Method for constructing protein stability prediction model, prediction method and system
The protein stability prediction model, constructed using natural language processing and machine learning algorithms, overcomes the limitations of sample size and technical approach, achieving high-performance sequence-based protein stability prediction, and is suitable for studies with sparse structural information.
Patent Information
- Application Number
- CN202311378716.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2022-12-28
- Filing Date
- 2023-10-23
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2043-10-23
AI Technical Summary
Existing protein stability prediction methods suffer from insufficient sample size, low sample quality, and limited technical approaches, leading to overfitting or underfitting of prediction models. Furthermore, their reliance on structural information restricts their applicability.
A pre-trained model using natural language processing is used to extract features from protein sequences. Combined with machine learning and deep learning algorithms, a protein stability prediction model is constructed. The number of samples is expanded by reversing the training set and supervised training is performed to achieve stability prediction based on sequence information.
It improves the performance and robustness of the prediction model, making it suitable for genome-scale studies with sparse structural information, avoiding overfitting problems, and expanding the applicability of the model.
Smart Images

Figure CN117476104B_ABST
Abstract
Description
[0001] Cross-references
[0002] This application claims priority to Chinese application No. 202211708226.8 filed on December 28, 2022, the entire contents of which are incorporated herein by reference. Technical Field
[0003] The present specification relates to the field of bioinformatics, and in particular to a method for constructing a protein stability prediction model, and a method and system for predicting protein stability after mutation. Background Art
[0004] The natural evolution of organisms has produced countless proteins with rich functions and complex structures, but natural proteins often cannot meet the needs of industrial production and theoretical research in terms of stability, tolerance, selectivity, etc., which has prompted humans to explore efficient protein modification methods. Protein stability prediction is an important link in the protein modification process. Protein stability is the basis for proteins to realize their functions. Proteins can only play specific functions through specific three-dimensional structures, and the ability to fold into corresponding three-dimensional structures is affected by protein stability (Ittisoponpisan et al. 2019; Ofoegbu et al. 2019).
[0005] The design and modification strategies of proteins have gone through a development process from primary rational design, directed evolution, semi-rational design, to computer-aided design. Using artificial intelligence technology to guide protein modification is one of the new strategies of computer-aided design that has gradually developed in recent years. It has shown unique advantages in many aspects such as protein structure prediction, function prediction, stability prediction and antibody affinity prediction. Natural language processing is one of the hottest fields of artificial intelligence at present, and is often used for text generation, question answering and translation. Natural language processing itself is suitable for analyzing the structure and meaning of text, and the most basic nucleic acids and proteins in life are composed of sequences, which has an intrinsic connection with text. The main concept behind these methods is to interpret protein sequences as sentences and their components - amino acids - as single words. There is still a certain gap in how to apply natural language processing methods to the field of life sciences, especially protein stability prediction.
[0006] At present, the application of artificial intelligence algorithms in protein stability prediction is limited by many problems. On the one hand, the above-mentioned application of artificial intelligence algorithms is limited by data set related issues. Artificial intelligence algorithms rely heavily on data, and the quantity and quality of initial data determine the generalization performance of the trained model. The data sets of existing protein stability prediction methods generally have problems with insufficient sample quantity or low quality. For example, existing protein stability prediction methods usually extract data sets from the ProTherm database, which contains less than 3,000 useful experimental mutation information data, which makes the prediction model prone to overfitting or underfitting. And the data sets of existing protein stability prediction methods generally have bias problems. For example, the bias of protein evolution and experiments leads to more than 70% of the samples in the data set being unstable mutations, and this bias will affect the performance of the prediction model. On the other hand, the above-mentioned application of artificial intelligence algorithms is limited by technical routes. Existing protein stability prediction methods are mainly structure-based prediction methods, such as FoldX (Guerois, Nielsen and Serrano 2002) and Rosetta (Leman et al. 2020), which have been commercialized, while the number of proteins with available three-dimensional structures in existing protein databases is very small. For example, only about 0.2% of proteins in the UniProt database have available three-dimensional structures, which directly limits the scope of use of structure-based prediction methods. There are still a small number of sequence-based prediction methods in the existing protein stability prediction methods, such as I-Mutant-seq (Capriotti, Fariselli and Casadio 2005), which only uses sequence information near the mutation site as a feature and uses support vector machine learning methods to predict changes in protein stability caused by mutations; DDGun-seq (Montanucci et al. 2022), which uses evolutionary scores and energy differences before and after mutations as features, and predicts changes in protein stability through formula fitting methods. The technical routes of sequence-based prediction methods are limited and need further development and research. Summary of the invention
[0007] One or more embodiments of the present specification provide a method for constructing a protein stability prediction model. The method includes: obtaining a training set, the training set includes information of multiple training samples, and the information of each training sample includes amino acid sequence information before and after the sample protein mutation and a mutation stability label; obtaining a pre-trained model, based on the information of the multiple training samples, performing feature extraction through the pre-trained model to obtain mutation features of the multiple training samples; and obtaining a classifier, based on the mutation features and the mutation stability label, training the classifier to obtain a prediction model, the prediction model includes the pre-trained model and the trained classifier.
[0008] In some embodiments, the mutation signature is used to indicate changes in a mutation region of an amino acid sequence before and after a sample protein mutation; the mutation region at least includes the amino acid corresponding to the mutation site.
[0009] In some embodiments, the mutation region is the amino acid corresponding to the mutation site.
[0010] In some embodiments, the mutation region comprises the amino acid corresponding to the mutation site and a plurality of amino acids adjacent to the mutation site; the plurality of amino acids adjacent to the mutation site are preferably adjacent 1-10 amino acids.
[0011] In some embodiments, the extracting features through the pre-training model to obtain the mutation features of the multiple training samples includes: extracting features through the pre-training model based on the amino acid sequence information of the sample protein before and after mutation of each of the training samples to obtain the pre-mutation features and post-mutation features of the training samples; and splicing the pre-mutation features and the post-mutation features to obtain the mutation features of the training samples.
[0012] In some embodiments, the amino acid sequence information of the sample protein of the training sample before and after mutation includes the amino acid sequence before mutation; the method for extracting the pre-mutation feature includes: obtaining N amino acids of the amino acid sequence before mutation; extracting at least one layer of output vectors through the pre-training model as the feature vector corresponding to the amino acid sequence before mutation, preferably the last layer, wherein the layer of output vectors includes N representation vectors and N position codes corresponding to the N amino acids of the amino acid sequence before mutation; and extracting the feature vector corresponding to the mutation region as the pre-mutation feature.
[0013] In some embodiments, the amino acid sequence information of the sample protein of the training sample before and after mutation includes the amino acid sequence after mutation; the method for extracting the post-mutation feature includes: obtaining N amino acids of the amino acid sequence after mutation; extracting at least one layer of output vectors through the pre-training model as the feature vector corresponding to the amino acid sequence after mutation, preferably the last layer, wherein the layer of output vectors includes N representation vectors and N position codes corresponding to the N amino acids of the amino acid sequence after mutation; and extracting the feature vector corresponding to the mutation region as the post-mutation feature.
[0014] In some embodiments, the pre-trained model is implemented using 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; more preferably, the pre-trained model is ProtT5-XL or ProtT5-XXL; more preferably, the pre-trained model is Prot T5-XL.
[0015] In some embodiments, the training set includes an initial training set and a reverse training set; the reverse training set is obtained by reversing the initial training set based on the thermodynamic characteristics of the protein, so that the number of training samples carrying different mutation stability labels in the training set is equal.
[0016] In some embodiments, the classifier is implemented by a combination of one or more of the following algorithms: linear regression algorithm, support vector machine algorithm, k-nearest neighbor algorithm, logistic regression algorithm, decision tree algorithm, k-means algorithm, random forest algorithm, naive Bayes algorithm, gradient boosting algorithm, ensemble learning algorithm, convolutional neural network, recurrent neural network, recursive neural network and long short-term memory network; more preferably, the classifier is implemented by a combination of one or more of the following algorithms: logistic regression algorithm, random forest algorithm, AdaBoost algorithm, support vector machine algorithm, gradient boosting decision tree algorithm, convolutional neural network and bidirectional long short-term memory network; more preferably, the classifier is implemented by one of the following algorithms: random forest algorithm, convolutional neural network and bidirectional long short-term memory network.
[0017] In some embodiments, before obtaining the classifier, the method further includes: performing dimensionality reduction processing on the mutation feature to obtain the mutation feature after dimensionality reduction; and the training of the classifier is based on the mutation feature after dimensionality reduction and the mutation stability label.
[0018] In some embodiments, the dimensionality reduction processing is implemented by a combination of one or more of the following algorithms: principal component analysis algorithm, discriminant analysis algorithm, multidimensional scaling transformation algorithm, isometric feature mapping algorithm and random neighbor embedding algorithm; preferably, the dimensionality reduction processing is implemented by a principal component analysis algorithm.
[0019] One or more embodiments of the present specification provide a method for predicting protein stability after mutation. The method includes: obtaining information of a sample to be predicted, the information of the sample to be predicted includes amino acid sequence information before and after the mutation of the protein to be tested; using a prediction model obtained by the above-mentioned method for constructing a protein stability prediction model to perform the following processing on the information of the sample to be predicted, the prediction model includes a pre-trained model and a trained classifier: based on the amino acid sequence information of the protein to be tested before and after the mutation of the sample to be predicted, using the pre-trained model to perform feature extraction to obtain the mutation characteristics of the sample to be predicted; and based on the mutation characteristics, using the trained classifier to perform prediction to determine the prediction result of the stability of the sample to be predicted after mutation.
[0020] One or more embodiments of the present specification also provide a system for constructing a stability prediction model for proteins after mutation. The system includes: an acquisition module for acquiring a training set, the training set includes information of multiple training samples, and the information of each training sample includes amino acid sequence information and mutation stability labels of the sample protein before and after mutation; a pre-training module for acquiring a pre-training model, based on the information of the multiple training samples, performing feature extraction through the pre-training model to obtain mutation features of the multiple training samples; and a training module for acquiring a classifier, based on the mutation features and the mutation stability labels, training the classifier to obtain a prediction model, the prediction model includes the pre-training model and the trained classifier.
[0021] One or more embodiments of the present specification also provide a system for predicting protein stability after mutation. The system includes: an acquisition module for acquiring information of a sample to be predicted, the information of the sample to be predicted includes amino acid sequence information before and after the mutation of the protein to be tested; a processing module for performing the following processing on the information of the sample to be predicted using a prediction model constructed by the above-mentioned protein stability prediction model construction method, the prediction model includes a pre-trained model and a trained classifier: based on the amino acid sequence information of the protein to be tested before and after the mutation of the sample to be predicted, the pre-trained model is used to perform feature extraction to obtain the mutation characteristics of the sample to be predicted; and based on the mutation characteristics, the trained classifier is used to perform prediction to determine the prediction result of the stability of the sample to be predicted after mutation.
[0022] One or more embodiments of the present specification also provide a device. The device includes at least one storage medium and at least one processor, wherein the at least one storage medium is used to store computer instructions; and the at least one processor is used to execute the computer instructions to implement the above-mentioned method for constructing a protein stability prediction model or the above-mentioned method for predicting protein stability after mutation.
[0023] One or more embodiments of the present specification also provide a computer-readable storage medium, characterized in that the storage medium stores computer instructions, and when the computer instructions are executed by a processor, the method for constructing a protein stability prediction model as described above or the method for predicting protein stability after mutation as described above is implemented. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:
[0025] Figure 1 It is a schematic diagram of the prediction model construction and prediction scenario of protein stability after mutation according to some embodiments of this specification;
[0026] Figure 2 is a schematic diagram of a process for constructing a protein stability prediction model according to some embodiments of this specification;
[0027] Figure 3 is a schematic diagram of a process for extracting mutation features of training samples according to some embodiments of this specification;
[0028] Figure 4 It is a schematic diagram of a process for extracting pre-mutation features and post-mutation features according to some embodiments of this specification;
[0029] Figure 5 It is a schematic diagram of a process for predicting protein stability after mutation according to some embodiments of this specification;
[0030] Figure 6 is an exemplary flow chart of the prediction model in different application scenarios according to some embodiments of this specification;
[0031] Figure 7 It is a module diagram of a system for constructing a protein stability prediction model according to some embodiments of this specification;
[0032] Figure 8 It is a module diagram of the system for predicting protein stability after mutation according to some embodiments of the present specification. DETAILED DESCRIPTION
[0033] In order to more clearly illustrate the technical solutions of the embodiments of this specification, the following is a brief introduction to the drawings required for the description of the embodiments. Obviously, the drawings described below are only some examples or embodiments of this specification. For ordinary technicians in this field, this specification can also be applied to other similar scenarios based on these drawings without creative work. Unless it is obvious from the language environment or otherwise explained, the same reference numerals in the figures represent the same structure or operation.
[0034] It should be understood that the "system", "device", "unit" and / or "module" used herein are a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the words can be replaced by other expressions.
[0035] As shown in this specification and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "comprise" and "include" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.
[0036] Flowcharts are used in this specification to illustrate the operations performed by the system according to the embodiments of this specification. It should be understood that the preceding or following operations are not necessarily performed precisely in order. Instead, the steps may be processed in reverse order or simultaneously. At the same time, other operations may also be added to these processes, or one or more operations may be removed from these processes.
[0037] Predicting the effect of amino acid mutations on protein stability plays an important role in production and research in many fields. Given that artificial intelligence technology has shown unique advantages in guiding protein modification in protein structure prediction, function prediction, stability prediction, and antibody affinity prediction, research and development of new artificial intelligence algorithm-based protein stability prediction methods after amino acid mutation will have great application value. At present, there are many problems in the application of artificial intelligence algorithms in protein stability prediction, such as insufficient sample size, low sample quality, or sample bias.
[0038] This specification provides a method for constructing a protein stability prediction model. The method uses a pre-trained model (PTMs) of natural language processing (NLP) as a feature extractor to extract features from training samples. The pre-trained model has learned biological field-related knowledge 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 supervised training is performed on the initial classifier based on the extracted mutation-related features, and only a small amount of labeled training samples are needed to optimize the classifier parameters. The prediction model constructed by the above method can predict the effect of amino acid mutations on protein stability based on sequence information, can be applied to genome-scale research and applications with sparse structural information, and has good model performance.
[0039] Figure 1 The following is a schematic diagram of the prediction model construction and prediction scenario of protein stability after mutation according to some embodiments of this specification. Figure 1 As shown, scenario 100 may include a processing device 110 , a storage device 120 , a user terminal 130 , a network 140 , and a data source 150 .
[0040] The processing device 110 can be used for processing and analyzing relevant information and / or data. In some embodiments, the processing device 110 can obtain relevant information and / or data from other devices or other components of the scene 100, and execute program instructions based on this information and / or data to perform one or more functions described in this specification. For example, the processing device 110 can obtain a training set from the data source 150, and obtain a pre-trained model and a classifier from the storage device 120, and build a model based on the training set, the pre-trained model and the classifier to obtain a prediction model. For another example, the processing device 110 can obtain information of a sample to be predicted from the data source 150, and use the prediction model obtained by the storage device 120 to perform relevant operations on the information of the sample to be predicted to obtain a stability prediction result after mutation. In some embodiments, the processing device 110 can be a server or a central processing unit.
[0041] The storage device 120 can be used to store data and / or information. In some embodiments, the storage device 120 can store data and / or information obtained from the processing device 110, the user terminal 130 and / or the data source 150, such as training sets, pre-trained models, classifiers, model optimization parameters, etc. In some embodiments, the storage device 120 can store instructions for the processing device 110 to execute or data and / or information used, and the processing device 110 can implement the exemplary method described in this specification by executing the instructions or using the data and / or information. In some embodiments, the storage device 120 can be connected to the network 140 to achieve communication with one or more other components in the scene 100 (for example, the processing device 110, the user terminal 130, the data source 150, etc.). The storage device 120 may include one or more storage components, each of which may be an independent device or part of another device. In some embodiments, the storage device 120 may include a random access memory (RAM), a read-only memory (ROM), a mass storage device, a removable memory, a volatile read-write memory, etc., or any combination thereof. Exemplarily, the mass storage may include a magnetic disk, an optical disk, a solid-state disk, etc. In some embodiments, the storage device 120 may be implemented on a cloud platform.
[0042] The user terminal 130 can be used for user interaction between the user and the scene 100. For example, the user terminal 130 can receive an instruction from the processing device 110 to predict the stability of the protein after the amino acid mutation based on the information of the sample to be predicted input by the user, and display the stability prediction result of the sample to be predicted to the user. In some embodiments, the user terminal 130 can be connected to the processing device 110, the storage device 120 and / or the data source 150 and / or communicate with them. In some embodiments, the user terminal 130 can include one or any combination of a smart phone 130-1, a tablet computer 130-2, a laptop computer 130-3, etc. or other devices with input and / or output functions.
[0043] The network 140 can provide information and / or data exchange. The network 140 connects the components of the scene 100 so that the components can communicate with each other. In some embodiments, the network 140 can be any one or more of a wired network or a wireless network. For example, the network 140 may include a cable network, a fiber optic network, etc. In some embodiments, the network 140 can be a point-to-point, shared, centralized, or a combination of multiple topologies. In some embodiments, the network 140 may include one or more network access points. For example, through entry and exit points such as a base station and / or one or more network exchange points, one or more components of the scene 100 can be connected to the network 140 to exchange data and / or information.
[0044] Data source 150 may include one or more sources that provide information for scenario 100. For example, data source 150 may be used to provide a training set for building a prediction model of protein stability for scenario 100. Data source 150 may exist in the form of a single central server or in the form of multiple servers connected via a network.
[0045] Figure 2 2 is a schematic diagram of a process for constructing a protein stability prediction model according to some embodiments of the present specification. In some embodiments, process 200 includes steps 210 to 250. At least a portion of the steps in process 200 (e.g., steps 210 to 250) can be performed by a computing device (e.g., Figure 1 For example, at least a portion of the steps in process 200 (eg, steps 210 to 250) may be implemented as an instruction (eg, an application) stored in storage device 120. Figure 1 The processing device 110 in the embodiment may execute the instruction, and when executing the instruction, the processing device 110 may be configured to perform process 200. The operations of the process shown below are for illustration purposes only. In some embodiments, process 200 may be completed using one or more additional operations not described and / or one or more operations not described. In addition, Figure 2 The order in which the operations of the processes are illustrated and described below is not intended to be limiting.
[0046] In step 210, a training set may be obtained, the training set including information of multiple training samples, each of the training samples including amino acid sequence information of the sample protein before and after mutation and a mutation stability label. In some embodiments, step 210 may be performed by a computing device (e.g., processing device 110, acquisition module 710).
[0047] As used herein, the information of the training sample refers to the information and / or data related to the sample protein mutation used to train the prediction model. In some embodiments, the sample protein may include an artificially induced mutated protein and / or a naturally mutated protein. As used herein, the amino acid sequence information before and after the sample protein mutation refers to the relevant information of the amino acid sequence of the sample protein before and after the mutation. Among them, the amino acid sequence information after the mutation can be determined by the amino acid sequence information before the mutation and the mutation site information. In some embodiments, the amino acid sequence information before and after the sample protein mutation may include the amino acid sequence information before the mutation and the amino acid sequence information after the mutation of the sample protein.
[0048] Mutation stability labels can be used to reflect the trend of the influence of amino acid mutations on the thermal stability of proteins. In some embodiments, mutation stability labels can be determined based on the folding free energy change value (ΔΔG) of the mutant of the protein. The folding free energy change value (ΔΔG) caused by the mutation is closely related to the thermal stability of the mutant, wherein ΔΔG = folding free energy of the protein before mutation - folding free energy of the protein after mutation; ΔΔG less than 0 indicates that the stability of the protein after mutation is improved, and vice versa, it indicates that the mutation will cause a decrease in protein stability. Exemplarily, mutation stability labels can be unstable mutations and stable mutations, wherein unstable mutations include decreased stability or loss of stability, and stable mutations include improved stability or flat stability. It can be understood that using mutation stability information as a label for training samples allows the prediction model obtained through training to output prediction results related to the effect of amino acid mutations on protein thermal stability.
[0049] In some embodiments, the information of the training sample can be obtained from one or more components of the scene 100 (e.g., the storage device 120, the network 140, the data source 150, and the user terminal 130). For example, the acquisition module 710 can obtain experimental data of one or more existing protein databases, such as the UniProt database and the ProTherm database, from the data source 150, and extract the amino acid sequence information before the mutation, the amino acid sequence information after the mutation, and the folding free energy change value (ΔΔG) of one or more sample proteins as training sample information. The protein databases that can be connected and used include, but are not limited to, UniProt, ProThermDB, MPTherm, ThermoMutDB, FireProtDB, etc. For another example, the acquisition module 710 can obtain the experimental data before and after the protein mutation input by the user from the user terminal 130, and extract the amino acid sequence information before the mutation, the amino acid sequence information after the mutation, and the folding free energy change value (ΔΔG) of one or more sample proteins as training sample information.
[0050] The training set is the basis for training the prediction model, and the quality of the training set affects the performance of the prediction model. In some embodiments, the training set includes an initial training set and a reversed training set. The reversed training set can be obtained by reversing the initial training set to expand the number of training samples in the training set. In some embodiments, the initial training set can be reversed based on the thermodynamic characteristics of the protein to obtain a reversed training set. Specifically, according to the thermodynamic theory of proteins, under the same environmental conditions, the corresponding folding free energy change values (ΔΔG) of protein mutation processes in opposite directions are opposite values. For example, for the process of mutation from protein A to protein B and the process of mutation from protein B to protein A, ΔΔG(AB)=-ΔΔG(BA). In some embodiments, the training set includes an initial training set and a reversed training set obtained by reversing the initial training set, and the number of training samples carrying different mutation stability labels in the training set can be equalized by reversal. Specifically, in the above-mentioned training set inversion process, the initial training samples of unstable mutations in the initial training set can be inverted into inverted training samples of corresponding stable mutations, and the initial training samples of stable mutations in the initial training set can be inverted into inverted training samples of corresponding unstable mutations, thereby balancing and expanding the training set, making the prediction model obtained based on the training set more robust.
[0051] In step 230, a pre-trained model may be obtained, and based on the information of the plurality of training samples, feature extraction may be performed through the pre-trained model to obtain mutation features of the plurality of training samples. In some embodiments, step 230 may be performed by a computing device (such as processing device 110, feature extraction module 720).
[0052] The pre-trained model can be a pre-trained language model trained using unlabeled data in an existing protein database. In some embodiments, the pre-trained model can extract features related to mutations from multiple training samples. In some embodiments, the pre-trained model can be a pre-trained language model based on an attention mechanism, for example, a Transformer model, a GPT (Generative Pre-Training) model, a BERT (Bidirectional Encoder Representation from Transformers) model, a T5 (Transfer Text-to-Text Transformer) model, or the like, or a combination thereof. In some embodiments, the pre-trained model can learn from unlabeled data in an existing protein database, and then use a small amount of labeled data to optimize the parameters of the model, thereby obtaining a pre-trained model that can be used in the biological field. In some embodiments, the pre-trained model can be a pre-trained model based on a natural language processing method, including ESM-1b, ProtTXL, ProtBert, ProtXLNet, ProtAlbert, ProtElectra, T5, UniRep, TAPE, ProteinBERT, ProtGPT2, or the like, or a combination thereof. In some preferred embodiments, the pre-trained model may be a T5 model, such as ProtT5-XL or ProtT5-XXL. In some more preferred embodiments, the pre-trained model is ProtT5-XL.
[0053] In some embodiments, the mutation signature is used to indicate the change in the mutation region of the amino acid sequence of the sample protein before and after the mutation. In some embodiments, the mutation region at least includes the amino acid corresponding to the mutation site, for example, it may only include the amino acid corresponding to the mutation site. In other embodiments, the mutation region also includes the adjacent region of the amino acid mutation position, for example, it may include the amino acid corresponding to the mutation site and multiple amino acids (such as 1, 2, 3, 4, 5, 6, 7, 8, 9 or 10) adjacent to the amino acid at the mutation site. In other embodiments, the mutation region may include the amino acid corresponding to the mutation site and the first 3 amino acids and the last 3 amino acids adjacent to the amino acid at the mutation site.
[0054] In some embodiments, after the amino acid sequence information before and after the mutation of the training sample is input into the pre-training model, the pre-training model performs vector conversion on the amino acid sequence information before and after the mutation to obtain the mutation feature corresponding to the training sample. The mutation feature can be used to represent one or more information contained in the training sample, for example, the amino acid feature of the mutation site itself, the amino acid position feature of the mutation site, the amino acid feature before and after the mutation site, etc. It can be understood that the feature extraction makes the extracted mutation feature of the training sample integrate the features of each amino acid itself, the features of each amino acid position, and the amino acid features before and after each amino acid, and the feature information is richer and deeper, which helps to obtain more accurate stability prediction results through the classifier based on the mutation feature of the training sample.
[0055] For more information on feature extraction, see Figure 3 and Figure 4 The relevant description will not be repeated here.
[0056] In some embodiments, the mutation features of the multiple training samples obtained in step 230 can be directly used to perform classifier training in step 250. However, in model application, the more features, the better. The more features, the more additional information. This information will increase the complexity of the model, which greatly increases the calculation time. In addition, the model may become overfitted due to the increased complexity of the model. In some embodiments, before performing the classifier training in step 250, the mutation features of the multiple training samples can also be subjected to dimensionality reduction processing. Dimensionality reduction processing can further perform feature extraction on the mutation features of the training samples, extract features with relevance and effectiveness as much as possible, ensure the effectiveness of the prediction model obtained based on the mutation feature training after dimensionality reduction, avoid the problem of overfitting in the prediction model, and simplify the prediction model, reduce the complexity of the calculation and shorten the calculation time, so that the subsequent steps are more efficient.
[0057] In some embodiments, the above-mentioned dimensionality reduction processing method may include but is not limited to: principal component analysis (PCA), discriminant analysis (LDA), multidimensional scaling (MDS), isometric mapping (ISOMAP), local linear embedding (LLE), stochastic neighbor embedding (SNE), kernel principle component analysis (KPCA), etc., or a combination thereof. In some embodiments, the above-mentioned dimensionality reduction processing is implemented by principal component analysis (PCA).
[0058] In step 250, a classifier may be obtained, and the classifier may be trained based on the mutation signature and the mutation stability label to obtain a prediction model, wherein the prediction model includes the pre-trained model and the trained classifier. In some embodiments, step 250 may be performed by a computing device (e.g., processing device 110, training module 730).
[0059] In some embodiments, the classifier can be a machine learning model and / or a deep learning model. In some embodiments, the machine learning model may include a linear regression algorithm (Linear Regression) model, a support vector machine algorithm (Support Vector Machine, SVM) model, a k-nearest neighbor algorithm (KNN) model, a logistic regression algorithm (Logistic Regression, LR) model, a decision tree algorithm (Decision Tree) model, a k-means algorithm (K-Means) model, a random forest algorithm (Random Forest) model, a naive Bayes algorithm (Naive Bayes) model, a gradient boosting algorithm (Gradient Boosting) model, an ensemble learning algorithm (EnsembleLearning) model, etc., or a combination thereof. The deep learning model may include a convolutional neural network (Convolutional Neural Network, CN N) model, a recurrent neural network (Recurrent Neural Network) model, a recursive neural network (Recursive Neural Network) model, a long short-term memory network (long-short term memory, LSTM) model, a bidirectional long short-term memory network (BiLSTM) model, etc., or a combination thereof. In some preferred embodiments, the classifier may include one or more of a logistic regression algorithm model, a random forest algorithm model, an ensemble learning algorithm model such as an AdaBoost algorithm model, a support vector machine algorithm model, a decision tree algorithm model such as a gradient boosting decision tree algorithm (Gradient Boosting Decision Tree, GBDT) model, a convolutional neural network model, and a bidirectional long short-term memory network model. In some more preferred embodiments, the classifier may be one of a random forest algorithm model, a convolutional neural network model, and a bidirectional long short-term memory network model.
[0060] It is understandable that in other embodiments, the classifier may also be other types of machine learning models and / or deep learning models that can implement the prediction model function.
[0061] In some embodiments, the classifier can be supervised trained based on the mutation features and mutation stability labels of multiple training samples to obtain a trained classifier. Supervised training is the process of adjusting the parameters of the classifier using a set of samples of known categories to achieve the required performance. In some embodiments, a loss function, such as a cross entropy loss function, can be constructed to reflect the difference between the post-mutation stability prediction result predicted by the classifier and the actual mutation stability label of the training sample, and then the parameters of the classifier are adjusted according to the difference to obtain the model optimization parameters of the classifier. When the loss function meets the predetermined conditions, such as the loss function converges, or the loss function value is less than the preset value, the classifier training is completed. In some embodiments, the classifier parameters can be adjusted using a parameter search algorithm, for example, a search algorithm such as a grid search, a Bayesian search, etc. In some embodiments, supervised training is the process of training a prediction model based on a stability label.
[0062] In some embodiments, the input of the trained classifier is the mutation feature vector of the sample to be predicted, and the output is the post-mutation stability prediction result of the sample to be predicted (such as mutation stability label, stability mutation probability, etc.).
[0063] The method for constructing a protein stability prediction model shown in some embodiments of this specification can use a pre-training model based on a natural language processing method to extract mutation features from the amino acid sequence information of the training sample protein before and after mutation, and then use the mutation features and mutation stability labels to supervise the classifier to obtain a prediction model including the pre-training model and the trained classifier. This method applies the natural language processing model to construct a prediction model for the effect of amino acid mutation on protein stability, and the performance of the prediction model finally obtained is better than the model constructed by the structure and sequence-based method in the prior art. On the one hand, the above method obtains a training set with amplified sample quantity and balanced sample types by reversing the initial training set, which can solve the problems of small number and low quality of training sets, avoid overfitting or underfitting problems in the prediction model, improve the performance of the prediction model after training, and make the prediction model obtained based on the training set more robust. On the other hand, the prediction model obtained by the above training method can predict stability after mutation based on the protein sequence, does not rely on structural information, has a wide range of applications, and has good model performance.
[0064] Figure 3 300 is a flowchart of extracting mutation features of training samples according to some embodiments of the present specification. In some embodiments, process 300 may include step 310 and step 330, and process 300 may be performed for each training sample in a plurality of training samples. At least a portion of the steps in process 300 (e.g., steps 310 to 330) may be performed by a computing device (e.g., Figure 1For example, at least a portion of the steps in process 300 may be implemented as an instruction (eg, an application) stored in storage device 120. Figure 1 The processing device 110 in the embodiment may execute the instruction, and when executing the instruction, the processing device 110 may be configured to perform the process 300. The operations of the process shown below are for illustration purposes only. In some embodiments, the process 300 may be completed using one or more additional operations not described and / or one or more operations not described. In addition, Figure 3 The order in which the operations of the processes are illustrated and described below is not intended to be limiting.
[0065] In step 310, based on the amino acid sequence information of the sample protein before and after mutation of each training sample, feature extraction can be performed through the pre-training model to obtain the pre-mutation feature and post-mutation feature of the training sample. In some embodiments, step 310 can be performed by a computing device (such as processing device 110, feature extraction module 720).
[0066] In some embodiments, the sample protein sequence information before and after the mutation of the training sample includes the amino acid sequence information before the mutation, which can be used to reflect the amino acid arrangement order, the position characteristics of each amino acid in the amino acid sequence before the mutation, etc. In some embodiments, the sample protein sequence information after the mutation of the training sample includes the amino acid sequence information after the mutation, which can be used to reflect the amino acid arrangement order, the position characteristics of each amino acid in the amino acid sequence after the mutation, etc.
[0067] In some embodiments, the pre-mutation feature can be used to represent the feature of the mutation region of the amino acid sequence before mutation. The pre-mutation feature can be obtained by extracting the feature based on the amino acid sequence information before mutation of the training sample. In some embodiments, the post-mutation feature can be used to represent the feature of the mutation region of the amino acid sequence after mutation. The post-mutation feature can be obtained by extracting the feature based on the amino acid sequence information after mutation of the training sample.
[0068] In some embodiments, the pre-trained model includes at least one encoding network. The encoding network can be used to encode the amino acid sequence to generate a corresponding encoded feature vector. In some embodiments, the encoding network includes an embedding layer and an encoding layer. The embedding layer can be used to perform word embedding processing on each amino acid in the amino acid sequence to obtain a representation vector of each amino acid, and to obtain the position of each amino acid in the amino acid sequence to generate a position code of each amino acid. Among them, the representation vector of the amino acid can be used to characterize the characteristics of the amino acid itself, and the position code of the amino acid can be used to characterize the position of the amino acid in the amino acid sequence. The encoding layer can be used to combine the representation vector and position code of each amino acid in the obtained amino acid sequence to obtain a feature vector of each amino acid in the amino acid sequence.
[0069] In some embodiments, the pre-mutation features of the amino acid sequence information before mutation are extracted by a pre-training model, and the feature vectors of each amino acid in the amino acid sequence before mutation can be obtained, and the feature vector matrix generated based on the feature vectors of at least part of the amino acids in the amino acid sequence before mutation (for example, the amino acids corresponding to the mutation region) can be used to represent the pre-mutation features. In some embodiments, the post-mutation features of the amino acid sequence information after mutation are extracted by a pre-training model, and the feature vectors of each amino acid in the amino acid sequence before mutation can be obtained, and the feature vector matrix generated based on the feature vectors of at least part of the amino acids in the amino acid sequence after mutation (for example, the amino acids corresponding to the mutation region) can be used to represent the post-mutation features.
[0070] For more information on extracting pre- and post-mutation features, see Figure 4 The relevant description will not be repeated here.
[0071] In step 330, the pre-mutation feature and the post-mutation feature may be concatenated to obtain the mutation feature of the training sample. In some embodiments, step 330 may be performed by a computing device (eg, processing device 110, feature extraction module 720).
[0072] In some embodiments, the mutation features of the training sample can be obtained by splicing the pre-mutation features and the post-mutation features, so that the mutation features simultaneously reflect the changes in the mutation regions of the amino acid sequences before and after the mutation. For example, the pre-mutation features extracted by the pre-mutation model based on the amino acid sequence information of the training sample are represented by a first eigenvector matrix of A×B dimensions, which is generated by the first eigenvectors of A amino acids in the mutation region on the amino acid sequence before the mutation, and each first eigenvector has a length of 1×B; the post-mutation features extracted by the pre-mutation model based on the amino acid sequence of the training sample are represented by a second eigenvector matrix of A×B dimensions, which is generated by the second eigenvectors of A amino acids in the mutation region on the amino acid sequence after the mutation, and each second eigenvector has a length of 1×B; the first eigenvector matrix and the second eigenvector matrix can be spliced to form a third eigenvector matrix of 2A×B dimensions, and the third eigenvector matrix can be used to represent the mutation features of the training sample.
[0073] It is understandable that in other embodiments, the mutation feature of the training sample may be obtained by summing, subtracting or averaging the pre-mutation feature and the post-mutation feature, and this embodiment does not limit this.
[0074] Figure 4 4 is a flow chart of extracting pre-mutation features and post-mutation features according to some embodiments of the present specification. In some embodiments, process 400 may include steps 410 to 450, and process 400 may be performed for each training sample in a plurality of training samples. At least a portion of the steps in process 400 (e.g., steps 410 to 450) may be performed by a computing device (e.g., Figure 1 For example, at least a portion of the steps in process 400 may be implemented as an instruction (eg, an application) stored in storage device 120. Figure 1 The processing device 110 in the embodiment may execute the instruction, and when executing the instruction, the processing device 110 may be configured to perform the process 400. The operations of the process shown below are for illustration purposes only. In some embodiments, the process 400 may be completed using one or more additional operations not described and / or one or more operations not described. In addition, Figure 4 The order in which the operations of the processes are illustrated and described below is not intended to be limiting.
[0075] In step 410, N amino acids of the amino acid sequence of the training sample before mutation are obtained; and N amino acids of the amino acid sequence of the training sample after mutation are obtained.
[0076] In step 430, at least one layer of output vectors can be extracted through the pre-training model as the feature vector corresponding to the amino acid sequence before mutation, wherein the layer of output vectors includes N representation vectors and N position codes corresponding to the N amino acids in the amino acid sequence before mutation; at least one layer of output vectors can be extracted through the pre-training model as the feature vector corresponding to the amino acid sequence after mutation, wherein the layer of output vectors includes N representation vectors and N position codes corresponding to the N amino acids in the amino acid sequence after mutation.
[0077] In some embodiments, the representation vector and position code corresponding to each amino acid in the amino acid sequence can be obtained through the embedding layer of at least one encoding network of the pre-trained model. For example, for the amino acids AA1, AA2, AA3, ..., AA N , by specifying the embedding layer of the encoding network to generate the corresponding representation vectors E1, E2, E3, ..., E N , and generate corresponding position codes 1, 2, 3, ..., N by specifying the embedding layer of the encoding network. In some embodiments, the above functions of the embedding layer of the encoding network can be implemented by any type of embedding method.
[0078] In some embodiments, the feature vector corresponding to the amino acid sequence can be obtained through the encoding layer of at least one encoding network of the pre-trained model, that is, the output vector of the encoding layer of at least one encoding network is used as the feature vector corresponding to the amino acid sequence. For example, the characterization vectors E1, E2, E3, ..., E4, E5, E6, E7, E8, E9, E10, E111, E12, E13, ..., E14, E15, E16, E17, E18, E19, E20, E210, E30, ..., E40, E50, E60, E70, E80, E90, E10, E111, E12, E13, ..., E14, E15, E16, E17, E18, E19, E20, E21 N and position codes 1, 2, 3, ..., N, by specifying the coding layer combination (or coding) of the coding network to obtain the feature vector corresponding to the amino acid sequence before mutation. In a preferred embodiment, the output vector of a layer corresponding to the coding layer of the last coding network of the pre-trained model can be specified as the feature vectors F1, F2, F3, ..., F corresponding to each amino acid in the amino acid sequence. N For the extraction of the feature vector of the amino acid sequence after mutation, please refer to the above description of the extraction of the feature vector of the amino acid sequence before mutation, which will not be repeated here.
[0079] In step 450, the feature vector corresponding to the mutation region in the feature vector of the amino acid sequence before mutation can be extracted as the pre-mutation feature; the feature vector corresponding to the mutation region in the feature vector of the amino acid sequence after mutation can be extracted as the post-mutation feature.
[0080] In some embodiments, the range of the mutation region can be determined based on a pre-set setting. For example, the pre-set range of the mutation region can be a region within N (e.g., 0, 1, 2, 3, 4, 5, 12, 14, 16, 18, or 20) amino acids from the mutation site, and the mutation region may include 2N+1 amino acids including the amino acids corresponding to the mutation site. It is understood that in other embodiments, the range of the mutation region can be determined based on user input, and this embodiment is not limited. Exemplarily, the feature vector corresponding to the amino acid sequence before mutation includes F1, F2, F3, ..., F N , the characteristic vector corresponding to the amino acid at the mutation site is F M , extract the feature vector F corresponding to the mutation area M-5 To the eigenvector F M+5 And generate a feature vector matrix, and the mutated features of the training sample can be represented by the feature vector matrix.
[0081] Some embodiments of the present specification use at least one encoding network of a pre-trained model to extract pre-mutation features and mutation features of the pre-mutation amino acid sequence information and post-mutation amino acid sequence information of the training samples, respectively. The obtained feature vectors of the pre-mutation / post-mutation amino acid sequences reflect the characteristics of each amino acid itself, the position characteristics, and the characteristics of the preceding and following adjacent amino acids. The feature information is richer and deeper, which helps to obtain more accurate stability prediction results based on the feature vectors through a classifier.
[0082] Figure 5 5 is a schematic diagram of a process for predicting protein stability after mutation according to some embodiments of the present specification. In some embodiments, process 500 includes steps 510 to 550. At least a portion of the steps in process 500 (e.g., steps 510 to 550) can be performed by a computing device (e.g., Figure 1 For example, at least a portion of the steps in process 500 may be implemented as an instruction (eg, an application) stored in storage device 120. Figure 1 The processing device 110 in the embodiment may execute the instruction, and when executing the instruction, the processing device 110 may be configured to perform the process 500. The operations of the process shown below are for illustration purposes only. In some embodiments, the process 500 may be completed using one or more additional operations not described and / or one or more operations not described. In addition, Figure 5 The order in which the operations of the processes are illustrated and described below is not intended to be limiting.
[0083] In step 510, information of the sample to be predicted may be obtained, and the information of the sample to be predicted includes amino acid sequence information before and after the mutation of the protein to be tested. In some embodiments, step 510 may be performed by a computing device (eg, processing device 110, acquisition module 810).
[0084] In some embodiments, the amino acid sequence information of the protein to be tested before and after mutation includes the original amino acid sequence information and the target amino acid sequence information corresponding to the mutation. The original amino acid sequence information refers to the amino acid sequence information before the protein to be tested mutates, and the target amino acid sequence information refers to the amino acid sequence information after the protein to be tested mutates.
[0085] Steps 530 and 550 process the information of the sample to be predicted using the prediction model obtained by the method for constructing a protein stability prediction model shown in some embodiments of this specification, wherein the prediction model includes a pre-trained model and a trained classifier.
[0086] In step 530, the pre-trained model can be used to extract features based on the amino acid sequence information of the protein to be tested of the sample to be predicted before and after the mutation, to obtain the mutation feature of the sample to be predicted. In some embodiments, step 530 can be performed by a computing device (such as processing device 110, processing module 820).
[0087] For more information about extracting mutation features of samples to be predicted, please refer to Figure 2 , Figure 3 and Figure 4 The relevant description of the extraction of mutation features of training samples in is not repeated here.
[0088] In step 550, the trained classifier may be used to make a prediction based on the mutation signature to determine the post-mutation stability prediction result of the sample to be predicted. In some embodiments, step 550 may be performed by a computing device (such as processing device 110, processing module 820). In some embodiments, the input of the trained classifier is the mutation signature of the sample to be predicted, and the output of the trained classifier is the post-mutation stability prediction result of the sample to be predicted.
[0089] In some embodiments, before step 550, process 500 further includes performing dimensionality reduction processing on the mutation features of the sample to be predicted extracted in step 530. Step 550 can use the trained classifier to perform prediction based on the mutation features of the sample to be predicted after dimensionality reduction to determine the prediction result of the stability after mutation of the sample to be predicted.
[0090] For more information about the dimensionality reduction of the mutation features of the samples to be predicted, please refer to Figure 2 The description of the dimensionality reduction processing of the mutation features of the training samples in is not repeated here.
[0091] In some embodiments, the prediction of the stability after mutation of the sample to be predicted may include single-point mutation prediction and saturation mutation prediction. As used herein, a single-point mutation refers to the amino acid at a specified site in the original amino acid sequence mutating into the other 19 amino acids in sequence; a saturation mutation refers to the amino acid at each site in the original amino acid sequence mutating into the other 19 amino acids in sequence. In some embodiments, the processing device 110 can adjust the prediction result of the stability after mutation of the sample to be predicted according to the specific prediction instruction of the user. For example, based on the single-point mutation prediction instruction of the user, the processing device 110 uses the prediction model to perform corresponding processing on the sample to be predicted, and can output the corresponding mutation stability label of the specified mutation site of the sample to be predicted to mutate into the other 19 amino acids, or can output the total stability mutation probability of the specified mutation site of the sample to be predicted.
[0092] For more information about the application process of prediction models in different scenarios, please refer to Figure 6 Related description.
[0093] Figure 6 is an exemplary flow chart of the prediction model in different application scenarios according to some embodiments of the present specification. Process 600 includes steps 610 to 670. At least a portion of the steps in process 600 (e.g., steps 610 to 670) can be performed by a computing device (e.g., Figure 1 The processing device 110 in is completed.
[0094] In step 610 , the processing device 110 may obtain information of the sample to be predicted based on the prediction instruction of the user.
[0095] In the application scenario of single-point mutation stability prediction, the processing device 110 receives a single-point mutation prediction instruction from the user. Based on the original amino acid sequence information and mutation site information input by the user, the target amino acid sequence information is determined to obtain multiple single-point mutation samples to be predicted; wherein the target amino acid sequence includes multiple corresponding amino acid sequences obtained after the specified mutation site of the original amino acid sequence mutates to other 19 amino acids. The amino acid sequence information of the protein to be tested before and after the mutation of each single-point mutation sample to be predicted corresponds to a target amino acid sequence formed by the associated original amino acid sequence and the single-point mutation.
[0096] In the application scenario of saturation mutation stability prediction, the processing device 110 receives a saturation mutation prediction instruction from the user. Based on the original amino acid sequence information input by the user, the target amino acid sequence information is determined to obtain multiple samples to be predicted for saturation mutations; wherein the target amino acid sequence information includes multiple corresponding amino acid sequences obtained after each amino acid site of the original amino acid sequence is mutated to other 19 amino acids. The amino acid sequence information of the protein to be tested before and after the mutation of each sample to be predicted for saturation mutation corresponds to a target amino acid sequence formed by associating the original amino acid sequence and the saturation mutation.
[0097] In step 630, the processing device 110 may use the prediction model to extract features from the amino acid sequence information of the tested protein before and after the mutation of the multiple samples to be predicted, and obtain the mutation features of the multiple samples to be predicted. Specifically, based on the original amino acid sequence information and the target amino acid sequence information of each sample to be predicted, the pre-mutation features and the post-mutation features are extracted using the pre-trained model, and based on the pre-mutation features and the post-mutation features, the mutation features of the sample to be predicted are spliced to obtain.
[0098] In step 650 , the processing device 110 may perform dimensionality reduction processing on the mutation features of the plurality of samples to be predicted.
[0099] In step 670 , the processing device 110 may use the prediction model to perform prediction based on the mutation features of the multiple samples to be predicted after dimensionality reduction, and obtain a prediction result of post-mutation stability.
[0100] In the application scenario of single-point mutation stability prediction, the trained classifier predicts based on the mutation features of multiple single-point mutation samples after dimensionality reduction, and outputs the single-point mutation prediction results. The single-point mutation prediction results can include all mutation stability labels or all stability mutation probabilities of the original amino acid sequence's specified mutation site mutating to other 19 amino acids.
[0101] In the application scenario of saturation mutation stability prediction, the trained classifier predicts based on the mutation features of multiple saturation mutation samples after dimensionality reduction, and outputs the saturation mutation prediction results. The saturation mutation prediction results can include all mutation stability labels or all stability mutation probabilities of each amino acid site in the original amino acid sequence mutated to other 19 amino acids, and can also include the ranking results of all stability mutation probabilities.
[0102] The process 600 may also include a display step. The display step includes: the processing device 110 sends the post-mutation stability prediction result to the user terminal 130 based on the user's display instruction, and displays the result in images and / or texts through the user terminal 130; or the processing device 110 displays the result in images and / or texts through other display devices connected to it based on the user's display instruction.
[0103] Figure 7 is a module diagram of a system for constructing a protein stability prediction model according to some embodiments of this specification. Figure 7 As shown, the system 700 includes an acquisition module 710 , a feature extraction module 720 and a training module 730 .
[0104] The acquisition module 710 can be used to acquire a training set, which includes information of multiple training samples, and the information of each training sample includes amino acid sequence information before and after the sample protein mutation and a mutation stability label.
[0105] In some embodiments, the training set includes an initial training set and a reverse training set; the reverse training set is obtained by reversing the initial training set based on the thermodynamic characteristics of the protein, so that the number of training samples carrying different mutation stability labels in the training set is equal.
[0106] The feature extraction module 720 may be used to obtain a pre-training model, and based on the information of the plurality of training samples, perform feature extraction through the pre-training model to obtain mutation features of the plurality of training samples.
[0107] In some embodiments, the feature extraction module 720 can perform feature extraction through the pre-training model based on the amino acid sequence information of the sample protein before and after mutation of each of the training samples to obtain the pre-mutation features and post-mutation features of the training samples; and splice the pre-mutation features and the post-mutation features to obtain the mutation features of the training samples.
[0108] In some embodiments, the mutation signature is used to indicate changes in the mutation region of the amino acid sequence before and after the mutation; the mutation region at least includes the amino acid corresponding to the mutation site.
[0109] The training module 730 can be used to obtain a classifier, train the classifier based on the mutation feature and the mutation stability label, and obtain a prediction model, wherein the prediction model includes the pre-trained model and the trained classifier. In some embodiments, the training module 730 can adjust the model parameters of the classifier based on the difference between the prediction result output by the classifier and the mutation stability label, so as to minimize the difference between the prediction result and the mutation stability label.
[0110] In some embodiments, the classifier can be a machine learning model and / or a deep learning model. In some embodiments, the machine learning model may include a linear regression algorithm (Linear Regression) model, a support vector machine algorithm (SVM) model, a k-nearest neighbor algorithm (KNN) model, a logistic regression algorithm (Logistic Regression, LR) model, a decision tree algorithm (Decision Tree) model, a k-means algorithm (K-Means) model, a random forest algorithm (Random Forest) model, a naive Bayes algorithm (Naive Bayes) model, a gradient boosting algorithm (Gradient Boosting) model, an ensemble learning algorithm (Ensemble Learning) model, etc., or a combination thereof. The deep learning model may include a convolutional neural network (Convolutional Neural Network, CN N) model, a recurrent neural network (Recurrent Neural Network) model, a recursive neural network (Recursive Neural Network) model, a long short-term memory network (long-short term memory, LSTM) model, a bidirectional long short-term memory network (BiLSTM) model, etc., or a combination thereof. In some embodiments, the classifier is preferably one of a random forest algorithm model, a convolutional neural network model, and a bidirectional long short-term memory network model.
[0111] In some embodiments, the system 700 further includes a dimensionality reduction module 740. The dimensionality reduction module 740 can be used to perform dimensionality reduction processing on the mutation features of the training samples. In some embodiments, the dimensionality reduction module 740 can achieve feature dimensionality reduction by combining one or more algorithms of principal component analysis (PCA), discriminant analysis (LDA), multidimensional scaling (MDS), isometric mapping (ISOMAP), and stochastic neighbor embedding (SNE).
[0112] For more information about how each module of system 700 implements its functions, see Figures 2 to 4 The relevant description will not be repeated here.
[0113] Figure 8is a module diagram of a system for predicting protein stability after mutation according to some embodiments of this specification. Figure 8 As shown, the system 800 includes an acquisition module 810 and a processing module 820 .
[0114] The acquisition module 810 can be used to acquire information of the sample to be predicted, where the information of the sample to be predicted includes amino acid sequence information before and after the mutation of the protein to be tested.
[0115] The processing module 820 may be used to process information of the sample to be predicted using the prediction model to obtain a prediction result of the post-mutation stability of the sample to be predicted.
[0116] In some embodiments, the processing module 820 may include a feature extraction unit 821 and a prediction unit 822. The feature extraction unit 821 may be used to extract features using a pre-trained model based on the amino acid sequence information of the protein to be tested of the sample to be predicted before and after the mutation, and obtain the mutation features of the sample to be predicted. The prediction unit 822 may be used to predict based on the mutation features of the sample to be predicted, using a trained classifier to determine the post-mutation stability prediction result of the sample to be predicted.
[0117] In some embodiments, the processing module 820 may further include a dimensionality reduction unit 823 for performing dimensionality reduction processing on the mutation features of the sample to be predicted, so that the trained classifier can make predictions based on the mutation features of the sample to be predicted after dimensionality reduction.
[0118] For more information about how each module of System 800 implements its functions, see Figures 2 to 6 The relevant description will not be repeated here.
[0119] It should be understood that Figure 7 and Figure 8The system and its modules shown can be implemented in various ways. For example, in some embodiments, the system and its modules can be implemented by hardware, software or a combination of software and hardware. Among them, the hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or a dedicated design hardware. Those skilled in the art will understand that the above methods and systems can be implemented using computer executable instructions and / or included in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the system and its modules of this specification. Not only can the hardware circuits such as ultra-large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc. be implemented, it can also be implemented with software such as executed by various types of processors, and it can also be implemented by a combination of the above hardware circuits and software (for example, firmware).
[0120] It should be noted that the above description of the construction of a protein stability prediction model and the system and its modules for predicting protein stability after mutation is only for convenience of description and cannot limit this specification to the scope of the embodiments. It can be understood that for those skilled in the art, after understanding the principle of the system, it is possible to arbitrarily combine the various modules or form a subsystem to connect with other modules without deviating from this principle. For example, in some embodiments, the system 700 and the system 800 can be two independent systems or integrated into one system. For example, in some embodiments, the acquisition module 710, the feature extraction module 720 and the training module 730 can be different modules in one system, or one module can realize the functions of the above two or more modules. For example, the acquisition module 710 and the acquisition module 810 can be two modules, or one module can have the functions of acquiring training sets and samples to be predicted at the same time. For example, each module can share a storage module, or each module can have its own storage module. Such variations are all within the scope of protection of this specification.
[0121] Example
[0122] The methods are all implemented based on Python, where the pre-trained model is based on the Pytorch framework and the classifier is based on the scikit-learn framework.
[0123] 1. Build a dataset
[0124] In order to evaluate the method provided in this specification and verify the model performance of the prediction model provided in this specification, this example uses the following protein databases for testing: MPTherm (https: / / www.iitm.ac.in / bioinfo / mptherm / ); ProThermDB (https: / / web.iitm.ac.in / bioinfo2 / prothermdb / index.html); ThermoMutDB (http: / / biosig.unimelb.edu.au / thermomutdb / ); and FireProtDB (https: / / loschmidt.chemi.muni.cz / fireprotdb / ).
[0125] Based on the experimental data of the above protein database, a total of 5917 training set experimental data that meet the requirements were screened, including 1451 stable mutation samples and 4466 unstable mutation samples; then based on the method of reversing the initial training set shown in some embodiments of this specification, a reversal training set was constructed, and the training data set finally obtained included 11834 single-point mutation training samples of 328 different proteins, among which the number of unstable mutation training samples and stable mutation training samples was the same; the test data set that met the requirements was screened to include 298 test samples, among which the number of unstable mutation test samples and stable mutation test samples was the same. There are no repeated samples between the training data set and the test data set. The experimental data screening criteria for constructing the training data set and the test data set are shown in Table 1.
[0126] Table 1 Experimental data screening criteria
[0127]
[0128] Note: Tm: refers to the midpoint temperature of thermal denaturation of protein, which indicates the temperature corresponding to 50% unfolding of protein.
[0129] 2. Build a prediction model based on deep learning algorithms
[0130] 2.1. Constructing the prediction model CNN+
[0131] 2.1.1 Model Structure
[0132] The prediction model CNN+ uses the pre-trained natural language model ProtT5-XL (Elnaggar, Ahmed et al. 2021. "ProtTrans: Towards Cracking the Language of Lifes Code Through Self-Supervised Deep Learning and High Performance Computing". IEEE Transactions on Pattern Analysis and Machine Intelligence: 1–1.) as a feature extractor and a convolutional neural network (CNN) model as a classifier.
[0133] 2.1.2 Model Training
[0134] ProtT5-XL is used to extract mutation features from the training data set. ProtT5-XL contains 24 layers of attention structure, each layer of attention structure consists of 32 self-attention heads and 1024 neurons, and the last layer of attention structure is selected as the output layer. Specifically, the encoder of ProtT5-XL encodes the N amino acids of the amino acid sequence before / after mutation of the training sample, and outputs N feature vectors of length 1024 of the amino acid sequence before / after mutation of the training sample. The mutation site and the range of 3 amino acids before and after the mutation site (7 amino acids in total) are selected as the mutation region, and 7 feature vectors corresponding to the mutation region are extracted from the N feature vectors of the amino acid sequence before / after mutation of the training sample. The 7 feature vectors corresponding to the mutation region of the amino acid sequence before mutation and the 7 feature vectors corresponding to the mutation region of the amino acid sequence after mutation are spliced to obtain a 14×1024-dimensional feature matrix, which is used as the mutation feature of the training sample.
[0135] The initial convolutional neural network is supervisedly trained using the mutation features of the training samples to obtain the trained convolutional neural network. The trained convolutional neural network is combined with ProtT5-XL to obtain the prediction model CNN+. Among them, the convolutional neural network body used in this embodiment includes two one-dimensional convolutional layers, two pooling layers and two fully connected layers. The specific parameters are set as follows: the number of convolution kernels of the first convolutional layer is 32, the convolution kernel size is 3, the convolution step is 1, the padding is the same, and ReLU is used as the activation function; the first pooling layer is connected, the pooling method is maximum pooling, the pooling kernel size is 3, and the step is 2; the number of convolution kernels of the second convolutional layer is 16, and the remaining parameters are the same as the first convolutional layer; the second pooling layer is connected, and the parameters are the same as the first pooling layer; the number of neurons in the two fully connected layers is 32 and 2 respectively, ReLU and softmax are used as activation functions, and the Dropout function is used, and the discard probability is 50%. The loss function is used for the initial convolutional neural network training. The loss function is set to cross entropy loss, and Adam is used as the optimizer.
[0136] 2.2. Building the prediction model BiLSTM+
[0137] 2.2.1 Model Structure
[0138] The prediction model BiLSTM+ uses the pre-trained natural language model ProtT5-XL as a feature extractor and the bidirectional long short-term memory network (BiLSTM) model as a classifier.
[0139] 2.2.2 Model Training
[0140] Refer to the method in 2.1.2 to extract mutation features from the training samples in the training dataset.
[0141] The initial bidirectional long short-term memory network is supervisedly trained using the mutation features of the training samples to obtain a trained bidirectional long short-term memory network. The trained bidirectional long short-term memory network is combined with ProtT5-XL to obtain the prediction model BiLSTM+. The main body of the bidirectional long short-term memory network used in this embodiment includes two BiLSTM layers and one fully connected layer. The specific parameters are set as follows: the number of neurons in the two layers of BiLSTM is 16 and 4 respectively, the Boolean value (return_sequences) is set to True, ReLU is used as the activation function, and the Dropout function is used with a drop probability of 50%; the number of neurons in the fully connected layer is 2, and softmax is used as the activation function. The loss function is used to train the initial bidirectional long short-term memory network, the loss function is set to cross entropy loss, and Adam is used as the optimizer.
[0142] 3. Build a prediction model based on machine learning algorithms
[0143] 3.1. Constructing the prediction model RF+
[0144] 3.1.1 Model structure
[0145] The prediction model RF+ uses the pre-trained natural language model ProtT5-XL as a feature extractor and the Random Forest (RF) model as a classifier.
[0146] 3.1.2 Model Training
[0147] ProtT5-XL is used to extract mutation features from the training data set. ProtT5-XL contains 24 layers of attention structure, each layer of attention structure consists of 32 self-attention heads and 1024 neurons, and the last layer of attention structure is selected as the output layer. Specifically, the encoder of ProtT5-XL encodes the N amino acids of the amino acid sequence before / after mutation of the training sample, and outputs N feature vectors of length 1024 of the amino acid sequence before / after mutation of the training sample. The amino acid at the mutation site is selected as the mutation region, and one feature vector corresponding to the mutation region is extracted from the N feature vectors of the amino acid sequence before / after mutation of the training sample. The one feature vector corresponding to the mutation region of the amino acid sequence before mutation and the one feature vector corresponding to the mutation region of the amino acid sequence after mutation are concatenated to obtain a 1×2048-dimensional feature matrix, which is used as the mutation feature of the training sample.
[0148] The principal component analysis (Principal Component Analysis) algorithm is used to reduce the dimension of the mutation features of the training samples, and the extracted 1×2048-dimensional mutation features are reduced to 1×10-dimensional mutation features.
[0149] The initial random forest model is supervisedly trained using the mutation features of the training sample to obtain the trained random forest model. The trained random forest model is combined with ProtT5-XL to obtain the prediction model RF+. During the training process, a network search algorithm is used to determine the optimal parameter combination of the random forest model.
[0150] 3.2. Constructing the prediction model Adaboost+
[0151] 3.2.1 Model Structure
[0152] The prediction model Adaboost+ uses the pre-trained natural language model ProtT5-XL as a feature extractor and the Adaboost model as a classifier.
[0153] 3.2.2 Model Training
[0154] Refer to the method in 3.1.2 to extract mutation features and reduce feature dimension of the training samples in the training data set.
[0155] The initial Adaboost model is supervisedly trained using the mutation features of the training samples to obtain the trained Adaboost model. The trained Adaboost model is combined with ProtT5-XL to obtain the prediction model Adaboost+. During the training process, a network search algorithm is used to determine the optimal parameter combination of the Adaboost model.
[0156] 3.3. Constructing the prediction model SVM+
[0157] 3.3.1 Model structure
[0158] The prediction model SVM+ uses the pre-trained language model ProtT5-XL as a feature extractor and the Support Vector Machine (SVM) model as a classifier.
[0159] 3.3.2 Model Training
[0160] Refer to the method in 3.1.2 to extract mutation features and reduce feature dimension of the training samples in the training data set.
[0161] The initial support vector machine model is supervisedly trained using the mutation features of the reduced dimension of the training samples to obtain the trained support vector machine model. The trained support vector machine model is combined with ProtT5-XL to obtain the prediction model SVM+. During the training process, a network search algorithm is used to determine the optimal parameter combination of the SVM model.
[0162] 3.4. Constructing the prediction model LR+
[0163] 3.4.1 Model Structure
[0164] The prediction model LR+ uses the pre-trained language model ProtT5-XL as the feature extractor and the logistic regression (LR) model as the classifier.
[0165] 3.4.2 Model Training
[0166] Refer to the method in 3.1.2 to extract mutation features and reduce feature dimension of the training samples in the training data set.
[0167] The initial logistic regression model is supervisedly trained using the reduced-dimensional mutation features of the training samples to obtain a trained logistic regression model. The trained logistic regression model is combined with ProtT5-XL to obtain the prediction model LR+. During the training process, a network search algorithm is used to determine the optimal parameter combination of the logistic regression model.
[0168] 3.5. Building the prediction model GBDT+
[0169] 3.5.1 Model Structure
[0170] The prediction model GBDT+ uses the pre-trained language model ProtT5-XL as the feature extractor and the Gradient Boosting Decision Tree (GBDT) model as the classifier.
[0171] 3.5.2 Model Training
[0172] Refer to the method in 3.1.2 to extract mutation features and reduce feature dimension of the training samples in the training data set.
[0173] The initial gradient boosting decision tree model is supervisedly trained using the mutation features of the training sample's dimensionality reduction to obtain a trained gradient boosting decision tree model. The trained gradient boosting decision tree model is combined with ProtT5-XL to obtain the prediction model GBDT+. During the training process, a network search algorithm is used to determine the optimal parameter combination of the gradient boosting decision tree model.
[0174] 4. Model performance verification
[0175] In order to measure the performance of the prediction model trained by the method provided in this specification, this embodiment introduces the following performance indicators for model performance evaluation:
[0176] The area under the curve (AUC) refers to the area under the receiver operating characteristic curve (ROC);
[0177] 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 judged as the positive class, and FP (False Positive) is the negative sample judged as the positive class.
[0178] Recall refers to the ratio of the number of correctly predicted positive samples to the total number of positive samples. The calculation formula for recall is: recall = TP / (TP+FN), where FN (False Negative) is the positive sample that is judged as negative.
[0179] The F1 score is defined as the harmonic mean of precision and recall. The calculation formula of the F1 score is: F1 = 2×(precision×rec all) / (precision+recall).
[0180] 4.1. Five-fold cross-validation
[0181] Based on the training data set constructed in Section 1, the prediction model CNN+ constructed in 2.1, the prediction model BiLSTM+ constructed in 2.2, the prediction model RF+ constructed in 3.1, the prediction model Adaboost+ constructed in 3.2, the prediction model SVM+ constructed in 3.3, the prediction model LR+ constructed in 3.4, and the prediction model GBDT+ constructed in 3.5 are trained. In the five-fold cross-validation, the performance indicators of each prediction model are shown in Table 2.
[0182] Table 2 Five-fold cross-validation performance indicators
[0183]
[0184] As shown in Table 2, the performance of the prediction model CNN+, prediction model BiLSTM+, prediction model RF+, prediction model Adaboost+, prediction model SVM+, prediction model LR+, and prediction model GBDT+ on the training data set is relatively good. Among them, the performance indicators of the prediction model CNN+, prediction model BiLSTM+, and prediction model RF+ are better than the other four prediction models.
[0185] 4.2 Comparative evaluation of model performance
[0186] In order to verify the performance difference between the prediction method provided in this manual and the existing prediction methods, 7 commonly used protein stability prediction methods were selected for comparative study, including 2 sequence-based prediction methods (I-Mutant-seq and DDGun-seq) and 5 structure-based methods (ProtsPOM, Thermnet, DDGun-3D, FoldX and Rosetta). The specific information is as follows:
[0187] FoldX(Guerois, Raphael, Jens Erik Nielsen and Luis Serrano. 2002. "Predicting Changes in the Stability of Proteins and Protein Complexes: A Study of More than 1000 Mutations". Journal of Molecular Biology 320(2): 369–87.);
[0188] Rosetta (Leman, Julia Koehler et al. 2020. "Macromolecular Modeling and Design in Rosetta: Recent Methods and Frameworks". Nature Methods 17(7): 665–80.);
[0189] ProtsPOM (Banerjee, Anupam, and Pralay Mitra. 2020. "Estimating the Effect of Single-Point Mutations on Protein Thermodynamic Stability and Analyzing the Mutation Landscape of the P53 Protein". Journal of Chemical Information and Modeling 60(6): 3315–23.);
[0190] Thermnet (Ofoegbu, Tochukwu C. et al. 2019. "PhyreRisk: A Dynamic Web Application to Bridge Genomics, Proteomics and 3D Structural Data to Guide Interpretation of Human Genetic Variants". Journal of Molecular Biology 431(13): 2460–66.);
[0191] DDGun-3D(Montanucci, Ludovica et al. 2022. "DDGun: An Untrained Predictor ofProtein Stability Changes upon Amino Acid Variants". Nucleic Acids Research 50(W1):W222–27.);
[0192] I-Mutant-seq (Capriotti, E., P. Fariselli and R. Casadio. 2005. "I-Mutant2.0: Predicting Stability Chang es upon Mutation from the Protein Sequence or Structure". Nucleic Acids Research 33 (Web Server): W306–10.);
[0193] DDGun-seq (Montanucci, Ludovica et al. 2022. "DDGun: An Untrained Predictor ofProtein Stability Changes upon Amino Acid Variants". Nucleic Acids Research 50(W1):W222–27.).
[0194] The test data set constructed in Section 1 was used to compare and evaluate the prediction model CNN+, the prediction model BiLSTM+, the prediction model RF+ and the common protein stability prediction methods in the prior art. The performance indicators of the comparative evaluation are shown in Table 3. The above test data set does not appear in the training set of the prediction method provided in this specification and the common protein stability prediction method in the prior art.
[0195] Table 3. Performance indicators for comparative evaluation
[0196]
[0197] As shown in Table 3, the performance indicators of the prediction model CNN+, the prediction model BiLSTM+ and the prediction model RF+ are significantly better than the common protein stability prediction methods in the prior art.
[0198] 4.3 Comparative evaluation of model prediction speed
[0199] The prediction model CNN+, the prediction model BiLSTM+, the prediction model RF+ and the common protein stability prediction methods in the prior art were used to predict the stability of the test sample after the protein single point mutation, and the running time required for the prediction was statistically calculated. The test sample was a protein with a length of 100 amino acids.
[0200] Table 4. Running time statistics of single point mutation prediction
[0201]
[0202] Note: The hardware platform used for the test is: CPU Intel(R)Xeon(R)Bronze 3206R CPU@1.90GHz, memory 256G, GPU: A100*2.
[0203] The prediction speed test results are shown in Table 4. From the test results, it can be seen that the prediction speeds of the prediction model CNN+, the prediction model BiLSTM+ and the prediction model RF+ are significantly better than the common protein stability prediction methods in the prior art.
[0204] 4.4 Application effect verification
[0205] In order to verify the practical application effect of the prediction method provided in this specification, this example selects a protein sample with a single point mutation that has a thermal stability test result in the literature as a test sample. The test sample does not appear in the training data set and test data set mentioned above in this specification. The prediction model CNN+, the prediction model BiLSTM+ and the prediction model RF+ are used to perform a protein stability prediction test after a single point mutation on the test sample. The test results are shown in Table 5. The prediction results are consistent with the test results. The prediction results of the prediction method provided in this specification can provide an effective reference for actual production and research work.
[0206] Table 5 Test results
[0207]
[0208] The basic concepts have been described above. Obviously, for those skilled in the art, the above detailed disclosure is only for example and does not constitute a limitation of this specification. Although not explicitly stated here, those skilled in the art may make various modifications, improvements and corrections to this specification. Such modifications, improvements and corrections are suggested in this specification, so such modifications, improvements and corrections still belong to the spirit and scope of the exemplary embodiments of this specification.
[0209] At the same time, this specification uses specific words to describe the embodiments of this specification. For example, "one embodiment", "an embodiment", and / or "some embodiments" refer to a certain feature, structure or characteristic related to at least one embodiment of this specification. Therefore, it should be emphasized and noted that "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or more in different places in this specification does not necessarily refer to the same embodiment. In addition, certain features, structures or characteristics in one or more embodiments of this specification can be appropriately combined.
[0210] In some embodiments, numbers describing the number of components and attributes are used. It should be understood that such numbers used in the description of the embodiments are modified by the modifiers "about", "approximately" or "substantially" in some examples. Unless otherwise specified, "about", "approximately" or "substantially" indicate that the numbers are allowed to vary by ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which can be changed according to the required features of individual embodiments. In some embodiments, the numerical parameters should take into account the specified significant digits and adopt the general method of retaining the digits. Although the numerical domains and parameters used to confirm the breadth of the range in some embodiments of this specification are approximate values, in specific embodiments, the setting of such numerical values is as accurate as possible within the feasible range.
[0211] Each patent, patent application, patent application publication, and other materials, such as articles, books, specifications, publications, documents, etc., cited in this specification are hereby incorporated by reference in their entirety. Except for application history documents that are inconsistent with or conflicting with the contents of this specification, documents that limit the broadest scope of the claims of this specification (currently or later attached to this specification) are also excluded. It should be noted that if the descriptions, definitions, and / or use of terms in the materials attached to this specification are inconsistent or conflicting with the contents described in this specification, the descriptions, definitions, and / or use of terms in this specification shall prevail.
[0212] Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, as an example and not a limitation, alternative configurations of the embodiments of this specification may be considered consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly introduced and described in this specification.
Claims
1. A method for constructing a protein stability prediction model, 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 amino acid sequence information of the sample protein before and after mutation and a mutation stability label; Obtaining a pre-training model, and extracting features based on the amino acid sequence information of each training sample before and after the protein mutation by using the pre-training model to obtain the pre-mutation features and post-mutation features of the training sample, wherein the amino acid sequence information includes the pre-mutation amino acid sequence and the post-mutation amino acid sequence; The method for extracting the pre-mutation feature comprises: Obtaining N amino acids of the amino acid sequence before mutation; Extracting at least one layer of output vectors through the pre-training model as a feature vector corresponding to the amino acid sequence before mutation, wherein the layer of output vectors includes N representation vectors and N position codes corresponding to N amino acids of the amino acid sequence before mutation; Extracting a feature vector corresponding to the mutation region as the pre-mutation feature; The method for extracting the post-mutation feature comprises: Obtaining N amino acids of the mutated amino acid sequence; Extracting at least one layer of output vectors through the pre-trained model as a feature vector corresponding to the mutated amino acid sequence, wherein the layer of output vectors includes N representation vectors and N position codes corresponding to N amino acids of the mutated amino acid sequence respectively; and Extracting a feature vector corresponding to the mutation region as the post-mutation feature; wherein The mutation region at least includes the amino acid corresponding to the mutation site; splicing the pre-mutation feature and the post-mutation feature to obtain mutation features of the multiple training samples; Performing dimensionality reduction processing on the mutation feature to obtain the mutation feature after dimensionality reduction; and Obtaining a classifier, and training the classifier based on the mutation features after dimensionality reduction and the mutation stability labels to obtain a prediction model; The prediction model includes the pre-trained model and the trained classifier, the pre-trained model is ProtT5-XL, and the classifier is implemented using one of the following algorithms: random forest algorithm, convolutional neural network, and bidirectional long short-term memory network.
2. The method according to claim 1, characterized in that The mutation signature is used to indicate changes in the mutation region of the amino acid sequence of the sample protein before and after mutation.
3. The method according to claim 2, characterized in that The mutation region is the amino acid corresponding to the mutation site.
4. The method according to claim 2, characterized in that The mutation region includes the amino acid corresponding to the mutation site and a plurality of amino acids adjacent to the mutation site.
5. The method according to claim 4, characterized in that The multiple amino acids adjacent to the mutation site are adjacent 1-10 amino acids.
6. The method according to claim 1, characterized in that When at least one layer of output vectors is extracted through the pre-training model as the feature vector corresponding to the amino acid sequence before mutation, or when at least one layer of output vectors is extracted through the pre-training model as the feature vector corresponding to the amino acid sequence after mutation, the at least one layer is the last layer.
7. The method according to any one of claims 1 to 4, characterized in that The training set includes an initial training set and a reverse training set; the reverse training set is obtained by reversing the initial training set based on the thermodynamic characteristics of the protein, so that the number of training samples carrying different mutation stability labels in the training set is equal.
8. The method according to claim 1, characterized in that The dimensionality reduction process is achieved by a combination of one or more of the following algorithms: principal component analysis algorithm, discriminant analysis algorithm, multidimensional scaling algorithm, isometric feature mapping algorithm and random nearest neighbor embedding algorithm.
9. A method for predicting protein stability after mutation, characterized in that: The method comprises: Acquiring information of a sample to be predicted, wherein the information of the sample to be predicted includes amino acid sequence information before and after the mutation of the protein to be tested; The prediction model constructed by the method according to any one of claims 1 to 8 is used to perform the following processing on the information of the sample to be predicted, wherein the prediction model includes a pre-trained model and a trained classifier, the pre-trained model is ProtT5-XL, and the classifier is implemented by one of the following algorithms: random forest algorithm, convolutional neural network and bidirectional long short-term memory network: Based on the amino acid sequence information of the protein to be tested of each sample to be predicted before and after the mutation, the pre-trained model is used to perform feature extraction to obtain the pre-mutation feature and the post-mutation feature of the sample to be predicted, and the pre-mutation feature and the post-mutation feature are spliced to obtain the mutation feature of the sample to be predicted; Performing dimensionality reduction processing on the mutation feature to obtain the mutation feature after dimensionality reduction; and Based on the mutation features after dimensionality reduction, the trained classifier is used to perform prediction to determine the post-mutation stability prediction result of the sample to be predicted.
10. A system for constructing a protein stability prediction model obtained by the method according to any one of claims 1 to 8, characterized in that: The system comprises: An acquisition module, used to acquire a training set, wherein the training set includes information of multiple training samples, and the information of each training sample includes amino acid sequence information and mutation stability label of the sample protein before and after mutation; A pre-training module, used to obtain a pre-training model, based on the amino acid sequence information of the protein before and after mutation of each of the multiple training samples, perform feature extraction through the pre-training model to obtain pre-mutation features and post-mutation features of the multiple training samples, and splice the pre-mutation features and the post-mutation features to obtain mutation features of the multiple training samples; Performing dimensionality reduction processing on the mutation feature to obtain the mutation feature after dimensionality reduction; and A training module is used to obtain a classifier, and the classifier is trained based on the mutation characteristics after dimensionality reduction and the mutation stability label to obtain a prediction model, wherein the prediction model includes the pre-trained model and the trained classifier, the pre-trained model is ProtT5-XL, and the classifier is implemented using one of the following algorithms: random forest algorithm, convolutional neural network, and bidirectional long short-term memory network.
11. A system for predicting protein stability after mutation, characterized in that: The system comprises: An acquisition module, used to acquire information of a sample to be predicted, wherein the information of the sample to be predicted includes amino acid sequence information before and after the mutation of the protein to be tested; A processing module, used to perform the following processing on the information of the sample to be predicted using the prediction model constructed by the method according to any one of claims 1 to 8, wherein the prediction model includes a pre-trained model and a trained classifier, wherein the pre-trained model is ProtT5-XL, and the classifier is implemented using one of the following algorithms: a random forest algorithm, a convolutional neural network, and a bidirectional long short-term memory network: Based on the amino acid sequence information of the protein to be tested of each sample to be predicted before and after the mutation, the pre-trained model is used to perform feature extraction to obtain the pre-mutation feature and the post-mutation feature of the sample to be predicted, and the pre-mutation feature and the post-mutation feature are spliced to obtain the mutation feature of the sample to be predicted; Performing dimensionality reduction processing on the mutation feature to obtain the mutation feature after dimensionality reduction; and Based on the mutation features after dimensionality reduction, the trained classifier is used to perform prediction to determine the post-mutation stability prediction result of the sample to be predicted.
12. A device, characterized in that: It comprises at least one storage medium and at least one processor, wherein the at least one storage medium is used to store computer instructions; the at least one processor is used to execute the computer instructions to implement the method for constructing a protein stability prediction model as described in any one of claims 1 to 8 or the method for predicting protein stability after mutation as described in claim 9.
13. A computer-readable storage medium, characterized in that: The storage medium stores computer instructions, and when the computer instructions are executed by the processor, the method for constructing a protein stability prediction model according to any one of claims 1 to 8 or the method for predicting protein stability after mutation according to claim 9 is implemented.
