Prediction for influence of mutation on protease activity

Abstract

The present disclosure relates to a prediction for the influence of a mutation on protease activity. A method for predicting the influence of a mutation on protease activity comprises: on the basis of sequence information of a protease sample to be subjected to prediction, performing feature extraction by using a pre-training model, so as to obtain a sequence feature of said protease sample (S110); on the basis of three-dimensional structure information of said protease sample, obtaining a structural feature of said protease sample (S120); and on the basis of the sequence feature and the structural feature, performing prediction by using a prediction model based on a graph neural network, and determining an activity change prediction result after said protease sample mutates (S130). Further disclosed is a method for constructing a prediction model used for predicting the influence of a mutation on protease activity. By means of the method in the present disclosure, prediction is performed on the basis of an advanced natural language processing model and a graph neural network, in combination with a sequence and a three-dimensional structure, and by means of an artificial intelligence model, showing a performance superior to that of other existing methods, and proving the wide application prospects of the method in protease engineering.

Description

Predicting the effects of mutations on protease activity

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to Chinese patent application number 202311028534.0 filed with the Chinese Patent Office on August 15, 2023, entitled “Predicting the Effect of Mutations on Protease Activity,” and incorporates its entire contents into this application by reference. Technical Field

[0003] The present disclosure relates to protease optimization, and more particularly to predicting the effects of amino acid mutations on protease activity. Background Art

[0004] Enzymes are important biocatalysts. Due to their high selectivity, biocompatibility, and mild reaction properties, they have broad application prospects in industrial biocatalysis, especially in the production of pharmaceutical intermediates. The vast majority of enzymes are proteins, and the molecular structure and function of proteins are affected by factors such as temperature, pH, and activators. Therefore, the catalytic activity of proteases is only manifested under certain conditions. Engineering modifications are often required to adapt proteases to different production environments and functional requirements.

[0005] Directed evolution and rational design are two common strategies for protease optimization. Directed evolution has been successfully applied to modify protease properties such as activity, stability, substrate specificity, and stereoselectivity, earning American scientist Frances H. Arnold the 2018 Nobel Prize in Chemistry for this work. However, directed evolution requires the construction of large-scale mutant libraries and the establishment of high-throughput screening methods, which consumes significant human, material, and financial resources. Directed evolution also struggles to comprehensively search sequence space, a fundamental flaw. Rational design relies on an understanding of the structure-function relationship of proteases to predict potential mutants, then constructs mutants in the target gene through site-directed mutagenesis. In essence, rational design is more efficient for modifying proteases than directed evolution. Furthermore, rational design methods are universally applicable; an effective rational design strategy can be applied to modify a wide range of proteases. However, the accuracy of current rational design methods is generally low, and their application is far less extensive than that of directed evolution. In recent years, artificial intelligence-assisted protein engineering has gradually developed into a new and efficient protein design strategy, showing unique advantages in many aspects such as protein structure prediction, function prediction, stability prediction and antibody affinity prediction, becoming another technological wave after rational design and directed evolution.

[0006] In recent years, with the rapid development of computer technology, various neural network models are being rapidly updated and iterated in a more accurate and efficient direction. Artificial intelligence (AI)-assisted enzyme engineering has gradually developed into a new and efficient protease design strategy. In order to cope with the explosive growth of protein sequence databases brought about by large-scale gene sequencing, the number of parameters of the current largest protein language model has reached 15 billion (see Lin, Zeming et al. 2023. “Evolutionary-scale prediction of atomic-level protein structure with a language model”. Science 379(6637):1123–30. The entire contents of the above-mentioned documents are hereby incorporated into the present disclosure by reference and become part of the contents of the present disclosure). The continuous development of high-performance computing and natural language processing (NLP) has enabled researchers to use large protein databases to enhance the prediction of sequence attributes or annotations of relatively small experimental data sets.

[0007] After decades of development, there are dozens of AI-based methods for predicting protein stability, and their performance has surpassed traditional rational design methods. Unlike the stability prediction of proteases, currently only the SCANEER (Sequence Co-Evolutionary Analysis To Control Efficiency Of Enzyme Reactions) method is based on sequence co-evolutionary analysis to find evolutionarily available alternative amino acids to improve protease activity (see Kim, Donghyo et al. 2022. “Enzyme activity engineering based on sequence co-evolution analysis”. Metabolic Engineering 74: 49–60. The entire contents of the above-mentioned documents are incorporated into this disclosure by reference, making them part of the contents of this disclosure). SCANEER evaluates the co-evolutionary relationship of amino acid pairs in multiple sequence alignments (MSA). Amino acid pairs that are not observed or rarely observed in MSA, i.e., amino acid replacements, cannot be predicted, ultimately resulting in the number of predictable mutations for each protease accounting for only 47% of all possible mutations, which will greatly limit the scope of application of this method. There is still a gap in how to apply the current advanced AI methods to the prediction method of protease activity.

[0008] Currently, the application of natural language processing to the prediction of protease activity encounters many problems. First, because artificial intelligence algorithms rely heavily on data, the quantity and quality of the initial data determine the generalization performance of the trained model (see Usmanova, Dinara R et al. 2018. “Self-consistency test reveals systematic bias in programs for prediction change of stability upon mutation”. Bioinformatics 34(21): 3653–58. The entire content of the above-mentioned document is incorporated into the present disclosure by reference, making it a part of the content of the present disclosure). Insufficient number of samples in the data set or too low quality will cause the model to overfit or underfit, which may be one of the important factors limiting the prediction of protease activity based on AI. At the same time, due to the bias of protease evolution and experiments, the data set is unbalanced, which in turn affects the performance of the prediction model. Secondly, for different substrates under different experimental conditions such as temperature and pH, the same mutation will have different effects on the activity of the protease. This multi-label data also brings great difficulties to the prediction work.

[0009] Summary of the Invention

[0010] Stability and catalytic activity are the two most important properties of proteases, among which stability is the most studied. Dozens of AI-based methods have been developed to predict the effect of mutations on protease stability and have excellent performance. However, catalytic activity has been limited by problems with data sets and methods and has stagnated to date. Only the SCANEER method (as cited above) can predict the effect of single-point mutations on protease activity. However, this method is limited because it can only predict mutations that have occurred in evolution based on co-evolutionary information and has many limitations and cannot be widely used. Based on the above-mentioned defects in the current methods for predicting the effect of mutations on protease activity, the present disclosure proposes a method for predicting the effect of mutations on protease activity. It is a deep learning method based on new natural language processing methods and graph structured data. It uses a T5 (i.e., Transfer Text-to-Text Transformer) pre-training model to extract sequence features and a graph neural network model (such as a graph attention network, Graph Attention Network (GAT)) to extract structural features. The protein language model and the graph attention network model are successfully used to predict changes in protease activity caused by single-point missense mutations.

[0011] According to a first aspect of the present disclosure, a method for predicting the effect of a mutation on protease activity is provided. The method may include the following steps: extracting features using a pre-trained model based on sequence information of a protease sample to obtain sequence features of the protease sample; obtaining structural features of the protease sample based on three-dimensional structural information of the protease sample; and performing a prediction using a graph neural network-based prediction model based on the sequence features and structural features to determine a predicted result of the activity change of the protease sample after the mutation.

[0012] In the method according to the first aspect of the present disclosure, the sequence information of the protease sample to be predicted may include the amino acid sequence information of the protease sample to be predicted before mutation and the amino acid sequence information of the protease sample to be predicted after mutation.

[0013] On the other hand, the sequence information of the protease sample to be predicted may include the amino acid sequence information of the protease sample to be predicted before mutation and the designated mutation information. Thus, based on the amino acid sequence information of the protease sample to be predicted before mutation and the designated mutation information, the amino acid sequence information of the protease sample to be predicted after mutation can be obtained.

[0014] Preferably, the specified mutation information includes a specified mutation site and / or a specified amino acid type after mutation.

[0015] In the method according to the first aspect of the present disclosure, the method of extracting features based on the sequence information of the protease sample to be predicted using a pre-trained model to obtain the sequence features of the protease sample to be predicted may further include: extracting features from the amino acid sequence information of the protease sample to be predicted before mutation and the amino acid sequence information after mutation using the pre-trained model 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 protease sample to be predicted.

[0016] 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.

[0017] In the method according to the first aspect of the present disclosure, the three-dimensional structural information of the protease sample to be predicted may include the three-dimensional structural information of the protease 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.

[0018] In the method according to the first aspect of the present disclosure, the three-dimensional structural information of the protease sample to be predicted may include interaction network, secondary structure, amino acid residue distance or physical environment. Preferably, the three-dimensional structural information of the protease sample to be predicted includes interaction network.

[0019] Preferably, the three-dimensional structural information of the protease sample to be predicted is obtained by: obtaining the coordinates of the three-dimensional structure of the protease sample to be predicted; calculating the distance between the amino acid residue pairs in the protease three-dimensional structure based on the coordinates of the obtained three-dimensional structure; 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 itself is recorded as 0, thereby obtaining the adjacency matrix of the entire protease sample to be predicted.

[0020] Correspondingly, in the method according to first aspect of the present disclosure, the described three-dimensional structure information based on protease sample to be predicted, obtaining the structural characteristic of described protease sample to be predicted can comprise: take the mutation site as center, from the adjacency matrix of described whole protease sample to be predicted, intercept the adjacency matrix of specified size, as the structural characteristic of described protease sample to be predicted.In some embodiments, take the mutation site as center, from the adjacency matrix of whole protease sample to be predicted, intercept the adjacency matrix of N*N (N is an integer greater than or equal to 1). From the adjacency matrix of whole protease sample to be predicted, intercepting the adjacency matrix of N*N can be based on the sequence of mutation site left and right sides to intercept, also can be based on and the distance between mutation site to intercept.In one embodiment, take the mutation site as center, select the amino acid less than 10 dusts in the adjacency matrix of whole protease sample to be predicted with mutation site distance, as the structural characteristic of protease sample to be predicted.

[0021] In the method according to the first aspect of the present disclosure, preferably, the prediction based on the sequence features and the structural features using a prediction model based on a graph neural network to determine the prediction result of the activity change of the protease sample to be predicted after mutation can further include: inputting the sequence features and the structural features into the graph neural network to obtain probability features about the activity change; combining the probability features of the activity change with supplementary features from other prediction methods, and inputting them into a decision tree model to finally obtain a classification about the activity change.

[0022] Preferably, the step of inputting the sequence features and the structural features into a graph neural network to obtain probabilistic features regarding activity changes may include: inputting the sequence features and the structural features into a 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 the several layers of graph network structures; compressing the features using a pooling operation; classifying the features according to activity changes according to the output layer classification function, and outputting probabilistic features regarding activity changes. Preferably, the pooling operation here may be average pooling, maximum pooling, or K-max pooling. The probabilistic features of activity changes here may be the probability value of increased activity after a protease sample mutates.

[0023] 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 three-dimensional structural information of the sample protease, amino acid sequence information of the sample protease before and after mutation, and a mutation activity change label; based on the amino acid sequence information of the sample protease before and after mutation, feature extraction is performed through a pre-training model to obtain sequence features of the multiple training samples; based on the three-dimensional structural information of the sample protease, the structural features of the multiple training samples are obtained; based on the sequence features, the structural features and the mutation activity change labels, the graph neural network is trained to obtain the prediction model based on the graph neural network.

[0024] Preferably, the graph neural network-based prediction model is a classifier.

[0025] Preferably, the graph neural network is a graph convolutional network or a graph attention network.

[0026] Preferably, the training of the graph neural network based on the sequence features, the structural features and the mutation activity change labels to obtain the prediction model based on the graph neural network may further include: inputting the sequence features and the structural features into the graph neural network to obtain probability features about activity changes; combining the probability features of the activity changes with supplementary features from other prediction methods, and inputting them into a decision tree model to obtain a classification about activity changes; and adjusting the parameters of the prediction model with the mutation activity change label as the output target.

[0027] Preferably, the decision tree model is a gradient boosted decision tree (GBDT) model. More preferably, the GBDT model is implemented using the LightGBM framework.

[0028] 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.

[0029] According to a second aspect of the present disclosure, a method for constructing a prediction model for predicting the effect of mutations on protease activity is provided. The method may include the following steps: obtaining a training set, the training set including information of multiple training samples, the information of each training sample including three-dimensional structural information of the sample protease, amino acid sequence information of the sample protease before and after mutation, and a mutation activity change label; based on the amino acid sequence information of the sample protease before and after mutation, performing feature extraction using a pre-training model to obtain sequence features of the multiple training samples; based on the three-dimensional structural information of the sample protease, obtaining structural features of the multiple training samples; based on the sequence features, the structural features, and the mutation activity change labels, training a graph neural network to obtain a prediction model for predicting the effect of mutations on protease activity.

[0030] 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 protease before and after mutation is performed through a pre-training model to obtain the sequence features of the multiple training samples may include: using the pre-training model to extract features of the amino acid sequence information of the sample protease 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.

[0031] 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.

[0032] In the method according to the second aspect of the present disclosure, the three-dimensional structural information of the sample protease may include the three-dimensional structural information of the sample protease obtained through at least one of the following databases or prediction software: PDB database, AlphaFold2, I-TASSER, RoseTTAFold, Modeller and Swiss-model.

[0033] In the method according to the second aspect of the present disclosure, the three-dimensional structural information of the sample protease may include an interaction network, a secondary structure, an amino acid residue distance or a physical environment. Preferably, the three-dimensional structural information of the sample protease includes an interaction network.

[0034] Preferably, the three-dimensional structural information of the sample protease can be obtained in the following manner: obtaining the coordinates of the three-dimensional structure of the sample protease; calculating the distance between the amino acid residue pairs in the three-dimensional structure of the sample protease based on the obtained coordinates of the three-dimensional structure; 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 itself is recorded as 0, thereby obtaining the adjacency matrix of the entire sample protease.

[0035] Correspondingly, in the method according to second aspect of the present disclosure, the three-dimensional structural information based on the sample protease, obtaining the structural characteristics of multiple training samples can include: taking the mutation site as the center, intercepting the adjacency matrix of specified size from the adjacency matrix of the whole sample protease as the structural characteristics of the sample protease. In some embodiments, taking the mutation site as the center, intercepting the adjacency matrix of N*N (N is an integer greater than or equal to 1) from the adjacency matrix of the whole sample protease. Intercepting the adjacency matrix of N*N from the adjacency matrix of the whole sample protease can be intercepted based on the sequence on the left and right sides of the mutation site, or can be intercepted based on the distance between the mutation site. In one embodiment, taking the mutation site as the center, select the amino acid less than 10 angstroms in the adjacency matrix of the whole sample protease from the mutation site distance as the structural characteristics of the sample protease.

[0036] In the method according to the second aspect of the present disclosure, the training of the graph neural network based on the sequence features, the structural features and the mutation activity change labels to obtain a prediction model for predicting the effect of mutations on protease activity may further include: inputting the sequence features and the structural features into the graph neural network to obtain probability features about activity changes; combining the probability features of activity changes with supplementary features from other prediction methods and inputting them into a decision tree model to obtain a classification about activity changes; and adjusting the parameters of the prediction model using the mutation activity change label as the output target.

[0037] Preferably, inputting the sequence features and the structural features into a graph neural network to obtain probabilistic features regarding activity changes may include: inputting the sequence features and the structural features into a graph neural network; updating and outputting the features using a plurality of graph network layers through matrix multiplication and aggregation operations; merging the features outputted by the plurality of graph network layers; compressing the features using a pooling operation; and classifying the features according to activity changes according to an output layer classification function, and outputting probabilistic features regarding activity changes. Preferably, the pooling operation may be average pooling, maximum pooling, or K-max pooling.

[0038] Preferably, the prediction model is a classifier.

[0039] Preferably, the graph neural network is a graph convolutional network or a graph attention network.

[0040] Preferably, the decision tree model is a GBDT model. More preferably, the GBDT model is implemented using the LightGBM framework.

[0041] 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.

[0042] According to the third aspect of the present disclosure, a system for predicting the effect of mutations on protease activity is provided. The system may include: an acquisition module for acquiring sequence information and three-dimensional structural information of a protease sample to be predicted; a processing module for obtaining a prediction result of the activity change of the protease sample to be predicted after mutation by inputting the sequence information and three-dimensional structural information of the protease 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 protease sample to be predicted to obtain the sequence features of the protease sample to be predicted; a structural feature submodule for obtaining the structural features of the protease sample to be predicted based on the three-dimensional structural information of the protease sample to be predicted; and 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 prediction result of the activity change of the protease sample to be predicted after mutation.

[0043] According to the third aspect of the present disclosure, the system for predicting the effect of mutation on protease activity can be implemented by a computer. Preferably, the following operations can be implemented by executing a computer program: obtaining the sequence information and three-dimensional structure information of the protease sample to be predicted; obtaining the activity change prediction result of the protease sample to be predicted after mutation by inputting the sequence information and three-dimensional structure information of the protease sample to be predicted, and the operation further includes the following sub-operations: based on the sequence information of the protease sample to be predicted, using a pre-trained model to perform feature extraction to obtain the sequence characteristics of the protease sample to be predicted; based on the three-dimensional structure information of the protease sample to be predicted, obtaining the structural characteristics of the protease sample to be predicted; based on the sequence characteristics and the structural characteristics, using a prediction model based on a graph neural network to perform prediction to determine the activity change prediction result of the protease sample to be predicted after mutation.

[0044] 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 the effect of a mutation on protease activity according to the first aspect of the present disclosure or the method for constructing a prediction model for predicting the effect of a mutation on protease activity according to the second aspect of the present disclosure.

[0045] 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 the effect of a mutation on protease activity according to the first aspect of the present disclosure or the method for constructing a prediction model for predicting the effect of a mutation on protease activity according to the second aspect of the present disclosure.

[0046] The above method is characterized by establishing a connection between the primary sequence and three-dimensional structure of the protease through a graph neural network, particularly a graph attention network (GAT), and then predicting the mutation effect. Therefore, it is possible to provide new insights into the annotation challenges brought by the increasing sequence and structure data. For the problem of low quality and bias in experimental data, the present invention extracts and filters out 5449 data containing experimental mutation information from the D3DistalMutation database, and reverses the data set (see below for the specific operation description of the reversal), which means that the data set finally used for training contains 10998 mutation information, and the number of mutations that cause reduced activity is equal to the number of mutations that cause increased activity. The model finally obtained can not only predict mutations at any position in the specified protease sequence, but also exceed expectations in the performance of the experimental test set. This method can be combined with a method for predicting protein stability after mutation to solve this negative correlation between protease stability and activity, the so-called activity-stability trade-off mechanism. It can not only improve the stability of the protease, but also solve the problem that the activity may decrease, and has a wide range of applications.

[0047] The beneficial technical effects of the method disclosed herein include:

[0048] 1. Highly innovative: Based on a self-constructed dataset, the latest AI method is used to predict the impact of mutations on protease activity, filling a gap in this field.

[0049] 2. Superior performance: As can be seen from the comparison results between the method in the examples and the only SCANEER in the following text, the method disclosed in the present invention has better performance.

[0050] 3. Broad Application Prospects: This method can predict the effects of mutations on protease activity that are currently unpredictable using SCANEER. These unpredictable mutations are particularly valuable because SCANEER's predictions overlap significantly with those of rationally designed methods, making the resulting results susceptible to limitations imposed by other patents or literature. However, the disclosed method can identify beneficial mutations that are difficult to detect using evolutionary data. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] 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:

[0052] FIG1 is a flow chart of a method for predicting the effect of a mutation on protease activity according to an embodiment of the present disclosure.

[0053] 2 is a schematic diagram illustrating the association between a method for predicting the effect of a mutation on protease activity and a method for constructing a prediction model for predicting the effect of a mutation on protease activity according to an embodiment of the present disclosure.

[0054] FIG3 is a flow chart of the method for constructing a prediction model used in the present disclosure for predicting the effect of mutations on protease activity.

[0055] Figure 4 is a schematic diagram of the adjacency matrix.

[0056] FIG5 is a schematic diagram of forming a node feature matrix.

[0057] FIG6 is a schematic diagram of the construction of the structure-based adjacency matrix.

[0058] FIG7 is a schematic diagram of constructing a sequence-based node feature matrix.

[0059] FIG8 is a schematic diagram of the framework of the prediction model.

[0060] FIG9 shows a schematic block diagram of a system for predicting the effect of mutations on protease activity according to the present disclosure. DETAILED DESCRIPTION

[0061] 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.

[0062] 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.

[0063] From the overall perspective of the prediction method, FIG1 shows a flow chart of a method for predicting the effect of a mutation on protease activity according to an embodiment of the present disclosure.

[0064] As shown in FIG1 , the method 100 for predicting the effect of mutations on protease activity according to an embodiment of the present disclosure begins at step S110 . In this step, based on the sequence information of the protease sample to be predicted, a pre-trained model is used to perform feature extraction to obtain the sequence features of the protease sample to be predicted.

[0065] The sequence information of the protease sample to be predicted mentioned herein may include the amino acid sequence information of the protease sample to be predicted before mutation and the amino acid sequence information of the protease sample to be predicted after mutation.

[0066] In another case, the sequence information of the protease sample to be predicted may also include the amino acid sequence information of the protease sample to be predicted before mutation and the specified mutation information. In this case, the amino acid sequence information of the protease sample to be predicted after mutation can be obtained based on the amino acid sequence information of the protease sample to be predicted before mutation and the specified mutation information. The specified mutation information may include a specified mutation site and / or a specified amino acid type after mutation. For example, if the specified mutation information is a specified mutation site, the amino acid at the site may be changed to form the amino acid sequence information of the protease sample to be predicted after mutation. That is, if the specified mutation information does not include the specified 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. If the specified mutation information also includes the specified amino acid type after mutation, the amino acid sequence information after mutation may be a single amino acid sequence. Another case is that the specified mutation information only includes the specified amino acid type after mutation but does not include the mutation site. In this case, each site of the sequence before mutation needs to be replaced with the specified amino acid type in sequence 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. The "amino acid types" mentioned here refer to the 20 different amino acids that make up the main units of proteins in living organisms, including glycine, alanine, valine, leucine, isoleucine, methionine (methionine), proline, tryptophan, serine, tyrosine, cysteine, phenylalanine, asparagine, glutamine, threonine, aspartic acid, glutamic acid, lysine, arginine and histidine.

[0067] Those skilled in the art should understand that in subsequent steps, it is necessary to extract features of the amino acid sequences before and after mutation respectively, and finally splice them into features about the amino acid sequence. Therefore, no matter what the protease sequence information initially obtained is, when executing step S110, it is necessary to form the amino acid sequences before and after mutation.

[0068] Specifically, in step S110, the amino acid sequence information of the protease sample to be predicted before and after 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 protease sample to be predicted.

[0069] 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.

[0070] 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.

[0071] The purpose of step S110 is to obtain the sequence characteristics of the protease sample to be predicted, while the purpose of step S120 to be described next is to obtain the structural characteristics of the protease sample to be predicted.

[0072] In step S120 , the structural features of the protease sample to be predicted are obtained based on the three-dimensional structural information of the protease sample to be predicted.

[0073] In addition to obtaining relevant features through pre-training models, relevant features can also be obtained by combining the three-dimensional structural information of proteases for supplementary information. The three-dimensional structural information can be obtained through the PDB database (https: / / www.rcsb.org / ) (see Berman, Helen M et al. 2000. "The protein data bank". Nucleic acids research 28 (1): 235–42. The entire contents of the above document are hereby incorporated into the present disclosure by reference, making them part of the contents of the present disclosure) or obtained through prediction software, such as the following prediction software (reference introduction or source is in brackets):

[0074] AlphaFold2(https: / / alphafold.ebi.ac.uk / );

[0075] I-TASSER(https: / / zhanggroup.org / I-TASSER / );

[0076] RoseTTAFold (see Minkyung Baek et al. "Accurate prediction of protein structures and interactions using a three-track neural network".Science 373,871-876(2021).DOI:10.1126 / science.abj8754);

[0077] Modeller(https: / / salilab.org / modeller / );

[0078] Swiss-model (https: / / swissmodel.expasy.org / ) etc.

[0079] The available three-dimensional structural information for proteases typically includes secondary structure, interaction networks, amino acid residue distances, physical environment, and energy information derived from the three-dimensional structure. The interaction network, which contains contact information between all residue pairs, is one of the most commonly used features in structure-based prediction methods.

[0080] 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 a prediction model based on a graph neural network (GNN) and its various variants is selected.

[0081] 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.

[0082] 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 activity change prediction result of the protease sample to be predicted after mutation.

[0083] The prediction model based on graph neural network is actually a classifier.

[0084] 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. 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.

[0085] In a preferred embodiment of the present disclosure, the graph neural network is a graph convolutional network or a graph attention network.

[0086] In addition, in a preferred embodiment of the present disclosure, in step S130, the sequence features and the structural features can be first input into a graph neural network to obtain probabilistic features about activity changes; then, the probabilistic features of activity changes are combined with supplementary features from other prediction methods (see detailed description below) and input into a decision tree model to finally obtain a classification about activity changes. In a more preferred embodiment, the decision tree model used is a gradient boosting decision tree model. The decision tree model is implemented using the LightGBM framework. A more detailed explanation will be given later in the description of the embodiments.

[0087] For the method of constructing the prediction model, please refer to the detailed description below.

[0088] As mentioned above, stability and catalytic activity are the two most important properties of proteases. Stability has been the subject of extensive research, and dozens of AI-based methods have been developed to predict the impact of mutations on protease stability with excellent performance. However, catalytic activity has been limited by dataset and methodological issues and has stagnated to date. Only the SCANEER method can predict the effect of single-point mutations on protease activity. However, this method, based on co-evolutionary information, can only predict mutations that have occurred during evolution, which has many limitations and cannot be widely used.

[0089] The difference between the technology provided by the present disclosure and the above-mentioned prior art is that the technology of the present disclosure introduces a graph neural network, particularly a graph convolutional network or a graph attention network, into the prediction of the activity change after the protease mutation, and is combined with a mature protein natural language model (pre-training model) in the prior art to act together on the determination of the prediction result. In other words, the sequence characteristics (obtained according to the primary sequence) of the protease sample to be predicted are combined with the structural characteristics (obtained according to the three-dimensional structure), and are jointly used as the input of the prediction model based on the graph neural network, so as to output the prediction result of the activity change after the protease mutation. The prediction result determined in this way is better than the prediction result determined by the method of the prior art.

[0090] FIG2 shows a schematic diagram 200 illustrating the association between a method for predicting the effect of a mutation on protease activity and a method for constructing a prediction model for predicting the effect of a mutation on protease activity according to an embodiment of the present disclosure. In the schematic diagram 200 shown in FIG2 , the left side of the dotted line shows a method for constructing a prediction model based on a graph neural network, while the right side of the dotted line shows a method for performing predictions using a prediction model based on a graph neural network.

[0091] The process on the right side of the dotted line in Figure 2 is actually the method for predicting the effect of mutations on protease activity according to an embodiment of the present disclosure shown in Figure 1. The process on the left side of the dotted line in Figure 2 will be described in detail below (for example, Figure 3 and its corresponding text description). As can be seen from Figure 2, both model construction and the actual prediction process require the use of two feature extraction steps, namely, the extraction of sequence features and structural features. For the training sample, it includes the three-dimensional structural information of the sample protease, the amino acid sequence information of the sample protease before and after mutation, and also includes a mutation activity change label. As described later, the mutation activity change label can be a change in the activity of the protease that increases or decreases (increases or decreases). Therefore, in the model construction process, the sequence information in the training sample is used to extract sequence features. The structural features are extracted using the three-dimensional structural information. Using the sequence features and structural features as input data and the mutation activity change 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 prediction model for predicting the effect of mutations on protease activity. With the constructed prediction model, it can be put into the actual work of predicting the change in protease mutation activity, i.e., the process on the right side of the dotted line in Figure 2. In addition, it should be noted that in Figure 2, for sequence information and sequence features, hollow arrows are used to represent their signal transmission process; while for three-dimensional structure information and structural features, different solid arrows are used to represent their signal transmission process; in addition, other signal flows and control flows are also represented by solid arrows.

[0092] The following is a detailed description of the process of the method for constructing a prediction model for predicting the effect of mutations on protease activity according to an embodiment of the present disclosure.

[0093] FIG3 is a flow chart of a method 300 for constructing a prediction model for predicting the effect of mutations on protease activity used in the present disclosure.

[0094] 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 of multiple training samples. Specifically, the information of each training sample includes the three-dimensional structure information of the sample protease, the amino acid sequence information of the sample protease before and after mutation, and a mutation activity change signature.

[0095] Artificial intelligence (AI) algorithms rely heavily on data. The quantity and quality of the 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 samples in the dataset can lead to overfitting or underfitting of the model. To address this problem, the most effective method is to obtain more experimental data as a training set. The main source of experimental data currently available is public databases. For sequence-based methods to predict the impact of mutations on protease catalytic activity, the dataset can be doubled in size by inverting the data in the database. Inversion means that under the same environmental conditions, opposite mutations at the same position have opposite activity effects. According to the above definition, merging the original data with the inverted data can yield twice the original data volume. By using this inversion method, the original dataset is expanded while also being balanced, making the final prediction model more robust.

[0096] It should be noted that this method of expanding the dataset by reversing the original dataset is only implemented for the training set.

[0097] In the specific embodiments below, the acquisition of the training data set will be described in more detail.

[0098] The mutation activity change label can be used to reflect the influence trend of amino acid mutation on the catalytic activity of protease. For example, the mutation activity change label can be that catalytic activity becomes stronger or weaker (activity rises or falls). More specifically, the description about building a training data set can be found in the specific embodiments below. It is understandable that, using the mutation activity change information as the label of the training sample, the prediction model obtained by training can output the prediction result about the influence of amino acid mutation on the catalytic activity of protease. Preferably, the protease mutation mentioned in the present application can be a single point mutation of the protease amino acid sequence.

[0099] In step S320, based on the amino acid sequence information of the sample protease before and after mutation, feature extraction is performed using a pre-training model to obtain sequence features of the multiple training samples.

[0100] Specifically, in step S320, the amino acid sequence information of the sample protease before and after mutation is subjected to feature extraction using the pre-trained model to obtain sequence feature information before and after mutation. Then, also in step S320, the sequence feature information before and after mutation is concatenated to obtain sequence features of the multiple training samples.

[0101] 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.

[0102] 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.

[0103] In step S330 , structural features of a plurality of training samples are obtained based on the three-dimensional structural information of the sample protease.

[0104] As mentioned above, the three-dimensional structure information of the sample protease 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.

[0105] The available three-dimensional structural information for proteases typically includes secondary structure, interaction networks, amino acid residue distances, physical environment, and energy information derived from the three-dimensional structure. The interaction network, which contains contact information between all residue pairs, is one of the most commonly used features in structure-based prediction methods.

[0106] 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 prediction model based on GNN is selected.

[0107] 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.

[0108] In step S340, a graph neural network is trained based on the sequence features, the structural features, and the mutation activity change labels to obtain a prediction model for predicting the effect of mutations on protease activity.

[0109] As mentioned earlier, this prediction model is actually a classifier.

[0110] In a preferred embodiment of the present disclosure, the graph neural network is a graph convolutional network (GCN) or a graph attention network (GAT).

[0111] For the method of constructing the prediction model, please refer to the detailed description in the examples below.

[0112] In addition, it should be noted that: for sequence features, by placing the processed protease activity 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.

[0113] During model training, sequence and structural features are used as input data for the prediction model, while mutation activity change signatures are used as output targets. The prediction model is fully trained. As the prediction model parameters are continuously optimized, the output data will align with the prediction results indicated by the mutation activity change signatures.

[0114] In a preferred embodiment of the present disclosure, step S340 may further include: inputting sequence features and structural features into a graph neural network to obtain probabilistic features about activity changes; then, combining the probabilistic features of activity changes with supplementary features from other prediction methods, and inputting them into a decision tree model to obtain a classification about activity changes; and adjusting the parameters of the prediction model using the mutation activity change label as the output target.

[0115] More specifically, in some embodiments of the present disclosure, can be based on the sequence characteristics, structural characteristics and the sudden change activity change label of a plurality of training samples, predictive model (classifier) ​​is carried out supervised training, obtain the classifier after training. Supervised training is arranged to utilize the parameters of one group of known class sample adjustment classifier, make it reach the process of required performance. In certain embodiments, can construct loss function, such as cross entropy loss function, reflect the difference of the catalytic activity change prediction result and the actual sudden change activity change label of training sample after the protease mutation predicted by classifier through loss function, and then adjust the parameter of classifier according to this difference, obtain the model optimization parameter of classifier. When loss function meets predetermined condition, such as loss function convergence, or loss function value is less than preset value, or training set and test set loss function value drop to minimum point simultaneously or training set loss function value is to minimum point, classifier training is completed. In certain embodiments, parameter search algorithm can be adopted to adjust classifier parameters, for example, search algorithm, such as grid search, Bayesian search etc. In certain embodiments, supervised training is arranged to be the process according to sudden change activity change label training predictive model.

[0116] 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 .

[0117] Specific example embodiments will be given below.

[0118] Example methods and materials

[0119] 1. Training Dataset

[0120] In order to overcome the problems caused by poor data quality, 5449 experimental samples containing activity change information were screened from the D3DistalMutation database published in 2021 (see Wang, Xiaoyu et al. 2021. “D3DistalMutation: a database to explore the effect of distal mutations on enzyme activity”. Journal of Chemical Information and Modeling 61(5):2499–2508. The entire contents of the above document are incorporated into the present disclosure by reference and become part of the content of the present disclosure). Among them, 5279 samples have mutations that reduce activity and 220 samples have mutations that increase activity. The specific screening criteria are as follows:

[0121] (1) Remove all multi-point mutation samples and limit the training dataset to single-point mutation dataset;

[0122] (2) neutral mutations (mutations that have no effect on protease activity) were removed, and mutations that reduced or lost activity were uniformly defined as activity-reducing mutations;

[0123] (3) Remove samples with different labels for the same mutation. The reason for removing samples with different labels is that samples with different labels for the same mutation may be caused by different experimental conditions or experimental errors, which makes it difficult to define labels for such data. If such samples are not removed, they may introduce noise to the model and affect its accuracy.

[0124] (4) Reverse and merge the data.

[0125] Finally, the dataset used to train the model contains 10,898 samples.

[0126] 2. Test Dataset

[0127] The test data comes from the BRENDA database (https: / / brenda-enzymes.org / news.php?) (see Chang, Antje et al. 2021. “BRENDA, the ELIXIR core data resource in 2021: new developments and updates”. Nucleic acids research 49(D1): D498–508. The entire contents of the above documents are incorporated into the present disclosure by reference, making them part of the contents of the present disclosure), literature (for example, see Amorosi, Clara J et al. 2021. “Massively parallel characterization of CYP2C9 variant enzyme activity and abundance”. The American Journal of Human Genetics 108(9): 1735–51, Cheng, Ya-Shan et al. 2015. “Improving the catalytic performance of a GH11 xylanase by rational protein engineering”. Applied microbiology and biotechnology 99:9503–10, You, Chun, Hanying Yuan, Qiang Huang and Hong Lu. 2010. “Substrate molecule enhances the thermostability of a mutant of a family 11 xylanase from Neocallimastix patriciarum”. African Journal of Biotechnology 9(9). The entire contents of the above documents are hereby incorporated into the present disclosure by reference and become part of the contents of the present disclosure) and patents (WO2020234200A1; EP3854872A1; WO2018195850; WO2017109262A1). The screening criteria are as follows:

[0128] (1) Remove non-single-point mutation samples and limit the test dataset to single-point mutation datasets;

[0129] (2) neutral mutation samples were removed and mutations with reduced or lost activity were uniformly defined as activity-reducing mutations;

[0130] (3) Remove samples with multiple labels. Samples with multiple labels have different reported protease activities at the same mutation site. Samples with this mutation need to be deleted to ensure data accuracy.

[0131] (4) Remove samples that are repeated in the training set.

[0132] Finally, the dataset used to test the model’s effectiveness contained 6,103 samples, of which 4,950 were mutations that reduced activity, and 1,153 were mutations that increased activity.

[0133] 2.1 S3459

[0134] Due to algorithmic limitations, SCANEER can only predict mutations that have occurred during evolution. Of the 6,103 data points in the test set, SCANEER was able to predict 3,459 samples, 1,797 of which were from Deep Mutational Scanning (DMS). DMS introduces random mutations into protein genes and then evaluates their effects on protein function through methods such as stress screening or high-throughput sequencing. The remaining 1,662 samples were non-DMS data.

[0135] 2.2 S2644

[0136] Among all the test data, SCANEER failed to predict mutations in 2644 samples. Similar to 2.1, this data was also divided into DMS data and non-DMS data, with sample sizes of 1492 and 1152, respectively.

[0137] 3. Pre-trained Model

[0138] 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: Toward understanding the language of life through self-supervised learning”. IEEE transactions on pattern analysis and machine intelligence 44(10): 7112–27. 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.

[0139] 4. Graph Convolutional Neural Network Model Construction

[0140] This preferred embodiment is based on a variant of the Graph Convolutional Network (GCN): Graph Attention Network (GAT) (see Petar et al. 2017. “Graph attention networks”. arXiv preprint arXiv:1710.10903. The entire contents of the above document are hereby incorporated into the present disclosure by reference, making them a part of the content of the present disclosure).

[0141] 4.1 GAT-1 model construction

[0142] 4.1.1 Adjacency Matrix

[0143] The three-dimensional structure of each protease was obtained through prediction software such as the PDB database, AlphaFold2, I-TASSER, RoseTTAFold, Modeller and Swiss-model. The distances between pairs of amino acid residues in the three-dimensional structure of the protease were calculated based on the coordinates of the obtained three-dimensional structure of the protease. 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 protease.

[0144] In this example, 10 amino acids on either side of the mutation site, ie, 10 consecutive amino acids, were selected as the residue sequence, and a 21*21 adjacency matrix was intercepted from the adjacency matrix of the entire protease.

[0145] FIG4 is a schematic diagram of an adjacency matrix. As shown in the figure, an adjacency matrix of 100 amino acids in length is used. Assuming that the mutation site is residue 12, the box is a 21*21 adjacency matrix of the residue. The McBASC Covariance algorithm (see McLachlan, Andrew D. 1971. "Tests for comparing related amino-acid sequences. Cytochrome c and cytochrome c551". Journal of molecular biology 61 (2): 409–24. The entire contents of the above document are incorporated into the present disclosure by reference and become part of the contents of the present disclosure) is used to calculate the coevolution scores between the corresponding 21 amino acids in the adjacency matrix to form a 21*21 coevolution score matrix, which is multiplied by the adjacency matrix to obtain an updated 21*21 adjacency matrix.

[0146] 4.1.2 Node Feature Matrix

[0147] After a protease sequence is input into the pre-trained model, the model first encodes each amino acid in the protease sequence through an encoder. Each encoded amino acid corresponds to a 1024-dimensional vector, meaning that a protease sequence of length N will generate an N*1024-dimensional vector. 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 are extracted for both the wild-type and mutant forms. The wild-type (i.e., before mutation) and mutant (i.e., after mutation) feature vectors are then concatenated to form a 2048-dimensional feature matrix. Using a sliding window size of 21, the mutation site is centered, and the 10 amino acid residues before and after are concatenated. This results in a 21*2048-dimensional feature matrix, which serves as the node feature matrix. The structure is shown in Figure 5.

[0148] 4.1.3 Model Construction

[0149] 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, the node features of the three-layer graph network structure are updated to 32-, 64-, and 64-dimensional feature vectors, respectively. Four attention heads are used in each layer, and the node features output by the three-layer graph network are merged into a 21*640-dimensional feature vector through a residual network structure. Average pooling is then used to compress the updated feature matrix into a 1*640-dimensional feature vector, representing the feature representation of the 21 amino acid residue sequence. Specific parameters are as follows: the output dimensions of the three-layer graph convolution are 32, 64, and 64, respectively; the number of neurons in the fully connected layer is 64 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 classification function for the output layer; the l2 regularization coefficient is 0.05, the dropout rate is 0.3, and the learning rate is 0.001.

[0150] The aggregation operation mentioned here is achieved through the attention mechanism, in which each node has an attention weight distribution. The feature matrix and the attention coefficient are weighted and summed to obtain the node's aggregated feature vector. Through this aggregation operation, each node can use the feature information of its neighboring nodes to update features and transfer information.

[0151] The residual network (RNN) described here is a deep convolutional neural network architecture that uses residual connections to construct skip connections between network layers. That is, the input of each network layer is not only the output of the previous layer, but also includes a direct skip connection. This skip connection adds the input directly to the output, forming a residual. Its purpose is to address the vanishing gradient and difficulty training deep networks.

[0152] The average pooling operation mentioned here is a commonly used technique for reducing the feature dimensionality of the convolutional layer output. It is achieved by averaging the features output by the convolutional layer. Its purpose is to effectively reduce network parameters and prevent overfitting. Those skilled in the art will understand that the average pooling operation here can also be replaced by any other pooling operation, such as max pooling or K-max pooling.

[0153] 4.2 GAT-2 model construction

[0154] The matrices generated in 4.1.1 and 4.1.2 are directly input into the graph convolutional neural network. Through matrix multiplication and aggregation operations, each layer of the three-layer graph network structure updates the node features into a 128-dimensional feature vector. Each layer uses four attention heads, and the node features output by the three-layer graph network are merged into a 21*1536-dimensional feature vector through a residual network structure. The updated feature matrix is ​​then compressed into a 1*1536-dimensional feature vector using average pooling, representing the feature representation of the 21 amino acid residue sequence. 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 the fully connected layer is 64 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 classification function for the output layer; the l2 regularization coefficient is 0.01, the dropout rate is 0.5, and the learning rate is 0.001.

[0155] 4.3 GAT-3 model construction

[0156] 4.3.1 First-order adjacency matrix

[0157] First, obtain the adjacency matrix of the entire protease 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 and calculate the adjacency matrix composed of these amino acids.

[0158] 4.3.2 Node Feature Matrix

[0159] After a protease sequence is input into the pre-trained model, the model first encodes each amino acid in the protease sequence through an encoder. Each amino acid is encoded into a vector of length 1024, that is, a protease sequence of length N will obtain a vector of length N*1024. The model has a 24-layer structure. In order to obtain the most information about the mutation, 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 4.3.1 should be kept consistent. The number of amino acids obtained from different proteases may be different. f As a benchmark, the insufficient amount was filled with 0 so that all proteases could obtain matrices of the same size.

[0160] 4.3.3 Model Construction

[0161] 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, the two-layer graph network structure updates the node features into 128- and 256-dimensional feature vectors respectively. Each layer uses 4 attention heads, and the node features output by the two-layer graph network are merged into N through the residual network structure. f *1536-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 as follows: the output dimensions of the two layers of graph convolution are 128 and 256 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 of the output layer; the l2 regularization coefficient is 0.05, the dropout rate is 0.3, and the learning rate is 0.001.

[0162] 5. LightGBM model construction

[0163] Gradient Boosting Decision Tree (GBDT) (see Friedman, Jerome H. 2001. “Greedy function approximation: a gradient boosting machine”. Annals of statistics: 1189–1232. The entire contents of the above-mentioned document are hereby incorporated into the present disclosure by reference, making it a part of the contents of the present disclosure) is a commonly used model in machine learning. Its main idea is to use weak classifiers (decision trees) for iterative training to obtain the optimal model. LightGBM (Light Gradient Boosting Machine, or LGBM for short) is a framework for implementing the GBDT algorithm (see Ke, Guolin et al. 2017. “Lightgbm: A highly efficient gradient boosting decision tree”. Advances in neural information processing systems 30. The entire contents of the above-mentioned document are hereby incorporated into the present disclosure by reference, making it a part of the contents of the present disclosure). The LightGBM used in the method of the present disclosure is version 3.3.3.

[0164] 5.1 LGBM-1 model construction

[0165] 5.1.1 Feature Extraction

[0166] The probability value (representing the probability of increased protease activity) obtained from the GAT-1 model constructed in 4.1 was used as a one-dimensional feature, and the feature extraction described by the PremPS (see 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 the above document are incorporated into the present disclosure by reference and become part of the contents of the present disclosure) and SCANEER (as described above) method was used to supplement the feature 11 dimensions. After splicing, a total of 12 dimensions of features were obtained. Among them, 10 dimensions were supplemented by PremPS and 1 dimension was supplemented by SCANEER.

[0167] 5.1.2 Model Construction

[0168] The obtained 12-dimensional features are input into the LightGBM framework for training, with the learning rate set to 0.05, extra_trees set to True, and other parameters using the default values.

[0169] 5.2 LGBM-2 model construction

[0170] 5.2.1 Feature Extraction

[0171] As in 5.1.1, the probability values ​​(representing the probability of increased protease activity) are obtained from the GAT-2 and GAT-3 models constructed in 4.2 and 4.3 as two-dimensional features. The feature extraction described by the PremPS and SCANEER methods is used to supplement the 11-dimensional features, and a total of 13-dimensional features are obtained after splicing.

[0172] 5.2.2 Model Construction

[0173] The obtained 13-dimensional features are input into the LightGBM framework for training, with the learning rate set to 0.05, extra_trees set to True, and other parameters using the default values.

[0174] Here are some summaries.

[0175] Figure 6 shows a schematic diagram of the construction of a structure-based adjacency matrix. Figure 7 shows a schematic diagram of the construction of a sequence-based node feature matrix. Figure 8 shows a schematic diagram of the framework of the prediction model.

[0176] 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 protease can be obtained by: obtaining the coordinates of the three-dimensional structure of the sample protease; calculating the distances between amino acid residue pairs in the three-dimensional structure of the sample protease based on the obtained three-dimensional 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 adjacency matrix of the entire sample protease is recorded as 0, thereby obtaining an adjacency matrix for the entire sample protease. Based on the three-dimensional structural information of the sample protease, obtaining the structural features of multiple training samples may include: intercepting the adjacency matrix of a specified size from the adjacency matrix of the entire sample protease with the mutation site as the center, as the structural feature of the sample protease. Training a graph neural network based on sequence features, structural features, and mutation activity change labels to obtain a prediction model for predicting the effect of mutations on protease activity may further include: inputting the sequence features and the structural features into the graph neural network to obtain a probability feature regarding activity change; combining the probability feature of activity change with supplementary features from other prediction methods and inputting them into a decision tree model to obtain a classification regarding activity change; and adjusting the parameters of the prediction model using the mutation activity change label as the output target. Inputting the sequence features and the structural features into a graph neural network to obtain probabilistic features regarding activity changes may further include: inputting the sequence features and the structural features into a graph neural network; updating and outputting the features using a plurality of graph network layers through matrix multiplication and aggregation operations; merging the features outputted by the plurality of graph network layers; compressing the features using a pooling operation; and classifying the features according to activity changes based on an output layer classification function, and outputting probabilistic features regarding activity changes. The pooling operation referred to herein may be average pooling, maximum pooling, or K-max pooling.

[0177] Accordingly, in the process of applying the prediction model to predict activity changes after protease mutation, the three-dimensional structural information of the protease sample to be predicted is obtained by: obtaining the coordinates of the three-dimensional structure of the protease sample to be predicted; calculating the distances between amino acid residue pairs in the protease 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 recorded as 0, and the adjacency matrix of the entire protease sample to be predicted is obtained. Based on the three-dimensional structural information of the protease sample to be predicted, the structural characteristics of the protease sample to be predicted are obtained by: intercepting the adjacency matrix of a specified size from the adjacency matrix of the entire protease sample to be predicted, centered on the mutation site, as the structural characteristics of the protease sample to be predicted. Based on the sequence characteristics and the structural characteristics, using the prediction model based on the graph neural network to predict the activity change of the protease sample to be predicted can further include: inputting the sequence characteristics and the structural characteristics into the graph neural network to obtain a probability feature regarding the activity change; combining the probability feature of the activity change with supplementary features from other prediction methods and inputting them into a decision tree model to obtain a classification regarding the activity change. More specifically, the step of inputting the sequence features and the structural features into a graph neural network to obtain probabilistic features regarding activity changes includes: inputting the sequence features and the structural features into a graph neural network; updating and outputting the features using a plurality of graph network layers through matrix multiplication and aggregation operations; merging the features outputted by the plurality of graph network layers; compressing the features using a pooling operation; and classifying the features according to activity changes according to an output layer classification function, thereby outputting probabilistic features regarding activity changes. Similarly, the pooling operation referred to herein may be average pooling, maximum pooling, or K-max pooling.

[0178] 6. Performance indicators

[0179] This example uses the following metrics for evaluation: area under the curve (AUC), precision, and recall. The specific definitions are as follows:

[0180] True Positive (TP)

[0181] False Positive (FP)

[0182] True Negative (TN)

[0183] False Negative (FN)

[0184] 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.

[0185] 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.

[0186] 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.

[0187] 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.

[0188] Result evaluation

[0189] 1. Model Performance

[0190] The model was trained on a self-designed balanced dataset containing the effects of 10,898 single-point mutations on enzyme activity for 1,303 proteases from the D3DistalMutation database. The performance of the model on the test data in a five-fold cross-validation is shown in Table 1 below.

[0191] Table 1: Five-fold cross-validation performance indicators of the prediction model of the present disclosure on the training set

[0192] 2. Comparison with other methods

[0193] In order to evaluate the difference between the method of the present invention and the SCANEER method, the model of the present invention was compared with the only prediction method SCANEER. SCANEER evaluates the co-evolutionary relationship of amino acid pairs in a multiple sequence alignment (MSA). Amino acid pairs that are not observed or rarely observed in the MSA, that is, amino acid substitutions, cannot be predicted, which ultimately results in the number of predictable mutations for each protease accounting for only 47% of all possible mutations (the authors of SCANEER assume that amino acid substitutions that are not observed or rarely observed in the MSA will lead to loss of protease function and have been eliminated during the evolutionary process. Therefore, the authors exclude positions with a gap of more than 20% or that are not in the top N co-evolving amino acid pairs. That is, in the MSA, positions with a gap of more than 20% or that are completely conserved, as well as amino acids at positions that are not in the top N co-evolving amino acid pairs, cannot calculate their mutation effects (N = protease length * 2)), which will greatly limit the scope of application of this method. Therefore, out of the 6103 samples in the experimental test set, SCANEER was able to predict 3459 samples, of which 2855 and 604 were activity-reducing and activity-enhancing mutations, respectively. Of these 3459 samples, 1797 were DMS and 1662 were non-DMS data. The comparison results are shown in Table 2 below.

[0194] Table 2: Performance comparison of the disclosed model and the SCANEER method on comparable experimental test sets

[0195] As can be seen from the above table, on S3459 and S1662, the method disclosed in the present invention is slightly better than or basically the same as SCANEER in terms of precision and recall, while on S1797, it has a great advantage in precision.

[0196] The S2644 dataset is not predictable by SCANEER, and contains 2095 samples with activity-reducing and 549 samples with activity-enhancing mutations. Within this dataset, the number of DMS and non-DMS data samples is 1492 and 1152, respectively. The performance of the disclosed method on this dataset is shown in Table 3 below.

[0197] Table 3: Performance of the disclosed model on experimental test data that SCANEER cannot predict

[0198] The results show that the method disclosed herein can still achieve good performance on experimental test data that SCANEER cannot predict, which shows the advantages of the method disclosed herein.

[0199] Prediction Process

[0200] Here, two cases are taken as examples to describe the process of using a prediction model to predict changes in protease activity.

[0201] One scenario involves inputting a single protease sequence and mutation site information. First, based on the input protease sequence and mutation site information, mutation site preprocessing is performed. Specifically, the input protease 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, 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 mutated protease 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 protease embedding layer feature network and a protease residue interaction feature network. This feature network is applied to a GAT-based prediction model to predict the probability of protease activity change before and after amino acid mutation. Based on this probability, LightGMB is used to predict the change in protease activity before and after amino acid mutation. Finally, the prediction results are output and displayed. In this example scenario, single-point mutation prediction results can be output.

[0202] Another scenario involves the input protease sequence information consisting of only a single protease sequence. First, preprocessing of the protease mutation sites is performed based on the input protease sequence. Specifically, the input protease sequence is recorded as the original sequence, and each amino acid at each site in the protease sequence is sequentially mutated to 19 other amino acids, representing 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 protease 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 neural network is constructed, requiring a protease embedding layer feature network and a protease residue interaction feature network. This feature network is applied to a GAT-based prediction model to predict the probability of protease activity change before and after amino acid mutation. Based on these probabilities, LightGMB is used to predict the change in protease activity before and after amino acid mutation. Finally, the prediction results are output and displayed. For this example case, a table of all mutation probabilities that improve enzyme activity and / or a ranking of all mutation possibilities that improve enzyme activity may be output as the final prediction result.

[0203] in conclusion

[0204] Protease mutants with higher activity are essential for academia and industry to explore more applications for these proteases. This disclosure proposes a new method for discovering the effects of amino acid mutations on protease activity. Based on advanced natural language processing and graph neural network models, combined with sequence and three-dimensional structure, this method can be applied to various biological fields and demonstrates superior performance to other methods, demonstrating its broad application prospects in protease engineering.

[0205] The following discusses the reasons why the method disclosed herein is superior to the prior art. As can be seen from the above description:

[0206] 1. In general, the method disclosed herein is based on the tertiary structure and amino acid sequence information of proteases, and a prediction model is constructed through AI training.

[0207] 2. The prediction model constructed in the present disclosure can be applied to predict the activity of protease mutations.

[0208] 3. After the training data set used in the method of the present invention is inverted, the original data set is expanded and the data set is balanced at the same time, making the final prediction model more robust.

[0209] Those skilled in the art will appreciate that, based on the disclosed method for predicting the effect of a mutation on protease activity, a system for predicting the effect of a mutation on protease activity can be developed. FIG9 shows a schematic block diagram of a system for predicting the effect of a mutation on protease activity according to the disclosed method. Specifically, the system 900 for predicting the effect of a mutation on protease activity comprises an acquisition module 910 and a processing module 920.

[0210] The acquisition module 910 is used to obtain the sequence information and three-dimensional structure information of the protease sample to be predicted.

[0211] The processing module 920 is used to obtain a prediction result of the activity change of the protease sample to be predicted after mutation by inputting the sequence information and three-dimensional structure information of the protease sample to be predicted. As shown in Figure 9, the processing module 920 can further include: a sequence feature submodule 921, which is used to extract features based on the sequence information of the protease sample to be predicted using a pre-trained model to obtain the sequence features of the protease sample to be predicted; a structural feature submodule 922, which is used to obtain the structural features of the protease sample to be predicted based on the three-dimensional structure information of the protease sample to be predicted; and a model prediction submodule 923, which is used to predict based on the sequence features and the structural features using a prediction model based on a graph neural network to determine the prediction result of the activity change of the protease sample to be predicted after mutation.

[0212] Those skilled in the art will appreciate that the system for predicting the effect of mutations on protease activity 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 protease sample to be predicted; inputting the sequence information and three-dimensional structural information of the protease sample to be predicted to obtain a predicted result of activity change of the protease 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 protease sample to be predicted to obtain sequence features of the protease sample to be predicted; obtaining structural features of the protease sample to be predicted based on the three-dimensional structural information of the protease 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 predicted result of activity change of the protease sample to be predicted after mutation.

[0213] That is to say, although in the system for predicting the effect of mutations on protease activity 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, cause the electronic device to implement the method for predicting the effect of a mutation on protease activity as described above.

[0216] In addition, according to the 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 above-described method for predicting the effect of a mutation on protease activity.

[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 prediction model for predicting the effect of mutations on protease activity as described above.

[0218] In addition, according to the 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 above-described method for constructing a prediction model for predicting the effect of mutations on protease activity.

[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 the effect of a mutation on protease activity, characterized in that: The method comprises: Based on the sequence information of the protease sample to be predicted, a pre-trained model is used to perform feature extraction to obtain the sequence features of the protease sample to be predicted; Based on the three-dimensional structural information of the protease sample to be predicted, obtaining the structural characteristics of the protease 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 predicted results of the activity changes of the protease sample to be predicted after mutation.

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

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

5. The method according to claim 3 or 4, 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 protease sample to be predicted using a pre-trained model to obtain the sequence features of the protease sample to be predicted further comprises: The amino acid sequence information of the protease sample to be predicted before mutation and the amino acid sequence information after mutation are respectively subjected to feature extraction using the pre-trained model 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 protease 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 structural information of the protease sample to be predicted includes the three-dimensional structural information of the protease 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 of the protease sample to be predicted includes an interaction network, a secondary structure, an amino acid residue distance or a physical environment; preferably, the three-dimensional structural information of the protease sample to be predicted includes an interaction network.

10. The method according to claim 8 or 9, characterized in that: The three-dimensional structure information of the protease sample to be predicted is further obtained by the following method: Obtaining the coordinates of the three-dimensional structure of the protease sample to be predicted; Calculating the distances between pairs of amino acid residues in the three-dimensional structure of the protease 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 protease sample to be predicted.

11. The method according to claim 10, characterized in that The method of obtaining the structural features of the protease sample to be predicted based on the three-dimensional structural information of the protease sample to be predicted includes: Taking the mutation site as the center, an adjacency matrix of a specified size is intercepted from the adjacency matrix of the entire protease sample to be predicted as the structural feature of the protease 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 predicted activity change result of the protease sample to be predicted after mutation further includes: Inputting the sequence features and the structural features into a graph neural network to obtain probability features regarding activity changes; The probability features of the activity changes are combined with supplementary features from other prediction methods and input into a decision tree model to obtain a classification of the activity changes.

13. The method according to claim 12, characterized in that The step of inputting the sequence features and the structural features into the graph neural network to obtain the probability features regarding the activity change 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; The activity changes of the features are classified according to the output layer classification function, and the probability features of the activity changes are output.

14. 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 protease, amino acid sequence information of the sample protease before and after mutation, and a mutation activity change label; Based on the amino acid sequence information of the sample protease 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 protease, obtaining structural features of the plurality of training samples; Based on the sequence features, the structural features and the mutation activity change labels, the graph neural network is trained to obtain the graph neural network-based prediction model.

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

16. The method according to claim 14, characterized in that The graph neural network is a graph convolutional network or a graph attention network.

17. The method according to claim 14, characterized in that The training of the graph neural network based on the sequence features, the structural features and the mutation activity change labels to obtain the prediction model based on the graph neural network further includes: Inputting the sequence features and the structural features into a graph neural network to obtain probability features regarding activity changes; Combining the probability features of the activity change with supplementary features from other prediction methods and inputting them into a decision tree model to obtain a classification of the activity change; The mutation activity change label is used as the output target to adjust the parameters of the prediction model.

18. The method according to claim 12 or 17, characterized in that The decision tree model is a gradient boosted decision tree (GBDT) model.

19. The method according to claim 18, characterized in that The LightGBM framework is used to implement the GBDT model.

20. The method according to claim 14, 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.

21. A method for constructing a prediction model for predicting the effect of mutation on protease activity, 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 protease, amino acid sequence information of the sample protease before and after mutation, and a mutation activity change label; Based on the amino acid sequence information of the sample protease 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 protease, the structural features of multiple training samples are obtained; Based on the sequence features, the structural features and the mutation activity change labels, a graph neural network is trained to obtain a prediction model for predicting the effect of mutations on protease activity.

22. The method according to claim 21, characterized in that The method of extracting features based on the amino acid sequence information of the sample protease 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 protease 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.

23. The method according to claim 21 or 22, 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.

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

25. The method according to claim 21, characterized in that The three-dimensional structural information of the sample protease includes an interaction network, a secondary structure, an amino acid residue distance or a physical environment; preferably, the three-dimensional structural information of the sample protease includes an interaction network.

26. The method according to claim 24 or 25, characterized in that The three-dimensional structural information of the sample protease is further obtained by: Obtaining coordinates of the three-dimensional structure of the sample protease; Calculating the distances between pairs of amino acid residues in the three-dimensional structure of the sample protease 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 to obtain the adjacency matrix of the entire sample protease.

27. The method according to claim 26, characterized in that The method of obtaining the structural features of multiple training samples based on the three-dimensional structural information of the sample protease comprises: Taking the mutation site as the center, an adjacency matrix of a specified size is intercepted from the adjacency matrix of the entire sample protease as the structural feature of the sample protease.

28. The method according to claim 21, characterized in that The training of the graph neural network based on the sequence features, the structural features and the mutation activity change labels to obtain a prediction model for predicting the effect of mutations on protease activity further includes: Inputting the sequence features and the structural features into a graph neural network to obtain probability features regarding activity changes; Combining the probability features of the activity change with supplementary features from other prediction methods and inputting them into a decision tree model to obtain a classification of the activity change; The mutation activity change label is used as the output target to adjust the parameters of the prediction model.

29. The method according to claim 28, characterized in that The step of inputting the sequence features and the structural features into the graph neural network to obtain the probability features regarding the activity change 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; The activity changes of the features are classified according to the output layer classification function, and the probability features of the activity changes are output.

30. The method according to claim 21, characterized in that The predictive model is a classifier.

31. The method according to claim 21, characterized in that The graph neural network is a graph convolutional network or a graph attention network.

32. The method according to claim 28, characterized in that The decision tree model is a GBDT model.

33. The method according to claim 32, characterized in that The LightGBM framework is used to implement the GBDT model.

34. The method according to claim 21, 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.

35. A system for predicting the effect of mutations on protease activity, characterized in that: The system comprises: An acquisition module is used to obtain sequence information and three-dimensional structure information of the protease sample to be predicted; The processing module is used to obtain the activity change prediction result of the protease sample to be predicted after mutation by inputting the sequence information and three-dimensional structure information of the protease 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 protease sample to be predicted using a pre-trained model to obtain the sequence features of the protease sample to be predicted; A structural feature submodule, used for obtaining the structural features of the protease sample to be predicted based on the three-dimensional structural information of the protease sample to be predicted; The model prediction submodule is used to make predictions based on the sequence features and the structural features using a prediction model based on a graph neural network to determine the predicted results of the activity changes of the protease sample to be predicted after mutation.

36. A computer-implemented system for predicting the effect of a mutation on protease activity, the computer-implemented system implementing the following operations by executing a computer program: Obtaining sequence information and three-dimensional structure information of the protease sample to be predicted; By inputting the sequence information and three-dimensional structure information of the protease sample to be predicted, the activity change prediction result of the protease sample to be predicted after mutation is obtained, and the operation further includes the following sub-operations: Based on the sequence information of the protease sample to be predicted, using a pre-trained model to perform feature extraction to obtain sequence features of the protease sample to be predicted; Based on the three-dimensional structural information of the protease sample to be predicted, obtaining the structural characteristics of the protease 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 predicted results of the activity changes of the protease sample to be predicted after mutation.

37. A non-transitory computer-readable storage medium for storing a computer program, wherein the computer program includes instructions that, when executed by a processor of an electronic device, cause the electronic device to implement the method for predicting the effect of a mutation on protease activity as described in any one of claims 1-20 or the method for constructing a predictive model for predicting the effect of a mutation on protease activity as described in any one of claims 21-34.

38. A computer system, the 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 the effect of mutations on protease activity as described in any one of claims 1-20 or the method for constructing a predictive model for predicting the effect of mutations on protease activity as described in any one of claims 21-34.

Images