Predicting stability of protein on the basis of graph neural network

Abstract

A method for predicting stability of protein on the basis of a graph neural network is provided, in particular, a method for predicting stability of protein after mutation is provided. The method comprises: on the basis of sequence information of a protein sample to be predicted, using a pre-trained model to perform feature extraction, so as to obtain a sequence feature of said protein sample; on the basis of three-dimensional structure information of said protein sample, obtaining a structural feature of said protein sample; and, on the basis of the sequence feature and the structural feature, using a graph neural network based prediction model to predict, so as to determine a stability prediction result of said protein sample after mutation. Further provided is a method for constructing a stability prediction model for protein after mutation. Combining a natural language processing model with a graph neural network for predicting stability changes caused by mutation presents a performance superior to that of earlier methods.

Description

Predicting protein stability based on graph neural network

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to Chinese patent application number 202310905338.0, filed with the Chinese Patent Office on July 21, 2023, entitled “Predicting Protein Stability Based on Graph Neural Networks,” and incorporates its entire contents into this application by reference. Technical Field

[0003] The present disclosure relates to protein stability research, and more specifically to protein stability prediction based on graph neural network (GNN). Background Art

[0004] Proteins can only perform specific functions by adopting specific three-dimensional structures, and their ability to fold into corresponding three-dimensional structures is affected by protein stability. (See Ittisoponpisan, Sirawit et al., 2019, “Can Predicted Protein 3D Structures Provide Reliable Insights into Whether Missense Variants Are Disease Associated?”, Journal of Molecular Biology 431(11):2197–2212 and Ofoegbu, Tochukwu C. et al., 2019, “PhyreRisk: A Dynamic Web Application to Bridge Genomics, Proteomics and 3D Structural Data to Guide Interpretation of Human Genetic Variants”, Journal of Molecular Biology 431(13):2460–66. The entire contents of the above documents are hereby incorporated into the present disclosure by reference, making them a part of the contents of the present disclosure.) Accurately predicting the impact of mutations on protein stability (folding free energy) is crucial in many fields, from biotechnology to medicine. For example, in protein engineering, targeted redesign of proteins makes it possible to optimize the biotechnology and biopharmaceutical processes involved in these proteins. (See Coluzza, Ivan, 2017, “Computational protein design: a review”, Journal of Physics: Condensed Matter 29(14):143001 and Korendovych, Ivan V., and William F. DeGrado, 2020, “De Novo Protein Design, a Retrospective”, Quarterly Reviews of Biophysics 53:e3. The entire contents of the above documents are hereby incorporated into the present disclosure by reference, making them a part of the content of the present disclosure.) Stability prediction also plays a key role in interpreting the effects of human genetic variation, which can help people better understand how these variations lead to disease.(See Gunning, Adam C et al., 2021, “Assessing Performance of Pathogenicity Predictors Using Clinically Relevant Variant Datasets”, Journal of Medical Genetics 58(8): 547–55 and Kopanos, Christos et al., 2019, “VarSome: The Human Genomic Variant Search Engine” ed. Jonathan Wren, Bioinformatics 35(11): 1978–80. The entire contents of the above-mentioned documents are hereby incorporated into the present disclosure by reference, making them part of the contents of the present disclosure.) Directed evolution and rational design are two common protein modification strategies. Directed evolution strategies have been successfully applied to the modification of enzyme properties such as enzyme activity, stability, substrate specificity, and stereoselectivity. American scientist Frances H. Arnold won the 2018 Nobel Prize in Chemistry for this. However, directed evolution requires the construction of a large-scale mutant library and the establishment of high-throughput screening methods, which consumes a lot of manpower, material resources, and financial resources. Directed evolution methods also struggle to fully search sequence space and suffer from fundamental flaws. Rational design strategies rely on understanding the relationship between protein structure and function to predict potential mutations, which are then constructed in the target gene through site-directed mutagenesis. Essentially, rational design is more efficient for protein engineering than directed evolution. Furthermore, rational design methods are universally applicable; an effective rational design strategy can be applied to a wide range of protein modifications. However, the accuracy of current rational design methods is generally low, and their application scope is far less extensive than that of directed evolution. Over the past few decades, numerous studies have focused on developing newer and more efficient methods, with artificial intelligence-based prediction of the effects of mutations on protein stability being a recent development. In recent years, artificial intelligence-assisted protein engineering has emerged as a highly effective new protein design strategy, demonstrating unique advantages in predicting protein structure, function, stability, and antibody affinity, becoming a new technological trend following rational design and directed evolution.

[0005] Over the past decade, artificial intelligence (AI) has significantly improved various protein prediction methods, from predicting residue interactions at the protein sequence embedding level to predicting protein structure and function. Natural language processing (NLP) is currently one of the hottest areas in AI, commonly used for text generation, question answering, and translation. Natural language processing (NLP) is inherently applicable to analyzing the structure and meaning of text, and the most fundamental nucleic acids and proteins in life are composed of sequences, which is intrinsically linked to text. The key concept behind these methods is to interpret protein sequences as sentences and their component amino acids as individual words. These methods have been able to extract specific features directly from protein amino acid sequences, overcoming the limitations of standard feature-based machine learning (ML) methods. A variety of protein-based NLP models have been developed to characterize protein sequence features and are used for downstream protein property prediction, demonstrating excellent performance. (See Rao, Roshan et al., 2019, “Evaluating Protein Transfer Learning with TAPE”, Advances in Neural Information Processing Systems 32:9689–9701. The entire contents of this reference are hereby incorporated by reference into this disclosure.)

[0006] In contrast, geometric deep learning methods, particularly graph neural networks (GNNs), can predict various protein properties through convolution operations on protein structures. If the three-dimensional structure of a protein is viewed as a network composed of amino acids directly connected according to specific relationships, then the connections between each node represent the interactions between an amino acid and its surrounding amino acids, which precisely reflect the characteristics of the protein structure. This feature is generally referred to as the edge feature of the graph, and the weight can be expressed by the distance between nodes. For example, in the structure of a drug molecule, the weight of the edge between two atoms can be determined by the distance between them or the type of chemical bond. By considering edge features and weights, GNNs can more accurately capture the relationships and distances between nodes, thereby improving the model's expressiveness and performance. GNNs have achieved great success in a variety of problems, from learning features for quantitative structure-activity relationship models, to predicting drug biochemical activity, to predicting interfaces between protein pairs.

[0007] Summary of the Invention

[0008] Based on the above-mentioned problems with current methods for predicting the effects of mutations on protein stability, the present disclosure proposes a fast and accurate method for predicting the effects of mutations on protein stability. It is based on the latest natural language processing methods, using the T5 (i.e., Transfer Text-to-Text Transformer) pre-trained model to extract sequence features, combined with structure-based features, and input into a graph neural network model, preferably a graph convolutional neural network model, to predict changes in folding free energy caused by amino acid substitutions, and obtain prediction results for protein stability. The results show that the performance has been further improved compared to existing methods.

[0009] According to a first aspect of the present disclosure, a method for predicting protein stability after mutation is provided. The method may include the following steps: extracting features using a pre-trained model based on the sequence information of the protein sample to be predicted, thereby obtaining sequence features of the protein sample to be predicted; obtaining structural features of the protein sample to be predicted based on the three-dimensional structure information of the protein sample to be predicted; and performing prediction using a graph neural network-based prediction model based on the sequence features and the structural features to determine a stability prediction result for the protein sample to be predicted after mutation.

[0010] In the method according to the first aspect of the present disclosure, the sequence information of the protein sample to be predicted may include amino acid sequence information of the protein sample to be predicted before mutation and amino acid sequence information of the protein sample to be predicted after mutation.

[0011] On the other hand, the sequence information of the protein sample to be predicted may include the amino acid sequence information of the protein sample to be predicted before mutation and the designated mutation information. Thus, based on the amino acid sequence information of the protein sample to be predicted before mutation and the designated mutation information, the amino acid sequence information of the protein sample to be predicted after mutation can be obtained.

[0012] Preferably, the specified mutation information includes a specified mutation site and / or a specified amino acid type after mutation.

[0013] In the method according to the first aspect of the present disclosure, the method of extracting features based on the sequence information of the protein sample to be predicted using a pre-trained model to obtain the sequence features of the protein sample to be predicted may further include: extracting features from the amino acid sequence information of the protein sample to be predicted before mutation and the amino acid sequence information after mutation using the pre-trained model to obtain sequence feature information before mutation and sequence feature information after mutation; and concatenating the sequence feature information before mutation and the sequence feature information after mutation to obtain the sequence features of the protein sample to be predicted.

[0014] In the method according to the first aspect of the present disclosure, the pre-trained model can be implemented by a combination of one or more of the following models: ESM-1b, UniRef, ProteinBert, TAPE, ProtGPT2, ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, ProtT5-XL and ProtT5-XXL; preferably, the pre-trained model is ProtT5-XL or ProtT5-XXL; more preferably, the pre-trained model is ProtT5-XL.

[0015] In the method according to the first aspect of the present disclosure, the three-dimensional structure information of the protein sample to be predicted may include the three-dimensional structure information of the protein sample to be predicted obtained through at least one of the following databases or prediction software: PDB database, AlphaFold2, I-TASSER, RoseTTAFold, Modeller and Swiss-model.

[0016] In the method according to the first aspect of the present disclosure, the three-dimensional structural information may include an interaction network, a secondary structure, an amino acid residue distance or a physical environment. Preferably, the three-dimensional structural information includes an interaction network.

[0017] Preferably, the three-dimensional structural information of the protein sample to be predicted is obtained by: obtaining the coordinates of the three-dimensional structure of the protein sample to be predicted; calculating the distance between amino acid residue pairs in the three-dimensional structure of the protein based on the obtained coordinates of the three-dimensional structure; when the alpha carbon atom distance between two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1; otherwise, it is 0, and the node itself is recorded as 0, thereby obtaining an adjacency matrix of the entire protein sample to be predicted.

[0018] Accordingly, in the method according to the first aspect of the present disclosure, the three-dimensional structural information of the protein sample to be predicted, obtaining the structural characteristics of the protein sample to be predicted may include: taking the mutation site as the center, intercepting an adjacency matrix of a specified size from the adjacency matrix of the entire protein sample to be predicted as the structural characteristics of the protein sample to be predicted; or, based on the amino acids in the adjacency matrix of the entire protein sample to be predicted, calculating the distance between the amino acid residue pairs, when the alpha carbon atom distance between the two residue pairs is less than 10 angstroms, indicating that there is an edge relationship between the nodes, recorded as 1, otherwise it is 0, and itself is recorded as 0, obtaining the second-order adjacency matrix of the protein sample to be predicted. In some embodiments, taking the mutation site as the center, intercepting an N*N (N is an integer greater than or equal to 1) adjacency matrix from the adjacency matrix of the entire protein. Intercepting an N*N adjacency matrix from the adjacency matrix of the entire protein can be intercepted based on the sequences on the left and right sides of the mutation site, or it can be intercepted based on the distance between the mutation site. In one embodiment, with the mutation site as the center, amino acids whose distance from the mutation site is less than 10 angstroms in the adjacency matrix of the entire protein sample to be predicted are selected as structural features of the protein sample to be predicted.

[0019] In the method according to the first aspect of the present disclosure, the prediction based on the sequence features and the structural features using a graph neural network-based prediction model to determine the stability prediction result of the protein sample to be predicted after mutation can further include: inputting the sequence features and the structural features into the graph neural network-based prediction model; updating the features using several layers of graph network structures through matrix multiplication and aggregation operations and then outputting them; merging the features outputted by the several layers of graph network structures; compressing the features using a pooling operation; and classifying the features according to the stability of the output layer classification function. Preferably, the pooling operation can be average pooling, maximum pooling, or K-max pooling.

[0020] In the method according to the first aspect of the present disclosure, the prediction model based on the graph neural network can be constructed by the following steps: obtaining a training set, the training set including information of multiple training samples, the information of each training sample including the three-dimensional structural information of the sample protein, the amino acid sequence information of the sample protein before and after mutation, and the mutation stability label; based on the amino acid sequence information of the sample protein before and after mutation, feature extraction is performed through a pre-training model to obtain the sequence features of the multiple training samples; based on the three-dimensional structural information of the sample protein, the structural features of the multiple training samples are obtained; based on the sequence features, the structural features and the mutation stability label, the graph neural network is trained to obtain the prediction model based on the graph neural network.

[0021] Preferably, the graph neural network-based prediction model is a classifier.

[0022] Preferably, the graph neural network is a graph convolutional network.

[0023] More preferably, the prediction model based on graph neural network includes one of the following models: MGC model, MGC_F model, MGC_S model, GINConv model and SageConv model.

[0024] Preferably, experimental data from a public database can be obtained as a training set. At the same time, the training set can be doubled in size by inverting the experimental data from the public database.

[0025] According to a second aspect of the present disclosure, a method for constructing a protein stability prediction model after mutation is provided. The method may include: obtaining a training set, the training set including information of multiple training samples, each of the training samples including three-dimensional structural information of the sample protein, amino acid sequence information of the sample protein before and after mutation, and a mutation stability signature; extracting features using a pre-trained model based on the amino acid sequence information of the sample protein before and after mutation to obtain sequence features of the multiple training samples; obtaining structural features of the multiple training samples based on the three-dimensional structural information of the sample protein; and training a graph neural network based on the sequence features, structural features, and mutation stability signature to obtain a prediction model based on the graph neural network.

[0026] In the method according to the second aspect of the present disclosure, the feature extraction based on the amino acid sequence information of the sample protein before and after mutation through a pre-training model to obtain the sequence features of the multiple training samples can further include: using the pre-training model to extract features of the amino acid sequence information of the sample protein before and after mutation, respectively, to obtain sequence feature information before mutation and sequence feature information after mutation; and splicing the sequence feature information before mutation and the sequence feature information after mutation to obtain the sequence features of the multiple training samples.

[0027] In the method according to the second aspect of the present disclosure, the pre-trained model can be implemented by a combination of one or more of the following models: ESM-1b, UniRef, ProteinBert, TAPE, ProtGPT2, ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, ProtT5-XL and ProtT5-XXL; preferably, the pre-trained model is ProtT5-XL or ProtT5-XXL; more preferably, the pre-trained model is ProtT5-XL.

[0028] In the method according to the second aspect of the present disclosure, the three-dimensional structural information of the sample protein may include the three-dimensional structural information of the sample protein obtained through at least one of the following databases or prediction software: PDB database, AlphaFold2, I-TASSER, RoseTTAFold, Modeller and Swiss-model.

[0029] In the method according to the second aspect of the present disclosure, the three-dimensional structural information may include an interaction network, a secondary structure, an amino acid residue distance or a physical environment. Preferably, the three-dimensional structural information includes an interaction network.

[0030] Preferably, the three-dimensional structural information of the sample protein can be obtained in the following manner: obtaining the coordinates of the three-dimensional structure of the sample protein; calculating the distance between amino acid residue pairs in the three-dimensional structure of the sample protein based on the obtained coordinates of the three-dimensional structure; when the alpha carbon atom distance between two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1; otherwise, it is 0, and the node itself is recorded as 0, thereby obtaining the adjacency matrix of the entire sample protein.

[0031] Accordingly, in the method according to the second aspect of the present disclosure, obtaining the structural features of multiple training samples based on the three-dimensional structural information of the sample protein may include: intercepting an adjacency matrix of a specified size from the adjacency matrix of the entire sample protein with the mutation site as the center, as the structural feature of the sample protein; or, calculating the distance between the amino acid residue pairs based on the amino acids in the adjacency matrix of the entire protein sample to be predicted, when the alpha carbon atom distance between the two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1, otherwise it is 0, and it itself is recorded as 0, thereby obtaining the second-order adjacency matrix of the protein sample to be predicted.

[0032] In the method according to the second aspect of the present disclosure, the training of a graph neural network based on the sequence features, the structural features, and the mutation stability labels to obtain a prediction model based on the graph neural network may include: inputting the sequence features and the structural features into the graph neural network; updating and outputting the features using several layers of graph network structures through matrix multiplication and aggregation operations; merging the features outputted by several layers of graph network structures; compressing the features using a pooling operation; classifying the features for stability according to the output layer classification function; and adjusting the network parameters of the graph neural network using the mutation stability label as the output target. Preferably, the pooling operation may be average pooling, maximum pooling, or K-max pooling.

[0033] Preferably, the graph neural network-based prediction model is a classifier.

[0034] Preferably, the graph neural network is a graph convolutional network.

[0035] More preferably, the prediction model based on graph neural network includes one of the following models: MGC model, MGC_F model, MGC_S model, GINConv model and SageConv model.

[0036] Preferably, experimental data from a public database can be obtained as a training set. At the same time, the training set can be doubled in size by inverting the experimental data from the public database.

[0037] According to the third aspect of the present disclosure, a system for predicting the stability of a protein after mutation is provided. The system may include: an acquisition module for acquiring sequence information and three-dimensional structural information of a protein sample to be predicted; a processing module for obtaining a stability prediction result of the protein sample to be predicted after mutation by inputting the sequence information and three-dimensional structural information of the protein sample to be predicted. The processing module further includes the following submodules: a sequence feature submodule for extracting features using a pre-trained model based on the sequence information of the protein sample to be predicted to obtain the sequence features of the protein sample to be predicted; a structural feature submodule for obtaining the structural features of the protein sample to be predicted based on the three-dimensional structural information of the protein sample to be predicted; a model prediction submodule for performing prediction based on the sequence features and the structural features using a prediction model based on a graph neural network to determine the stability prediction result of the protein sample to be predicted after mutation.

[0038] According to the third aspect of the present disclosure, the system for predicting protein stability based on graph neural network can be implemented by a computer. Preferably, the following operations can be implemented by executing a computer program: obtaining sequence information and three-dimensional structure information of the protein sample to be predicted; obtaining the stability prediction result of the protein sample to be predicted after mutation by inputting the sequence information and three-dimensional structure information of the protein sample to be predicted, and the operation further includes the following sub-operations: based on the sequence information of the protein sample to be predicted, using a pre-trained model to perform feature extraction to obtain the sequence characteristics of the protein sample to be predicted; based on the three-dimensional structure information of the protein sample to be predicted, obtaining the structural characteristics of the protein sample to be predicted; based on the sequence characteristics and the structural characteristics, using a prediction model based on graph neural network to perform prediction to determine the stability prediction result of the protein sample to be predicted after mutation.

[0039] According to a fourth aspect of the present disclosure, a non-transitory computer-readable storage medium is provided for storing a computer program. The computer program includes instructions. When executed by a processor of an electronic device, the instructions cause the electronic device to implement the method for predicting protein stability after mutation according to the first aspect of the present disclosure or the method for constructing a protein stability prediction model after mutation according to the second aspect of the present disclosure.

[0040] According to a fifth aspect of the present disclosure, a computer system is provided. The computer system comprises: a processor, a memory, and a computer program. The computer program is stored in the memory and is configured to be executed by the processor. The computer program comprises instructions for implementing the method for predicting protein stability after mutation according to the first aspect of the present disclosure or the method for constructing a protein stability prediction model after mutation according to the second aspect of the present disclosure.

[0041] The above methods are based on advanced natural language processing models combined with graph neural networks, can be applied to different biological fields, and have shown better performance than earlier methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The present disclosure will be more fully understood from the following detailed description taken in conjunction with the accompanying drawings, in which like elements are numbered in a similar manner, and in which:

[0043] FIG1 is a flowchart of a method for predicting protein stability after mutation according to an embodiment of the present disclosure.

[0044] FIG2 is a schematic diagram illustrating the association between a method for predicting protein stability after mutation and a method for constructing a protein stability prediction model after mutation according to an embodiment of the present disclosure.

[0045] FIG3 is a flowchart of the method for constructing a protein stability prediction model after mutation used in the present disclosure.

[0046] Figure 4 is a schematic diagram of the adjacency matrix.

[0047] FIG5 is a schematic diagram of forming a node feature matrix.

[0048] FIG6 is a schematic diagram of the construction of the structure-based adjacency matrix.

[0049] FIG7 is a schematic diagram of constructing a sequence-based node feature matrix.

[0050] FIG8 is a schematic diagram of the framework of the prediction model.

[0051] FIG9 is a schematic diagram combining FIG6-8.

[0052] FIG10 shows a schematic block diagram of a system for predicting protein stability after mutation according to the present disclosure. DETAILED DESCRIPTION

[0053] Unless otherwise defined, technical and scientific terms used in the present disclosure have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs.

[0054] The technical solutions of the present disclosure are further described in detail below through examples and in conjunction with the accompanying drawings. Unless otherwise stated, the methods and materials of the embodiments described below are conventional products that can be purchased on the market. Those skilled in the art will understand that the methods and materials described below are merely exemplary and should not be construed as limiting the scope of the present disclosure.

[0055] At present, the methods for predicting the effect of mutations on protein stability are mainly divided into structure-based and sequence-based methods. The performance of structure-based methods is generally slightly better than that of sequence-based methods, while sequence-based methods are not limited by the lack of structure. Although only about 0.2% of proteins in the UniProt database have available three-dimensional structures (see Quan, Lijun, Qiang Lv and Yang Zhang, 2016, “STRUM: Structure-Based Prediction of Protein Stability Changes upon Single-Point Mutation”, Bioinformatics 32(19):2936–46, which is hereby incorporated by reference into the present disclosure in its entirety, making it a part of the present disclosure), the emergence of software tools such as AlphaFold2 has greatly compensated for the limitations of the application of structure-based methods due to the lack of three-dimensional structure.

[0056] At this stage, the input features generated by the natural language model only contain sequence information to obtain excellent performance (for example, see Chinese patent application CN202211708226.8, which is incorporated into the present disclosure by reference herein in its entirety to form part of the present disclosure), and using deep neural networks to extract structural features from the three-dimensional structure of proteins can also achieve good performance. However, there are still great difficulties in further improving it. Therefore, introducing graph neural networks into protein stability prediction and combining them with protein natural language models will be a promising method.

[0057] From the overall perspective of the prediction method, FIG1 shows a flowchart of a method for predicting protein stability after mutation according to an embodiment of the present disclosure.

[0058] As shown in FIG1 , the method 100 for predicting protein stability after mutation according to an embodiment of the present disclosure begins at step S110 . In this step, based on the sequence information of the protein sample to be predicted, a pre-trained model is used to perform feature extraction to obtain the sequence features of the protein sample to be predicted.

[0059] The sequence information of the protein sample to be predicted mentioned herein may include amino acid sequence information of the protein sample to be predicted before mutation and amino acid sequence information of the protein sample to be predicted after mutation.

[0060] In another case, the sequence information of the protein sample to be predicted may also include the amino acid sequence information of the protein sample to be predicted before mutation and the designated mutation information. In this case, the amino acid sequence information of the protein sample to be predicted after mutation can be obtained based on the amino acid sequence information of the protein sample to be predicted before mutation and the designated mutation information. The designated mutation information may include a designated mutation site and / or a designated amino acid type after mutation. For example, if the designated mutation information specifies a designated mutation site, the amino acid at that site may be changed to form the amino acid sequence information of the protein sample to be predicted after mutation. If the designated mutation information also includes a designated amino acid type after mutation, the amino acid sequence information after mutation may be a single amino acid sequence. If the designated mutation information does not include a designated amino acid type after mutation, the amino acid sequence information after mutation may be 19 amino acid sequences in which the amino acid at the mutation site is replaced with 19 other amino acids. Another case is that the designated mutation information only includes the designated amino acid type after mutation. In this case, the amino acid at each site in the pre-mutation sequence needs to be changed sequentially to form N sequences after mutation (N is the number of sites in the sequence). In addition, another extreme case is that the specified mutation sites are multiple sites or even all sites, or the specified amino acid types after mutation are multiple types or all other 19 types. In this case, it is necessary to traverse all mutation possibilities to form a maximum of 19*N mutated sequences.

[0061] Those skilled in the art should understand that in subsequent steps, it is necessary to extract features from the sequences before and after mutation respectively, and finally splice them into features about the sequences. Therefore, no matter what the protein sequence information initially obtained is, it is necessary to form the sequences before and after mutation when executing step S110.

[0062] Specifically, in step S110, the amino acid sequence information of the protein sample to be predicted before and after the mutation is subjected to feature extraction using a pre-trained model to obtain pre-mutation sequence feature information and post-mutation sequence feature information. Then, also in step S110, the pre-mutation sequence feature information and post-mutation sequence feature information are concatenated to obtain the sequence features of the protein sample to be predicted.

[0063] In the field of natural language processing (NLP) research, with the continuous increase in computing power, a growing number of pre-trained models (PTMs) for general language representation have emerged. This is extremely helpful for downstream pre-processing tasks, enabling them to learn language knowledge from massive amounts of unlabeled data and then fine-tune on a smaller amount of labeled data, allowing downstream tasks to better learn language characteristics and task-specific knowledge. Pre-trained models suitable for biological applications can be obtained by training on protein or nucleic acid sequence corpora. Existing models trained on protein sequence corpora include ESM-1b, UniRef, ProteinBert, TAPE, ProtGPT2, ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, ProtT5-XL, and ProtT5-XXL. While it is also possible to obtain pre-trained models by training them on protein sequence corpora, the performance of these models is often not as good as that of the aforementioned models due to cost and time constraints.

[0064] As mentioned above, the pre-trained model is essentially a natural language model. In the biological field, a pre-trained model that has been improved and optimized based on the characteristics and uses of the biological field is generally used. This pre-trained model has learned relevant knowledge in the biological field from a large amount of unlabeled sequence data in the existing protein corpus and can be used to extract mutation-related features from the protein sequence information of the training samples; then, based on the extracted mutation-related features, training is performed, and only a small number of labeled training samples are required to optimize the model parameters. For example, the pre-trained model described in this disclosure can be implemented using a combination of one or more of the models mentioned above and listed again below: ESM-1b, UniRef, ProteinBert, TAPE, ProtGPT2, ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, ProtT5-XL, and ProtT5-XXL. Preferably, the pre-trained model can be ProtT5-XL or ProtT5-XXL. More preferably, the pre-trained model is ProtT5-XL. That is, the pre-training model used here is preferably a T5 (i.e., Transfer Text-to-Text Transformer) pre-training model based on the latest natural language processing method.

[0065] The purpose of step S110 is to obtain the sequence characteristics of the protein sample to be predicted, while the purpose of step S120 to be described next is to obtain the structural characteristics of the protein sample to be predicted.

[0066] In step S120 , the structural features of the protein sample to be predicted are obtained based on the three-dimensional structural information of the protein sample to be predicted.

[0067] In addition to obtaining relevant features through pre-trained models, relevant features can also be obtained by combining protein 3D structural information to supplement information. 3D structural information can be obtained from the PDB database (https: / / www.rcsb.org / ) or through prediction software, such as the following prediction software (references or sources are in parentheses):

[0068] AlphaFold2(https: / / alphafold.ebi.ac.uk / );

[0069] I-TASSER(https: / / zhanggroup.org / I-TASSER / );

[0070] RoseTTAFold (see Minkyung Baek et al., "Accurate prediction of protein structures and interactions using a three-track neural network", Science373, 871-876 (2021). DOI: 10.1126 / science.abj8754);

[0071] Modeller(https: / / salilab.org / modeller / );

[0072] Swiss-model (https: / / swissmodel.expasy.org / ) etc.

[0073] The available three-dimensional structural information of proteins generally includes secondary structure, interaction network, distance between amino acid residues, physical environment, and energy information obtained from the three-dimensional structure. Among them, the interaction network is one of the most commonly used features in structure-based prediction methods because it contains contact information between all residue pairs.

[0074] According to a preferred embodiment of the present disclosure, the three-dimensional structural information utilized is the interaction network. More specifically, it is based on the characteristics of the interaction network that the GNN, more specifically the graph convolutional network (GCN) prediction model, is selected.

[0075] Those skilled in the art should understand that the present disclosure does not limit the order of step S110 and step S120. In other words, the acquisition of sequence features and structural features can be performed in any order or simultaneously.

[0076] Finally, in step S130, based on the sequence features obtained in step S110 and the structural features obtained in step S120, a prediction model based on a graph neural network is used to perform a prediction to determine the stability prediction result of the protein sample to be predicted after mutation.

[0077] The prediction model based on graph neural network is actually a classifier.

[0078] In step S130, the sequence-based features and structure-based features are put into a classifier to predict a specific task. Classifiers are generally divided into machine learning methods and deep learning methods. Machine learning methods are generally used: Linear Regression algorithm, Support Vector Machine (SVM) algorithm, Nearest Neighbor / k-Nearest Neighbor (KNN) algorithm, Logistic Regression (LR) algorithm, Decision Tree algorithm, K-Means algorithm, Random Forest algorithm, Naive Bayes algorithm, Gradient Boosting algorithm, Ensemble Learning method. Deep learning methods generally include: Convolutional Neural Network, Recurrent Neural Network, Recursive Neural Network, Long Short-Term Memory Network (LSTM). Graph Neural Network (GNN) is a deep learning model based on graph structured data, used to process structured data of nodes and edges. The Graph Convolutional Network (GCN) is a GNN model based on the Convolutional Neural Network (CNN). It achieves effective representation and learning of node features by aggregating information from neighboring nodes. The Graph Attention Network (GAT) is a GNN model based on the attention mechanism. It achieves adaptive aggregation of node features by learning the attention weights of neighboring nodes, which can better handle the relationships between different nodes. GraphSAGE (Graph Sampling and Aggregation, SageConv) is a sampling-based GNN model. It achieves effective learning and representation of node features by sampling and aggregating neighboring nodes. GraphSAGE can process large-scale graph data while ensuring model efficiency and accuracy. The Graph Isomorphism Network (GINConv) is a GNN model based on graph isomorphism. It achieves effective learning and representation of node features by aggregating the features of neighboring nodes and combining the aggregation results with the node's own features.The Diffusion Convolutional Neural Network (DCNN) is a GNN model based on the diffusion process. It effectively learns and represents node features by diffusing and aggregating them. Multi-Graph Convolution (MGC) can perform graph convolution on different types of graphs, merging them together and fully utilizing information from different graph types. MGC-F represents a first-order adjacency matrix as the input, while MFC-S represents a second-order adjacency matrix as the input. Each of these methods has its own advantages and scope of application. The appropriate method can be selected based on the specific task and data characteristics.

[0079] In a preferred embodiment of the present disclosure, the graph neural network is a graph convolutional network.

[0080] In a more preferred embodiment of the present disclosure, the prediction model based on the graph neural network includes one of the following models: MGC model, MGC_F model, MGC_S model, GINConv model, and SageConv model. For more details, please refer to the detailed description below.

[0081] For the method of constructing the prediction model, please refer to the detailed description below.

[0082] As mentioned above, existing methods for predicting the effects of mutations on protein stability are primarily structure-based and sequence-based. Natural language models can achieve excellent performance by generating input features containing only sequence information, while deep neural networks can also achieve good performance by extracting structural features from the protein's three-dimensional structure.

[0083] The difference between the technology provided by the present disclosure and the above-mentioned prior art is that the technology disclosed in the present disclosure introduces graph neural networks, more specifically graph convolutional networks, into the prediction of protein stability after mutation, and combines them with mature protein natural language models (pre-training models) in the prior art to jointly determine the prediction results. In other words, the sequence features and structural features of the protein sample to be predicted are combined and used as inputs to the prediction model based on the graph neural network, thereby outputting the prediction results of the protein stability after mutation. The prediction results determined in this way are better than the prediction results determined by the methods of the prior art.

[0084] FIG2 shows a schematic diagram 200 illustrating the relationship between a method for predicting protein stability after mutation and a method for constructing a model for predicting protein stability after mutation according to an embodiment of the present disclosure. In the schematic diagram 200 shown in FIG2 , the left side of the dotted line shows the model construction method, while the right side of the dotted line shows the method for predicting using the model.

[0085] The process to the right of the dotted line in Figure 2 is actually the method for predicting protein stability after mutation according to an embodiment of the present disclosure shown in Figure 1. The process to the left of the dotted line in Figure 2 will be described in detail below (e.g., Figure 3 and its corresponding text description). As can be seen from Figure 2, both model construction and the actual prediction process require two feature extraction steps: sequence feature extraction and structural feature extraction. For the training sample, it includes the three-dimensional structural information of the sample protein, the amino acid sequence information of the sample protein before and after mutation, and the mutation stability label. As described later, the mutation stability label can be an unstable mutation or a stable mutation, where an unstable mutation includes a decrease in stability or a loss of stability, and a stable mutation includes an increase in stability or a constant stability. Therefore, during the model construction process, sequence information in the training sample is used to extract sequence features. Structural features are extracted using three-dimensional structural information. Using sequence features and structural features as input data and mutation stability labels in the training sample as output data, the GNN-based prediction model, i.e., the classifier, is fully trained to obtain the final GNN-based protein stability prediction model after mutation. Once the prediction model has been constructed, it can be put into practical application for predicting protein stability after mutation, as shown in the process to the right of the dotted line in Figure 2. It should also be noted that in Figure 2, hollow arrows represent the signaling process for sequence information and sequence features, while solid arrows represent the signaling process for three-dimensional structure information and structural features. Furthermore, solid arrows are used to represent other signal flows and control flows.

[0086] The following is a detailed description of the process of the method for constructing a protein stability prediction model after mutation according to an embodiment of the present disclosure.

[0087] FIG3 is a flowchart of a method 300 for constructing a protein stability prediction model after mutation used in the present disclosure.

[0088] Those skilled in the art will appreciate that, as shown in FIG3 , in step S310, a training set is first obtained, wherein the training set includes information about a plurality of training samples. Specifically, the information about each training sample includes the three-dimensional structure information of the sample protein, the amino acid sequence information of the sample protein before and after mutation, and a mutation stability signature.

[0089] Artificial intelligence (AI) algorithms rely heavily on data. The quantity and quality of initial data determine the generalization performance of the trained model, which is most directly reflected in the model's predictive accuracy. Insufficient or low-quality datasets can lead to overfitting or underfitting of the model. To address this issue, the most effective approach is to obtain more experimental data as a training set. Currently, the primary source of experimental data is public databases. For sequence-based methods predicting the effects of mutations on proteins, the dataset can be doubled in size by inverting the data in the database. Inversion means that under identical environmental conditions, the change in folding free energy (ΔΔG) is reverse-symmetric upon mutation. More precisely, if protein B is a mutant of protein A, then ΔΔG(A→B) = -ΔΔG(B→A), which conforms to the law of conservation of energy in physics. Based on this definition, merging the original data with the inverted data doubles the original dataset size. This inversion method not only expands the original dataset but also balances it, making the final prediction model more robust.

[0090] It should be noted that this method of expanding the dataset by reversing the original dataset is only implemented for the training set.

[0091] In the specific embodiments below, the acquisition of the training data set will be described in more detail.

[0092] Mutation stability labels can be used to reflect the trend of the impact of amino acid mutations on protein thermal stability. In some embodiments, the mutation stability label can be determined based on the folding free energy change value (ΔΔG) of the protein mutant. The folding free energy change value (ΔΔG) caused by the mutation is closely related to the thermal stability of the mutant, wherein ΔΔG = folding free energy of the protein before mutation - folding free energy of the protein after mutation; ΔΔG is less than 0, indicating that the stability of the protein is improved after the mutation, and vice versa, indicating that the mutation will cause a decrease in protein stability. Exemplarily, the mutation stability label can be an unstable mutation and a stable mutation, wherein the unstable mutation includes a decrease in stability or a loss of stability, and the stable mutation includes an increase in stability or a flat stability. It is understandable that using mutation stability information as a label for the training sample allows the prediction model obtained by training to output prediction results related to the impact of amino acid mutations on protein thermal stability.

[0093] In step S320, based on the amino acid sequence information of the sample protein before and after mutation, feature extraction is performed using a pre-training model to obtain sequence features of the multiple training samples.

[0094] Specifically, in step S320, the pre- and post-mutation amino acid sequence information of the sample protein is subjected to feature extraction using the pre- and post-mutation amino acid sequence information, respectively, to obtain pre- and post-mutation sequence feature information. Then, also in step S320, the pre- and post-mutation sequence feature information are concatenated to obtain sequence features for the multiple training samples.

[0095] As mentioned above, the pre-training model is implemented by a combination of one or more of the following models: ESM-1b, UniRef, ProteinBert, TAPE, ProtGPT2, ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, ProtT5-XL and ProtT5-XXL. Preferably, the pre-training model is ProtT5-XL or ProtT5-XXL. More preferably, the pre-training model is ProtT5-XL. That is, the pre-training model used here is preferably a T5 (i.e., Transfer Text-to-Text Transformer) pre-training model based on the latest natural language processing method.

[0096] The purpose of step S320 is to obtain the sequence characteristics of the training sample, while the purpose of step S330 to be described next is to obtain the structural characteristics of the training sample.

[0097] In step S330 , structural features of a plurality of training samples are obtained based on the three-dimensional structural information of the sample protein.

[0098] As mentioned above, the three-dimensional structure information of the sample protein can be obtained through at least one of the following databases or prediction software: PDB database, AlphaFold2, I-TASSER, RoseTTAFold, Modeller and Swiss-model, etc.

[0099] The available three-dimensional structural information of proteins generally includes secondary structure, interaction network, distance between amino acid residues, physical environment, and energy information obtained from the three-dimensional structure. Among them, the interaction network is one of the most commonly used features in structure-based prediction methods because it contains contact information between all residue pairs.

[0100] According to a preferred embodiment of the present disclosure, the three-dimensional structural information used here is the interaction network. More specifically, it is based on the characteristics of the interaction network that the GNN is selected, and more specifically the prediction model of the graph convolutional network (GCN).

[0101] Those skilled in the art should understand that the present disclosure does not limit the order of step S320 and step S330. In other words, the acquisition of sequence features and structural features can be performed in any order or simultaneously.

[0102] In step S340, a graph neural network is trained based on the sequence features, the structural features, and the mutation stability labels to obtain a prediction model based on the graph neural network.

[0103] As mentioned earlier, the prediction model based on graph neural network is actually a classifier.

[0104] In a preferred embodiment of the present disclosure, the graph neural network is a graph convolutional network (GCN).

[0105] In a more preferred embodiment of the present disclosure, the prediction model based on the graph neural network includes one of the following models: MGC model, MGC_F model, MGC_S model, GINConv model, and SageConv model. For more details, please refer to the detailed description below.

[0106] For the method of constructing the prediction model, please refer to the detailed description in the examples below.

[0107] In addition, it should be noted that: for sequence features, by placing the processed protein stability change dataset into the pre-trained model, the pre-trained model can be migrated to the new task through transfer learning. In transfer learning, there are two commonly used application methods: feature extraction and fine-tuning. In feature extraction, after the network structure is pre-trained, a simple classifier can be modified or added, and the pre-trained network on the original task can be used as a feature extractor for another target task. Only the last added classifier parameters are re-learned, and the pre-trained network parameters will not be modified or frozen. The preferred embodiment of the present disclosure adopts a transfer learning method of feature extraction.

[0108] During model training, sequence and structural features are used as input data for the GNN-based prediction model, and mutational stability labels are used as output targets for the prediction model. As model parameters are continuously optimized, the output data of the prediction model will align with the prediction results indicated by the mutational stability labels.

[0109] More specifically, in some embodiments of the present disclosure, a prediction model (classifier) ​​can be supervised trained based on the sequence features, structural features and mutation stability labels of multiple training samples to obtain a trained classifier. Supervised training is the process of adjusting the parameters of the classifier using a set of samples of known categories to achieve the required performance. In some embodiments, a loss function, such as a cross-entropy loss function, can be constructed to reflect the difference between the post-mutation stability prediction results predicted by the classifier and the actual mutation stability labels of the training samples, and then the parameters of the classifier are adjusted according to the difference to obtain the model optimization parameters of the classifier. When the loss function meets the predetermined conditions, such as the loss function converges, or the loss function value is less than the preset value, the classifier training is completed. In some embodiments, a parameter search algorithm can be used to adjust the classifier parameters, for example, a search algorithm such as grid search, Bayesian search, etc. In some embodiments, supervised training is the process of training a prediction model based on a stability label.

[0110] Those skilled in the art should understand that the prediction model obtained by, for example, the method of FIG. 3 can be used in the prediction method shown in FIG. 1 .

[0111] Specific example embodiments will be given below.

[0112] Example methods and materials

[0113] 1. Training Dataset

[0114] To overcome the problem of insufficient samples in the datasets used in previous methods, we used the recently published MPTherm (https: / / www.iitm.ac.in / bioinfo / mptherm / ) (see Kulandaisamy, A., R. Sakthivel and M. Michael Gromiha, 2021, “MPTherm: Database for Membrane Protein Thermodynamics for Understanding Folding and Stability”, Briefings in Bioinformatics 22(2):2119–25, the entire contents of which are incorporated into the present disclosure by reference and become part of the present disclosure), ProThermDB (https: / / web.iitm.ac.in / bioinfo2 / prothermdb / index.html) (see Nikam, Rahul et al., 2021, “ProThermDB: Thermodynamic Database for Proteins and Mutants Revisited after 15 Years”, Nucleic Acids Research 49(D1):D420–24, the entire contents of which are hereby incorporated by reference into the present disclosure and become part of the present disclosure), ThermoMutDB (http: / / biosig.unimelb.edu.au / thermomutdb / ) (see Xavier, Joicymara S et al., 2021, “ThermoMutDB: A Thermodynamic Database for Missense Mutations”, Nucleic Acids Research 49(D1):D475–79, the entire contents of which are hereby incorporated by reference into the present disclosure and become part of the present disclosure) and FireProtDB (https: / / loschmidt.chemi.muni.cz / fireprotdb / ) (Stourac, Jan et al., 2021, “FireProtDB: Database of Manually Curated Protein Stability Data”, Nucleic Acids Research 49(D1):D319–24, the entire contents of which are hereby incorporated by reference into this disclosure and become a part of this disclosure). 5,917 experimental samples containing thermal stability changes were re-screened from the experimental database, including 1,451 samples with stabilizing mutations and 4,466 samples with destabilizing mutations. The specific screening criteria are as follows:

[0115] 1) Remove all samples that do not contain the difference in folding free energy (ΔΔG) between the wild-type and mutant proteins

[0116] 2) Remove all samples with mutation information other than single-point mutations

[0117] 3) Remove duplicate samples

[0118] 4) Reverse and merge the data

[0119] The dataset used to train the model contains 11,834 samples.

[0120] 2. Test Dataset

[0121] 2.1 Tm141

[0122] The source of test data is consistent with the training set, and the screening criteria are as follows:

[0123] 1) Remove all experimental Tm changes (ΔTm) that do not include those before and after the mutation.

[0124] 2) Remove all samples with mutation information other than single-point mutations

[0125] 3) Remove duplicate samples

[0126] 4) |Δtm| is greater than or equal to 15 degrees Celsius. The dataset used to test the model contains 141 experimental data. It should be noted that there are no duplicate samples between the test dataset and the training set.

[0127] 2.2 S669

[0128] In order to rigorously test the differences between the method disclosed herein and other methods, a test dataset S669 was introduced, which contains 669 single-point mutations. When testing this dataset, duplicate data were removed from the training set. This test dataset is widely used in the following literature: Pancotti, Corrado et al., 2022, "Predicting Protein Stability Changes upon Single-Point Mutation: A Thorough Comparison of the Available Tools on a New Dataset", Briefings in Bioinformatics 23(2): bbab555. The entire contents of the above-mentioned literature are hereby incorporated into the present disclosure by reference, making it a part of the content of the present disclosure.

[0129] 3. Pre-trained Model

[0130] The ProtT5-XL (ProtT5-XL-U50 / ProtT5-XL-UniRef50) model used in this method was released by Ahmed Elnaggar et al. in 2021. The T5 model was first trained on the BFD dataset, and then the model was fine-tuned and optimized using the UniRef50 dataset to obtain a pre-trained protein model. The model contains a 24-layer attention structure, each layer consisting of 32 self-attention heads and 1024 neurons. See Elnaggar, Ahmed et al., 2021, “ProtTrans: Towards Cracking the Language of Lifes Code Through Self-Supervised Deep Learning and High Performance Computing”, IEEE Transactions on Pattern Analysis and Machine Intelligence: 1-1. By reference hereto, the entire contents of the above-mentioned documents are incorporated into the present disclosure to make it a part of the contents of the present disclosure.

[0131] 4. Graph Convolutional Neural Network Model Construction

[0132] This preferred embodiment is based on variants of graph convolutional networks (GCN): MGC, MGC_F, MGC_S, GINConv, and SageConv. The difference between the MGC method and GCN is the combination of different graph convolutional layers with different propagation rules. Compared with other graph neural network methods, the performance is excellent (see Vladimir et al., 2021, “Structure-Based Protein Function Prediction Using Graph Convolutional Networks”, Nature Communications 12(1): 3168. The entire contents of the above-mentioned document are hereby incorporated into the present disclosure by reference, making them a part of the contents of the present disclosure).

[0133] 4.1 MGC Model

[0134] 4.1.1 Adjacency Matrix

[0135] The three-dimensional structure of each protein was obtained through prediction software such as the PDB database, AlphaFold2, I-TASSER, RoseTTAFold, Modeller and Swiss-model. The distance between the amino acid residue pairs in the protein three-dimensional structure was calculated based on the coordinates of the obtained protein three-dimensional structure. When the alpha carbon atom distance between two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1, otherwise it is 0 (itself is 0), which is the adjacency matrix of the entire protein.

[0136] In this embodiment, three amino acids on each side of the mutation site are taken as the residue sequence, that is, 7 consecutive amino acids are taken as the residue sequence, and a 7*7 adjacency matrix is ​​intercepted from the adjacency matrix of the entire protein.

[0137] Figure 4 is a schematic diagram of an adjacency matrix. As shown in the figure, the adjacency matrix is ​​100 amino acids long. Assuming that the mutation site is residue 5, the box is the 7*7 adjacency matrix of this residue.

[0138] 4.1.2 Node Feature Matrix

[0139] After a protein sequence is input into the pre-trained model, the model first encodes each amino acid in the protein sequence through an encoder. Each encoded amino acid corresponds to a vector of length 1024, meaning that a protein sequence of length N will generate a vector of length N*1024. This model has a 24-layer structure. To maximize the acquisition of information about mutations, the last layer is selected as the output. 1024-dimensional feature vectors corresponding to the mutation position for both the wild-type and mutant forms are extracted, and the wild-type (i.e., before mutation) and mutant (i.e., after mutation) feature vectors are concatenated to form a 2048-dimensional feature matrix. Using a sliding window size of 7 and the mutation site as the center, the three amino acid residues before and after the mutation are concatenated, ultimately resulting in a 7*2048-dimensional feature matrix as the node feature matrix. The structure is shown in Figure 5.

[0140] 4.1.3 Model Construction

[0141] The matrices generated in 4.1.1 and 4.1.2 are input into the graph convolutional neural network as structural features and sequence features, respectively. Through matrix multiplication and aggregation operations, each layer of the graph network structure updates the node features into a 512-dimensional feature vector. The node features output by the three-layer graph network are merged into a 7*1536 (1536=512*3)-dimensional feature vector through the residual network structure. The updated feature matrix is ​​then compressed into a 1*1536-dimensional feature vector using an average pooling operation, representing the feature representation of the 7-amino acid residue sequence. The specific parameters are as follows: the output dimensions of the three-layer graph convolution are 512, 512, and 512, respectively; the number of neurons in the fully connected layer is 256 and 2; the optimizer uses Adam; the loss function is binary cross entropy loss; 'relu' is used as the activation function; 'softmax' is used as the output layer classification function; the l2 regularization coefficient is 0.0001, the dropout rate is 0.5, and the learning rate is 5e-05. Those skilled in the art should understand that the average pooling operation here can also be replaced by any pooling operation such as maximum pooling or K-max pooling.

[0142] 4.2 MGC_F model construction

[0143] 4.2.1 First-order adjacency matrix

[0144] Obtain the adjacency matrix of the entire protein according to 4.1.1. Then, with the mutation site residue as the center, select all amino acids that are less than 10 angstroms away from the mutation site to obtain the adjacency matrix composed of these amino acids.

[0145] 4.2.2 Node Feature Matrix

[0146] After a protein sequence is input into the pre-trained model, the model first encodes each amino acid in the protein sequence through an encoder. Each amino acid is encoded into a vector of length 1024, that is, a protein sequence of length N will generate a vector of length N*1024. The model has a 24-layer structure. In order to obtain the most information about mutations, this embodiment selects the last layer as the output and simultaneously extracts the 1024-dimensional feature vectors of the wild type and mutant corresponding to the mutation position. The feature vectors of the wild type and mutant are sequentially spliced ​​into 2048 dimensions, and finally N is obtained. f *2048-dimensional features as node feature matrix, where N f The number of corresponding amino acids obtained from 4.2.1 is the same as the order. The number of amino acids obtained from different proteins may be different. f As a benchmark, the insufficient number is filled with 0 so that all proteins can obtain a matrix of the same size.

[0147] 4.2.3 Model Construction

[0148] The matrices generated in 4.2.1 and 4.2.2 are input into the graph convolutional neural network as structural features and sequence features respectively. Through matrix multiplication and aggregation operations, each layer of the graph network structure updates the node features into a 512-dimensional feature vector, and the node features output by the three-layer graph network are merged into N through the residual network structure. f *1536 (1536=512*3) dimensional feature vector. Then the average pooling operation is used to compress the updated feature matrix into a 1*1536 dimensional feature vector represented as N f The specific parameters are: the output dimensions of the three-layer graph convolution are 512, 512, and 512, respectively; the number of neurons in the fully connected layer is 128 and 2; the Adam optimizer is used; the loss function is binary cross-entropy loss; 'relu' is used as the activation function; 'softmax' is used as the classification function for the output layer; the l2 regularization coefficient is 0.0005, the dropout rate is 0.5, and the learning rate is 0.0001. Those skilled in the art will appreciate that the average pooling operation here can also be replaced by any other pooling operation, such as max pooling or K-max pooling.

[0149] 4.3 MGC_S model construction

[0150] 4.3.1 Second-order adjacency matrix

[0151] Construct a first-order adjacency matrix as shown in method 4.2.1, and then take the amino acids in the first-order adjacency matrix and recalculate the first-order adjacency matrix. That is, recalculate the distance between the amino acid residue pairs. When the alpha carbon atom distance between the two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1. Otherwise, it is 0, and the node itself is recorded as 0. After removing duplicate amino acids in all matrices, a new matrix is ​​generated, namely the second-order adjacency matrix.

[0152] 4.3.2 Node Feature Matrix

[0153] After a protein sequence is input into the pre-trained model, the model first encodes each amino acid in the protein sequence through an encoder. Each amino acid is encoded into a vector of length 1024, that is, a protein sequence of length N will generate a vector of length N*1024. The model has a 24-layer structure. In order to obtain the most information about mutations, this embodiment selects the last layer as the output and simultaneously extracts the 1024-dimensional feature vectors of the wild type and mutant corresponding to the mutation position. The feature vectors of the wild type and mutant are sequentially spliced ​​into 2048 dimensions, and finally N is obtained. s *2048-dimensional features as node feature matrix, where N sThe number of corresponding amino acids obtained from 4.3.1 should be kept consistent. The number of amino acids obtained from different proteins may be different. s As a benchmark, the insufficient number is filled with 0 so that all proteins can obtain a matrix of the same size.

[0154] 4.3.3 Model Construction

[0155] The matrices generated in 4.3.1 and 4.3.2 are input into the graph convolutional neural network as structural features and sequence features respectively. Through matrix multiplication and aggregation operations, each layer of the graph network structure updates the node features into a 512-dimensional feature vector, and the node features output by the three-layer graph network are merged into N through the residual network structure. s *1536 (1536=512*3) dimensional feature vector. Then the average pooling operation is used to compress the updated feature matrix into a 1*1536 dimensional feature vector represented as N s The feature representation of the amino acid residue sequence is input into the fully connected layer. The network consists of a hidden layer with 256 neurons and an output layer with 2 neurons, and ultimately outputs the probabilities of being predicted as a stabilizing mutation and an unstable mutation, respectively. Specific parameters are: the output dimensions of the three-layer graph convolution are 512, 512, 512, respectively; the number of neurons in the fully connected layer is 256 and 2; the optimizer uses Adam; the loss function is binary cross-entropy loss; 'relu' is used as the activation function; 'softmax' is used as the output layer classification function; the l2 regularization coefficient is 0.0001, the dropout rate is 0.3, and the learning rate is 5e-05. Those skilled in the art will understand that the average pooling operation here can also be replaced by any pooling operation such as maximum pooling or K-max pooling.

[0156] 4.4 GINConv model construction

[0157] 4.4.1 First-order adjacency matrix

[0158] Same as 4.2.1 method.

[0159] 4.4.2 Node Feature Matrix

[0160] Same method as 4.2.2.

[0161] 4.4.3 Model Construction

[0162] The matrices generated in 4.4.1 and 4.4.2 are input into the graph convolutional neural network as structural features and sequence features respectively. Through matrix multiplication and aggregation operations, each layer of the graph network structure updates the node features into a 64-dimensional feature vector, and the node features output by the three-layer graph network are merged into N through the residual network structure. f*192-dimensional feature vector. Then, the updated feature matrix is ​​compressed into a 1*192-dimensional feature vector represented by N by using the addition and pooling operation. f The specific parameters are as follows: the output dimensions of the three-layer graph convolution are 64, 64, and 64 respectively; the number of neurons in each fully connected layer is 64, 64, and 2; the aggregator_type of GINConv is 'max'; the initial init_eps value is 0.01; learn_eps is set to False; the optimizer is Adam; the loss function is binary cross entropy loss; 'relu' is used as the activation function; 'softmax' is used as the classification function of the output layer; and the learning rate is 0.01.

[0163] 4.5 SageConv model construction

[0164] 4.5.1 First-order adjacency matrix

[0165] Same as 4.2.1 method.

[0166] 4.5.2 Node Feature Matrix

[0167] Same method as 4.2.2.

[0168] 4.5.3 Model Construction

[0169] The matrices generated in 4.5.1 and 4.5.2 are input into the graph convolutional neural network as structural features and sequence features respectively. Through matrix multiplication and aggregation operations, each layer of the graph network structure updates the node features into a 128-dimensional feature vector, and the node features output by the three-layer graph network are merged into N through the residual network structure. f *384-dimensional feature vector. Then, the updated feature matrix is ​​compressed into a 1*384-dimensional feature vector represented by N by using the addition and pooling operation. f The specific parameters are as follows: the output dimensions of the three-layer graph convolution are 128, 128, and 128, respectively; the number of neurons in each fully connected layer is 256, 256, and 2; in the SAGEConv model, the aggregator_type is 'gcn'; the optimizer is Adam; the loss function is binary cross-entropy loss; 'relu' is used as the activation function; 'softmax' is used as the classification function for the output layer; and the learning rate is 0.01 (SageConv(128,256,'gcn',0.01)).

[0170] Here are some summaries.

[0171] Figure 6 shows a schematic diagram of the structure-based adjacency matrix construction. Figure 7 shows a schematic diagram of the sequence-based node feature matrix construction. Figure 8 shows a schematic diagram of the prediction model framework. Figure 9 is a schematic diagram that combines Figures 6-8.

[0172] According to the above embodiments, it can be seen that in the process of constructing a prediction model, the three-dimensional structural information of the sample protein can be obtained by the following methods: obtaining the coordinates of the three-dimensional structure of the sample protein; calculating the distance between amino acid residue pairs in the three-dimensional structure of the sample protein based on the obtained three-dimensional structure coordinates; when the alpha carbon atom distance between two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1, otherwise it is 0, and the nodes themselves are recorded as 0, thereby obtaining an adjacency matrix for the entire sample protein. Based on the three-dimensional structural information of the sample protein, obtaining the structural features of multiple training samples may include: intercepting an adjacency matrix of a specified size from the adjacency matrix of the entire sample protein with the mutation site as the center, as the structural feature of the sample protein; or, based on the amino acids in the adjacency matrix of the entire protein sample to be predicted, calculating the distance between the amino acid residue pairs; when the alpha carbon atom distance between the two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1, otherwise it is 0, and the nodes themselves are recorded as 0, thereby obtaining a second-order adjacency matrix for the protein sample to be predicted. Training a graph neural network based on sequence features, structural features, and mutation stability labels to obtain a prediction model based on the graph neural network can further include: inputting the sequence features and the structural features into the graph neural network; updating and outputting the features using several layers of graph network structures through matrix multiplication and aggregation operations; merging the features output by the several layers of graph network structures; compressing the features using a pooling operation; classifying the features according to the stability of the output layer classification function; and adjusting the network parameters of the graph neural network using the mutation stability label as the output target. The pooling operation mentioned here can be average pooling, maximum pooling, or K-max pooling.

[0173] Accordingly, in the process of applying the prediction model to protein stability prediction, the three-dimensional structural information of the protein sample to be predicted is obtained by: obtaining the coordinates of the three-dimensional structure of the protein sample to be predicted; calculating the distances between amino acid residue pairs in the protein three-dimensional structure based on the obtained three-dimensional structure coordinates; when the alpha carbon atom distance between two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1; otherwise, it is 0, and the nodes themselves are recorded as 0, thereby obtaining an adjacency matrix for the entire protein sample to be predicted. Based on the three-dimensional structural information of the protein sample to be predicted, the structural characteristics of the protein sample to be predicted are obtained, including: intercepting an adjacency matrix of a specified size from the adjacency matrix of the entire protein sample to be predicted with the mutation site as the center, as the structural characteristics of the protein sample to be predicted; or, based on the amino acids in the adjacency matrix of the entire protein sample to be predicted, calculating the distances between the amino acid residue pairs; when the alpha carbon atom distance between the two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1; otherwise, it is 0, and the nodes themselves are recorded as 0, thereby obtaining a second-order adjacency matrix for the protein sample to be predicted. Based on the sequence features and the structural features, a prediction model based on a graph neural network is used to perform predictions, and determining the stability prediction results of the protein sample to be predicted after mutation further includes: inputting the sequence features and the structural features into the prediction model based on the graph neural network; updating and outputting the features using a plurality of layers of graph network structures through matrix multiplication and aggregation operations; merging the features outputted by the plurality of layers of graph network structures; compressing the features using an average pooling operation; and classifying the features according to the stability of the output layer classification function. Similarly, the pooling operation referred to here can be average pooling, maximum pooling, or K-max pooling.

[0174] 5. Performance indicators

[0175] This example uses the following metrics for evaluation: area under the curve (AUC), precision, and recall. The specific definitions are as follows:

[0176] True Positive (TP)

[0177] False Positive (FP)

[0178] True Negative (TN)

[0179] False Negative (FN)

[0180] Among them, TP (True Positive) and TN (True Negative) represent the number of correctly classified positive samples and negative samples, FP (False Positive) and FN (False Negative) represent the number of incorrectly classified positive samples and negative samples, and P and N represent the number of positive samples and negative samples.

[0181] AUC is a popular parameter-independent metric used to describe binary classifiers. It is the area under the receiver operating characteristic (ROC) curve. Larger values ​​indicate better classifier performance, with a maximum value of 1.

[0182] Precision refers to the proportion of correctly predicted results among all the results predicted as positive samples. The precision calculation formula is: precision = TP / (TP + FP), where TP (True Positive) is the positive sample that is judged to be correctly classified, and FP (False Positive) is the positive sample that is judged to be incorrectly classified.

[0183] Recall refers to the ratio of the number of correctly predicted positive samples to the total number of positive samples. The recall calculation formula is: recall = TP / (TP + FN), where FN (False Negative) is the negative sample that is judged to be misclassified.

[0184] Result evaluation

[0185] 1. Model Performance

[0186] The model was trained on a self-designed balanced dataset containing 11,834 ΔΔG values ​​from single-point mutation experiments for 328 different proteins from four databases. The model performance in five-fold cross-validation is shown in Table 1 below.

[0187] Table 1: Five-fold cross-validation performance indicators of the prediction model of the present disclosure on the training set

[0188] 2. Comparison with other methods

[0189] In order to evaluate the difference between the method of the present invention and other prediction methods, the model of the present invention was compared with other existing excellent prediction methods on the experimental test set, including FoldX (see Guerois, Raphael, Jens Erik Nielsen and Luis Serrano, 2002, "Predicting Changes in the Stability of Proteins and Protein Complexes: A Study of More than 1000 Mutations", Journal of Molecular Biology 320(2):369–87, the entire contents of which are incorporated into the present disclosure by reference, making them part of the contents of the present disclosure), Rosetta (Leman, Julia Koehler et al., 2020, "Macromolecular Modeling and Design in Rosetta: Recent Methods and Frameworks", Nature Methods 17(7):665–80, the entire contents of which are incorporated into the present disclosure by reference, making them part of the contents of the present disclosure), Thermnet (Ofoegbu, Tochukwu C. et al., 2019, “PhyreRisk: A Dynamic Web Application to Bridge Genomics, Proteomics and 3D Structural Data to Guide Interpretation of Human Genetic Variants”, Journal of Molecular Biology 431(13):2460–66, the entire contents of which are hereby incorporated by reference into the present disclosure to become a part of the present disclosure), PremPS (Chen, Yuting et al., 2020, “PremPS: Predicting the Impact of Missense Mutations on Protein Stability”, PLoS computational biology 16(12):e1008543, the entire contents of which are hereby incorporated by reference into the present disclosure to become a part of the present disclosure), I-Mutant (Capriotti, E., P. Fariselli and R. Casadio, 2005, “I-Mutant2.0: Predicting Stability Changes upon Mutation from the Protein Sequence or Structure”, Nucleic Acids Research 33(Web Server):W306–10, the entire contents of which are hereby incorporated by reference into the present disclosure and become a part of the present disclosure) and DDGun (Montanucci, Ludovica et al., 2022, “DDGun: An Untrained Predictor of Protein Stability Changes upon Amino Acid Variants”, Nucleic Acids Research 50(W1): W222–27, the entire contents of which are hereby incorporated by reference into the present disclosure and become a part of the present disclosure). The samples in the experimental test set did not appear in the training sets of this method and the above prediction methods. The results are shown in Table 2 below.

[0190] Table 2: Performance comparison of the disclosed model with other methods on the experimental test set Tm141

[0191] As can be seen from the above table, the graph-based method has a significant advantage in accuracy over the previous method, and the recall rate remains high.

[0192] Table 3: Performance comparison of the disclosed model and other methods on the experimental test set S669

[0193] The results show that the performance of the method disclosed in the present invention on the test set is outstanding compared with other methods.

[0194] 3. Experimental Validation of the Prediction Model

[0195] In order to verify the practical application of the model disclosed in this paper, additional verification was performed by predicting the experimental results available in the literature. The results are shown in Table 4 below:

[0196] Table 4: Test results of the method disclosed in the actual application

[0197] Prediction Process

[0198] Here, we take two cases as examples to describe the process of using a prediction model to predict protein stability.

[0199] In one scenario, the input protein sequence information includes a single protein sequence and mutation site information. First, based on the input single protein sequence and mutation site information, protein mutation site preprocessing is performed. Specifically, the input protein sequence is recorded as the original sequence, and the mutated sequence is obtained based on the mutation site information. Next, amino acid embedding layer extraction, which is the aforementioned sequence information extraction, is performed. Specifically, the original sequence is input into the natural language processing T5 pre-trained model to retrieve embedded features; the mutated protein sequence is input into the natural language processing T5 pre-trained model to retrieve embedded features; amino acid embedding features are extracted based on the mutation site, and the pre- and post-mutation features are concatenated to serve as input for subsequent models. Next, a graph network is constructed, which requires a protein embedding layer feature network and a protein residue interaction feature network. This feature network is applied to the MGC prediction model to predict changes in protein thermal stability before and after amino acid mutation. Finally, the prediction results are output and displayed. In this example, single-point mutation prediction results can be output.

[0200] Another scenario involves inputting protein sequence information that only includes a single protein sequence. First, based on the input single protein sequence, protein mutation site preprocessing is performed. Specifically, the input protein sequence is recorded as the original sequence, and the amino acid at each site in the protein sequence is sequentially mutated to 19 other amino acids, recording them as the post-mutation sequence. Next, amino acid embedding layer extraction, or the aforementioned sequence information extraction, is performed. Specifically, the original sequence is input into a pre-trained natural language processing T5 model to retrieve embedded features; the post-mutation protein sequence is input into a pre-trained natural language processing T5 model to retrieve embedded features. Amino acid embedding features are extracted based on the mutation site, and the pre- and post-mutation features are concatenated to serve as input for subsequent models. Next, a graph network is constructed, requiring a protein embedding layer feature network and a protein residue interaction feature network. This feature network is applied to the MGC prediction model to predict changes in protein thermal stability before and after amino acid mutation. Finally, the prediction results are output and displayed. For this example scenario, a table of all stability mutation probabilities and / or a ranking of all stability mutation likelihoods can be output as the final prediction result.

[0201] in conclusion

[0202] More stable protein mutants are essential for academia and industry to explore more applications for these proteins. This paper proposes a new method for discovering the effects of amino acid mutations on protein stability. Based on advanced natural language processing models combined with graph neural networks, this method can be applied across diverse biological fields and demonstrates superior performance to earlier methods. This method has been successfully applied to predict stability changes caused by mutations in a variety of experiments, demonstrating its applicability in protein engineering.

[0203] The reasons why the method of the present disclosure is superior to the prior art are discussed below.

[0204] From the above description we can see that:

[0205] 1. The method disclosed in this paper is based on the existing latest natural language processing model ProtT5-XL, combined with structure-based features, and successfully applied to protein stability prediction, achieving results that are superior to existing technologies and verified by experiments.

[0206] 2. The method disclosed in this paper is based on graph neural network, which greatly improves the prediction performance of the model.

[0207] 3. The training set of the method disclosed in this invention is collected from a variety of the latest databases, and the sample size is nearly twice as large as before.

[0208] 4. After the training data set used in the method of the present invention is inverted, the data set is expanded and the stable mutation and unstable mutation data are balanced, thereby ensuring the predictive robustness of the model.

[0209] Those skilled in the art will appreciate that, based on the disclosed method for predicting protein stability after mutation, a system for predicting protein stability after mutation can be developed. FIG10 shows a schematic block diagram of a system for predicting protein stability after mutation according to the disclosed method. Specifically, system 1000 for predicting protein stability after mutation includes an acquisition module 1010 and a processing module 1020.

[0210] The acquisition module 1010 is used to obtain the sequence information and three-dimensional structure information of the protein sample to be predicted.

[0211] Processing module 1020 is configured to obtain a stability prediction result for a protein sample to be predicted after mutation by inputting sequence information and three-dimensional structural information of the protein sample to be predicted. As shown in FIG10 , processing module 1020 may further include: a sequence feature submodule 1021 for extracting features using a pre-trained model based on the sequence information of the protein sample to be predicted to obtain sequence features of the protein sample to be predicted; a structural feature submodule 1022 for extracting structural features of the protein sample to be predicted based on the three-dimensional structural information of the protein sample to be predicted; and a model prediction submodule 1023 for performing a prediction using a graph neural network-based prediction model based on the sequence features and structural features to determine a stability prediction result for the protein sample to be predicted after mutation.

[0212] Those skilled in the art will appreciate that the system for predicting protein stability after mutation described above can be implemented by a computer. For example, a computer program can be executed to implement the following operations: obtaining sequence information and three-dimensional structural information of a protein sample to be predicted; inputting the sequence information and three-dimensional structural information of the protein sample to be predicted to obtain a stability prediction result for the protein sample to be predicted after mutation, wherein the operation further includes the following sub-operations: extracting features using a pre-trained model based on the sequence information of the protein sample to be predicted to obtain sequence features of the protein sample to be predicted; obtaining structural features of the protein sample to be predicted based on the three-dimensional structural information of the protein sample to be predicted; and performing prediction using a graph neural network-based prediction model based on the sequence features and structural features to determine a stability prediction result for the protein sample to be predicted after mutation.

[0213] That is to say, although in the system for predicting protein stability based on graph neural networks according to the present disclosure, the various operational functions corresponding to the prediction method are described as modules or sub-modules, those skilled in the art should understand that such modules or sub-modules may be circuit elements or other physical components, or may not be composed of circuit elements or other physical components, but may be functional modules constructed by computer programs.

[0214] In addition, those skilled in the art will recognize that the methods disclosed herein can be implemented as computer programs. As described above in conjunction with the accompanying drawings, the methods of the above embodiments are executed by one or more programs, and the instructions in the programs cause a computer or processor to execute the algorithms described in conjunction with the accompanying drawings. These programs can be stored and provided to a computer or processor using various types of non-transitory computer-readable media. Non-transitory computer-readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (such as floppy disks, magnetic tapes, and hard disk drives), magneto-optical recording media (such as magneto-optical disks), CD-ROMs (compact disk read-only memories), CD-Rs, CD-R / Ws, and semiconductor memories (such as ROMs, PROMs (programmable ROMs), EPROMs (erasable PROMs), flash ROMs, and RAMs (random access memories)). Furthermore, these programs can be provided to a computer using various types of transient computer-readable media. Examples of transient computer-readable media include electrical signals, optical signals, and electromagnetic waves. Transient computer-readable media can be used to provide programs to a computer via wired communication paths such as wires and optical fibers or wireless communication paths.

[0215] For example, according to one embodiment of the present disclosure, a non-transitory computer-readable storage medium may be provided for storing a computer program, wherein the computer program includes instructions that, when executed by a processor of an electronic device, enable the electronic device to implement the method for predicting protein stability after mutation as described above.

[0216] In addition, according to the present disclosure, a computer system may be provided, comprising: a processor; a memory; and a computer program. The computer program is stored in the memory and configured to be executed by the processor. The computer program includes instructions for implementing the method for predicting protein stability after mutation described above.

[0217] On the other hand, according to one embodiment of the present disclosure, a non-transitory computer-readable storage medium may be provided for storing a computer program, wherein the computer program includes instructions that, when executed by a processor of an electronic device, enable the electronic device to implement the method for constructing a protein post-mutation stability prediction model as described above.

[0218] In addition, according to the present disclosure, a computer system may be provided, comprising: a processor; a memory; and a computer program. The computer program is stored in the memory and configured to be executed by the processor. The computer program includes instructions for implementing the method for constructing a protein post-mutation stability prediction model described above.

[0219] The implementation methods of the present disclosure are not limited to those described in the above embodiments. Without departing from the spirit and scope of the present disclosure, ordinary technicians in this field can make various changes and improvements to the present disclosure in form and details, and these are all considered to fall within the scope of protection of the present disclosure.

Claims

1. A method for predicting protein stability after mutation, characterized in that: The method comprises: Based on the sequence information of the protein sample to be predicted, a pre-trained model is used to perform feature extraction to obtain the sequence features of the protein sample to be predicted; Based on the three-dimensional structure information of the protein sample to be predicted, obtaining the structural characteristics of the protein sample to be predicted; Based on the sequence features and the structural features, a prediction model based on a graph neural network is used to perform predictions to determine the stability prediction results of the protein sample to be predicted after mutation.

2. The method according to claim 1, characterized in that The sequence information of the protein sample to be predicted includes the amino acid sequence information of the protein sample to be predicted before mutation and the amino acid sequence information of the protein sample to be predicted after mutation.

3. The method according to claim 1, characterized in that The sequence information of the protein sample to be predicted includes the amino acid sequence information of the protein sample to be predicted before mutation and the designated mutation information.

4. The method according to claim 3, characterized in that The method further comprises: obtaining the amino acid sequence information of the protein sample to be predicted after the mutation according to the amino acid sequence information of the protein sample to be predicted before the mutation and the designated mutation information.

5. The method according to claim 3, characterized in that: The specified mutation information includes a specified mutation site and / or a specified amino acid type after mutation.

6. The method according to claim 2 or 4, characterized in that: The method of extracting features based on the sequence information of the protein sample to be predicted using a pre-trained model to obtain the sequence features of the protein sample to be predicted further includes: Using the pre-trained model to extract features from the amino acid sequence information of the protein sample to be predicted before mutation and the amino acid sequence information after mutation, respectively, to obtain sequence feature information before mutation and sequence feature information after mutation; The sequence feature information before the mutation and the sequence feature information after the mutation are spliced ​​to obtain the sequence feature of the protein sample to be predicted.

7. The method according to claim 1, characterized in that The pre-trained model is implemented by a combination of one or more of the following models: ESM-1b, UniRef, ProteinBert, TAPE, ProtGPT2, ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, ProtT5-XL and ProtT5-XXL; preferably, the pre-trained model is ProtT5-XL or ProtT5-XXL; more preferably, the pre-trained model is ProtT5-XL.

8. The method according to claim 1, characterized in that The three-dimensional structure information of the protein sample to be predicted includes the three-dimensional structure information of the protein sample to be predicted obtained through at least one of the following databases or prediction software: PDB database, AlphaFold2, I-TASSER, RoseTTAFold, Modeller and Swiss-model.

9. The method according to claim 1, characterized in that: The three-dimensional structural information includes interaction network, secondary structure, amino acid residue distance or physical environment; preferably, the three-dimensional structural information includes interaction network.

10. The method according to claim 9, characterized in that The three-dimensional structure information of the protein sample to be predicted is obtained by: Obtaining the coordinates of the three-dimensional structure of the protein sample to be predicted; Calculate the distances between pairs of amino acid residues in the three-dimensional structure of the protein based on the coordinates of the acquired three-dimensional structure; When the alpha carbon atom distance between two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1, otherwise it is 0, and the node itself is recorded as 0, thus obtaining the adjacency matrix of the entire protein sample to be predicted.

11. The method according to claim 10, characterized in that The step of obtaining the structural features of the protein sample to be predicted based on the three-dimensional structural information of the protein sample to be predicted includes: Taking the mutation site as the center, intercepting an adjacency matrix of a specified size from the adjacency matrix of the entire protein sample to be predicted as the structural feature of the protein sample to be predicted; or According to the amino acids in the adjacency matrix of the entire protein sample to be predicted, the distance between the amino acid residue pairs is calculated. When the alpha carbon atom distance between two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1; otherwise, it is 0, and it itself is recorded as 0, thus obtaining the second-order adjacency matrix of the protein sample to be predicted.

12. The method according to claim 1, characterized in that The prediction based on the sequence features and the structural features using a prediction model based on a graph neural network to determine the stability prediction result of the protein sample to be predicted after mutation further includes: Inputting the sequence features and the structural features into a prediction model based on a graph neural network; Through matrix multiplication and aggregation operations, the features are updated and output using several layers of graph network structures; Merge the features output by several layers of graph network structures; Use pooling operation to compress features; The features are classified for stability according to the output layer classification function.

13. The method according to claim 1, characterized in that The graph neural network-based prediction model is constructed through the following steps: Acquire a training set, wherein the training set includes information of a plurality of training samples, and the information of each training sample includes three-dimensional structure information of the sample protein, amino acid sequence information of the sample protein before and after mutation, and a mutation stability label; Based on the amino acid sequence information of the sample protein before and after mutation, feature extraction is performed through a pre-training model to obtain sequence features of the plurality of training samples; Based on the three-dimensional structural information of the sample protein, obtaining structural features of the plurality of training samples; Based on the sequence features, the structural features and the mutation stability labels, the graph neural network is trained to obtain the graph neural network-based prediction model.

14. The method according to claim 13, characterized in that The graph neural network-based prediction model is a classifier.

15. The method according to claim 13, characterized in that The graph neural network is a graph convolutional network.

16. The method according to claim 15, characterized in that The prediction model based on graph neural network includes one of the following models: MGC model, MGC_F model, MGC_S model, GINConv model and SageConv model.

17. The method according to claim 13, characterized in that The step of obtaining the training set includes: Obtain experimental data from a public database as a training set; The training set is doubled in size by inverting the experimental data in the public database.

18. A method for constructing a protein stability prediction model after mutation, characterized in that: The method comprises: Acquire a training set, wherein the training set includes information of a plurality of training samples, and the information of each training sample includes three-dimensional structure information of the sample protein, amino acid sequence information of the sample protein before and after mutation, and a mutation stability label; Based on the amino acid sequence information of the sample protein before and after mutation, feature extraction is performed through a pre-training model to obtain sequence features of the plurality of training samples; Based on the three-dimensional structural information of the sample protein, the structural features of multiple training samples are obtained; Based on the sequence features, the structural features and the mutation stability labels, a graph neural network is trained to obtain a prediction model based on the graph neural network.

19. The method according to claim 18, characterized in that The method of extracting features based on the amino acid sequence information of the sample protein before and after mutation by using a pre-training model to obtain sequence features of the plurality of training samples includes: The amino acid sequence information of the sample protein before and after mutation is respectively subjected to feature extraction using the pre-trained model to obtain sequence feature information before and after mutation; The sequence feature information before the mutation and the sequence feature information after the mutation are concatenated to obtain the sequence features of the multiple training samples.

20. The method according to claim 18, characterized in that The pre-trained model is implemented by a combination of one or more of the following models: ESM-1b, UniRef, ProteinBert, TAPE, ProtGPT2, ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, ProtT5-XL and ProtT5-XXL; preferably, the pre-trained model is ProtT5-XL or ProtT5-XXL; more preferably, the pre-trained model is ProtT5-XL.

21. The method according to claim 18, characterized in that The three-dimensional structural information of the sample protein includes the three-dimensional structural information of the sample protein obtained through at least one of the following databases or prediction software: PDB database, AlphaFold2, I-TASSER, RoseTTAFold, Modeller and Swiss-model.

22. The method according to claim 18, characterized in that The three-dimensional structural information includes interaction network, secondary structure, amino acid residue distance or physical environment; preferably, the three-dimensional structural information includes interaction network.

23. The method according to claim 22, characterized in that The three-dimensional structural information of the sample protein is obtained by: Obtain the coordinates of the three-dimensional structure of the sample protein; Calculate the distances between pairs of amino acid residues in the three-dimensional structure of the sample protein based on the coordinates of the acquired three-dimensional structure; When the alpha carbon atom distance between two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1, otherwise it is 0, and the node itself is recorded as 0, and the adjacency matrix of the entire sample protein is obtained.

24. The method according to claim 23, characterized in that The step of obtaining the structural features of multiple training samples based on the three-dimensional structural information of the sample protein includes: Taking the mutation site as the center, extracting an adjacency matrix of a specified size from the adjacency matrix of the entire sample protein as the structural feature of the sample protein; or According to the amino acids in the adjacency matrix of the entire protein sample to be predicted, the distance between the amino acid residue pairs is calculated. When the alpha carbon atom distance between two residue pairs is less than 10 angstroms, it indicates that there is an edge relationship between the nodes, which is recorded as 1; otherwise, it is 0, and it itself is recorded as 0, thus obtaining the second-order adjacency matrix of the protein sample to be predicted.

25. The method according to claim 18, characterized in that The training of the graph neural network based on the sequence features, the structural features and the mutation stability labels to obtain a prediction model based on the graph neural network includes: Inputting the sequence features and the structural features into a graph neural network; Through matrix multiplication and aggregation operations, the features are updated and output using several layers of graph network structures; Merge the features output by several layers of graph network structures; Use pooling operation to compress features; Classify the stability of features according to the output layer classification function; The mutation stability label is used as an output target to adjust the network parameters of the graph neural network.

26. The method according to claim 18, characterized in that The graph neural network-based prediction model is a classifier.

27. The method according to claim 18, characterized in that The graph neural network is a graph convolutional network.

28. The method according to claim 27, characterized in that The prediction model based on graph neural network includes one of the following models: MGC model, MGC_F model, MGC_S model, GINConv model and SageConv model.

29. The method according to claim 18, characterized in that The step of obtaining the training set includes: Obtain experimental data from a public database as a training set; The training set is doubled in size by inverting the experimental data in the public database.

30. A system for predicting protein stability after mutation, characterized in that: The system comprises: An acquisition module is used to obtain the sequence information and three-dimensional structure information of the protein sample to be predicted; The processing module is used to obtain the stability prediction result of the protein sample to be predicted after mutation by inputting the sequence information and three-dimensional structure information of the protein sample to be predicted. The processing module further includes the following submodules: A sequence feature submodule, used to extract features based on the sequence information of the protein sample to be predicted using a pre-trained model to obtain the sequence features of the protein sample to be predicted; A structural feature submodule, used for obtaining the structural features of the protein sample to be predicted based on the three-dimensional structural information of the protein sample to be predicted; The model prediction submodule is used to predict the stability of the protein sample to be predicted after mutation based on the sequence features and the structural features using a prediction model based on a graph neural network.

31. A computer-implemented system for predicting protein stability after mutation, wherein the computer-implemented system implements the following operations by executing a computer program: Obtaining the sequence information and three-dimensional structure information of the protein sample to be predicted; By inputting the sequence information and three-dimensional structure information of the protein sample to be predicted, the stability prediction result of the protein sample to be predicted after mutation is obtained. This operation further includes the following sub-operations: Based on the sequence information of the protein sample to be predicted, using a pre-trained model to perform feature extraction to obtain sequence features of the protein sample to be predicted; Based on the three-dimensional structural information of the protein sample to be predicted, obtaining the structural characteristics of the protein sample to be predicted; Based on the sequence features and the structural features, a prediction model based on a graph neural network is used to perform predictions to determine the stability prediction results of the protein sample to be predicted after mutation.

32. A non-transitory computer-readable storage medium for storing a computer program, wherein the computer program includes instructions, which, when executed by a processor of an electronic device, causes the electronic device to implement the method for predicting protein stability after mutation as described in any one of claims 1-17 or the method for constructing a protein stability prediction model after mutation as described in any one of claims 18-29.

33. A computer system, comprising: processor; Memory; and A computer program, wherein the computer program is stored in the memory and is configured to be executed by the processor, the computer program comprising instructions for implementing the method for predicting protein stability after mutation as described in any one of claims 1 to 17 or the method for constructing a protein stability prediction model after mutation as described in any one of claims 18 to 29.

Images