Training method of protein structure prediction model and protein structure prediction method

Abstract

The application provides a protein structure prediction model training method and a protein structure prediction method, and relates to the technical field of computers. The protein structure prediction method and the protein structure prediction method provided by the application are characterized in that each amino acid sequence sample in a training data set has a first feature matrix and a second feature matrix, the first feature matrix and the second feature matrix of each amino acid sequence sample are used as a basis, and a knowledge distillation method is adopted to jointly train a feature enhancement network and a structure prediction network contained in a protein structure prediction model by means of an auxiliary training classification network. The obtained protein structure prediction model can be used to predict the structure of a to-be-processed protein based on a low-quality feature representation matrix of the to-be-processed protein, and a protein result prediction result with high precision can be obtained, thereby improving the prediction accuracy of protein structure prediction.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of computer technology, and in particular to a protein structure prediction model training method and a protein structure prediction method. BACKGROUND

[0002] Protein structure prediction can help understand the role of proteins, understand how proteins perform their biological functions, and understand the interaction between proteins and other molecules, which is very important for biology, medicine and pharmacy. Protein structure prediction usually needs to predict the secondary structure of the protein first, and then obtain the complete protein structure according to the determined protein secondary structure.

[0003] The prediction accuracy of protein structure prediction, especially the prediction accuracy of protein secondary structure prediction, depends on the quality of the feature representation matrix, which is a score matrix used to represent the specific position of the amino acid sequence of the protein. The higher the quality of the feature representation matrix, the higher the prediction accuracy of the protein structure prediction. However, some proteins do not have a high-quality feature representation matrix, and the quality of their feature representation matrix is low. At present, the prediction accuracy is low when using a low-quality feature representation matrix for protein structure prediction. SUMMARY

[0004] To solve the technical problems in the related art, embodiments of the present application provide a protein structure prediction model training method and a protein structure prediction method, which can improve the prediction accuracy of protein structure prediction.

[0005] To achieve the above object, the technical scheme of the embodiments of the present application is as follows:

[0006] In a first aspect, the embodiments of the present application provide a protein structure prediction model training method, which comprises:

[0007] obtaining a training data set containing a plurality of amino acid sequence samples; the first representation matrix of each amino acid sequence sample is determined according to the associated homologous sequence set, and the number of homologous sequences in the homologous sequence set is not less than a set threshold; the second representation matrix of each amino acid sequence sample is determined according to the down-sampling result of the associated homologous sequence set;

[0008] based on the training data set, iteratively training the feature enhancement network and the structure prediction network contained in the protein structure prediction model until a set convergence condition is met, wherein one iteration training process comprises:

[0009] based on the amino acid sequence sample and the corresponding second representation matrix extracted from the training data set, obtaining a corresponding enhanced sample representation matrix through the feature enhancement network, and determining a first loss value according to the enhanced sample representation matrix and the first representation matrix of the amino acid sequence sample;

[0010] After splicing the enhanced sample representation matrix and the amino acid sequence sample, inputting them into the structure prediction network and the classification network for auxiliary training respectively, and based on the obtained output result, determining a second loss value;

[0011] According to the first loss value and the second loss value, respectively adjusting the parameters of the feature enhancement network and the structure prediction network.

[0012] In a second aspect, the embodiments of the present application also provide a protein structure prediction method, the method comprising:

[0013] inputting a protein information matrix corresponding to a to-be-processed protein into a trained protein structure prediction model to obtain a set structure category to which each amino acid in the to-be-processed protein belongs; the protein structure prediction model is obtained through the training method in any one of the first aspect; the protein information matrix is obtained based on the amino acid sequence of the to-be-processed protein;

[0014] determining the structure of the to-be-processed protein according to the set structure category to which each amino acid belongs.

[0015] In a third aspect, the embodiments of the present application also provide a training device of a protein structure prediction model, the device comprising:

[0016] a data acquisition unit configured to acquire a training data set containing a plurality of amino acid sequence samples; a first representation matrix of each amino acid sequence sample is determined according to an associated homologous sequence set, and the number of homologous sequences in the homologous sequence set is not less than a set threshold; a second representation matrix of each amino acid sequence sample is determined according to a down-sampling result of the associated homologous sequence set;

[0017] The model training unit is configured to iteratively train the feature enhancement network and the structure prediction network included in the protein structure prediction model based on the training data set until a preset convergence condition is met. In one iteration training process, based on an amino acid sequence sample and a corresponding second feature matrix extracted from the training data set, a corresponding enhanced sample feature matrix is obtained by the feature enhancement network, and a first loss value is determined according to the enhanced sample feature matrix and the first feature matrix of the amino acid sequence sample. The enhanced sample feature matrix and the amino acid sequence sample are input into the structure prediction network and a classification network used for auxiliary training after being spliced, and a second loss value is determined based on the obtained output result. The feature enhancement network and the structure prediction network are adjusted in parameters according to the first loss value and the second loss value, respectively.

[0018] In an optional embodiment, the data acquisition unit is specifically configured to:

[0019] Based on the amino acid sequence sample, a plurality of homologous sequences corresponding to the amino acid sequence sample are searched from an amino acid sequence database, and the plurality of homologous sequences are grouped into a homologous sequence set associated with the amino acid sequence sample.

[0020] The homologous sequence set associated with the amino acid sequence sample is down-sampled to obtain a down-sampling result of the homologous sequence set.

[0021] The second feature matrix of the amino acid sequence sample is determined according to the down-sampling result of the homologous sequence set.

[0022] In an optional embodiment, the model training unit is specifically configured to:

[0023] The amino acid sequence sample is input into a matrix generation network to obtain an initial sample feature matrix of the amino acid sequence sample. The matrix generation network is trained based on a natural language processing model.

[0024] The amino acid sequence sample, the second feature matrix of the amino acid sequence sample and the initial sample feature matrix are spliced to obtain a first protein information sample matrix.

[0025] The first protein information sample matrix is input into the feature enhancement network to obtain an enhanced sample feature matrix corresponding to the amino acid sequence sample.

[0026] In an optional embodiment, the model training unit is further configured to:

[0027] The enhanced sample feature matrix and the amino acid sequence sample are spliced to obtain a second protein information sample matrix.

[0028] inputting the second protein information sample matrix into the structure prediction network to obtain structure prediction data corresponding to the amino acid sequence sample;

[0029] inputting the second protein information sample matrix into the classification network to obtain structure prediction reference data corresponding to the amino acid sequence sample;

[0030] comparing the structure prediction data and the structure prediction reference data to determine the second loss value.

[0031] In an optional embodiment, the model training unit is further configured to:

[0032] concatenating the amino acid sequence sample and the first representation matrix of the amino acid sequence sample to obtain a first auxiliary training information matrix, and inputting the first auxiliary training information matrix into the classification network to obtain first auxiliary structure prediction data corresponding to the amino acid sequence sample;

[0033] concatenating the amino acid sequence sample and the second representation matrix of the amino acid sequence sample to obtain a second auxiliary training information matrix, and inputting the second auxiliary training information matrix into the classification network to obtain second auxiliary structure prediction data corresponding to the amino acid sequence sample;

[0034] concatenating the amino acid sequence sample and the initial sample representation matrix of the amino acid sequence sample to obtain a third auxiliary training information matrix, and inputting the third auxiliary training information matrix into the classification network to obtain third auxiliary structure prediction data corresponding to the amino acid sequence sample;

[0035] determining a third loss value based on the structure prediction reference data, the first auxiliary structure prediction data, the second auxiliary structure prediction data and the third auxiliary structure prediction data;

[0036] adjusting parameters of the feature enhancement network and the structure prediction network according to the first loss value, the second loss value and the third loss value.

[0037] In an optional embodiment, the model training unit is further configured to:

[0038] determining a joint loss value according to a weighted sum of the first loss value, the second loss value and the third loss value;

[0039] adjusting parameters of the feature enhancement network and the structure prediction network according to the joint loss value.

[0040] In a fourth aspect, the embodiments of the present application further provide a protein structure prediction device, the device comprising:

[0041] a category determining unit, configured to input a protein information matrix corresponding to a to-be-processed protein into a trained protein structure prediction model to obtain a set structure category to which each amino acid in the to-be-processed protein belongs; the protein structure prediction model is obtained by the training method in any one of the first aspect; the protein information matrix is obtained based on an amino acid sequence of the to-be-processed protein;

[0042] a structure determining unit, configured to determine a structure of the to-be-processed protein according to the set structure category to which each amino acid belongs.

[0043] In an optional embodiment, the apparatus further includes a matrix generating unit, configured to:

[0044] input the amino acid sequence into a matrix generating network, perform homology information representation on the amino acid sequence based on the matrix generating network, and obtain an initial homology information representation matrix of the to-be-processed protein; the matrix generating network is trained based on a natural language processing model;

[0045] splice the amino acid sequence and the initial homology information representation matrix to obtain the protein information matrix corresponding to the to-be-processed protein.

[0046] In an optional embodiment, the category determining unit is specifically configured to:

[0047] input the protein information matrix corresponding to the to-be-processed protein into a feature enhancement network of the protein structure prediction model, and determine an enhanced homology information representation matrix corresponding to the to-be-processed protein based on the feature enhancement network;

[0048] splice the amino acid sequence and the enhanced homology information representation matrix to obtain an enhanced protein information matrix;

[0049] input the enhanced protein information matrix into a structure prediction network of the protein structure prediction model, and determine the set structure category to which each amino acid in the to-be-processed protein belongs based on the structure prediction network.

[0050] In an optional embodiment, the feature enhancement network includes a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer and at least one fully connected layer; and the category determining unit is further configured to:

[0051] input the protein information matrix into the convolutional subnetwork, and obtain a basic feature of the protein information matrix based on the convolutional subnetwork;

[0052] input the protein information matrix into the long short-term memory subnetwork, and obtain time sequence features of the protein information matrix based on the long short-term memory subnetwork;

[0053] input the base features and the time sequence features into the concatenation layer, concatenate the base features and the time sequence features based on the concatenation layer, and obtain synthetic features;

[0054] input the synthetic features into the full connection layer, and obtain the enhanced homology information representation matrix corresponding to the to-be-processed protein based on the full connection layer.

[0055] In an optional embodiment, the structure prediction network comprises a convolution subnetwork, a long short-term memory subnetwork, a concatenation layer, and a multi-classification layer; and the category determination unit is further configured to:

[0056] input the enhanced protein information matrix into the convolution subnetwork, and obtain base features of the enhanced protein information matrix based on the convolution subnetwork;

[0057] input the enhanced protein information matrix into the long short-term memory subnetwork, and obtain time sequence features of the enhanced protein information matrix based on the long short-term memory subnetwork;

[0058] input the base features and the time sequence features into the concatenation layer, concatenate the base features and the time sequence features based on the concatenation layer, and obtain synthetic features;

[0059] input the synthetic features into the multi-classification layer, and obtain a set structure category to which each amino acid in the to-be-processed protein belongs based on the multi-classification layer.

[0060] In a fifth aspect, the embodiments of the present application further provide a computer readable storage medium, wherein a computer program is stored in the computer readable storage medium, and the computer program is executed by a processor to implement the training method of the protein structure prediction model in the first aspect.

[0061] In a sixth aspect, the embodiments of the present application further provide a computer readable storage medium, wherein a computer program is stored in the computer readable storage medium, and the computer program is executed by a processor to implement the protein structure prediction method in the second aspect.

[0062] In a seventh aspect, the embodiments of the present application further provide an electronic device, comprising a memory and a processor, wherein the memory stores a computer program capable of running on the processor, and when the computer program is executed by the processor, the processor implements the training method of the protein structure prediction model in the first aspect.

[0063] In an eighth aspect, the embodiments of the present application further provide an electronic device including a memory and a processor, the memory storing a computer program capable of running on the processor, and when the computer program is executed by the processor, the processor implements the protein structure prediction method of the second aspect.

[0064] The training method of the protein structure prediction model and the protein structure prediction method provided by the embodiments of the present application, each amino acid sequence sample in the training data set has a first feature matrix and a second feature matrix, the first feature matrix is a high-quality feature matrix determined according to the homologous sequence set associated with the amino acid sequence sample, and the second feature matrix is a low-quality feature matrix determined according to the down-sampling result of the homologous sequence set. The embodiments of the present application are based on the first feature matrix and the second feature matrix of each amino acid sequence sample, and the feature enhancement network and the structure prediction network contained in the protein structure prediction model are jointly trained by means of the auxiliary training classification network by using the knowledge distillation method. The obtained protein structure prediction model can predict the structure of the to-be-processed protein based on the low-quality feature matrix of the to-be-processed protein, obtain a protein result prediction result with high precision, and improve the prediction accuracy of the protein structure prediction. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0066] Figure 1 An application scenario diagram of the training method of the protein structure prediction model provided by the embodiments of the present application;

[0067] Figure 2 A flowchart of the training method of the protein structure prediction model provided by the embodiments of the present application;

[0068] Figure 3 A flowchart of another training method of the protein structure prediction model provided by the embodiments of the present application;

[0069] Figure 4 A flowchart of another training method of the protein structure prediction model provided by the embodiments of the present application;

[0070] Figure 5 A flowchart of the protein structure prediction method provided by the embodiments of the present application;

[0071] Figure 6A flowchart of another protein structure prediction method provided by the embodiment of the present application is shown in FIG. 1.

[0072] Figure 7 A structural diagram of a feature enhancement network provided by the embodiment of the present application is shown in FIG. 2.

[0073] Figure 8 A structural diagram of a structure prediction network provided by the embodiment of the present application is shown in FIG. 3.

[0074] Figure 9 A performance effect comparison bar chart of a protein structure prediction method provided by the embodiment of the present application and related methods is shown in FIG. 4.

[0075] Figure 10 A structural block diagram of a training device of a protein structure prediction model provided by the embodiment of the present application is shown in FIG. 5.

[0076] Figure 11 A structural block diagram of a protein structure prediction device provided by the embodiment of the present application is shown in FIG. 6.

[0077] Figure 12 A structural block diagram of another protein structure prediction device provided by the embodiment of the present application is shown in FIG. 7.

[0078] Figure 13 A structural diagram of an electronic device provided by the embodiment of the present application is shown in FIG. 8. DETAILED DESCRIPTION

[0079] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0080] It should be noted that the terms "include" and "have" and their conjugations involved in the documents of the present application are intended to cover the inclusions not exclusively, for example, a process, method, system, product or device including a series of steps or units need not be limited to those clearly listed steps or units, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0081] The following explains some terms in the embodiments of the present application to facilitate the understanding of those skilled in the art.

[0082] (1) Protein Secondary Structure Prediction (PSSP): Generally considered as the first step of protein structure prediction, which can be a sequence labeling problem. A protein sequence is composed of an amino acid chain, each position can be any one of the 20 basic amino acids, and the role of PSSP is to predict the label of each position.

[0083] (2) Multiple Sequence Alignment (MSA): Aligning two or more amino acid sequences, comparing the similarities and differences of amino acids contained in each sequence column by column, so that the amino acid sequences involved in the alignment have as many columns as possible with the same amino acid, and the same residue position is in the same column, so as to find the similar parts between different sequences, and to infer the structural and functional similarities of these amino acid sequences.

[0084] (3) Position-Specific Scoring Matrix (PSSM): It can be calculated by multiple sequence alignment MSA, and the quality of PSSM can determine the accuracy of PSSP. When the quality of PSSM is higher, the accuracy of PSSP is higher, and when the quality of PSSM is lower, the accuracy of PSSP is lower.

[0085] (4) Knowledge Distillation (KD): To improve the performance of student network by transferring the knowledge learned by teacher model to student network.

[0086] (5) Contrastive Learning (CL): To increase the inter-class distance of different class samples and reduce the intra-class distance between the same class samples by learning positive and negative samples.

[0087] The word "exemplary" used in the following means "serving as an example, an implementation, or an illustration". Any embodiment described as "exemplary" is not necessarily to be construed as superior or better than other embodiments.

[0088] The terms "first", "second" in the text are only for descriptive purposes, and cannot be understood as explicitly or implicitly indicating relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more features, and in the description of the embodiments of the present application, unless otherwise specified, the meaning of "multiple" is two or more.

[0089] Embodiments of the present application are related to artificial intelligence (AI) and machine learning technologies, and are designed based on computer vision (CV) technology, speech technology and machine learning (ML) in artificial intelligence.

[0090] Artificial intelligence is a theory, method, technology and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is the design principle and implementation method of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making. Artificial intelligence technology mainly includes computer vision technology, speech processing technology, and machine learning / deep learning and several other directions.

[0091] With the research and progress of artificial intelligence technology, artificial intelligence has been researched and applied in many fields, such as common smart home, image retrieval, video monitoring, smart sound box, intelligent marketing, unmanned driving, autonomous driving, unmanned aerial vehicle, robot, intelligent medical treatment, etc. It is believed that with the development of technology, artificial intelligence will be applied in more fields and play an increasingly important role.

[0092] Machine learning is a multi-disciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is a subject that studies how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, etc. Embodiments of the present application train a protein structure prediction model based on machine learning or deep learning to obtain a trained protein structure prediction model, and determine the set structure category to which each amino acid in the protein to be processed belongs through the protein structure prediction model based on machine learning or deep learning in the protein structure prediction process.

[0093] To better understand the technical solutions provided in the embodiments of this application, the following is a brief introduction to the application scenarios to which the technical solutions provided in the embodiments of this application are applicable. It should be noted that the application scenarios described below are only for illustrating the embodiments of this application and are not intended to limit the scope. In specific implementation, the technical solutions provided in the embodiments of this application can be flexibly applied according to actual needs.

[0094] The protein structure prediction method provided in this application embodiment can be applied to… Figure 1 In the application scenarios shown. See also Figure 1 As shown, this application scenario includes multiple terminal devices 11 and a server 12. Terminal devices 11 and server 12 can connect and transmit data via wired or wireless connections. For example, terminal devices 11 and server 12 can be connected via a data cable or a wired network; they can also be connected via an RF module, a WiFi module, or a wireless network.

[0095] The terminal device 11 can be a computer, laptop, PDA, tablet computer, etc. The server 12 can be a single server, a server cluster consisting of several servers, a cloud computing center, a virtualization platform, or a personal computer, large or medium-sized computer, or computer cluster. Depending on the implementation requirements, the application scenarios in this application embodiment can have any number of terminal devices and servers. This application does not impose any special limitations on this. The protein structure prediction method and protein structure prediction model training method provided in this application embodiment can be executed by the server 12, or by the terminal device 11 and the server 12 collaboratively.

[0096] For example, a medical research institution has a server 12 used to train a protein structure prediction model and predict protein structures. Each laboratory within the institution has a terminal device 11. Researchers can first train the protein structure prediction model on server 12 using a training dataset containing multiple amino acid sequence samples, resulting in a trained model. After obtaining the trained model, researchers can transmit the amino acid sequence of the protein to be processed to server 12 via their laboratory's terminal device 11. Upon receiving the amino acid sequence, server 12 can first determine the protein information matrix corresponding to the protein to be processed based on the sequence, then input the protein information matrix into the trained model to obtain the defined structure category of each amino acid in the protein to be processed. Finally, based on the defined structure category of each amino acid, the structure of the protein to be processed can be determined.

[0097] To further illustrate the technical solutions provided by the embodiments of the present application, the following will describe in detail in conjunction with the drawings and specific embodiments. Although the embodiments of the present application provide the method operation steps as described in the following embodiments or shown in the drawings, more or less operation steps can be included in the method based on conventional or non-creative labor. The execution order of the steps is not limited to the execution order provided by the embodiments of the present application in the steps that there is no necessary causal relationship in logic. The method can be executed in sequence or in parallel when the method is executed in actual processing or by the device, as shown in the embodiments or the drawings.

[0098] Figure 2 A flowchart of a method for training a protein structure prediction model is shown, which can be executed by the server 12 in Figure 1 , or by the terminal device 11 or other electronic devices. Exemplarily, the following describes the specific implementation process of the training method of the protein structure prediction model of the present application, taking a computer for training a protein structure prediction model as an execution subject. As shown in Figure 2 , the training method of the protein structure prediction model includes the following steps:

[0099] Step S201, obtaining a training data set containing a plurality of amino acid sequence samples.

[0100] The obtained training data set can include a plurality of amino acid sequence samples, and the first feature matrix of each amino acid sequence sample is determined according to the associated homologous sequence set, and the number of homologous sequences in the homologous sequence set is not less than a set threshold. In addition, the second feature matrix of each amino acid sequence sample is determined according to the down-sampling result of the associated homologous sequence set.

[0101] Based on the training data set, the feature enhancement network and the structure prediction network contained in the protein structure prediction model can be iteratively trained until the set convergence condition is met, wherein one iteration training process can include steps S202 to S204.

[0102] In an embodiment, based on the amino acid sequence sample, a plurality of homologous sequences corresponding to the amino acid sequence sample can be searched from the amino acid sequence database, and the plurality of homologous sequences can be combined into a homologous sequence set associated with the amino acid sequence sample. Then, the homologous sequence set associated with the amino acid sequence sample is down-sampled to obtain a down-sampling result of the homologous sequence set. Finally, according to the down-sampling result of the homologous sequence set, a second feature matrix of the amino acid sequence sample is determined. For example, based on the amino acid sequence sample D0, 200 homologous sequences corresponding to the amino acid sequence sample D0 can be searched from the amino acid sequence database, then 50 homologous sequences are randomly extracted from the 200 homologous sequences, and a second feature matrix of the amino acid sequence sample D0 is determined according to the 50 extracted homologous sequences.

[0103] In step S202, based on the amino acid sequence sample and the corresponding second feature matrix extracted from the training data set, a corresponding enhanced sample feature matrix is obtained through the feature enhancement network, and a first loss value is determined according to the enhanced sample feature matrix and the first feature matrix of the amino acid sequence sample.

[0104] The amino acid sequence sample can be first input into the matrix generation network to obtain an initial sample feature matrix of the amino acid sequence sample. The matrix generation network can be a pre-trained language model (Bidirectional Encoder Representations from Transformers, BERT) trained based on a bidirectional attention mechanism. The matrix generation network can also be trained based on other natural language processing models, which are not limited in the present application.

[0105] Then, the amino acid sequence sample, the second feature matrix of the amino acid sequence sample and the initial sample feature matrix are spliced to obtain a first protein information sample matrix, and the first protein information sample matrix is input into the feature enhancement network to obtain an enhanced sample feature matrix corresponding to the amino acid sequence sample.

[0106] According to the enhanced sample feature matrix and the first feature matrix of the amino acid sequence sample, a first loss value can be determined. In calculating the first loss value, a mean square error (MSE) loss function can be used to calculate the first loss value. The first loss value is used to determine the closeness between the actual output first feature matrix and the expected output enhanced sample feature matrix. The smaller the first loss value, the closer the actual output is to the expected output. The first loss value can also be calculated by using other loss functions, which are not limited in the present application.

[0107] In step S203, the enhanced sample representation matrix and the amino acid sequence sample are spliced and respectively input into the structure prediction network and the classification network for auxiliary training, and based on the obtained output results, a second loss value is determined.

[0108] The enhanced sample representation matrix and the amino acid sequence sample are spliced to obtain a second protein information sample matrix. The second protein information sample matrix is input into the structure prediction network to obtain structure prediction data corresponding to the amino acid sequence sample. The second protein information sample matrix is input into the classification network to obtain structure prediction reference data corresponding to the amino acid sequence sample. The structure prediction data and the structure prediction reference data are compared to determine the second loss value.

[0109] In calculating the second loss value, a knowledge distillation (KD) loss function can be used to calculate the second loss value. The second loss value is used to determine the closeness of the actual output structure prediction data to the expected output structure prediction reference data. The smaller the second loss value, the closer the actual output is to the expected output. The second loss value can also be calculated using other loss functions, which are not limited by the embodiments of the present application.

[0110] In step S204, the feature enhancement network and the structure prediction network are respectively adjusted according to the first loss value and the second loss value.

[0111] In an embodiment, a joint loss value can be determined only according to the weighted sum of the first loss value and the second loss value, and it is determined whether the joint loss value converges to a preset target value. If the joint loss value is less than or equal to the preset target value, or the change amplitude of the joint loss value obtained by continuous N times of training is less than or equal to the preset target value, it is considered that the joint loss value has converged to the preset target value, indicating that the joint loss value converges; otherwise, it is considered that the joint loss value has not converged. If the joint loss value has not converged, the parameters of the feature enhancement network and the structure prediction network are adjusted, and after the parameters are adjusted, the step of extracting the amino acid sequence sample from the training data set is executed again to continue the next round of training process. If the joint loss value converges, the current feature enhancement network and structure prediction network are taken as a trained protein structure prediction model. The joint loss value determined according to the weighted sum of the first loss value and the second loss value is used to train the protein structure prediction model, and the obtained protein structure prediction model can accurately predict the set structure category to which the amino acid of the protein structure belongs.

[0112] In another embodiment, the amino acid sequence sample and the first characterization matrix of the amino acid sequence sample are spliced to obtain a first auxiliary training information matrix, and the first auxiliary training information matrix is input into the classification network, so that the first auxiliary structure prediction data corresponding to the amino acid sequence sample can be obtained. The amino acid sequence sample and the second characterization matrix of the amino acid sequence sample are spliced to obtain a second auxiliary training information matrix, and the second auxiliary training information matrix is input into the classification network, so that the second auxiliary structure prediction data corresponding to the amino acid sequence sample can be obtained. The amino acid sequence sample and the initial sample characterization matrix of the amino acid sequence sample are spliced to obtain a third auxiliary training information matrix, and the third auxiliary training information matrix is input into the classification network, so that the third auxiliary structure prediction data corresponding to the amino acid sequence sample can be obtained. Based on the structure prediction reference data, the first auxiliary structure prediction data, the second auxiliary structure prediction data and the third auxiliary structure prediction data, a third loss value is determined. And according to the weighted sum of the first loss value, the second loss value and the third loss value, a joint loss value is determined, and the parameter adjustment of the feature enhancement network and the structure prediction network is carried out according to the joint loss value. According to the joint loss value determined by the first loss value, the second loss value and the third loss value, the training effect of the protein structure prediction model can be further improved, and the prediction accuracy of the obtained protein structure prediction model for predicting the set structure category to which the amino acid constituting the protein structure belongs can also be further improved.

[0113] The training method of the protein structure prediction model provided by the embodiments of the present application can obtain the protein structure prediction model by jointly training the feature enhancement network and the structure prediction network included in the protein structure prediction model based on the first characterization matrix and the second characterization matrix of each amino acid sequence sample in the training data set and by using the knowledge distillation method and the auxiliary training classification network. Therefore, the trained protein structure prediction model can be used to predict the structure of a to-be-processed protein based on the low-quality feature characterization matrix of the to-be-processed protein, so as to obtain a protein structure prediction result with high accuracy and improve the prediction accuracy of the protein structure prediction.

[0114] Figure 3 The detailed implementation process of the training method of the protein structure prediction model provided by the present application is shown in the following Figure 3 The training method of the protein structure prediction model provided by the present application is shown in the following

[0115] In step S301, a training data set is obtained.

[0116] The obtained training data set can include a plurality of amino acid sequence samples, and each amino acid sequence sample corresponds to a first feature matrix, a second feature matrix and an initial sample feature matrix. The first feature matrix corresponding to the amino acid sequence sample is determined according to the homologous sequence set associated with the amino acid sequence sample, the second feature matrix corresponding to the amino acid sequence sample is determined according to the down-sampling result of the homologous sequence set associated with the amino acid sequence sample, and the initial sample feature matrix corresponding to the amino acid sequence sample is obtained by inputting the amino acid sequence sample into the trained matrix generation network.

[0117] In step S302, an amino acid sequence sample is extracted from the training data set.

[0118] The training data set can be obtained in advance, and when the protein structure prediction model is trained, an amino acid sequence sample is extracted from the training data set as training sample data.

[0119] In step S303, the extracted amino acid sequence sample and the corresponding second feature matrix and initial sample feature matrix are input into the feature enhancement network to be trained to obtain a corresponding enhanced sample feature matrix.

[0120] When the training data set is used to train the protein structure prediction model, an amino acid sequence sample in the training data set can be extracted, and the extracted amino acid sequence sample and the corresponding second feature matrix and initial sample feature matrix are input into the feature enhancement network included in the protein structure prediction model to be trained. Based on the feature enhancement network, an enhanced sample feature matrix corresponding to the amino acid sequence sample can be obtained.

[0121] In step S304, a first loss value is determined according to the enhanced sample feature matrix and the first feature matrix of the amino acid sequence sample.

[0122] The first loss value can be determined by comparing the enhanced sample feature matrix and the first feature matrix of the amino acid sequence sample. In calculating the first loss value, the MSE loss function can be used to calculate the first loss value. The first loss value is used to determine the closeness between the actual output first feature matrix and the expected output enhanced sample feature matrix. The smaller the first loss value, the closer the actual output is to the expected output.

[0123] In step S305, the enhanced sample feature matrix and the amino acid sequence sample are spliced and input into the structure prediction network to be trained to obtain structure prediction data corresponding to the amino acid sequence sample.

[0124] The enhanced sample representation matrix and the amino acid sequence sample can be spliced to obtain a second protein information sample matrix, and the second protein information sample matrix is input into a structure prediction network included in the protein structure prediction model to be trained. Based on the structure prediction network, structure prediction data corresponding to the amino acid sequence sample can be obtained.

[0125] In step S306, the enhanced sample representation matrix and the amino acid sequence sample are spliced and input into a classification network for auxiliary training to obtain structure prediction reference data corresponding to the amino acid sequence sample.

[0126] The enhanced sample representation matrix and the amino acid sequence sample can be spliced to obtain a second protein information sample matrix, and the second protein information sample matrix is input into a classification network for auxiliary training, and the classification network is a trained classification network. Based on the classification network, structure prediction reference data corresponding to the amino acid sequence sample can be obtained.

[0127] In step S307, a second loss value is determined according to the structure prediction data corresponding to the amino acid sequence sample and the structure prediction reference data.

[0128] By comparing the structure prediction data corresponding to the amino acid sequence sample with the structure prediction reference data corresponding to the amino acid sequence sample, the second loss value can be determined. In calculating the second loss value, a KD loss function can be used to calculate the second loss value. The second loss value is used to determine the closeness of the actual output structure prediction data to the expected output structure prediction reference data. The smaller the second loss value, the closer the actual output is to the expected output.

[0129] In step S308, the amino acid sequence sample and the first representation matrix of the amino acid sequence sample are spliced and input into the classification network to obtain first auxiliary structure prediction data corresponding to the amino acid sequence sample.

[0130] The amino acid sequence sample and the first representation matrix of the amino acid sequence sample can be spliced to obtain a first auxiliary training information matrix, and the first auxiliary training information matrix is input into the trained classification network. Based on the classification network, first auxiliary structure prediction data corresponding to the amino acid sequence sample can be obtained.

[0131] In step S309, the amino acid sequence sample and the second representation matrix of the amino acid sequence sample are spliced and input into the classification network to obtain second auxiliary structure prediction data corresponding to the amino acid sequence sample.

[0132] The amino acid sequence sample and the second characterization matrix of the amino acid sequence sample can be spliced to obtain a second auxiliary training information matrix, and the second auxiliary training information matrix is input into the trained classification network. Based on the classification network, second auxiliary structure prediction data corresponding to the amino acid sequence sample can be obtained.

[0133] In step S310, the amino acid sequence sample and the initial sample characterization matrix of the amino acid sequence sample are spliced and input into the classification network to obtain third auxiliary structure prediction data corresponding to the amino acid sequence sample.

[0134] The amino acid sequence sample and the initial sample characterization matrix of the amino acid sequence sample are spliced to obtain a third auxiliary training information matrix, and the third auxiliary training information matrix is input into the trained classification network. Based on the classification network, third auxiliary structure prediction data corresponding to the amino acid sequence sample can be obtained.

[0135] In step S311, a third loss value is determined according to the structure prediction reference data, the first auxiliary structure prediction data, the second auxiliary structure prediction data and the third auxiliary structure prediction data.

[0136] The structure prediction reference data, the first auxiliary structure prediction data, the second auxiliary structure prediction data and the third auxiliary structure prediction data are compared to determine the third loss value. In calculating the third loss value, a contrastive learning (CL) loss function can be used to calculate the third loss value. The third loss value is used to determine the closeness between the actual output structure prediction reference data and the expected output first auxiliary structure prediction data, second auxiliary structure prediction data and third auxiliary structure prediction data. The smaller the third loss value, the closer the actual output is to the expected output. The third loss value can also be calculated by using other loss functions, which are not limited in the embodiments of the present application.

[0137] In an embodiment, when the prediction accuracy of the set structure category of the amino acid constituting the protein structure is not required to be high, the third loss value can not be determined, and only the first loss value and the second loss value are used to train the protein structure prediction model.

[0138] In step S312, a joint loss value is determined according to the weighted sum of the first loss value, the second loss value and the third loss value.

[0139] In step S313, it is determined whether the joint loss value converges to a preset target value; if not, step S314 is performed; if yes, step S315 is performed.

[0140] The first loss value, the second loss value and the third loss value can be subjected to a weighting and operation to obtain a joint loss value. It is determined whether the joint loss value converges to a preset target value. If the joint loss value is less than or equal to the preset target value, or the change amplitude of the joint loss value obtained by continuous N times of training is less than or equal to the preset target value, it is considered that the joint loss value has converged to the preset target value, indicating that the joint loss value converges. Otherwise, it is considered that the joint loss value has not converged.

[0141] In step S314, the parameters of the feature enhancement network and the structure prediction network to be trained are adjusted according to the determined joint loss value.

[0142] If the joint loss value has not converged, the parameters of the feature enhancement network and the structure prediction network included in the protein structure prediction model are adjusted respectively. After adjusting the model parameters, step S302 is returned to continue the next round of training process.

[0143] In step S315, the training is ended to obtain a trained protein structure prediction model.

[0144] If the joint loss value converges, the current obtained feature enhancement network and structure prediction network are taken as a trained protein structure prediction model.

[0145] Referring to Figure 4 As shown in the figure, a specific application scenario is used below to further illustrate the above embodiments:

[0146] Suppose that the training data set obtained contains an amino acid sequence sample S, a plurality of homologous sequences corresponding to the amino acid sequence sample S can be found from an amino acid sequence database, and the plurality of homologous sequences form a homologous sequence set associated with the amino acid sequence sample S. According to the homologous sequence set associated with the amino acid sequence sample S, a first feature matrix X h corresponding to the amino acid sequence sample S can be determined. By downsampling the homologous sequence set associated with the amino acid sequence sample S, a downsampled result of the homologous sequence set can be obtained, and according to the downsampled result of the homologous sequence set, a second feature matrix X l corresponding to the amino acid sequence sample S can be determined.

[0147] Based on the amino acid sequence sample S in the training data set, the feature enhancement network and the structure prediction network included in the protein structure prediction model can be iteratively trained until a set convergence condition is met. One iteration training process can include:

[0148] The amino acid sequence sample S is input into the trained matrix generation network, and based on the matrix generation network, an initial sample feature matrix X b. The amino acid sequence sample S, the second feature matrix X l , and the initial sample feature matrix X b are spliced to obtain a first protein information sample matrix (S, X b , X l ), and the first protein information sample matrix (S, X b , X l ) is input into a feature enhancement network F e to be trained, so that an enhanced sample feature matrix X e corresponding to the amino acid sequence sample S can be obtained based on the feature enhancement network F e . According to the first feature matrix X h corresponding to the amino acid sequence sample S and the enhanced sample feature matrix X e corresponding to the amino acid sequence sample S, an MSE loss value L h between the first feature matrix X e and the enhanced sample feature matrix X m can be calculated.

[0149] The amino acid sequence sample S and the enhanced sample feature matrix X e are spliced to obtain a second protein information sample matrix (S, X e ), and the second protein information sample matrix (S, X e ) is input into a structure prediction network F s to be trained, so that structure prediction data F s (S, X s ) corresponding to the amino acid sequence sample S can be obtained based on the structure prediction network F e . The second protein information sample matrix (S, X e ) is input into a classification network F t that has been trained, so that structure prediction reference data F t (S, X t ) corresponding to the amino acid sequence sample S can be obtained based on the classification network F e . According to the structure prediction data F s (S, X e ) corresponding to the amino acid sequence sample S and the structure prediction reference data F t (S, X e ) corresponding to the amino acid sequence sample S, a KD loss value L s between the structure prediction data F e (S, X t ) and the structure prediction reference data F e (S, X d ) can be calculated.

[0150] The amino acid sequence sample S and the first feature matrix Xh , the first auxiliary training information matrix (S, X h ) is obtained, and the first auxiliary training information matrix (S, X h ) is input into the trained classification network F t , and based on the classification network F t , the first auxiliary structure prediction data F t (S, X h ) corresponding to the amino acid sequence sample S can be obtained. The amino acid sequence sample S is spliced with the second representation matrix X l , the second auxiliary training information matrix (S, X l ) is obtained, and the second auxiliary training information matrix (S, X l ) is input into the trained classification network F t , and based on the classification network F t , the second auxiliary structure prediction data F t (S, X l ) corresponding to the amino acid sequence sample S can be obtained. The amino acid sequence sample S is spliced with the initial sample representation matrix X b , the third auxiliary training information matrix (S, X b ) is obtained, and the second auxiliary training information matrix (S, X b ) is input into the trained classification network F t , and based on the classification network F t , the third auxiliary structure prediction data F t (S, X b ) corresponding to the amino acid sequence sample S can be obtained. According to the structure prediction reference data F t (S, X e ), the first auxiliary structure prediction data F t (S, X h ), the second auxiliary structure prediction data F t (S, X l ) and the third auxiliary structure prediction data F t (S, X b ), the CL loss value L t between the structure prediction reference data F e (S, X t ), the first auxiliary structure prediction data F h (S, X t ), the second auxiliary structure prediction data F l (S, X t ) and the third auxiliary structure prediction data F b (S, X l ) can be calculated.

[0151] According to the MSE loss value Lm , a weighted sum of the KD loss value L d and the CL loss value L l , a joint loss value is determined, and then the parameter adjustment of the feature enhancement network and the structure prediction network is performed according to the joint loss value respectively.

[0152] Figure 5 A flowchart of a protein structure prediction method provided by an embodiment of the present application is shown, which can be executed by the server 12 in the computer system 10, or by the terminal device 11 or other electronic devices. Exemplarily, the specific implementation process of the protein structure prediction method of the present application is described below with the computer for predicting protein structure as the execution subject. As shown in FIG. 1, the protein structure prediction method includes the following steps: Figure 1 Figure 5 Step S501: inputting the protein information matrix corresponding to the to-be-processed protein into the trained protein structure prediction model to obtain the set structure category to which each amino acid in the to-be-processed protein belongs.

[0153] Step S501: inputting the protein information matrix corresponding to the to-be-processed protein into the trained protein structure prediction model to obtain the set structure category to which each amino acid in the to-be-processed protein belongs.

[0154] The amino acid sequence can be input into the matrix generation network, and the homologous information representation of the amino acid sequence is performed based on the matrix generation network to obtain the initial homologous information representation matrix of the to-be-processed protein. The matrix generation network is trained based on the BERT model. The amino acid sequence and the initial homologous information representation matrix are spliced to obtain the protein information matrix corresponding to the to-be-processed protein.

[0155] After obtaining the trained protein structure prediction model, the protein information matrix corresponding to the to-be-processed protein can be input into the feature enhancement network included in the trained protein structure prediction model, and the enhanced homologous information representation matrix corresponding to the to-be-processed protein can be determined based on the feature enhancement network. The amino acid sequence and the enhanced homologous information representation matrix are spliced to obtain the enhanced protein information matrix, and the enhanced protein information matrix is input into the structure prediction network included in the trained protein structure prediction model, and the set structure category to which each amino acid in the to-be-processed protein belongs can be determined based on the structure prediction network.

[0156] Step S502: determining the structure of the to-be-processed protein according to the set structure category to which each amino acid belongs.

[0157] According to the obtained set structure category to which each amino acid in the to-be-processed protein belongs, the structure of the to-be-processed protein can be determined.

[0158] ​The protein secondary structure refers to a specific conformation formed by the helix or folding of the polypeptide backbone atoms along an axis, i.e., the spatial position arrangement of the backbone atoms of the peptide chain, without involving the side chain of the amino acid residue. Moreover, the main forms of the protein secondary structure include α-helix, β-sheet, β-turn and random coil. The secondary structure of the protein to be processed can be determined according to the set structure category to which each amino acid in the protein to be processed belongs, i.e., it can be determined that each amino acid in the protein to be processed belongs to which one of the four main forms of α-helix, β-sheet, β-turn and random coil, and then the secondary structure of the protein is determined according to the form to which each amino acid belongs.

[0159] Specifically, as shown in Figure 6 , a specific application scenario can be used to further detail the above embodiments:

[0160] It is assumed that the protein to be processed contains an amino acid sequence S composed of three amino acids D0, D1 and D2.

[0161] In step S601, the amino acid sequence S is input into the matrix generation network to obtain an initial homologous information representation matrix of the protein to be processed.

[0162] The amino acid sequence S can be input into the matrix generation network, and the homologous information of the amino acid sequence is expressed based on the matrix generation network to obtain the initial homologous information representation matrix of the protein to be processed.

[0163] In step S602, the amino acid sequence S and the initial homologous information representation matrix are spliced to obtain a protein information matrix corresponding to the protein to be processed.

[0164] In step S603, the protein information matrix corresponding to the protein to be processed is input into the trained protein structure prediction model to determine that the amino acid D0 in the protein to be processed belongs to the α-helix structure category, the amino acid D1 belongs to the β-sheet structure category, and the amino acid D2 belongs to the β-turn structure category.

[0165] The protein information matrix corresponding to the protein to be processed is input into the trained protein structure prediction model, and based on the protein structure prediction model, it can be determined that the amino acid D0 in the protein to be processed belongs to the α-helix structure category, the amino acid D1 belongs to the β-sheet structure category, and the amino acid D2 belongs to the β-turn structure category.

[0166] In step S604, the structure of the protein to be processed is determined according to the structure category to which the amino acid D0, the amino acid D1 and the amino acid D2 belong.

[0167] According to the fact that the amino acid D0 belongs to the alpha-helix structure category, the amino acid D1 belongs to the beta-sheet structure category, and the amino acid D2 belongs to the beta-turn structure category, the secondary structure of the protein to be processed can be determined.

[0168] In one embodiment, as shown in FIG. 4, the feature enhancement network can include a convolutional subnetwork, a long short-term memory subnetwork, a concatenation layer, and at least one fully connected layer. The protein information matrix can be input into the convolutional subnetwork, and the base features of the protein information matrix can be obtained based on the convolutional subnetwork. The protein information matrix can be input into the long short-term memory subnetwork, and the time sequence features of the protein information matrix can be obtained based on the long short-term memory subnetwork. The base features and the time sequence features can be input into the concatenation layer, and the base features and the time sequence features can be concatenated based on the concatenation layer to obtain the synthetic features. The synthetic features can be input into the fully connected layer, and the enhanced homology information representation matrix corresponding to the protein to be processed can be obtained based on the fully connected layer. Figure 7 The convolutional subnetwork and the long short-term memory subnetwork in the feature enhancement network can also be replaced by other networks, which are not limited in the embodiments of the present application. For example, the transformer network can be used to replace the convolutional subnetwork and the long short-term memory subnetwork.

[0169] In another embodiment, as shown in FIG. 5, the structure prediction network includes a convolutional subnetwork, a long short-term memory subnetwork, a concatenation layer, and a multi-classification layer. The enhanced protein information matrix can be input into the convolutional subnetwork, and the base features of the enhanced protein information matrix can be obtained based on the convolutional subnetwork. The enhanced protein information matrix can be input into the long short-term memory subnetwork, and the time sequence features of the enhanced protein information matrix can be obtained based on the long short-term memory subnetwork. The base features and the time sequence features can be input into the concatenation layer, and the base features and the time sequence features can be concatenated based on the concatenation layer to obtain the synthetic features. The synthetic features can be input into the multi-classification layer, and the set structure category to which each amino acid in the protein to be processed belongs can be obtained based on the multi-classification layer.

[0170] Figure 8 In some embodiments, the protein structure prediction method proposed in the present application can be compared with the Real method and the Bagging method on all the key data, and when the set structure category is three classification, the comparison results can be as shown in Table 1:

[0171] Table 1

[0172] Table 1

[0173]

[0174] ​As shown in Table 1, MSA Counts is the number of multiple sequence alignment, BC40, CullPDB and CB513 are key datasets, Number is the number of amino acid sequences with MSA Counts found from the key datasets, Real and Bagging are related methods for comparison with the protein structure prediction method proposed in the present application. As can be seen from Table 1, when the structure category is set to three categories, the performance of the protein structure prediction method proposed in the present application is higher than that of the Real method and the Bagging method.

[0175] When the structure category is set to eight categories, the comparison results can be as shown in Table 2:

[0176] Table 2

[0177]

[0178] As shown in Table 2, when the structure category is set to eight categories, the performance of the protein structure prediction method proposed in the present application is also higher than that of the Real method and the Bagging method.

[0179] Figure 8 The bar chart for comparing the protein structure prediction method proposed in the present application with the Bagging method on all key data. As shown in Table 3, PSSM-Distill is the protein structure prediction method proposed in the present application, and the prediction accuracy of the PSSM-Distill method and the Bagging method can be compared respectively in the interval containing different MSA Count and MSA Meff. As can be seen from Table 3, the prediction accuracy of the PSSM-Distill method is much higher than that of the Bagging method. Figure 8 Figure 9

[0180] In other embodiments, the BERT, MSE and CL in the training method of the protein structure prediction model proposed in the present application can be stripped respectively, and the performance of the protein structure prediction method proposed in the present application and the protein structure prediction method obtained by stripping BERT, MSE and CL respectively can be compared. The comparison results can be as shown in Table 3:

[0181] Table 3

[0182]

[0183] ​​As shown in Table 3, Our is the protein structure prediction method proposed in the present application, w / o BERT is the protein prediction method obtained after stripping the BERT part, w / o CL is the protein prediction method obtained after stripping the CL part, and w / o MSE is the protein prediction method obtained after stripping the MSE part. As can be seen from Table 3, the performance of the protein structure prediction method proposed in the present application is higher than that of the protein structure prediction methods after stripping the BERT, CL and MSE parts.

[0184] With Figure 2 The training method of the protein structure prediction model shown in the present application is based on the same inventive concept, and the present application also provides a training device of a protein structure prediction model, which can be arranged in a server or a terminal device. Since the device is the device corresponding to the protein structure prediction model training method of the present application, and the principle of solving the problem of the device is similar to that of the method, the implementation of the device can be referred to the implementation of the above method, and the repeated parts will not be repeated.

[0185] Figure 10 The structure of the training device of the protein structure prediction model provided in the present application is shown in the structure diagram of the training device of the protein structure prediction model provided in the present application, as shown in Figure 10 The training device of the protein structure prediction model includes a data acquisition unit 1001 and a model training unit 1002.

[0186] The data acquisition unit 1001 is configured to acquire a training data set containing a plurality of amino acid sequence samples; the first feature matrix of each amino acid sequence sample is determined according to the associated homologous sequence set, and the number of homologous sequences in the homologous sequence set is not less than a set threshold; and the second feature matrix of each amino acid sequence sample is determined according to the down-sampling result of the associated homologous sequence set.

[0187] The model training unit 1002 is configured to iteratively train the feature enhancement network and the structure prediction network contained in the protein structure prediction model based on the training data set until a set convergence condition is met, wherein one iteration training process includes: based on the amino acid sequence sample and the corresponding second feature matrix extracted from the training data set, obtaining the corresponding enhanced sample feature matrix through the feature enhancement network, and determining the first loss value according to the enhanced sample feature matrix and the first feature matrix of the amino acid sequence sample; the enhanced sample feature matrix and the amino acid sequence sample are spliced and input into the structure prediction network and the classification network for auxiliary training, respectively, and based on the obtained output result, the second loss value is determined; and the feature enhancement network and the structure prediction network are adjusted in parameters according to the first loss value and the second loss value.

[0188] In an optional embodiment, the data acquisition unit 1001 is specifically configured to:

[0189] Based on the amino acid sequence sample, find a plurality of homologous sequences corresponding to the amino acid sequence sample from an amino acid sequence database, and group the plurality of homologous sequences into a homologous sequence set associated with the amino acid sequence sample;

[0190] Downsample the homologous sequence set associated with the amino acid sequence sample to obtain a downsampled result of the homologous sequence set;

[0191] According to the downsampled result of the homologous sequence set, determine a second feature matrix of the amino acid sequence sample.

[0192] In an optional embodiment, the model training unit 1002 is specifically configured to:

[0193] Input the amino acid sequence sample into a matrix generation network to obtain an initial sample feature matrix of the amino acid sequence sample; the matrix generation network is trained based on a natural language processing model;

[0194] Concatenate the amino acid sequence sample, the second feature matrix of the amino acid sequence sample, and the initial sample feature matrix to obtain a first protein information sample matrix;

[0195] Input the first protein information sample matrix into a feature enhancement network to obtain an enhanced sample feature matrix corresponding to the amino acid sequence sample.

[0196] In an optional embodiment, the model training unit 1002 is further configured to:

[0197] Concatenate the enhanced sample feature matrix and the amino acid sequence sample to obtain a second protein information sample matrix;

[0198] Input the second protein information sample matrix into a structure prediction network to obtain structure prediction data corresponding to the amino acid sequence sample;

[0199] Input the second protein information sample matrix into a classification network to obtain structure prediction reference data corresponding to the amino acid sequence sample;

[0200] Compare the structure prediction data with the structure prediction reference data to determine a second loss value.

[0201] In an optional embodiment, the model training unit 1002 is further configured to:

[0202] Concatenate the amino acid sequence sample and the first feature matrix of the amino acid sequence sample to obtain a first auxiliary training information matrix, and input the first auxiliary training information matrix into the classification network to obtain first auxiliary structure prediction data corresponding to the amino acid sequence sample;

[0203] The amino acid sequence sample and the second characterization matrix of the amino acid sequence sample are spliced to obtain a second auxiliary training information matrix, and the second auxiliary training information matrix is input into the classification network to obtain second auxiliary structure prediction data corresponding to the amino acid sequence sample;

[0204] The amino acid sequence sample and the initial sample characterization matrix of the amino acid sequence sample are spliced to obtain a third auxiliary training information matrix, and the third auxiliary training information matrix is input into the classification network to obtain third auxiliary structure prediction data corresponding to the amino acid sequence sample;

[0205] Based on the structure prediction reference data, the first auxiliary structure prediction data, the second auxiliary structure prediction data and the third auxiliary structure prediction data, a third loss value is determined;

[0206] According to the first loss value, the second loss value and the third loss value, the parameter adjustment is performed on the feature enhancement network and the structure prediction network.

[0207] In an optional embodiment, the model training unit 1002 is further configured to:

[0208] According to the weighted sum of the first loss value, the second loss value and the third loss value, a joint loss value is determined;

[0209] According to the joint loss value, the parameter adjustment is performed on the feature enhancement network and the structure prediction network.

[0210] With Figure 5 The protein structure prediction method shown in FIG. 1 is based on the same inventive concept. In the embodiments of the present application, a protein structure prediction device is also provided, which can be arranged in a server or a terminal device. Since the device is the device corresponding to the protein structure prediction method of the present application, and the principle of solving problems of the device is similar to that of the method, the implementation of the device can be referred to the implementation of the above-mentioned method, and the repeated parts will not be described here.

[0211] Figure 11 A structural schematic diagram of a protein structure prediction device provided by an embodiment of the present application is shown in FIG. 11, which includes a category determination unit 1101 and a structure determination unit 1102, as shown in FIG. 11. Figure 11

[0212] The category determination unit 1101 is configured to input a protein information matrix corresponding to a to-be-processed protein into a trained protein structure prediction model to obtain a set structure category to which each amino acid in the to-be-processed protein belongs; the protein structure prediction model is obtained by any one of the training methods of the protein structure prediction model; and the protein information matrix is obtained based on an amino acid sequence of the to-be-processed protein; ​

[0213] a structure determination unit configured to determine the structure of the to-be-processed protein according to the set structure category to which each amino acid belongs.

[0214] In an optional embodiment, as shown in Figure 12 The protein structure prediction apparatus can further include a matrix generation unit 1201 configured to:

[0215] input the amino acid sequence into a matrix generation network, perform homology information representation on the amino acid sequence based on the matrix generation network, and obtain an initial homology information representation matrix of the to-be-processed protein; the matrix generation network is trained based on a natural language processing model;

[0216] splice the amino acid sequence and the initial homology information representation matrix to obtain a protein information matrix corresponding to the to-be-processed protein.

[0217] In an optional embodiment, the category determination unit 1101 is specifically configured to:

[0218] input the protein information matrix corresponding to the to-be-processed protein into a feature enhancement network of the protein structure prediction model, and determine an enhanced homology information representation matrix corresponding to the to-be-processed protein based on the feature enhancement network;

[0219] splice the amino acid sequence and the enhanced homology information representation matrix to obtain an enhanced protein information matrix;

[0220] input the enhanced protein information matrix into a structure prediction network of the protein structure prediction model, and determine the set structure category to which each amino acid in the to-be-processed protein belongs based on the structure prediction network.

[0221] In an optional embodiment, the feature enhancement network includes a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer, and at least one fully connected layer; and the category determination unit 1101 is further configured to:

[0222] input the protein information matrix into the convolutional subnetwork, and obtain a basic feature of the protein information matrix based on the convolutional subnetwork;

[0223] input the protein information matrix into the long short-term memory subnetwork, and obtain a time sequence feature of the protein information matrix based on the long short-term memory subnetwork;

[0224] input the basic feature and the time sequence feature into the splicing layer, splice the basic feature and the time sequence feature based on the splicing layer, and obtain a synthetic feature;

[0225] input the synthetic feature into the fully connected layer, and obtain the enhanced homology information representation matrix corresponding to the to-be-processed protein based on the fully connected layer.

[0226] In an optional embodiment, the structure prediction network comprises a convolutional subnetwork, a long short-term memory subnetwork, a concatenation layer, and a multi-classification layer; the category determination unit 1101 is further configured to:

[0227] input the enhanced protein information matrix into the convolutional subnetwork, and obtain base features of the enhanced protein information matrix based on the convolutional subnetwork;

[0228] input the enhanced protein information matrix into the long short-term memory subnetwork, and obtain time sequence features of the enhanced protein information matrix based on the long short-term memory subnetwork;

[0229] input the base features and the time sequence features into the concatenation layer, and concatenate the base features and the time sequence features based on the concatenation layer to obtain synthetic features;

[0230] input the synthetic features into the multi-classification layer, and obtain the set structure category to which each amino acid in the to-be-processed protein belongs based on the multi-classification layer.

[0231] Based on the same inventive concept as the method embodiments and the device embodiments, an electronic device is further provided in the embodiments of the present application. The electronic device can be a server, such as the server 100 shown in Figure 1 In the embodiments, the structure of the electronic device can be as shown in Figure 13 The electronic device comprises a memory 1301, a communication module 1303, and one or more processors 1302.

[0232] The memory 1301 is configured to store computer programs executed by the processor 1302. The memory 1301 can mainly comprise a program storage area and a data storage area, wherein the program storage area can store an operating system and programs required for running instant messaging functions, etc.; and the data storage area can store various instant messaging information and operation instruction sets, etc.

[0233] The memory 1301 can be a volatile memory, such as a random-access memory (RAM); the memory 1301 can also be a non-volatile memory, such as a read-only memory, a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD), or the memory 1301 can be any other medium capable of carrying or storing desired program codes in the form of instructions or data structures and capable of being accessed by a computer, but is not limited to this. The memory 1301 can be a combination of the above memories.

[0234] The processor 1302 can include one or more central processing units (CPUs), or digital processing units, etc. The processor 1302 is configured to invoke a computer program stored in the memory 1301 to implement the training method of the protein structure prediction model and the protein structure prediction method.

[0235] The communication module 1303 is configured to communicate with a terminal device and other electronic devices. If the electronic device is a server, the server can receive an amino acid sequence sent by a terminal device through the communication module 1303.

[0236] The specific connection medium between the memory 1301, the communication module 1303, and the processor 1302 is not limited in the embodiments of the present application. In the embodiments of the present disclosure, Figure 13 the memory 1301 and the processor 1302 are connected through a bus 1304, and the bus 1304 is Figure 13 represented by a thick line in the embodiments of the present disclosure, and the connection mode between other components is only schematically illustrated and is not limited. The bus 1304 can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 13 only one thick line is used in the embodiments of the present disclosure, but it does not mean that there is only one bus or only one type of bus.

[0237] In another embodiment, the electronic device can also be a tablet computer, a desktop computer, a notebook computer, etc.

[0238] According to an aspect of the present application, a computer program product or computer program is provided, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to enable the computer device to perform the training method of the protein structure prediction model and the protein structure prediction method in the above embodiments. The program product can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium may, for example, be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or apparatus, or any combination of the above. More specific examples (non-exhaustive list) of the readable storage medium include an electrical connection having one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0239] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A training method for a protein structure prediction model, characterized in that, The method comprises: obtaining a training data set comprising a plurality of amino acid sequence samples; a first feature matrix of each amino acid sequence sample is determined according to an associated homologous sequence set, and the number of homologous sequences in the homologous sequence set is not less than a set threshold; a second feature matrix of each amino acid sequence sample is determined according to a down-sampling result of the associated homologous sequence set; based on the training data set, iteratively training a feature enhancement network and a structure prediction network included in the protein structure prediction model until a set convergence condition is met, wherein one iteration training process comprises: based on the amino acid sequence sample and the corresponding second feature matrix extracted from the training data set, obtaining a corresponding enhanced sample feature matrix through the feature enhancement network, and determining a first loss value according to the enhanced sample feature matrix and the first feature matrix of the amino acid sequence sample; after the enhanced sample feature matrix and the amino acid sequence sample are spliced, they are respectively input into the structure prediction network and a classification network used for auxiliary training, and a second loss value is determined based on the obtained output result; according to the first loss value and the second loss value, respectively adjusting the parameters of the feature enhancement network and the structure prediction network.

2. The method of claim 1, wherein, The second feature matrix of the amino acid sequence sample is obtained by the following way: based on the amino acid sequence sample, searching for a plurality of homologous sequences corresponding to the amino acid sequence sample from an amino acid sequence database, and grouping the plurality of homologous sequences into a homologous sequence set associated with the amino acid sequence sample; down-sampling the homologous sequence set associated with the amino acid sequence sample to obtain a down-sampling result of the homologous sequence set; determining the second feature matrix of the amino acid sequence sample according to the down-sampling result of the homologous sequence set.

3. The method of claim 1, wherein, The method comprises: obtaining a training data set comprising a plurality of amino acid sequence samples; a first feature matrix of each amino acid sequence sample is determined according to an associated homologous sequence set, and the number of homologous sequences in the homologous sequence set is not less than a set threshold; a second feature matrix of each amino acid sequence sample is determined according to a down-sampling result of the associated homologous sequence set; based on the training data set, iteratively training a feature enhancement network and a structure prediction network included in the protein structure prediction model until a set convergence condition is met, wherein one iteration training process comprises: based on the amino acid sequence sample and the corresponding second feature matrix extracted from the training data set, obtaining a corresponding enhanced sample feature matrix through the feature enhancement network, and determining a first loss value according to the enhanced sample feature matrix and the first feature matrix of the amino acid sequence sample; 4. The method of claim 3, wherein, after the enhanced sample feature matrix and the amino acid sequence sample are spliced, they are respectively input into the structure prediction network and a classification network used for auxiliary training, and a second loss value is determined based on the obtained output result; according to the first loss value and the second loss value, respectively adjusting the parameters of the feature enhancement network and the structure prediction network. The second feature matrix of the amino acid sequence sample is obtained by the following way: based on the amino acid sequence sample, searching for a plurality of homologous sequences corresponding to the amino acid sequence sample from an amino acid sequence database, and grouping the plurality of homologous sequences into a homologous sequence set associated with the amino acid sequence sample; down-sampling the homologous sequence set associated with the amino acid sequence sample to obtain a down-sampling result of the homologous sequence set; determining the second feature matrix of the amino acid sequence sample according to the down-sampling result of the homologous sequence set. The method comprises: obtaining a training data set comprising a plurality of amino acid sequence samples; a first feature matrix of each amino acid sequence sample is determined according to an associated homologous sequence set, and the number of homologous sequences in the homologous sequence set is not less than a set threshold; a second feature matrix of each amino acid sequence sample is determined according to a down-sampling result of the associated homologous sequence set; based on the training data set, iteratively training a feature enhancement network and a structure prediction network included in the protein structure prediction model until a set convergence condition is met, wherein one iteration training process comprises: based on the amino acid sequence sample and the corresponding second feature matrix extracted from the training data set, obtaining a corresponding enhanced sample feature matrix through the feature enhancement network, and determining a first loss value according to the enhanced sample feature matrix and the first feature matrix of the amino acid sequence sample; after the enhanced sample feature matrix and the amino acid sequence sample are spliced, they are respectively input into the structure prediction network and a classification network used for auxiliary training, and a second loss value is determined based on the obtained output result; according to the first loss value and the second loss value, respectively adjusting the parameters of the feature enhancement network and the structure prediction network. The second feature matrix of the amino acid sequence sample is obtained by the following way: based on the amino acid sequence sample, searching for a plurality of homologous sequences corresponding to the amino acid sequence sample from an amino acid sequence database, and grouping the plurality of homologous sequences into a homologous sequence set associated with the amino acid sequence sample; down-sampling the homologous sequence set associated with the amino acid sequence sample to obtain a down-sampling result of the homologous sequence set; determining the second feature matrix of the amino acid sequence sample according to the down-sampling result of the homologous sequence set. The method comprises: obtaining a training data set comprising a plurality of amino acid sequence samples; a first feature matrix of each amino acid sequence sample is determined according to an associated homologous sequence set, and the number of homologous sequences in the homologous sequence set is not less than a set threshold; a second feature matrix of each amino acid sequence sample is determined according to a down-sampling result of the associated homologous sequence set; based on the training data set, iteratively training a feature enhancement network and a structure prediction network included in the protein structure prediction model until a set convergence condition is met, wherein one iteration training process comprises: based on the amino acid sequence sample and the corresponding second feature matrix extracted from the training data set, obtaining a corresponding enhanced sample feature matrix through the feature enhancement network, and determining a first loss value according to the enhanced sample feature matrix and the first feature matrix of the amino acid sequence sample; after the enhanced sample feature matrix and the amino acid sequence sample are spliced, they are respectively input into the structure prediction network and a classification network used for auxiliary training, and a second loss value is determined based on the obtained output result; according to the first loss value and the second loss value, respectively adjusting the parameters of the feature enhancement network and the structure prediction network. inputting the second protein information sample matrix into the classification network to obtain structure prediction reference data corresponding to the amino acid sequence sample; comparing the structure prediction data with the structure prediction reference data to determine the second loss value.

5. The method of claim 4, wherein, The parameter adjustment of the feature enhancement network and the structure prediction network according to the first loss value and the second loss value comprises: splicing the amino acid sequence sample and the first representation matrix of the amino acid sequence sample to obtain a first auxiliary training information matrix, and inputting the first auxiliary training information matrix into the classification network to obtain first auxiliary structure prediction data corresponding to the amino acid sequence sample; splicing the amino acid sequence sample and the second representation matrix of the amino acid sequence sample to obtain a second auxiliary training information matrix, and inputting the second auxiliary training information matrix into the classification network to obtain second auxiliary structure prediction data corresponding to the amino acid sequence sample; splicing the amino acid sequence sample and the initial sample representation matrix of the amino acid sequence sample to obtain a third auxiliary training information matrix, and inputting the third auxiliary training information matrix into the classification network to obtain third auxiliary structure prediction data corresponding to the amino acid sequence sample; determining a third loss value based on the structure prediction reference data, the first auxiliary structure prediction data, the second auxiliary structure prediction data and the third auxiliary structure prediction data; adjusting the parameters of the feature enhancement network and the structure prediction network according to the first loss value, the second loss value and the third loss value.

6. The method of claim 5, wherein, The parameter adjustment of the feature enhancement network and the structure prediction network according to the first loss value, the second loss value and the third loss value comprises: determining a joint loss value according to the weighted sum of the first loss value, the second loss value and the third loss value; adjusting the parameters of the feature enhancement network and the structure prediction network according to the joint loss value.

7. A method of protein structure prediction, characterized by, The method comprises: inputting a protein information matrix corresponding to a to-be-processed protein into a trained protein structure prediction model to obtain a set structure category to which each amino acid in the to-be-processed protein belongs; the protein structure prediction model is obtained by the training method in any one of claims 1-6; the protein information matrix is obtained based on an amino acid sequence of the to-be-processed protein; determining the structure of the to-be-processed protein according to the set structure category to which each amino acid belongs.

8. The prediction method of claim 7, wherein, Before inputting the protein information matrix corresponding to the to-be-processed protein into the trained protein structure prediction model, the prediction method further comprises: inputting the amino acid sequence into a matrix generation network, and performing homologous information representation on the amino acid sequence based on the matrix generation network to obtain an initial homologous information representation matrix of the to-be-processed protein; the matrix generation network is trained based on a natural language processing model; The amino acid sequence and the initial homologous information representation matrix are spliced to obtain a protein information matrix corresponding to the to-be-processed protein.

9. The prediction method of claim 7, wherein, The protein information matrix corresponding to the to-be-processed protein is input into the feature enhancement network of the protein structure prediction model, and an enhanced homologous information representation matrix corresponding to the to-be-processed protein is determined based on the feature enhancement network. The amino acid sequence and the enhanced homologous information representation matrix are spliced to obtain an enhanced protein information matrix. The enhanced protein information matrix is input into the structure prediction network of the protein structure prediction model, and the set structure category to which each amino acid in the to-be-processed protein belongs is determined based on the structure prediction network. The feature enhancement network comprises a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer, and at least one fully connected layer.

10. The prediction method of claim 9, wherein, The protein information matrix is input into the convolutional subnetwork, and the basic features of the protein information matrix are obtained based on the convolutional subnetwork. The protein information matrix is input into the long short-term memory subnetwork, and the time sequence features of the protein information matrix are obtained based on the long short-term memory subnetwork. The basic features and the time sequence features are input into the splicing layer, and the basic features and the time sequence features are spliced based on the splicing layer to obtain synthetic features. The synthetic features are input into the fully connected layer, and the enhanced homologous information representation matrix corresponding to the to-be-processed protein is obtained based on the fully connected layer. The structure prediction network comprises a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer, and a multi-classification layer. The enhanced protein information matrix is input into the convolutional subnetwork, and the basic features of the enhanced protein information matrix are obtained based on the convolutional subnetwork.

11. The prediction method of claim 9, wherein, The enhanced protein information matrix is input into the long short-term memory subnetwork, and the time sequence features of the enhanced protein information matrix are obtained based on the long short-term memory subnetwork. The basic features and the time sequence features are input into the splicing layer, and the basic features and the time sequence features are spliced based on the splicing layer to obtain synthetic features. The synthetic features are input into the multi-classification layer, and the set structure category to which each amino acid in the to-be-processed protein belongs is obtained based on the multi-classification layer. The feature enhancement network comprises a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer, and at least one fully connected layer. The protein information matrix is input into the convolutional subnetwork, and the basic features of the protein information matrix are obtained based on the convolutional subnetwork. The protein information matrix is input into the long short-term memory subnetwork, and the time sequence features of the protein information matrix are obtained based on the long short-term memory subnetwork.

12. A training device for a protein structure prediction model, characterized in that, The basic features and the time sequence features are input into the splicing layer, and the basic features and the time sequence features are spliced based on the splicing layer to obtain synthetic features. The synthetic features are input into the fully connected layer, and the enhanced homologous information representation matrix corresponding to the to-be-processed protein is obtained based on the fully connected layer. The structure prediction network comprises a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer, and a multi-classification layer. The enhanced protein information matrix is input into the convolutional subnetwork, and the basic features of the enhanced protein information matrix are obtained based on the convolutional subnetwork. The enhanced protein information matrix is input into the long short-term memory subnetwork, and the time sequence features of the enhanced protein information matrix are obtained based on the long short-term memory subnetwork. The basic features and the time sequence features are input into the splicing layer, and the basic features and the time sequence features are spliced based on the splicing layer to obtain synthetic features. The synthetic features are input into the multi-classification layer, and the set structure category to which each amino acid in the to-be-processed protein belongs is obtained based on the multi-classification layer. The feature enhancement network comprises a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer, and at least one fully connected layer. The protein information matrix is input into the convolutional subnetwork, and the basic features of the protein information matrix are obtained based on the convolutional subnetwork. The protein information matrix is input into the long short-term memory subnetwork, and the time sequence features of the protein information matrix are obtained based on the long short-term memory subnetwork. The basic features and the time sequence features are input into the splicing layer, and the basic features and the time sequence features are spliced based on the splicing layer to obtain synthetic features. The synthetic features are input into the fully connected layer, and the enhanced homologous information representation matrix corresponding to the to-be-processed protein is obtained based on the fully connected layer. The structure prediction network comprises a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer, and a multi-classification layer. The enhanced protein information matrix is input into the convolutional subnetwork, and the basic features of the enhanced protein information matrix are obtained based on the convolutional subnetwork. The enhanced protein information matrix is input into the long short-term memory subnetwork, and the time sequence features of the enhanced protein information matrix are obtained based on the long short-term memory subnetwork. The basic features and the time sequence features are input into the splicing layer, and the basic features and the time sequence features are spliced based on the splicing layer to obtain synthetic features. The synthetic features are input into the multi-classification layer, and the set structure category to which each amino acid in the to-be-processed protein belongs is obtained based on the multi-classification layer. The feature enhancement network comprises a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer, and at least one fully connected layer. The protein information matrix is input into the convolutional subnetwork, and the basic features of the protein information matrix are obtained based on the convolutional subnetwork. The protein information matrix is input into the long short-term memory subnetwork, and the time sequence features of the protein information matrix are obtained based on the long short-term memory subnetwork. The basic features and the time sequence features are input into the splicing layer, and the basic features and the time sequence features are spliced based on the splicing layer to obtain synthetic features. The synthetic features are input into the fully connected layer, and the enhanced homologous information representation matrix corresponding to the to-be-processed protein is obtained based on the fully connected layer. The structure prediction network comprises a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer, and a multi-classification layer. The enhanced protein information matrix is input into the convolutional subnetwork, and the basic features of the enhanced protein information matrix are obtained based on the convolutional subnetwork. The enhanced protein information matrix is input into the long short-term memory subnetwork, and the time sequence features of the enhanced protein information matrix are obtained based on the long short-term memory subnetwork. The basic features and the time sequence features are input into the splicing layer, and the basic features and the time sequence features are spliced based on the splicing layer to obtain synthetic features. The synthetic features are input into the multi-classification layer, and the set structure category to which each amino acid in the to-be-processed protein belongs is obtained based on the multi-classification layer. The feature enhancement network comprises a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer, and at least one fully connected layer. The protein information matrix is input into the convolutional subnetwork, and the basic features of the protein information matrix are obtained based on the convolutional subnetwork. The protein information matrix is input into the long short-term memory subnetwork, and the time sequence features of the protein information matrix are obtained based on the long short-term memory subnetwork. The basic features and the time sequence features are input into the splicing layer, and the basic features and the time sequence features are spliced based on the splicing layer to obtain synthetic features. The synthetic features are input into the fully connected layer, and the enhanced homologous information representation matrix corresponding to the to-be-processed protein is obtained based on the fully connected layer. The structure prediction network comprises a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer, and a multi-classification layer. The enhanced protein information matrix is input into the convolutional subnetwork, and the basic features of the enhanced protein information matrix are obtained based on the convolutional subnetwork. The enhanced protein information matrix is input into the long short-term memory subnetwork, and the time sequence features of the enhanced protein information matrix are obtained based on the long short-term memory subnetwork. The basic features and the time sequence features are input into the splicing layer, and the basic features and the time sequence features are spliced based on the splicing layer to obtain synthetic features. The synthetic features are input into the multi-classification layer, and the set structure category to which each amino acid in the to-be-processed protein belongs is obtained based on the multi-classification layer. The feature enhancement network comprises a convolutional subnetwork, a long short-term memory subnetwork, a splicing layer, and at least one fully connected layer. The protein information matrix is input into the convolutional subnetwork, and the basic features of the protein information matrix are obtained based on the convolutional subnetwork. The protein information matrix is input into the long short-term memory subnetwork, and the time sequence features of the protein information matrix are obtained based on the long The data acquisition unit is configured to acquire a training data set containing a plurality of amino acid sequence samples; a first feature matrix of each amino acid sequence sample is determined according to an associated homologous sequence set, and the number of homologous sequences in the homologous sequence set is not less than a set threshold; and a second feature matrix of each amino acid sequence sample is determined according to a down-sampling result of the associated homologous sequence set; The model training unit is configured to iteratively train a feature enhancement network and a structure prediction network included in the protein structure prediction model based on the training data set until a set convergence condition is met, wherein one iteration training process includes: obtaining a corresponding enhanced sample feature matrix by the feature enhancement network based on an amino acid sequence sample and a corresponding second feature matrix extracted from the training data set, and determining a first loss value according to the enhanced sample feature matrix and a first feature matrix of the amino acid sequence sample; inputting the enhanced sample feature matrix and the amino acid sequence sample after splicing into the structure prediction network and a classification network used for auxiliary training respectively, and determining a second loss value based on obtained output results; and adjusting parameters of the feature enhancement network and the structure prediction network respectively according to the first loss value and the second loss value.

13. A protein structure prediction apparatus characterized by comprising: The method comprises: The category determination unit is configured to input a protein information matrix corresponding to a to-be-processed protein into a trained protein structure prediction model to obtain a set structure category to which each amino acid in the to-be-processed protein belongs; The protein structure prediction model is obtained by the training method in any one of claims 1 to 6; and the protein information matrix is obtained based on an amino acid sequence of the to-be-processed protein; The structure determination unit is configured to determine a structure of the to-be-processed protein according to the set structure category to which each amino acid belongs.

14. A computer readable storage medium having stored therein a computer program, characterized in that: The computer program is executed by the processor to implement the method in any one of claims 1 to 6 or 7 to 11.

15. An electronic device, comprising: The computer program is executed by the processor to implement the method in any one of claims 1 to 6 or 7 to 11. The computer program is executed by the processor to implement the method in any one of claims 1 to 6 or 7 to 11.

Images