

7. The training method according to claim 2, characterized in that, The training process of the second prediction model includes: The three-dimensional structural information of the mutant complex, the three-dimensional structural information of mutant protein 1, and the three-dimensional structural information of mutant protein 2 are respectively input into the protein message passing neural network to obtain the first embedding vector and the second complex affinity prediction value, the second embedding vector and the third affinity prediction value, and the third embedding vector and the fourth affinity prediction value output by the encoder of the protein message passing neural network. The residue interaction features of the mutant protein complex and the wild-type protein complex are fused with the first embedding vector to generate a second residual correction feature; the residue interaction features of the mutant protein 1 and the wild-type protein 1 are fused with the second embedding vector to generate a third residual correction feature; the residue interaction features of the mutant protein 2 and the wild-type protein 2 are fused with the third embedding vector to generate a fourth residual correction feature. Based on the predicted affinity of the second complex corrected by the second residual correction feature, the predicted affinity of the third complex corrected by the third residual correction feature, and the predicted affinity of the fourth complex corrected by the fourth residual correction feature, the predicted relative affinity of the mutant protein complex is output. Based on the predicted relative affinity of the mutant protein complex and the training labels, the loss is calculated and backpropagated to obtain the trained second prediction model.

8. The training method according to claim 1, characterized in that: The training data includes multiple clusters, and the mutant protein complexes in each cluster are clustered based on structural similarity, confidence of interface regions, and / or sequence homology. The training data includes a training set, a validation set, and a test set, wherein the training set, the validation set, and the test set have distinct clusters, and each cluster contains distinct mutant protein complexes.

9. The training method according to any one of claims 1-4 and 6-8, characterized in that, Before training the preset affinity prediction model based on the training data and the multimodal features, the method further includes: Acquire pre-training data, which includes the three-dimensional structural information of the mutant protein complex and pre-training labels, wherein the pre-training labels are exploitable data; Based on the pre-training data, the preset affinity prediction model is pre-trained to obtain a pre-trained preset affinity prediction model.

10. The training method according to claim 9, characterized in that, The exploitability data include at least one of the following: aggregation tendency of mutant protein complexes and / or mutant monomeric proteins, thermal stability, pharmacokinetic properties, solubility, and chemical stability.

11. A method for predicting the affinity of mutant protein complexes, characterized in that, include: Obtain the information to be predicted, which includes the target mutant three-dimensional structure information of the target mutant protein complex, or the target wild-type three-dimensional structure information and target mutation information of the target wild-type protein complex. Based on the information to be predicted, corresponding multimodal features are generated; the multimodal features include the molecular physical properties of the target mutant protein complex, and / or the residue interaction features of the target mutant protein complex and the corresponding residue interaction features of the target wild-type protein complex. Based on the information to be predicted and the multimodal features, the target affinity prediction model trained by any one of the training methods in claims 1 to 10 generates the relative affinity corresponding to the target mutant protein complex.

12. The affinity prediction method according to claim 11, characterized in that, The target affinity prediction model includes a first prediction model, which is trained based on the three-dimensional structural information of the mutant complex, the training label, and the molecular physical characteristics of the mutant protein complex. The step of generating the relative affinity corresponding to the target mutant protein complex based on the information to be predicted and the multimodal characteristics, using the trained target affinity prediction model, includes: Based on the information to be predicted and the molecular physical characteristics of the target mutant protein complex, a first predicted value of the relative affinity corresponding to the target mutant protein complex is generated using the first prediction model; or The target affinity prediction model includes a second prediction model, which is trained based on the three-dimensional structural information of the mutant complex, the training tag, the residue interaction characteristics of the mutant protein complex, and the residue interaction characteristics of the wild-type protein complex. The step of generating the relative affinity corresponding to the target mutant protein complex based on the information to be predicted and the multimodal characteristics, using the trained target affinity prediction model, includes: Based on the information to be predicted, the residue interaction characteristics of the target mutant protein complex, and the residue interaction characteristics of the target wild-type protein complex, a second predicted value of the relative affinity corresponding to the target mutant protein complex is generated using the second prediction model; or The target affinity prediction model includes a first prediction model and a second prediction model. The step of generating the relative affinity corresponding to the target mutant protein complex using the trained target affinity prediction model, based on the information to be predicted and the multimodal features, includes: Based on the information to be predicted and the molecular physical characteristics of the target mutant protein complex, a first predicted value of the relative affinity corresponding to the target mutant protein complex is generated through the first prediction model. Based on the information to be predicted, the residue interaction characteristics of the target mutant protein complex, and the residue interaction characteristics of the target wild-type protein complex, a second predicted value of the relative affinity corresponding to the target mutant protein complex is generated by the second prediction model. Calculate the average of the first predicted value and the second predicted value, and use the average value as the relative affinity corresponding to the target mutant protein complex.

13. A training device for a protein complex affinity prediction model, characterized in that, include: A data acquisition module is used to acquire training data, which includes the three-dimensional structural information of the mutant protein complex and training tags, wherein at least a portion of the mutant protein complex is obtained by mutation of the corresponding wild-type protein complex, and the training tags include the relative affinity of the mutant protein complex. The feature extraction module is used to obtain the multimodal features of the mutant protein complex based on the three-dimensional structural information of the mutant complex, wherein the multimodal features include the molecular physical properties of the mutant protein complex, and / or the residue interaction features of the mutant protein complex and the corresponding residue interaction features of the wild-type protein complex. The model training module is used to train the preset affinity prediction model based on the training data and the multimodal features to obtain the trained target affinity prediction model.

14. The training device according to claim 13, characterized in that, The target affinity prediction model includes a first prediction model and / or a second prediction model. The first prediction model is trained using the three-dimensional structural information of the mutant complex, the training tag, and the molecular physical characteristics of the mutant protein complex. The second prediction model is trained using the three-dimensional structural information of the mutant complex, the training tag, the residue interaction characteristics of the mutant protein complex, and the residue interaction characteristics of the wild-type protein complex.

15. An affinity prediction device for mutant protein complexes, characterized in that, include: The information acquisition module is used to acquire the information to be predicted, which includes the target mutant three-dimensional structure information of the target mutant protein complex, or the target wild-type three-dimensional structure information and target mutation information of the target wild-type protein complex. The feature encoding module is used to generate corresponding multimodal features based on the information to be predicted; the multimodal features include the molecular physical properties of the target mutant protein complex, and / or the residue interaction features of the target mutant protein complex and the corresponding residue interaction features of the target wild-type protein complex; An affinity prediction module is used to generate the relative affinity of the target mutant protein complex based on the information to be predicted and the multimodal features, using a target affinity prediction model trained by any one of the training methods in claims 1 to 10.

16. An electronic device, characterized in that, include: processor; as well as A memory having executable code stored thereon, which, when executed by the processor, causes the processor to perform the training method as described in any one of claims 1 to 10 or the affinity prediction method as described in claim 11 or 12.

17. A computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform a training method as claimed in any one of claims 1 to 10 or to perform an affinity prediction method as claimed in claim 11 or 12.

18. A computer program product, characterized in that, The computer program product includes computer instructions that, when executed by a processor, implement the training method as described in any one of claims 1 to 10 or perform the affinity prediction method as described in claim 11 or 12.

Images