Method and device for predicting protein mutation effect

Abstract

The invention discloses a method and a device for predicting a protein mutation effect in the field of bioinformatics. The technical problem to be solved by the invention is how to accurately predict the protein mutation effect. The method for predicting the protein mutation effect comprises the following steps: receiving structural data of model training protein, constructing a prediction model of occurrence probabilities of various amino acids, obtaining a mutation effect score of a target protein mutant, and obtaining a prediction function type of the target protein mutant according to the mutation effect score, obtaining the prediction function type of the verification sample according to the prediction function type, performing parameter updating through the distribution difference between the prediction function type and the verification function type of the verification sample to obtain an optimized prediction model, and obtaining an updating result of the occurrence probability of various amino acids in the target protein mutant based on the optimized prediction model. And obtaining the mutation effect of the target protein according to the updating result. The method can be used for rapidly screening a protein virtual mutation library and providing a target protein candidate mutation modification scheme.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method and device for predicting protein mutation effects. Background Art

[0002] Proteins are biological macromolecules that perform catalytic, transport and regulatory functions, and their functions are mainly determined by their amino acid sequences and spatial structures. Proteins with catalytic efficiency are the "chips" of the biomanufacturing industry and are widely used in important industrial fields such as medicine, food, materials and chemicals. Although long-term evolution and natural selection have given birth to hundreds of millions of natural proteins, their catalytic activity, stability and tolerance often cannot meet the needs of industrial production, making protein optimization and transformation a core technology that affects the large-scale application of industrial enzymes.

[0003] In recent years, AI-assisted design methods have developed rapidly and are expected to address many challenges faced by directed evolution and rational design strategies. By predicting the potential functional impact of amino acid mutations and identifying potentially beneficial mutations from a large number of mutants, AI-assisted design methods can reduce the huge human resources and experimental costs required for multiple rounds of random mutation, expression and screening, and reduce the designer's reliance on prior knowledge of the target protein structure and function.

[0004] Most of the classical AI-assisted design methods adopt supervised learning strategies, based on existing protein mutants and corresponding functional data, to train machine learning or deep learning models such as perceptrons, support vector machines, and deep neural networks, and to predict the functions of new mutants and build a "small and precise" candidate mutation library by fitting the complex nonlinear relationship between mutant sequence / structure and function, thus accelerating the protein optimization and transformation process. Although there are many successful cases of classical AI-assisted design methods, there are serious limitations in terms of versatility and accuracy, and the prediction methods developed for a specific type of protein may have little effect on other types of proteins, or the effectiveness of the prediction methods depends on the algorithm designer's prior knowledge, resulting in a large number of false positive results when predicting mutants.

[0005] Pre-trained language model is a deep learning paradigm that aims to use unsupervised strategies to implicitly learn the probability distribution of data, and then use the learned common knowledge to promote downstream classification, labeling, generation and other tasks. At present, the progress of data, computing power and algorithms jointly drives the rapid development of pre-trained language models represented by BERT and GPT. Due to the similarity between protein amino acid sequences and human natural language, pre-trained language model technology has also been transferred to the field of life sciences, and pre-trained language models such as ProteinBERT, ESM and ProtTrans have emerged, and have made significant progress in tasks such as protein structure prediction, protein function annotation and amino acid mutation effect prediction. At present, pre-trained language models have been initially applied in protein optimization and modification, but they still face bottlenecks in the application of small sample data and multimodal information fusion.

[0006] First, the application ability of small sample data is insufficient. Existing protein language models usually use a zero-sample mutation effect prediction method. Without relying on the experimental data of the protein to be modified, they directly predict the fitness score of the mutation, and use this as a basis to screen favorable mutations to modify the target protein. However, a large amount of functional data of protein mutants has been generated in existing public literature and public databases. Directly predicting mutation fitness without considering existing data may lead to limited model prediction accuracy. On the other hand, in the process of designing and modifying the target protein, some new mutant functional data are often generated, but the existing language model lacks the ability to use the feedback results of wet experiments and cannot dynamically adjust the prediction results.

[0007] Secondly, the ability to fuse protein sequence and structural multimodal information is insufficient. Existing pre-trained language models are usually trained only based on single-modal sequence data, and three-dimensional structural data that is more closely related to protein function has not been fully utilized. Compared with sequences, protein three-dimensional structures are more conservative during evolution, and contain amino acid spatial contact information that is difficult to directly reflect in sequences, which is of great value for further improving the prediction accuracy of amino acid mutation effects. With the continuous development of protein structure prediction methods such as Alphafold2, high-precision protein three-dimensional structure data has grown rapidly. DeepMind and Meta have released more than 800 million predicted structures in 2022, providing a new opportunity for simultaneously utilizing protein sequence and structural multimodal information, accurately predicting mutation effects, and guiding protein design and modification.

[0008] Due to the lack of effective small sample data application capabilities and multimodal information fusion capabilities, existing protein language models have problems with prediction accuracy and effectiveness when actually used to guide protein design and modification. In recent years, the FSFP method of Shanghai Jiao Tong University and the uFormer model of the Microsoft team have combined meta-learning or reinforcement learning technology to optimize the mutation effect prediction accuracy using wet experiment feedback results, which has improved the small sample data application capabilities of the language model to a certain extent. However, such methods only use protein sequence information, and the bottleneck of multimodal information fusion capabilities including three-dimensional structure has not yet been solved. The ProMEP mutation effect prediction method of the team of the Academy of Military Medical Sciences significantly improves the multimodal information fusion capabilities of large protein models for sequence and structure, but lacks the ability to apply small sample data. How to simultaneously improve the prediction accuracy of amino acid mutation effects based on small sample data and multimodal information, and then effectively guide protein design and modification, still faces huge challenges. Summary of the invention

[0009] The technical problem to be solved by the present invention is how to accurately predict the effect of amino acid mutation and / or how to improve the prediction accuracy of amino acid mutation effect based on the fusion of multimodal information and / or how to use the multimodal information of protein sequence and structure to accurately predict the effect of protein mutation and / or how to accurately design and transform proteins to obtain protein sequences with target functions and / or how to combine the multimodal information of protein amino acid sequence and structure to accurately transform proteins.

[0010] In order to solve the above technical problems, the present invention first provides a method for (accurately) predicting the effect of protein mutations, which may include the following steps: B1) Obtaining training sample data: receiving structural data of model training proteins, extracting amino acid sequence information and structural information of model training proteins based on the structural data, and obtaining training sample data; B2) Model construction: embedding the amino acid sequence information of the model training protein into a vector space to obtain a sequence feature vector, embedding the structural information of the model training protein into a vector space to obtain a structural feature vector, adding the sequence feature vector and the structural feature vector to obtain a fusion feature vector of the model training protein; inputting the fusion feature vector of the model training protein into an encoder for data encoding to obtain an encoded protein feature vector; using the encoded protein feature vector as input to train a multi-layer perceptron to obtain a prediction model for the probability of occurrence of various amino acids at each amino acid position of the protein; B3) Calculation of mutation effect of target protein mutants: construct a mutant library of target protein composed of target protein mutants, input the feature vector of the encoded mutant of the target protein mutant in the mutant library into the prediction model described in B2) to obtain the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculate the mutation effect score of the target protein mutant according to the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; The characteristic vector of the encoded mutant can be obtained by the following steps: extracting the target protein mutant amino acid sequence information and the target protein mutant structure information based on the structural data of the target protein mutant, embedding the target protein mutant amino acid sequence information into the vector space to obtain a mutant sequence characteristic vector, embedding the target protein mutant structure information into the vector space to obtain a mutant structure characteristic vector, adding the mutant sequence characteristic vector and the mutant structure characteristic vector to obtain a mutant fusion characteristic vector; inputting the mutant fusion characteristic vector into the encoder for data encoding to obtain an encoded mutant feature vector; B4) Classification of mutant functional types: setting a threshold for the mutation effect score, and obtaining the predicted functional type of the target protein mutant according to the threshold; B5) Small sample data reception: receiving the verification function type of the verification sample (small sample) in the mutant library; B6) Model optimization: based on the predicted function type in step B4), the predicted function type of the verification sample is obtained, and the loss value of the prediction model is calculated by comparing the distribution difference between the predicted function type of the verification sample and the verification function type of the verification sample obtained in step B5), and the parameters of the prediction model are updated based on the loss value to obtain an optimized prediction model; B7) Secondary calculation of mutation effect: input the feature vector of the encoded mutant of the target protein mutant in the mutant library into the optimized prediction model described in B6) to obtain an updated result of the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculate the optimized mutation effect score of the target protein mutant according to the updated result to obtain the mutation effect of the target protein.

[0011] In the above method, the encoder in B2) may be a Transformer encoder, and the Transformer encoder may include 33 layers; updating the parameters of the prediction model in B6) may include freezing the first 29 layers of the Transformer encoder, and iteratively updating the Transformer parameters of the remaining 4 layers of the Transformer encoder except the 29 layers.

[0012] The vector space mentioned above may be a vector space with 1280 dimensions.

[0013] The functional types mentioned above may include function enhancement (favorable mutation), function unchanged (neutral mutation) and function reduction (unfavorable mutation).

[0014] The prediction model of the occurrence probability of various amino acids at each amino acid position of the protein described above can be obtained by training a multi-layer perceptron.

[0015] The loss value described in B6) can be obtained by calculating the gradient using the cross entropy loss function.

[0016] The training cycle for the parameter update may be 5 cycles.

[0017] The validation samples mentioned above may come from more than one sample in the mutant library, and the validation function type of the sample may come from literature collection, database and / or experimentally obtained experimentally validated function type (function enhancement, unchanged function or weakened function). The more than one sample may be greater than or equal to one sample.

[0018] In order to solve the above technical problems, the present invention also provides a device for (accurately) predicting the effect of protein mutations, which may include the following modules: A1) a training sample data acquisition module: used to receive the structural data of the model training protein, extract the amino acid sequence information and the structural information of the model training protein based on the structural data, and obtain the training sample data; A2) Model building module: used for embedding the amino acid sequence information of the model training protein into a vector space to obtain a sequence feature vector, embedding the structural information of the model training protein into a vector space to obtain a structural feature vector, adding the sequence feature vector and the structural feature vector to obtain a fusion feature vector of the model training protein; inputting the fusion feature vector of the model training protein into an encoder for data encoding to obtain an encoded protein feature vector; using the encoded protein feature vector as input data for training to obtain a prediction model for the occurrence probability of various amino acids at each amino acid position of the protein; A3) Target protein mutant mutation effect calculation module: used to construct a mutant library of target protein composed of target protein mutants, input the feature vector of the encoded mutant of the target protein mutant in the mutant library into the prediction model in A2) to obtain the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculate the mutation effect score of the target protein mutant according to the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; The characteristic vector of the encoded mutant can be obtained by the following steps: extracting the target protein mutant amino acid sequence information and the target protein mutant structure information based on the structural data of the target protein mutant, embedding the target protein mutant amino acid sequence information into the vector space to obtain a mutant sequence characteristic vector, embedding the target protein mutant structure information into the vector space to obtain a mutant structure characteristic vector, adding the mutant sequence characteristic vector and the mutant structure characteristic vector to obtain a mutant fusion characteristic vector; inputting the mutant fusion characteristic vector into the encoder for data encoding to obtain an encoded mutant feature vector; A4) Mutant function type classification module: used to set a threshold for the mutation effect score, and obtain the predicted function type of the target protein mutant according to the threshold; A5) Small sample data receiving module: used for receiving the verification function type of the verification sample (small sample) in the mutant library; A6) Model optimization module: used to obtain the prediction function type of the verification sample based on the prediction function type in step A4), calculate the loss value of the prediction model by comparing the distribution difference between the prediction function type of the verification sample and the verification function type of the verification sample obtained in step A5), and update the parameters of the prediction model based on the loss value to obtain an optimized prediction model; A7) Mutation effect secondary calculation module: used to input the feature vector of the encoded mutant of the target protein mutant in the mutant library into the optimized prediction model in A6) to obtain an updated result of the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculate and obtain the optimized mutation effect score of the target protein mutant according to the updated result to obtain the mutation effect of the target protein.

[0019] In the above device, the encoder in A2) may be a Transformer encoder, and the Transformer encoder may include 33 layers; updating the parameters of the prediction model in A6) may include freezing the first 29 layers of the Transformer encoder and iteratively updating the Transformer parameters of the remaining 4 layers of the Transformer encoder except the 29 layers.

[0020] The vector space mentioned above may be a vector space with 1280 dimensions.

[0021] The functional types mentioned above may include function enhancement (favorable mutation), function unchanged (neutral mutation) and function reduction (unfavorable mutation).

[0022] The loss value described in A6) can be obtained by calculating the gradient using the cross entropy loss function.

[0023] The training cycle for the parameter update may be 5 cycles.

[0024] The validation samples mentioned above may come from more than one sample in the mutant library, and the validation function type of the sample may come from literature collection, database and / or experimentally obtained experimentally validated function type (function enhancement, unchanged function or weakened function). The more than one sample may be greater than or equal to one sample.

[0025] The prediction model of the occurrence probability of various amino acids at each amino acid position of the protein described above can be obtained by training a multi-layer perceptron.

[0026] In order to solve the above technical problem, the present invention further provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor can execute the computer program to implement the following steps: S1) Obtaining training sample data: receiving structural data of a model training protein, extracting amino acid sequence information and structural information of the model training protein based on the structural data to obtain training sample data; S2) Model construction: embedding the amino acid sequence information of the model training protein into a vector space to obtain a sequence feature vector, embedding the structural information of the model training protein into a vector space to obtain a structural feature vector, adding the sequence feature vector and the structural feature vector to obtain a fusion feature vector of the model training protein; inputting the fusion feature vector of the model training protein into an encoder for data encoding to obtain an encoded protein feature vector; using the encoded protein feature vector as input data for training to obtain a prediction model for the probability of occurrence of various amino acids at each amino acid position of the protein; S3) Calculation of mutation effect of target protein mutants: construct a mutant library of target protein composed of target protein mutants, input the feature vector of the encoded mutant of the target protein mutant in the mutant library into the prediction model in S2) to obtain the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculate the mutation effect score of the target protein mutant according to the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; The characteristic vector of the encoded mutant is obtained by the following steps: extracting the target protein mutant amino acid sequence information and the target protein mutant structure information based on the structural data of the target protein mutant, embedding the target protein mutant amino acid sequence information into the vector space to obtain a mutant sequence characteristic vector, embedding the target protein mutant structure information into the vector space to obtain a mutant structure characteristic vector, adding the mutant sequence characteristic vector and the mutant structure characteristic vector to obtain a mutant fusion characteristic vector; inputting the mutant fusion characteristic vector into the encoder for data encoding to obtain an encoded mutant feature vector; S4) Classification of mutant functional types: setting a threshold for the mutation effect score, and obtaining the predicted functional type of the target protein mutant according to the threshold; S5) receiving small sample data: receiving the verification function type of the verification sample (small sample) in the mutant library; S6) Model optimization: based on the prediction function type described in step S4), the prediction function type of the verification sample is obtained, and the loss value of the prediction model is calculated by comparing the distribution difference between the prediction function type of the verification sample and the verification function type of the verification sample obtained in step S5), and the parameters of the prediction model are updated based on the loss value to obtain an optimized prediction model.

[0027] In the above-mentioned computer device, the encoder in S2) may be a Transformer encoder, and the Transformer encoder may include 33 layers; updating the parameters of the prediction model in S6) includes freezing the first 29 layers of the Transformer encoder and iteratively updating the Transformer parameters of the remaining 4 layers of the Transformer encoder except the 29 layers.

[0028] The vector space mentioned above may be a vector space with 1280 dimensions.

[0029] The functional types mentioned above may include function enhancement (favorable mutation), function unchanged (neutral mutation) and function reduction (unfavorable mutation).

[0030] The loss value described in S6) can be obtained by calculating the gradient using the cross entropy loss function.

[0031] The training cycle for the parameter update may be 5 cycles.

[0032] The validation samples mentioned above may come from more than one sample in the mutant library, and the validation function type of the sample may come from literature collection, database and / or experimentally obtained experimentally validated function type (function enhancement, unchanged function or weakened function). The more than one sample may be greater than or equal to one sample.

[0033] The prediction model of the occurrence probability of various amino acids at each amino acid position of the protein described above can be obtained by training a multi-layer perceptron.

[0034] In the above-mentioned computer device, the steps may further include: S7) Secondary calculation of mutation effect: inputting the feature vector of the encoded mutant of the target protein mutant in the mutant library into the optimized prediction model in S6) to obtain an updated result of the occurrence probability of various amino acids at each amino acid position in the target protein mutant; calculating and obtaining the optimized mutation effect score of the target protein mutant according to the updated result, and obtaining the mutation effect of the target protein.

[0035] The above steps may further include: S8) target mutant output: sorting the target protein mutants from high to low according to the optimized mutation effect scores, and outputting the target modified mutants of the target protein according to the sorting results.

[0036] In order to solve the above technical problems, the present invention also provides a method for (precisely) modifying a protein, which may include the following steps: C1) Obtaining training sample data: receiving structural data of a model training protein, extracting amino acid sequence information and structural information of the model training protein based on the structural data, and obtaining training sample data; C2) Model construction: embedding the amino acid sequence information of the model training protein into a vector space to obtain a sequence feature vector, embedding the structural information of the model training protein into a vector space to obtain a structural feature vector, adding the sequence feature vector and the structural feature vector to obtain a fusion feature vector of the model training protein; inputting the fusion feature vector of the model training protein into an encoder for data encoding to obtain an encoded protein feature vector; using the encoded protein feature vector as input data for training to obtain a prediction model for the probability of occurrence of various amino acids at each amino acid position of the protein; C3) Calculation of mutation effect of target protein mutants: construct a mutant library of target protein composed of target protein mutants, input the feature vector of the encoded mutant of the target protein mutant in the mutant library into the prediction model in C2) to obtain the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculate the mutation effect score of the target protein mutant according to the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; The characteristic vector of the encoded mutant is obtained by the following steps: extracting the target protein mutant amino acid sequence information and the target protein mutant structure information based on the structural data of the target protein mutant, embedding the target protein mutant amino acid sequence information into the vector space to obtain a mutant sequence characteristic vector, embedding the target protein mutant structure information into the vector space to obtain a mutant structure characteristic vector, adding the mutant sequence characteristic vector and the mutant structure characteristic vector to obtain a mutant fusion characteristic vector; inputting the mutant fusion characteristic vector into the encoder for data encoding to obtain an encoded mutant feature vector; C4) Classification of mutant functional types: setting a threshold for the mutation effect score, and obtaining the predicted functional type of the target protein mutant according to the threshold; C5) Small sample data reception: receiving the verification function type of the verification sample (small sample) in the mutant library; C6) Model optimization: based on the predicted function type in step C4), the predicted function type of the verification sample is obtained, and the loss value of the prediction model is calculated by comparing the distribution difference between the predicted function type of the verification sample and the verification function type of the verification sample obtained in step C5), and the parameters of the prediction model are updated based on the loss value to obtain an optimized prediction model; C7) Secondary calculation of mutation effect: inputting the feature vector of the encoded mutant of the target protein mutant in the mutant library into the optimized prediction model in C6) to obtain an updated result of the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculating and obtaining the optimized mutation effect score of the target protein mutant according to the updated result, and obtaining the mutation effect of the target protein; C8) Output of target mutants: sorting the target protein mutants from high to low according to the optimized mutation effect scores, and outputting the target modified mutants of the target protein according to the sorting results.

[0037] In the above method, the encoder in C2) may be a Transformer encoder, and the Transformer encoder may include 33 layers; in C6), updating the parameters of the prediction model includes freezing the first 29 layers of the Transformer encoder and iteratively updating the Transformer parameters of the remaining 4 layers of the Transformer encoder except the 29 layers.

[0038] The vector space mentioned above may be a vector space with 1280 dimensions.

[0039] The functional types mentioned above may include function enhancement (favorable mutation), function unchanged (neutral mutation) and function reduction (unfavorable mutation).

[0040] The prediction model of the occurrence probability of various amino acids at each amino acid position of the protein described above can be obtained by training a multi-layer perceptron.

[0041] The loss value described in C6) can be obtained by calculating the gradient using the cross entropy loss function.

[0042] The training cycle for the parameter update may be 5 cycles.

[0043] The validation samples mentioned above may come from more than one sample in the mutant library, and the validation function type of the sample may come from literature collection, database and / or experimentally obtained experimentally validated function type (function enhancement, unchanged function or weakened function). The more than one sample may be greater than or equal to one sample.

[0044] In order to solve the above technical problems, the present invention also provides a computer-readable storage medium on which a computer program is stored. The computer program may be a program for (accurately) predicting the effects of protein mutations and / or a program for (accurately) designing and modifying proteins. When the program for (accurately) predicting the effects of protein mutations is executed by a processor, the steps of the method described above can be implemented; when the program for (accurately) designing and modifying proteins is executed by a processor, the steps of the method described above can be implemented.

[0045] The calculation formula for the mutation effect score mentioned above can be as follows (7): Mutation Effect Score = Formula (7); In formula (7), T It is the set of mutations contained in a certain protein mutant; t is any mutation in the mutation set; x t is the location of the mutation, mt is the probability of the mutant amino acid appearing, wt is the probability of occurrence of the wild-type amino acid, is the sequence context information, C It is the structural context information. and C All are included in the protein's fusion feature vector.

[0046] The structural data of the protein used for the model training described above can be derived from a database.

[0047] The various amino acids mentioned above may be 20 kinds of (common) amino acids, and the 20 kinds of (common) amino acids may specifically be glycine (G), alanine (A), valine (V), leucine (L), isoleucine (I), methionine (methionine) (M), proline (P), tryptophan (W), serine (S), tyrosine (Y), cysteine ​​(C), phenylalanine (F), asparagine (N), glutamine (Q), threonine (T), aspartic acid (D), glutamic acid (E), lysine (K), arginine (R) and histidine (H).

[0048] Due to the lack of effective small sample data application capabilities and multimodal information fusion capabilities, existing protein language models have problems with prediction accuracy and effectiveness when actually used to guide protein design and modification. To address this problem, the present invention proposes a protein mutation effect prediction method based on small sample data and multimodal information, which uses target protein mutant functional data collected from literature, downloaded from databases, or fed back from wet experiments, and deeply integrates protein sequence context and structural context to accurately predict protein mutation effects, and supports dynamic tuning of the method in combination with experimental verification results. On this basis, the present invention completes an intelligent protein design and modification method and system, which supports the accurate and efficient realization of protein mutant design in combination with experimental feedback results.

[0049] The present invention designs a mutation effect prediction method based on small sample data and multimodal information. By making full use of a small amount of mutant functional data of the protein to be modified, and deeply integrating protein sequence and structure information, the mutation effect prediction accuracy is dynamically optimized to achieve efficient and intelligent protein mutant design and modification. The protein design and modification method and system based on small sample data and multimodal information proposed by the present invention have a series of significant advantages.

[0050] First, it has dynamic optimization capabilities. Compared with ordinary protein language model methods, the method of the present invention can dynamically adjust model parameters based on a small amount of mutant function data fed back by wet experiments, improve the prediction accuracy of mutation effects, and thus achieve more effective guidance of protein design and modification. Second, it supports multimodal information. Existing protein language models usually only use protein sequence information, while the method of the present invention further utilizes structural information that is more closely related to protein function to accurately predict the functional effects caused by amino acid mutations. Third, it is easy to use. Users do not need to understand the functional and structural characteristics of the target protein molecule. They can directly use the method of the present invention to quickly and accurately screen the virtual mutation library and provide candidate mutation modification plans, which significantly reduces the difficulty of use.

[0051] In order to verify the advanced nature of the method designed by the present invention, the predictive validity of the prediction model in the method of the present invention is first verified using the ProteinGym test benchmark, which covers 42 proteins and 1.19 million mutant function data in humans, other eukaryotes and prokaryotes. All mutant function data are measured by wet experimental methods such as deep mutation scanning, and the evaluation index is the Spearman correlation coefficient between the model-predicted mutant function data and the actual mutant function data. The comparison method is the current optimal mutation effect prediction method ProMEP. Figure 2 As shown in the figure, compared with ProMEP (n=0), the mutation effect prediction method ProMEP-SLWF of the present invention can effectively use a small amount of mutant data fed back by wet experiments to improve the prediction accuracy of mutation effects, and the improvement is positively correlated with the amount of data used by the model. Under 50 pairs of positive and negative samples, the average Spearman correlation coefficient between the model prediction value and the real mutant data increased from 0.367 (n=0) to 0.447 (n=50). Figure 3 and Figure 4 The number of training cycles and training strategies of ProMEP-SLWF under small sample data (50 pairs of positive and negative samples) were also compared. In general, the performance improvement of the model gradually reached saturation after 3 cycles of training under small sample data, and the effect was optimal when training for 5 cycles. In terms of training strategy, freezing some model parameters will bring better performance improvement. For example, when fine-tuning all model parameters, the average Spearman correlation coefficient is 0.397; as the number of frozen layers increases, the model's ability to predict mutation effects continues to improve, and reaches the best at 28 / 29 layers, with an average Spearman correlation coefficient of 0.447.

[0052] Secondly, ProMEP-SLWF was used to guide the design and modification of TnpB-ISAam1 protein. TnpB protein is considered to be the ancestor of V-type CAS protein and has gene editing activity, while TnpB-ISAam1 is a homologous protein of TnpB protein. The overall modification process is divided into two rounds. In the first round, ProMEP-SLWF directly predicted the mutant fitness scores without relying on any prior knowledge and experimental data, and screened the top 10 mutants for modification and verified their editing activity. In the second round, based on the editing activity results fed back from the first round of wet experiments, the prediction model in ProMEP-SLWF was dynamically tuned, and after the tuning was completed, the mutant fitness scores were re-predicted, and the top 10 mutants were screened for modification and their editing activity was verified. As shown in the attached figure. Figure 4The test results show that in the first round of predictions, the mutant positive rate predicted by ProMEP-SLWF was 10%, that is, only one mutant actually enhanced editing activity compared to the wild type. In the second round of predictions, the mutant positive (favorable mutation) rate predicted by ProMEP-SLWF reached 80%, and 8 mutants actually enhanced editing activity compared to the wild type, and the strongest mutant C266V increased its activity by about 1.4 times compared to the wild type (WT).

[0053] The computer program product of the present invention may be a software product that implements its solution mainly through a computer program.

[0054] The computer-readable storage medium refers to a carrier for storing data, which may be a magnetic tape, magnetic disk, floppy disk, optical disk, magneto-optical disk, ROM, PROM, VCD, DVD, hard disk, flash memory, U disk, CF card, SD card, MMC card, SM card, memory stick (Memory Stick) or xD card, etc.

[0055] the term: Multilayer Perceptron (MLP) is an artificial neural network model consisting of an input layer, multiple hidden layers, and an output layer. Each hidden layer contains multiple neurons, each neuron is connected to all neurons in the previous layer, and each connection has a weight. MLP can perform tasks such as classification or regression on input data by learning weights.

[0056] Normalized exponential function: also known as Softmax function, is a generalization of the logic function. It can "compress" a K-dimensional vector z containing any real number into another K-dimensional real vector σ(z), so that the range of each element is between (0,1) and the sum of all elements is 1. This function is often used in multi-classification problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic diagram of the process of the present invention.

[0058] Figure 2 The figure shows a comparison of the mutation effect prediction capabilities of the prediction model in ProMEP-SLWF of the present invention under different amounts of fine-tuning data.

[0059] Figure 3 The figure shows a comparison of the mutation effect prediction capabilities of the prediction model in ProMEP-SLWF of the present invention under different fine-tuning cycles.

[0060] Figure 4 This is a comparison of the mutation effect prediction capabilities of the prediction model in the ProMEP-SLWF method of the present invention under different fine-tuning strategies.

[0061] Figure 5 This is a comparison of the enzyme activity of the TnpB-ISAam1 protein mutants screened by the ProMEP-SLWF method of the present invention before and after the optimization of its prediction model. The upper figure shows the mutant enzyme activity results before the prediction model is optimized; the lower figure shows the mutant enzyme activity results after the prediction model is optimized. DETAILED DESCRIPTION

[0062] The present invention is further described in detail below in conjunction with specific embodiments, and the examples provided are only for illustrating the present invention, rather than for limiting the scope of the present invention. The examples provided below can be used as a guide for further improvements by those of ordinary skill in the art, and do not constitute a limitation of the present invention in any way.

[0063] The experimental methods in the following examples, unless otherwise specified, are all conventional methods, and are performed according to the techniques or conditions described in the literature in the field or according to the product instructions. The materials, reagents, etc. used in the following examples, unless otherwise specified, can all be obtained from commercial channels.

[0064] The 20 (common) amino acids in the present invention are glycine (G), alanine (A), valine (V), leucine (L), isoleucine (I), methionine (methionine) (M), proline (P), tryptophan (W), serine (S), tyrosine (Y), cysteine ​​(C), phenylalanine (F), asparagine (N), glutamine (Q), threonine (T), aspartic acid (D), glutamic acid (E), lysine (K), arginine (R) and histidine (H).

[0065] Example 1. Based on the protein mutation effect combined with small sample data and multimodal information, practical guidance of TnpB-ISAam1 protein design and modification.

[0066] 1. Construction of virtual mutant library of TnpB-ISAam1 protein.

[0067] According to the different settings for the number of mutant amino acids in the wild-type protein, a protein virtual mutation library is constructed (such as a single-point mutation library in which each mutant contains only one amino acid mutation, a double-point mutation library in which each mutant contains only two amino acid mutations, etc.). The relationship between the size of the virtual mutation library and the number of mutations is calculated as follows: Formula (6); In formula (6), is the number of mutated amino acids, L is the length of the protein sequence, Represents a temporary variable, with a value of 1 to , for The factorial of .

[0068] Taking the TnpB-ISAam1 protein as an example, the amino acid sequence length of the protein is 369aa, and the total number of single-point mutants is 7011.

[0069] 2. Calculation of mutation effects of mutants in the TnpB-ISAam1 protein virtual mutation library.

[0070] 2.1 Construction of a prediction model for the probability of common amino acids occurring at each position in the amino acid sequence of a protein.

[0071] The model training of the present invention uses 160 million high-precision protein structures predicted by AlphaFold2 as training data, and systematically learns the sequence distribution and spatial folding rules of massive proteins: protein sequence and structural information are used as input; the network structure encoder based on the multi-layer Transformer fuses protein sequence features and structural features to further learn the distribution and spatial folding rules of amino acid sequences; and outputs the probability of occurrence of 20 common amino acids at each position in the amino acid sequence.

[0072] 2.1.1 Multimodal feature fusion.

[0073] The amino acid sequence information and structural information of the protein extracted from the protein structure information data are embedded into a 1280-dimensional vector space, and then the sequence feature vector and the structural feature vector are fused and added to obtain the fused feature vector of the protein. For a protein of length L, the module finally outputs a fused feature vector of L×1280, which contains the sequence context and structural context information of the entire protein.

[0074] Sequence embedding module: A multilayer perceptron (MLP) model is used to dynamically embed the amino acid sequences of multiple proteins. Each amino acid in the amino acid sequence of each protein is embedded as a vector of length 1280, represented as , that is, converting the amino acid sequence of each protein into a 1280-dimensional vector representation to obtain the protein sequence feature vector.

[0075] Structural embedding module: A 1-layer SE(3)-Transformer model is used to dynamically embed protein structure information. Each amino acid residue in the protein structure is also embedded as a vector of length 1280 (1280 dimensions), represented as , that is, converting the amino acid sequence in each protein structure into a 1280-dimensional vector representation to obtain the structural feature vector of the protein.

[0076] Fusion feature vector: The sequence feature vector of the protein and the structural feature vector of the protein are combined and added to obtain the fusion feature vector of each protein, which includes 1280 feature vectors for each amino acid fusion. The embedding vector (fused feature vector) of the entire protein is expressed as: ; In formula (1), H x represents the embedding vector of all amino acids in the input protein; T represents transposition, n represents the last amino acid number of the protein (for a protein of length m, n=m-1), There are 1280 feature vectors fused to each amino acid in the protein.

[0077] 2.1.2 Encoder module.

[0078] A multi-layer Transformer encoder is used to further encode the fusion features of each protein for further learning of the amino acid sequence distribution and spatial folding rules. The encoding module takes the fusion feature vector of the protein obtained in step 2.1.1 as input, uses the attention mechanism to mine the interactive relationship between the 1280-dimensional feature vectors of each residue of the protein, and finally outputs the feature vector of the further encoded protein: a feature vector of L×1280 dimensions.

[0079] Specifically, the protein embedding representation is further updated. The encoder module of the present invention includes 33 layers, and each layer of the encoder includes two main steps: 1) Self-attention calculation (self-attention).

[0080] First, perform self-attention calculation: ; In formula (2), Represents the encoder at The temporary output of the layer, Represents the encoder at The final output of the layer is also the encoder's The input of the layer, " represents the nesting of functions, and MultiHead represents multi-head attention operation.

[0081] 2) Feedforward neural network calculation.

[0082] Then use a feedforward neural network to calculate: ; In formula (3), Represents the encoder at The final output of the layer is , , Represents a learnable parameter. Assuming that the length of a protein is m, after encoding, the encoding representation of each protein is: Formula (4); In formula (4), Represents the entire protein The feature vectors of all amino acids obtained after the layer encoder, represents the number of encoder layers, a represents the input protein, T represents transposition, Indicates passing The feature vector of each amino acid obtained after the layer encoder.

[0083] The encoding module further encodes the fusion features of each protein to obtain the further encoding result of each protein: the encoded protein fusion feature vector, that is, the feature vector of L×1280 dimensions.

[0084] 2.1.3. Prediction module.

[0085] A multi-layer perceptron and a normalized exponential function (Softmax function) are used to train and calculate the prediction model for the occurrence probability of various (20 common) amino acids at each amino acid position in the amino acid sequence of the protein based on the encoded protein fusion feature vector obtained by the encoding module, which contains the encoding results of each amino acid position: ; In formula (5), The fused feature vector obtained by the encoder represents the sequence; a represents the input protein; and Represents the weight parameter of the prediction module; The representative model predicts the probability of occurrence of 20 common amino acids at each position in the amino acid sequence of a protein.

[0086] 2.2 Calculation of mutation effect score for each mutant in the mutation library.

[0087] The structural information data of the TnpB-ISAam1 protein (nuclease, WP066322673.1) of the present invention and the TnpB-ISAam1 protein mutants in the mutant library are derived from the prediction results of the AlphaFold2 protein structure prediction model based on the amino acid sequence, and can also be searched through protein structure databases such as PDB or other protein structure prediction models or websites for protein structure prediction results.

[0088] According to the structural information of the TnpB-ISAam1 protein mutant, the encoder module of step 2.1.2 is used to obtain the encoded (fused) feature vector of the TnpB-ISAam1 protein mutant (the amino acid sequence information of the TnpB-ISAam1 protein mutant is embedded in the vector space to obtain the mutant sequence feature vector, the structural information of the TnpB-ISAam1 protein mutant is embedded in the vector space to obtain the mutant structure feature vector, the mutant sequence feature vector and the mutant structure feature vector are added to obtain the mutant fusion feature vector; the mutant fusion feature vector is input into the Transformer encoder for further data encoding to obtain the encoded mutant feature vector); according to the prediction model of formula (5) obtained by the prediction module of step 2.1.3, the encoded feature vector of the TnpB-ISAam1 protein mutant is input to calculate the occurrence probability of various amino acids at each position in the amino acid sequence of the TnpB-ISAam1 protein mutant; then the mutation effect score of each TnpB-ISAam1 mutant is calculated by the following formula (7): Mutation Effect Score = ; In formula (7), T It is the set of mutations contained in a certain protein mutant; t is any mutation in the mutation set; x t is the location of the mutation, mt is the probability of the mutant amino acid appearing, wt is the probability of occurrence of the wild-type amino acid, is the sequence context information, C It is the structural context information. and C All are included in the protein's fusion feature vector.

[0089] For a double-point mutant, for example, a double-point mutant is R2A and S10A, which contains mutations at two different positions, the 2nd and 10th positions of the amino acid sequence. Then the mutation set T can be expressed as [2, 10]. Each time the mutation score is calculated, a mutation is taken from this set, such as taking position 2 for the first time and calculating the score; taking position 10 for the second time and calculating the score. Finally, the score of the double-point mutant is the sum of the two scores.

[0090] According to the mutation effect score of the mutant, the mutation is divided into a favorable mutation, a neutral mutation or an unfavorable mutation by setting a score threshold. The score threshold set for the TnpB-ISAam1 protein in the present invention is 1.05, and the mutant whose editing activity (enzyme activity) is more than 1.05 times the editing activity of the wild-type TnpB-ISAam1 protein is a favorable mutation (function enhancement mutation), the mutant whose editing activity is between 0.95 times and 1.05 times the editing activity of the wild-type TnpB-ISAam1 protein is a neutral mutation (function unchanged mutation), and the mutant whose editing activity is less than 0.95 times the editing activity of the wild-type TnpB-ISAam1 protein is an unfavorable mutation (function weakening mutation).

[0091] 3. First round of mutant screening.

[0092] According to the mutant fitness score, all mutants in the TnpB-ISAam1 protein virtual mutation library are sorted from high to low, and the top X (such as 10) mutants are output as the first round of mutant screening results of TnpB-ISAam1, that is, the top 10 TnpB-ISAam1 protein mutants are used as templates to verify the downstream transformation of TnpB-ISAam1, mainly including: K155S, D327G, V280A, S5A, N104S, V2A, V280S, F137I, T56K and N271D. K155S represents the mutant obtained by mutating the 155th amino acid residue of the amino acid sequence of the TnpB-ISAam1 protein from lysine K to serine S.

[0093] 4. Verification of mutant editing activity.

[0094] Ten recombinant encoding plasmids corresponding to the ten mutants (verified mutants) of the TnpB-ISAam1 protein obtained in step 3 were constructed by standard molecular cloning techniques, amplified using pST1374 vector and Phanta Max ultra-high-fidelity DNA polymerase, and assembled by Golden Gate cloning to form an array of TnpB-ISAam1 mutant recombinant encoding plasmids.

[0095] Secondly, HEK293T cells were used to culture and transfect 10 kinds of TnpB-ISAam1 mutant recombinant coding plasmids, and cells were added with FBS and antibiotics in DMEM culture medium to maintain cell growth. During transfection, the recombinant plasmid constructed was introduced into cells using the transfection reagent of Vazyme. After plasmid transfection, cells were treated with selective antibiotics, and samples were finally collected for deep sequencing to evaluate the editing efficiency of TnpB-ISAam1 protein mutants. Targeted deep sequencing was performed using specific primers to obtain mutant data, and then data analysis was performed using CRISPResso2 to obtain the editing effect of TnpB-ISAam1 protein mutants at the target site in the genome. The target site used in the present embodiment is the AGBL1 endogenous site, and the editing activity of TnpB-ISAam1 protein mutants is the average value of three independent biological experiment results. The results of the first round of verification showed that the mutant positivity rate (favorable mutation ratio) predicted before optimization in the method of the present invention was 10%, that is, compared with the wild type, only one mutant (N271D) actually enhanced the editing activity of the TnpB-ISAam1 protein (function-enhancing mutation).

[0096] 5. Small sample data processing.

[0097] The functional type data (enzyme activity) of the 10 TnpB-ISAam1 mutants fed back from the first round of experiments in step 4 at the AGBL1 endogenous site were divided into function (enzyme activity) enhancing mutations (N271D) and function (enzyme activity) weakening mutations (K155S, D327G, V280A, S5A, N104S, V2A, V280S, F137I and T56K), and this category information (functional type) was input into the prediction model obtained in step 2.1 for optimization.

[0098] 6. Dynamic tuning of protein modification systems.

[0099] According to step 2.1, the occurrence probability of various amino acids at each amino acid position in the amino acid sequence of the TnpB-ISAam1 protein is obtained; at the same time, the mutation effect score is calculated based on step 2.2, and the predicted functional type (mutation type) of the verification mutant (small sample) in steps 4 and 5 is determined by the mutation effect score, and the distribution difference between it and the verification functional type (whether it is a functional enhancement mutation) of the verification mutant obtained by experimental verification in steps 4 and 5 is compared, and the gradient is calculated based on this through the cross entropy loss function to obtain the loss value of the prediction model in step 2.1. Afterwards, the prediction model is optimized based on the loss value: the parameters of the first 29 layers of Transformer in the ProMEP-SLWF encoder module are frozen, and the parameters of the remaining 4 layers of Transformer in the encoder module are iteratively updated using the full parameter fine-tuning technology, and the training cycle is set to 5 cycles to obtain the optimized prediction model.

[0100] 7. Mutation effect scores are recalculated.

[0101] Using the prediction model optimized in step 6, according to the process of step 2, the optimized mutation effect score of the target protein is recalculated using formula (7), and the mutation effect of the target protein is obtained, which is the process of the protein mutation effect prediction method (ProMEP-SLWF) of the present invention. ProMEP-SLWF is used to recalculate the optimized mutation effect scores of 7011 single-point mutants of the TnpB-ISAam1 protein.

[0102] 8. Second round of mutant screening and verification.

[0103] According to the optimized mutation effect scores of the mutants obtained in step 7, all mutants in the TnpB-ISAam1 protein virtual mutation library are ranked from high to low, and the top X mutants (TnpB-ISAam1 mutants) are used as the TnpB-ISAam1 protein modification scheme, that is, TnpB-ISAam1 is modified using X TnpB-ISAam1 mutants as templates.

[0104] The present invention again selected the top 10 TnpB-ISAam1 mutants for downstream transformation verification, mainly including: S348I (representing a mutant obtained by mutating the 348th amino acid residue of the amino acid sequence of the TnpB-ISAam1 protein from serine S to isoleucine I), S123K, R200K, N271D, P176K, A201E, M363L, A151K, C266V and T181K. Afterwards, according to the steps described in step 4, the functional type (enzyme activity) of the TnpB-ISAam1 protein mutant was evaluated again. The second round of screening prediction results showed that the mutant positive rate (favorable mutation ratio) predicted by ProMEP-SLWF reached 80%. Compared with the wild-type TnpB-ISAam1 protein, 8 mutants (S348I, S123K, R200K, N271D, P176K, M363L, C266V and T181K) actually enhanced the enzyme activity, and the strongest mutant C266V increased the functional activity by about 1.4 times compared with the wild-type TnpB-ISAam1 protein (WT). The enzyme activity results of the TnpB-ISAam1 protein mutants screened in this example are shown in Figure 5 .

[0105] Example 2. Detection of the effect of the method of the present invention and comparison of the results with those of the prior art.

[0106] In order to verify the advancedness of the method designed by the present invention, the present invention uses the functional data of 42 proteins and 1.19 million mutants covering humans, other eukaryotes and prokaryotes in the ProteinGym test benchmark (data source: https: / / proteingym.org / ) to verify the small sample data application ability and dynamic tuning strategy of ProMEP-SLWF. All mutant functional data are measured by wet experimental methods such as deep mutation scanning, and the evaluation index is the Spearman correlation coefficient between the model-predicted mutant functional data and the actual mutant functional data. The comparison method is the best mutation effect prediction method ProMEP reported in the current literature (reference: Cheng, P., Mao, C., Tang, J. et al. Zero-shot prediction of mutation effects with multimodal deep representation learningguides protein engineering. Cell Res 34, 630–647 (2024)). Figure 2The test results show that compared with ProMEP (n=0), ProMEP-SLWF can effectively use a small amount of mutant data fed back by wet experiments to improve the prediction accuracy of mutation effects, and the degree of improvement is positively correlated with the amount of data (number of samples) used by the model. Under 100 pairs of positive and negative samples, the average Spearman correlation coefficient between the model prediction value and the real mutant data increased from 0.367 (n=0) to 0.447 (n=50). Figure 3 and Figure 4 The number of training cycles and training strategies of ProMEP-SLWF under small sample data (50 pairs of positive and negative samples) were also compared. In general, the performance improvement of the model gradually reached saturation after 3 cycles of training under small sample data, and the effect was optimal when training for 5 cycles. In terms of training strategy, freezing some model parameters will bring better performance improvement. For example, when fine-tuning all model parameters, the average Spearman correlation coefficient is 0.397; as the number of frozen layers increases, the model's ability to predict mutation effects continues to improve, and reaches the best at 28 / 29 layers, with an average Spearman correlation coefficient of 0.447.

[0107] The present invention has been described in detail above. For those skilled in the art, without departing from the purpose and scope of the present invention, and without the need to carry out unnecessary experimental conditions, the present invention can be implemented in a wide range under equivalent parameters, concentrations and conditions. Although the present invention provides specific embodiments, it should be understood that the present invention can be further improved. In a word, according to the principles of the present invention, the application is intended to include any changes, uses or improvements to the present invention, including departure from the disclosed scope in the application, and changes made with conventional techniques known in the art.

Claims

1. A method for predicting the effect of a protein mutation, characterized in that: The method comprises the following steps: B1) Obtaining training sample data: receiving structural data of model training proteins, extracting amino acid sequence information and structural information of model training proteins based on the structural data, and obtaining training sample data; B2) Model construction: embedding the amino acid sequence information of the model training protein into a vector space to obtain a sequence feature vector, embedding the structural information of the model training protein into a vector space to obtain a structural feature vector, adding the sequence feature vector and the structural feature vector to obtain a fusion feature vector of the model training protein; inputting the fusion feature vector of the model training protein into an encoder for data encoding to obtain an encoded protein feature vector; using the encoded protein feature vector as input data for training to obtain a prediction model for the probability of occurrence of various amino acids at each amino acid position of the protein; B3) Calculation of mutation effect of target protein mutants: construct a mutant library of target protein composed of target protein mutants, input the feature vector of the encoded mutant of the target protein mutant in the mutant library into the prediction model described in B2) to obtain the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculate the mutation effect score of the target protein mutant according to the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; The characteristic vector of the encoded mutant is obtained by the following steps: extracting the target protein mutant amino acid sequence information and the target protein mutant structure information based on the structural data of the target protein mutant, embedding the target protein mutant amino acid sequence information into the vector space to obtain a mutant sequence characteristic vector, embedding the target protein mutant structure information into the vector space to obtain a mutant structure characteristic vector, and adding the mutant sequence characteristic vector and the mutant structure characteristic vector to obtain a mutant fusion characteristic vector; Inputting the mutant fusion feature vector into the encoder for data encoding to obtain an encoded mutant feature vector; B4) Classification of mutant functional types: setting a threshold for the mutation effect score, and obtaining the predicted functional type of the target protein mutant according to the threshold; B5) Small sample data reception: receiving the verification function type of the verification sample in the mutant library; B6) Model optimization: based on the predicted function type in step B4), the predicted function type of the verification sample is obtained, and the loss value of the prediction model is calculated by comparing the distribution difference between the predicted function type of the verification sample and the verification function type of the verification sample obtained in step B5), and the parameters of the prediction model are updated based on the loss value to obtain an optimized prediction model; B7) Secondary calculation of mutation effect: input the feature vector of the encoded mutant of the target protein mutant in the mutant library into the optimized prediction model described in B6) to obtain an updated result of the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculate the optimized mutation effect score of the target protein mutant according to the updated result to obtain the mutation effect of the target protein.

2. The method according to claim 1, characterized in that: The encoder in B2) is a Transformer encoder, and the Transformer encoder includes 33 layers; in B6), updating the parameters of the prediction model includes freezing the first 29 layers of the Transformer encoder and iteratively updating the Transformer parameters of the remaining 4 layers of the Transformer encoder except the 29 layers.

3. A device for predicting the effects of protein mutations, characterized in that: The device comprises the following modules: A1) a training sample data acquisition module: used to receive the structural data of the model training protein, extract the amino acid sequence information and the structural information of the model training protein based on the structural data, and obtain the training sample data; A2) Model building module: used for embedding the amino acid sequence information of the model training protein into a vector space to obtain a sequence feature vector, embedding the structural information of the model training protein into a vector space to obtain a structural feature vector, adding the sequence feature vector and the structural feature vector to obtain a fusion feature vector of the model training protein; inputting the fusion feature vector of the model training protein into an encoder for data encoding to obtain an encoded protein feature vector; using the encoded protein feature vector as input data for training to obtain a prediction model for the occurrence probability of various amino acids at each amino acid position of the protein; A3) Target protein mutant mutation effect calculation module: used to construct a mutant library of target protein composed of target protein mutants, input the feature vector of the encoded mutant of the target protein mutant in the mutant library into the prediction model in A2) to obtain the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculate the mutation effect score of the target protein mutant according to the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; The characteristic vector of the encoded mutant is obtained by the following steps: extracting the target protein mutant amino acid sequence information and the target protein mutant structure information based on the structural data of the target protein mutant, embedding the target protein mutant amino acid sequence information into the vector space to obtain a mutant sequence characteristic vector, embedding the target protein mutant structure information into the vector space to obtain a mutant structure characteristic vector, and adding the mutant sequence characteristic vector and the mutant structure characteristic vector to obtain a mutant fusion characteristic vector; Inputting the mutant fusion feature vector into the encoder for data encoding to obtain an encoded mutant feature vector; A4) Mutant function type classification module: used to set a threshold for the mutation effect score, and obtain the predicted function type of the target protein mutant according to the threshold; A5) Small sample data receiving module: used for receiving the verification function type of the verification sample in the mutant library; A6) Model optimization module: used to obtain the prediction function type of the verification sample based on the prediction function type in step A4), calculate the loss value of the prediction model by comparing the distribution difference between the prediction function type of the verification sample and the verification function type of the verification sample obtained in step A5), and update the parameters of the prediction model based on the loss value to obtain an optimized prediction model; A7) Mutation effect secondary calculation module: used to input the feature vector of the encoded mutant of the target protein mutant in the mutant library into the optimized prediction model in A6) to obtain an updated result of the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculate and obtain the optimized mutation effect score of the target protein mutant according to the updated result to obtain the mutation effect of the target protein.

4. The device according to claim 3, characterized in that: The encoder in A2) is a Transformer encoder, and the Transformer encoder includes 33 layers; in A6), updating the parameters of the prediction model includes freezing the first 29 layers of the Transformer encoder and iteratively updating the Transformer parameters of the remaining 4 layers of the Transformer encoder except the 29 layers.

5. A computer device comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the following steps: S1) Obtaining training sample data: receiving structural data of a model training protein, extracting amino acid sequence information and structural information of the model training protein based on the structural data to obtain training sample data; S2) Model construction: embedding the amino acid sequence information of the model training protein into a vector space to obtain a sequence feature vector, embedding the structural information of the model training protein into a vector space to obtain a structural feature vector, adding the sequence feature vector and the structural feature vector to obtain a fusion feature vector of the model training protein; inputting the fusion feature vector of the model training protein into an encoder for data encoding to obtain an encoded protein feature vector; using the encoded protein feature vector as input data for training to obtain a prediction model for the probability of occurrence of various amino acids at each amino acid position of the protein; S3) Calculation of mutation effect of target protein mutants: construct a mutant library of target protein composed of target protein mutants, input the feature vector of the encoded mutant of the target protein mutant in the mutant library into the prediction model in S2) to obtain the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculate the mutation effect score of the target protein mutant according to the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; The characteristic vector of the encoded mutant is obtained by the following steps: extracting the target protein mutant amino acid sequence information and the target protein mutant structure information based on the structural data of the target protein mutant, embedding the target protein mutant amino acid sequence information into the vector space to obtain a mutant sequence characteristic vector, embedding the target protein mutant structure information into the vector space to obtain a mutant structure characteristic vector, and adding the mutant sequence characteristic vector and the mutant structure characteristic vector to obtain a mutant fusion characteristic vector; Inputting the mutant fusion feature vector into the encoder for data encoding to obtain an encoded mutant feature vector; S4) Classification of mutant functional types: setting a threshold for the mutation effect score, and obtaining the predicted functional type of the target protein mutant according to the threshold; S5) receiving small sample data: receiving the verification function type of the verification sample in the mutant library; S6) Model optimization: based on the prediction function type described in step S4), the prediction function type of the verification sample is obtained, and the loss value of the prediction model is calculated by comparing the distribution difference between the prediction function type of the verification sample and the verification function type of the verification sample obtained in step S5), and the parameters of the prediction model are updated based on the loss value to obtain an optimized prediction model.

6. The computer device according to claim 5, characterized in that: The encoder in S2) is a Transformer encoder, and the Transformer encoder includes 33 layers; in S6), updating the parameters of the prediction model includes freezing the first 29 layers of the Transformer encoder and iteratively updating the Transformer parameters of the remaining 4 layers of the Transformer encoder except the 29 layers.

7. The computer device according to claim 5, characterized in that: The steps also include: S7) Secondary calculation of mutation effect: inputting the feature vector of the encoded mutant of the target protein mutant in the mutant library into the optimized prediction model in S6) to obtain an updated result of the occurrence probability of various amino acids at each amino acid position in the target protein mutant; calculating and obtaining the optimized mutation effect score of the target protein mutant according to the updated result, and obtaining the mutation effect of the target protein.

8. A method for modifying a protein, characterized in that: The method comprises the following steps: C1) Obtaining training sample data: receiving structural data of a model training protein, extracting amino acid sequence information and structural information of the model training protein based on the structural data, and obtaining training sample data; C2) Model construction: embedding the amino acid sequence information of the model training protein into a vector space to obtain a sequence feature vector, embedding the structural information of the model training protein into a vector space to obtain a structural feature vector, adding the sequence feature vector and the structural feature vector to obtain a fusion feature vector of the model training protein; inputting the fusion feature vector of the model training protein into an encoder for data encoding to obtain an encoded protein feature vector; using the encoded protein feature vector as input data to train a prediction model for the probability of occurrence of various amino acids at each amino acid position of the protein; C3) Calculation of mutation effect of target protein mutants: construct a mutant library of target protein composed of target protein mutants, input the feature vector of the encoded mutant of the target protein mutant in the mutant library into the prediction model in C2) to obtain the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculate the mutation effect score of the target protein mutant according to the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; The characteristic vector of the encoded mutant is obtained by the following steps: extracting the target protein mutant amino acid sequence information and the target protein mutant structure information based on the structural data of the target protein mutant, embedding the target protein mutant amino acid sequence information into the vector space to obtain a mutant sequence characteristic vector, embedding the target protein mutant structure information into the vector space to obtain a mutant structure characteristic vector, and adding the mutant sequence characteristic vector and the mutant structure characteristic vector to obtain a mutant fusion characteristic vector; Inputting the mutant fusion feature vector into the encoder for data encoding to obtain an encoded mutant feature vector; C4) Classification of mutant functional types: setting a threshold for the mutation effect score, and obtaining the predicted functional type of the target protein mutant according to the threshold; C5) Small sample data reception: receiving the verification function type of the verification sample in the mutant library; C6) Model optimization: based on the predicted function type in step C4), the predicted function type of the verification sample is obtained, and the loss value of the prediction model is calculated by comparing the distribution difference between the predicted function type of the verification sample and the verification function type of the verification sample obtained in step C5), and the parameters of the prediction model are updated based on the loss value to obtain an optimized prediction model; C7) Secondary calculation of mutation effect: inputting the feature vector of the encoded mutant of the target protein mutant in the mutant library into the optimized prediction model in C6) to obtain an updated result of the probability of occurrence of various amino acids at each amino acid position in the target protein mutant; calculating and obtaining the optimized mutation effect score of the target protein mutant according to the updated result, and obtaining the mutation effect of the target protein; C8) Output of target mutants: sorting the target protein mutants from high to low according to the optimized mutation effect scores, and outputting the target modified mutants of the target protein according to the sorting results.

9. The method according to claim 8, characterized in that: The encoder in C2) is a Transformer encoder, and the Transformer encoder includes 33 layers; the parameter updating of the prediction model in C6) includes freezing the first 29 layers of the Transformer encoder and iteratively updating the Transformer parameters of the remaining 4 layers of the Transformer encoder except the 29 layers.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is a program for predicting the effects of protein mutations and / or a program for designing and modifying proteins. When the program for predicting the effects of protein mutations is executed by a processor, the steps of the method described in claim 1 or 2 are implemented; when the program for designing and modifying proteins is executed by a processor, the steps of the method described in claim 8 or 9 are implemented.

Images