Protein three-dimensional structure distribution generation method and device and electronic equipment

By acquiring the target protein sequence and its multiple sequence comparison data, and using the Alpha folding and diffusion models to generate the three-dimensional structural distribution of the protein, the problem of low efficiency in molecular dynamics simulation is solved, and efficient generation of protein three-dimensional structural distribution and capture of dynamic change features are achieved.

CN121171320APending Publication Date: 2025-12-19SHANGHAI SHUHUI BIOLOGICAL CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510678312.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2024-06-17
Filing Date
2025-05-23
Publication Date
2025-12-19

AI Technical Summary

Technical Problem

Existing technologies that use molecular dynamics simulations to generate protein three-dimensional structure distributions are inefficient and struggle to effectively capture the dynamic changes in protein three-dimensional structures.

Method used

By acquiring the target protein sequence and its multiple sequence alignment data, the Alpha folding model is used to extract residue embedding parameters and residue pair embedding parameters. Combined with the diffusion model, multiple predictions are made to generate the three-dimensional structural distribution of the protein, thus avoiding complex molecular dynamics simulations.

Benefits of technology

It achieves efficient generation of protein three-dimensional structure distribution, simplifies the generation process, improves generation efficiency, and can accurately capture various three-dimensional structural changes of proteins under different conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121171320A_ABST
    Figure CN121171320A_ABST
Patent Text Reader

Abstract

The invention discloses a protein three-dimensional structure distribution generation method and device and electronic equipment. Relates to the technical field of biological information. The method comprises the following steps: acquiring a target protein sequence and corresponding multi-sequence comparison data; obtaining a target residue embedding parameter and a target residue pair embedding parameter based on the target protein sequence and the corresponding multi-sequence comparison data; based on the multiple groups of initial random main chain structure parameters, the target residue embedding parameters and the target residue pair embedding parameters corresponding to the target protein sequence, adopting a diffusion model to carry out multiple times of prediction to obtain multiple groups of predicted main chain structure parameters corresponding to the target protein sequence; and generating three-dimensional structure distribution of the target protein sequence based on the multiple groups of predicted main chain structure parameters. According to the invention, the technical problem of low protein three-dimensional structure distribution generation efficiency of a method for generating protein three-dimensional structure distribution based on a molecular dynamics simulation mode in the prior art is solved.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Cross-references to related applications

[0002] This application claims priority to Chinese Patent Application No. 202410784309.8, filed on June 17, 2024, entitled “Method, Apparatus and Electronic Equipment for Generating Three-Dimensional Structure Distribution of Proteins”, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This invention relates to the field of bioinformatics, and more specifically, to a method, apparatus, and electronic device for generating the three-dimensional structure distribution of proteins. Background Technology

[0004] Artificial intelligence methods in related technologies have largely solved the major challenge of predicting protein crystal structures using artificial intelligence. However, in the real world, the three-dimensional structure of proteins is dynamically changing. Therefore, exploring the dynamic characteristics of protein three-dimensional structure changes from the perspective of these changes becomes particularly important. Obtaining the dynamic three-dimensional structure distribution of proteins in related technologies mainly relies on long-term molecular dynamics simulations. However, long-term molecular dynamics simulations are inefficient and slow, resulting in low efficiency in generating protein three-dimensional structure distributions.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This invention provides a method, apparatus, and electronic device for generating the three-dimensional structure distribution of proteins, thereby at least solving the technical problem of low efficiency in generating the three-dimensional structure distribution of proteins based on molecular dynamics simulation in related technologies.

[0007] According to an aspect of some embodiments of the present application, there is provided a method for generating a protein three-dimensional structure distribution, comprising: obtaining a target protein sequence, and a plurality of sequence comparison data corresponding to the target protein sequence, wherein the plurality of sequence comparison data is used to indicate comparison results between the target protein sequence and a plurality of related proteins corresponding thereto; obtaining a target residue embedding parameter corresponding to the target protein sequence and a target residue pair embedding parameter based on the target protein sequence and the plurality of sequence comparison data corresponding to the target protein sequence, wherein the target residue embedding parameter is used to describe characteristics of a plurality of amino acid residues included in the target protein sequence, and the target residue pair embedding parameter is used to describe interaction and distance relationship between the plurality of amino acid residues under influence of the plurality of sequence comparison data corresponding to the target protein sequence; determining a plurality of groups of initial random main chain structure parameters corresponding to the target protein sequence, wherein the initial random main chain structure parameters are used to describe spatial arrangement of each residue in a main chain structure of the target protein sequence, and the initial random main chain structure parameters are randomly generated; performing a plurality of predictions based on the plurality of groups of initial random main chain structure parameters, the target residue embedding parameter, and the target residue pair embedding parameter by using a diffusion model to obtain a plurality of groups of predicted main chain structure parameters corresponding to the target protein sequence, wherein the plurality of groups of initial random main chain structure parameters and the plurality of groups of predicted main chain structure parameters correspond to each other, and the diffusion model is pre-learned to associate residue embedding parameters and residue pair embedding parameters of a corresponding protein sequence with changes in protein three-dimensional structure; and generating a three-dimensional structure distribution of the target protein sequence based on the plurality of groups of predicted main chain structure parameters, wherein the three-dimensional structure distribution is used to indicate a plurality of different three-dimensional structures in which the target protein sequence exists.

[0008] According to a further aspect of the embodiments of the present application, there is also provided a device for generating a protein three-dimensional structure distribution, comprising: a data obtaining module configured to obtain a target protein sequence and a plurality of sequence comparison data corresponding to the target protein sequence, wherein the plurality of sequence comparison data is configured to indicate comparison results between the target protein sequence and a plurality of related proteins corresponding to the target protein sequence; a vector generating module configured to obtain a target residue embedding parameter corresponding to the target protein sequence and a target residue pair embedding parameter based on the target protein sequence and the plurality of sequence comparison data corresponding to the target protein sequence, wherein the target residue embedding parameter is configured to describe characteristics of a plurality of amino acid residues included in the target protein sequence, and the target residue pair embedding parameter is configured to describe interactions and distance relationships between the plurality of amino acid residues under an influence of the plurality of sequence comparison data corresponding to the target protein sequence; an initial parameter determining module configured to determine a plurality of sets of initial random main chain structure parameters corresponding to the target protein sequence, wherein the initial random main chain structure parameters are configured to describe spatial arrangements of respective residues in a main chain structure of the target protein sequence, and the initial random main chain structure parameters are randomly generated; a parameter predicting module configured to obtain a plurality of sets of predicted main chain structure parameters corresponding to the target protein sequence by performing a plurality of predictions based on the plurality of sets of initial random main chain structure parameters, the target residue embedding parameter and the target residue pair embedding parameter using a diffusion model, wherein the plurality of sets of initial random main chain structure parameters and the plurality of sets of predicted main chain structure parameters correspond to each other, and the diffusion model is pre-learned to associate the residue embedding parameter and the residue pair embedding parameter of a corresponding protein sequence with a change in a three-dimensional structure of the protein; and a three-dimensional structure distribution generating module configured to generate a three-dimensional structure distribution of the target protein sequence based on the plurality of sets of predicted main chain structure parameters, wherein the three-dimensional structure distribution is configured to indicate a plurality of different three-dimensional structures in which the target protein sequence exists.

[0009] According to a further aspect of the embodiments of the present application, there is also provided a non-transitory storage medium having a plurality of instructions stored therein, the instructions being adapted to be loaded and executed by a processor to implement any of the methods for generating a protein three-dimensional structure distribution.

[0010] According to a further aspect of the embodiments of the present application, there is also provided an electronic device comprising one or more processors and a memory, the memory being configured to store one or more programs, wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement any of the methods for generating a protein three-dimensional structure distribution.

[0011] In the embodiment of the present application, by acquiring a target protein sequence and multi-sequence comparison data corresponding to the target protein sequence, wherein the multi-sequence comparison data is used to indicate comparison results between the protein sequence and a plurality of related proteins corresponding thereto; based on the target protein sequence and the multi-sequence comparison data corresponding to the target protein sequence, target residue embedding parameters corresponding to the target protein sequence and target residue pair embedding parameters are obtained, wherein the target residue embedding parameters are used to describe characteristics of a plurality of amino acid residues included in the target protein sequence, and the target residue pair embedding parameters are used to describe interactions and distance relationships between the plurality of amino acid residues under the influence of the multi-sequence comparison data corresponding to the target protein sequence; a plurality of groups of initial random main chain structure parameters corresponding to the target protein sequence are determined, wherein the initial random main chain structure parameters are used to describe spatial arrangements of each residue in a main chain structure of the target protein sequence; based on the plurality of groups of initial random main chain structure parameters, the target residue embedding parameters and the target residue pair embedding parameters, a diffusion model is used for multiple times of prediction to obtain a plurality of groups of predicted main chain structure parameters corresponding to the target protein sequence, wherein the plurality of groups of initial random main chain structure parameters and the plurality of groups of predicted main chain structure parameters correspond to each other, and the diffusion model is pre-learned to associate the residue embedding parameters and the residue pair embedding parameters of the corresponding protein sequence with changes in the three-dimensional structure of the protein; and based on the plurality of groups of predicted main chain structure parameters, a three-dimensional structure distribution of the target protein sequence is generated, wherein the three-dimensional structure distribution is used to indicate a plurality of different three-dimensional structures of the target protein sequence, so as to achieve the purpose of obtaining the residue embedding parameters, the residue pair embedding parameters and the randomly generated main chain structure parameters of the target protein, using the diffusion model for multiple times of prediction to obtain a plurality of three-dimensional structures of the protein under different conditions, forming the three-dimensional structure distribution of the protein, and realizing efficient prediction of the three-dimensional structure distribution of the protein without using a complex molecular dynamics simulation method, thereby realizing the technical effects of simplifying the generation process of the three-dimensional structure distribution of the protein and improving the generation efficiency of the three-dimensional structure distribution of the protein, and further solving the technical problem of low generation efficiency of the three-dimensional structure distribution of the protein in the related art based on the molecular dynamics simulation method. BRIEF DESCRIPTION OF DRAWINGS

[0012] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0013] Figure 1 is a flowchart of a generation method of a three-dimensional structure distribution of a protein according to an embodiment of the present application;

[0014] Figure 2is a structure and data interaction schematic diagram of an optional neural network model according to an embodiment of the present application;

[0015] Figure 3 is a structure schematic diagram of an optional single-layer transformer module according to an embodiment of the present application;

[0016] Figure 4 is an optional amino acid fluctuation comparison schematic diagram according to an embodiment of the present application;

[0017] Figure 5 is an optional protein three-dimensional structure comparison diagram according to an embodiment of the present application;

[0018] Figure 6 is an optional Kras protein amino acid displacement correlation matrix comparison schematic diagram according to an embodiment of the present application;

[0019] Figure 7 is a Kras protein amino acid displacement correlation curve comparison diagram according to an embodiment of the present application;

[0020] Figure 8 is a schematic diagram of a protein three-dimensional structure distribution generation device according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to enable persons skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative labor should fall within the scope of protection of the present application.

[0022] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0023] First, for the convenience of understanding the embodiments of the present application, the following will explain and describe some terms or nouns involved in the present application:

[0024] Uniref90 is a component of the UniProt database, which is a comprehensive protein knowledge database containing protein sequence, structure and function information. Uniref90 is formed by clustering protein sequences from the UniProt sequence database based on sequence similarity. Sequences in each cluster have at least 90% similarity.

[0025] The Big Fantastic Database (BFD) is a database that maintains and provides known protein sequence information. The database contains a large amount of protein sequence data, including protein amino acid sequences, structural information, functional annotations, etc.

[0026] The Microbial Genomics and Metagenomics Information System (MGnify) is a comprehensive database and analysis platform focused on metagenomic data, aiming to help researchers analyze and understand the structure and function of metagenomes.

[0027] According to an embodiment of the present application, a method for generating a protein three-dimensional structure distribution is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in a different order.

[0028] Figure 1 The flowchart of the method for generating a protein three-dimensional structure distribution according to an embodiment of the present application is shown in FIG. 1, which includes the following steps: Figure 1

[0029] Step S102, obtaining a target protein sequence and multiple sequence alignment data corresponding to the target protein sequence;

[0030] Optionally, the multiple sequence alignment data is used to indicate the comparison result between the protein sequence and the corresponding multiple related proteins, which can be searched on Uniref90, MGnify, and BFD to obtain the multiple sequence alignment (MSA) data corresponding to the target protein sequence.

[0031] ​In step S104, target residue embedding parameters corresponding to the target protein sequence and target residue pair embedding parameters are obtained based on the target protein sequence and the multiple sequence alignment data corresponding to the target protein sequence.

[0032] Optionally, the target residue embedding parameters are used to describe the characteristics of the plurality of amino acid residues included in the target protein sequence, and the target residue pair embedding parameters are used to describe the interaction and distance relationship between the plurality of amino acid residues under the influence of the multiple sequence alignment data corresponding to the target protein sequence; wherein the target residue embedding parameters and the target residue pair embedding parameters can be in the form of vectors, i.e., based on the target protein sequence and the multiple sequence alignment data corresponding to the target protein sequence, target residue embedding vectors corresponding to the target protein sequence and target residue pair embedding vectors can be obtained.

[0033] In an optional embodiment, based on the target protein sequence and the multiple sequence alignment data corresponding to the target protein sequence, the target residue embedding parameters corresponding to the target protein sequence and the target residue pair embedding parameters are obtained, including: using a pre-trained alphafold model to encode the target protein sequence and the multiple sequence alignment data corresponding to the target protein sequence to obtain the target residue embedding parameters and the target residue pair embedding parameters.

[0034] Optionally, the pre-trained alphafold model (i.e., Alphafold model) can be used to encode the protein sequence and its multiple sequence alignment data to obtain the target residue embedding parameters and the target residue pair embedding parameters as input information of the diffusion model. Specifically, the EVOformer module in the alphafold model can be used to encode the protein sequence and its multiple sequence alignment data to obtain the target residue embedding parameters and the target residue pair embedding parameters as input of the diffusion model. The EVOformer module mainly uses the evolutionary information in the protein sequence to assist in predicting its three-dimensional structure. Based on the evolutionary coupling effect of the EVOformer module, the mutual influence of some positions of amino acid residues in the target protein sequence in the evolution process can be mined, while the stability of the function and structure of the target protein sequence is maintained, and more accurate and reliable target residue embedding parameters and target residue pair embedding parameters are obtained.

[0035] Optionally, the alpha fold model can extract meaningful features from the input target protein sequence and the corresponding multiple sequence alignment data, which are implicitly represented in the embedding vectors. These embedding vectors can be considered as an abstract representation of the input data. It can better capture the key features in the input data and provide a more effective input representation for the model. This representation of embedding vectors can better support the diffusion model to predict and understand the protein structure and properties. The target residue embedding vector is obtained by converting each amino acid residue in the target protein sequence into a high-dimensional vector to represent the characteristics of the residue in the target protein sequence structure. The characteristics of the amino acid residue, such as hydrophobicity, charge, etc. can be captured through the target residue embedding vector. Based on the multiple sequence alignment data corresponding to the target protein sequence, the alpha fold model can predict the interaction and distance between residues. These residue pair embedding vectors represent the relationship between different residues, which helps the subsequent diffusion model better understand the spatial position and interaction between residues in the target protein sequence.

[0036] It should be noted that the alpha fold model is a deep learning model for protein structure prediction, which effectively captures the complex relationship between sequence and structure through deep learning technology to predict the three-dimensional structure of the protein. Based on this, the present embodiment encodes the target protein sequence and its corresponding multiple sequence alignment data based on the alpha fold model to effectively capture the characteristics of the complex relationship between sequence and structure, to better understand the structural characteristics of the protein, and thus accurately predict the three-dimensional structure of the protein.

[0037] In an optional embodiment, when the multiple sequence alignment data corresponding to the target protein sequence is multiple, based on the target protein sequence and the multiple sequence alignment data corresponding to the target protein sequence, the target residue embedding parameter and the target residue pair embedding parameter corresponding to the target protein sequence are obtained, including: clustering and grouping the multiple sequence alignment data corresponding to the target protein sequence to obtain multiple groups of multiple sequence alignment data; based on the target protein sequence and the multiple groups of multiple sequence alignment data, obtaining multiple residue embedding parameters and corresponding multiple residue pair embedding parameters corresponding to the target protein sequence, wherein the number of multiple residue embedding parameters corresponds to the number of groups of multiple sequence alignment data; performing average calculation on the multiple residue embedding parameters to obtain the target residue embedding parameter; performing average calculation on the multiple residue pair embedding parameters to obtain the target residue pair embedding parameter.

[0038] Optionally, at the inference end of the model, the multiple pieces of multiple sequence alignment data corresponding to the target protein sequence can be clustered and grouped, for example, clustered according to 50% similarity, and a plurality of groups of multiple sequence alignment data with the largest number of multiple sequence alignment data (for example, 5) are selected. Each group is used together with the target protein sequence to extract residue embedding parameters (such as residue embedding vectors) and residue pair embedding parameters (such as residue pair embedding vectors) using the alpha fold model, to obtain a plurality of residue embedding parameters and a plurality of corresponding residue pair embedding parameters. After averaging the plurality of obtained residue embedding parameters, a new residue embedding vector is obtained as the target residue embedding parameter, and after averaging the plurality of obtained residue pair embedding parameters, a new residue pair embedding vector is obtained as the target residue pair embedding vector. The above method can expand the distribution of the generated structure of the model for a protein sequence with a wide conformation distribution.

[0039] Step S106, determining a plurality of groups of initial random main chain structure parameters corresponding to the target protein sequence, wherein the initial random main chain structure parameters are used to describe the spatial arrangement of each residue in the main chain structure of the target protein sequence, and the initial random main chain structure parameters are randomly generated;

[0040] Optionally, the plurality of groups of initial random main chain structure parameters corresponding to the target protein sequence can be randomly generated based on a prior distribution such as a normal distribution noise. That is, a plurality of initial random main chain structure parameters are randomly generated based on the prior distribution in advance, as the prediction of the main chain structure parameters corresponding to the target protein sequence based on the diffusion model subsequently.

[0041] Step S108, based on the plurality of groups of initial random main chain structure parameters, the target residue embedding parameters, and the target residue pair embedding parameters, a diffusion model is used to perform multiple predictions to obtain a plurality of groups of predicted main chain structure parameters corresponding to the target protein sequence, wherein the plurality of groups of initial random main chain structure parameters correspond one-to-one to the plurality of groups of predicted main chain structure parameters;

[0042] Optionally, the diffusion model learns the association between the residue embedding parameters and the residue pair embedding parameters corresponding to the protein sequence and the change in the three-dimensional structure of the protein in advance. With the initial random main chain structure parameters randomly generated based on the prior distribution and the target residue embedding parameters and the target residue pair embedding parameters representing the characteristics of the target protein sequence as inputs, the diffusion model is used to perform repeated multiple predictions, and a plurality of groups of predicted main chain structure parameters are obtained, which are used to construct the three-dimensional structure distribution of the target protein subsequently, to reflect the dynamic change characteristics of the three-dimensional structure of the target protein sequence.

[0043] It should be noted that the main chain structure parameter is used to represent the relative position parameter between residues in the protein main chain structure, which can be but not limited to using the (p, R) parameter as the main chain structure parameter to describe the spatial arrangement of each residue in the protein main chain structure. The protein main chain three-dimensional structure can be represented as the (x, y, z) coordinates p of the alpha carbon atom C α in each residue, and the rotation matrix R (3X3 matrix) of the triangle composed of (N, C α , C) three atoms, wherein N represents a nitrogen atom and C represents a carbon atom. The initial random main chain structure parameter can correspondingly use the three-dimensional coordinates p of the alpha carbon atom C α in each residue to conform to the standard normal distribution, and the initial rotation matrix R conforms to the uniform distribution on the rotation group (SO(3)).

[0044] In an optional embodiment, based on multiple sets of initial random main chain structure parameters, target residue embedding parameters, and target residue pair embedding parameters, a diffusion model is used for multiple times of prediction to obtain multiple sets of predicted main chain structure parameters corresponding to the target protein sequence, including: performing code masking on the embedding vectors corresponding to the predetermined amino acid residues in the target residue embedding parameters and the target residue pair embedding parameters to obtain processed residue embedding parameters and processed residue pair embedding parameters; based on the multiple sets of initial random main chain structure parameters, the processed residue embedding parameters, and the processed residue pair embedding parameters, the diffusion model is used for multiple times of prediction to obtain multiple sets of predicted main chain structure parameters.

[0045] Optionally, the code masking manner can be to set the corresponding embedding vector values to 0, and the predetermined amino acid residues can be determined in a random manner, for example, a predetermined proportion (such as 70%) of amino acid residues are randomly selected, and their corresponding residue embedding parameters and target residue pair embedding parameters are replaced with 0. It should be noted that the predetermined proportion is different in the training stage and the inference stage (i.e. the prediction stage) of the diffusion model. For example, the predetermined proportion selected in the inference stage can be higher than the proportion selected in the training stage (such as 70% in the inference stage and 50% in the training stage). The higher 0 proportion in the inference stage can improve the diversity of the structure generated by the model, and therefore the 0 proportion adopted in the inference stage can be used as an adjustable parameter to adjust the diversity of the structure generated by the diffusion model according to the specific generation target. Through the above manner, before predicting the main chain structure parameters based on the diffusion model, the residue embedding parameters and the target residue pair embedding parameters corresponding to the predetermined amino acid residues in the input are set to a predetermined value (such as 0), thereby improving the diversity of the prediction results of the diffusion model.

[0046] In an optional embodiment, the diffusion model comprises a first linear layer, a second linear layer and a transformer network, and the diffusion model is used to perform multiple predictions based on the multiple sets of initial random main chain structure parameters, the processed residue embedding parameters and the processed residue pair embedding parameters, to obtain multiple sets of predicted main chain structure parameters, comprising: taking the multiple sets of initial random main chain structure parameters as the first main chain structure parameters in turn, and obtaining the multiple sets of predicted main chain structure parameters by the following manner: starting from the last time step in the multiple time steps, taking the multiple time steps as the target time step in turn, obtaining the second main chain structure parameters by the following manner: processing the processed residue embedding parameters by the first linear layer to obtain a third embedding vector; processing the processed residue pair embedding parameters and the first main chain structure parameters by the second linear layer to obtain a fourth embedding vector; processing the third embedding vector and the fourth embedding vector by the transformer network to obtain the second main chain structure parameters, wherein the transformer network comprises multiple transformer layers; calculating a new first main chain structure parameter according to the second main chain structure parameters, and repeating the above operations until the first time step in the multiple time steps is reached; and taking the second main chain structure parameters output by the first time step as a set of predicted main chain structure parameters.

[0047] Optionally, the transformer network can be, but is not limited to, a Transformer network, and the Transformer network can be a three-dimensional rotationally covariant Transformer network. By setting the linear layers (i.e., the first linear layer and the second linear layer) in the diffusion model, linear transformation and feature extraction can be performed on the input data, and the linear layers can perform linear combination on the input features through matrix multiplication and bias addition, so that the new feature representations (i.e., the third embedding vector and the fourth embedding vector) obtained therefrom can capture higher-level features in the input data and provide richer and more complex representation capabilities for the model. Further inputting the obtained third embedding vector and fourth embedding vector into the transformer network, the transformer network can not only further extract features from the input third embedding vector and fourth embedding vector, but also capture long-term dependencies therein, integrate and model the context information in the input third embedding vector and fourth embedding vector, and the predicted main chain structure parameters output by the diffusion model better fuse the features of the input protein sequence (such as the target protein sequence) and are more accurate and reliable.

[0048] It should be noted that in the diffusion model, the number of time steps is used to represent different stages in the inference process. In the inference stage of the diffusion model, the number of time steps generally decreases from large to small, and the proportion of random noise contained in the output structure gradually decreases. Specifically, the initial random main chain structure parameters are randomly generated from the prior distribution, and are taken as the input (p T ,R T), the predicted main chain structure parameters are obtained through the kernel of the diffusion model According to the method of the general diffusion model, the main chain structure parameters (p ) at the t-th time step are calculated according to the main chain structure parameters (p T-1 ,R T-1 ) at the (t-1)-th time step as the input of the T-1 step, and the new main chain structure parameters (p T-1 ,R T-1 ) are predicted according to the main chain structure parameters (p ) at the t-th time step T-2 ,R T-2 ) are calculated (the inference stage start step of the diffusion model is the T step, and the steps are T-1, T-2, …, 1 in turn). According to the main chain structure parameters (p ) at the t-th time step, the specific implementation method of the main chain three-dimensional structure parameters (p t-1 ,R t-1 ) at the t-th (t=[T, T-1, …, 2, 1]) time step input is various. For example, according to the formula 【2】 and formula

[10] recorded in the reference (Watson, J. L. et al. De novo design of protein structure and function with RF diffusion. Nature 620, 1089-1100 (2023).), a feasible way is:

[0049] Wherein, ∈ is a random vector subject to standard normal distribution, β t is a model-defined parameter. α t =1-β t , β t represents a set of model-defined parameters including [β1, β2, …, β T ];

[0050]

[0051] Wherein, R t-1 represents the main chain three-dimensional structure parameters calculated at the (t-1)-th time step input, is a model-defined hyperparameter at the t-th time step, represents a model-defined hyperparameter at the (t-1)-th time step, is an isotropic Gaussian distribution on the rotation group, ∈ d is a random number subject to standard normal distribution, represents the gradient of the function with respect to the variable R t , is the gradient operator symbol, represents the transpose of the R t matrix. f d(d = 1, 2, 3) denote constant matrices, where ). In this way, the final step prediction is obtained as the final output of the diffusion model. Finally, the three-dimensional coordinates of the backbone atoms are calculated from the ideal structure of the residues. Different protein three-dimensional structures can be obtained by multiple iterations.

[0052] Step S110, based on the plurality of sets of predicted backbone structure parameters, generate a three-dimensional structure distribution of the target protein sequence, wherein the three-dimensional structure distribution is used to indicate a plurality of different three-dimensional structures of the target protein sequence.

[0053] Optionally, the (N, C α , C, O, C β ) atomic coordinates of each residue can be calculated according to (p, R). The coordinates of other side chain atoms can be filled in after the coordinates of the backbone atoms are determined using Open Molecular Mechanics (OpenMM) and other tools, thereby obtaining a complete protein three-dimensional structure. Based on this, based on the plurality of sets of predicted backbone structure parameters corresponding to the target protein sequence, a plurality of three-dimensional structures of the target protein sequence can be obtained, and based on the plurality of three-dimensional structures, a three-dimensional structure distribution can be obtained to represent the dynamic change characteristics of the target protein sequence.

[0054] In an alternative embodiment, before the plurality of sets of predicted backbone structure parameters corresponding to the target protein sequence are obtained by using the diffusion model to perform multiple predictions based on the plurality of sets of initial random backbone structure parameters, the target residue embedding parameters, and the target residue pair embedding parameters, the method further comprises: obtaining a plurality of protein sequences, a plurality of sequence comparison data corresponding to the plurality of protein sequences respectively, and a plurality of actual backbone structure parameters corresponding to the plurality of protein sequences respectively; based on the plurality of protein sequences and the plurality of sequence comparison data corresponding to the plurality of protein sequences respectively, obtaining residue embedding parameters and residue pair embedding parameters corresponding to the plurality of protein sequences respectively; adding noise to the actual backbone structure parameters corresponding to the plurality of protein sequences respectively to obtain noise backbone structure parameters corresponding to the plurality of protein sequences respectively; taking the noise backbone structure parameters corresponding to the plurality of protein sequences respectively, the residue embedding parameters and the residue pair embedding parameters corresponding to the plurality of protein sequences respectively as the input of the initial neural network model, and taking the actual backbone structure parameters corresponding to the plurality of protein sequences respectively as the target output of the initial neural network model, training the initial neural network model to obtain the diffusion model.

[0055] Optionally, based on the plurality of protein sequences and the plurality of multiple sequence alignment data corresponding to the plurality of protein sequences respectively, residue embedding parameters and residue pair embedding parameters corresponding to the plurality of protein sequences respectively are predicted by using the alpha fold model, and the specific implementation manner is the same as that of obtaining the foregoing target residue embedding parameters and target residue pair embedding parameters, which will not be described here. For each protein sequence in the plurality of protein sequences, the residue main chain position (p t ,R t ) mixed with noise, that is, the noise main chain structure parameter is input into the initial neural network model together with the corresponding initialized residue embedding and residue pair embedding, and the initial neural network model is trained by using the machine learning manner to obtain the diffusion model.

[0056] Optionally, in the model training stage of the initial neural network model, each data sample includes a protein sequence, residue embedding parameters and residue pair embedding parameters corresponding to the protein sequence, and the main chain structure parameter (p, R) of the protein sequence and the time step t randomly selected from [1, T]. One sequence in the training data may correspond to multiple structures, and the main chain structure parameter (p, R) of the protein sequence is randomly selected from the multiple structures. In the training process, a certain order of random noise is mixed into (p t ,R t ), and the kernel of the initial neural network model is required to predict the protein main chain structure parameter (p0, R0) before the noise is mixed. It should be noted that in the training stage, a part of amino acid residues in the protein sequence are also randomly selected, and the residue embedding parameters and residue pair embedding parameters corresponding to the amino acid residues are set to 0.

[0057] Optionally, in the training stage of the initial neural network model, the protein three-dimensional structure mixed with different proportions of noise and the embedding vector extracted from the alpha fold model are used as the input vector, and the initial neural network model is required to learn and predict the protein three-dimensional structure before the noise is mixed. The kernel of the initial neural network model in this embodiment is a Transformer network with three-dimensional rotation and translation covariance. The invariant point attention (IPA) module in the alpha fold model can be used to ensure that the network has three-dimensional rotation and translation covariance. The residue main chain position (p t ,R t), i.e. the noise main chain structure parameters are input into the initial neural network model together with the corresponding initialized residue embedding and residue pair embedding, the initial neural network model is trained by machine learning, and the diffusion model is obtained.

[0058] Specifically, Figure 2 is an optional structure and data interaction schematic diagram of a neural network model according to an embodiment of the present application, as shown in Figure 2 The input data of the initial neural network model is the noise main chain structure parameters of a protein sequence, residue embedding parameters and residue pair embedding parameters. First, the input residue embedding parameters of the protein sequence are input into the first linear layer in the initial neural network model, the residue embedding parameters are linearly transformed and feature extracted to obtain the fifth embedding vector. The residue main chain position (p t ,R t ) and residue pair embedding parameters of the protein sequence are input into the second linear layer in the initial neural network model, the residue main chain position (p t ,R t ) and residue pair embedding parameters are linearly transformed and feature extracted to obtain the sixth embedding vector. The fifth embedding vector and the sixth embedding vector are input into the converter network (i.e. the multi-layer Transformer module) in the initial neural network model, and the final output is the network predicted residue main chain position

[0059] In the model training stage, a loss function can be set, which is used to indicate the difference between the real protein main chain structure parameters (p0, R0) and the predicted main chain structure parameters , which is specifically represented as the root mean square error (rmsd) between the predicted coordinates and the real coordinates p0, and the L2 norm of . When training the model based on the corresponding data of multiple protein sequences, the initial neural network model can be fully trained using the machine-related data of multiple protein sequences obtained based on the PDB database and the Swiss-model repository, to obtain the trained network model. Then, the trained network model is fine-tuned using the structure data measured by the nuclear magnetic resonance (NMR) method from the related data obtained from the PDB database, to obtain the diffusion model.

[0060] Still as Figure 2As shown, in the multi-layer Transformer module of the converter network, the input of the first layer Transformer module is the fifth embedding vector h and the sixth embedding vector z corresponding to the protein sequence, and the noise backbone structure parameter (p t ,R t ); the input of the remaining layers of Transformer modules is the output of the previous layer, including two embedding vectors processed by the previous layer of Transformer module (one of which corresponds to the fifth embedding vector h of the input, and the other corresponds to the sixth embedding vector z of the input), and the backbone structure parameter output by the previous layer of Transformer module. Each layer of Transformer module includes an invariant point attention layer, an edge update layer and a backbone update layer. The invariant point attention layer is used to calculate the attention weight of the invariant point in the corresponding input data, so as to pay more attention to the processing of these invariant points in the subsequent processing, and update the residue embedding vector h; the edge update layer learns the interaction relationship between the residues, and updates the embedding vector z; the backbone update layer is used to update the backbone structure parameter.

[0061] Figure 3 is an optional single-layer Transformer module architecture according to an embodiment of the present application, as shown in Figure 3 , taking the current layer of Transformer module as the l+1 layer for example, the output of the previous layer of Transformer module includes: two embedding vectors h l and Z l output by the previous layer of Transformer module, and the backbone structure parameter (p l ,R l ) output by the previous layer of Transformer module. Input h l , Z l and (p l ,R l ) to the current layer of Transformer module, first process h l , Z l and (p l ,R l ) through the invariant point attention layer to obtain one of the embedding vectors h l+1 output by the current layer of Transformer module; process Z l and h l+1 through the edge update layer to obtain the other embedding vector Z l+1; by the backbone update layer l+1 and (p l , R l ) are processed to obtain main chain structure parameters (p l+1 , R l+1 ) output by the current layer of the Transformer module; the same operation as the current layer of the Transformer module is performed on each layer of the multiple layers of the Transformer module until the last layer of the Transformer module in the multiple layers of the Transformer module is reached; and the main chain structure parameters output by the last layer of the Transformer module are taken as predicted main chain structure parameters, which are the denoised main chain structure parameters.

[0062] In an alternative embodiment, the diffusion model is obtained by training an initial neural network model with the noise main chain structure parameters corresponding to the plurality of protein sequences, the residue embedding parameters and the residue pair embedding parameters corresponding to the plurality of protein sequences as inputs of the initial neural network model, and the actual main chain structure parameters corresponding to the plurality of protein sequences as target outputs of the initial neural network model. Specifically, the diffusion model is obtained by training the initial neural network model with the noise main chain structure parameters corresponding to the plurality of protein sequences, the residue embedding parameters and the residue pair embedding parameters corresponding to the plurality of protein sequences as inputs of the initial neural network model, and the actual main chain structure parameters corresponding to the plurality of protein sequences as target outputs of the initial neural network model, and then adjusting the parameters of the initial neural network model to obtain the diffusion model.

[0063] Optionally, the plurality of protein sequences are obtained from a protein data bank (PDB) and Swiss-model repository, multi-chain protein structures in the PDB database are disassembled into single chains, and then the data are sorted into pairs of protein sequences and corresponding structures, and one unique protein sequence usually corresponds to multiple structures. The plurality of protein sequences include a first number of first protein sequences and a second number of second protein sequences, wherein the first protein sequences can be protein sequences determined based on NMR technology obtained from a protein data bank (PDB), and the three-dimensional structure information of the part of protein sequences is more abundant than the structure determined based on X-ray crystal diffraction and cryo-EM. Each protein sequence usually contains 10-20 three-dimensional structures. The structure data corresponding to the first protein sequences determined based on NMR technology are taken out separately, and after the model is trained to convergence using all the data at the beginning of training, the model is fine-tuned only using the first number of first protein sequences at the end of training.

[0064] The Swiss-model repository is a protein structure database containing protein three-dimensional structure models generated by the Swiss-model automated modeling library (SWISS-MODEL) homology modeling process. A plurality of protein sequences corresponding to the protein three-dimensional structure models generated by Swiss-model are extracted from the Swiss-model repository, and structures with a template sequence similarity identity of 100% are removed.

[0065] Through the above steps S102 to S110, the residue embedding parameters, the residue pair embedding parameters, and the randomly generated main chain structure parameters of the target protein can be obtained by using the diffusion model to perform multiple predictions, and the protein three-dimensional structure distribution under different conditions can be obtained. The purpose of efficiently predicting the protein three-dimensional structure distribution can be achieved without using complex molecular dynamics simulation methods, thereby simplifying the protein three-dimensional structure distribution generation process and improving the efficiency of generating the protein three-dimensional structure distribution. Further, the technical problem of low efficiency of generating the protein three-dimensional structure distribution based on the molecular dynamics simulation method in the related art is solved.

[0066] Based on the above embodiments and optional embodiments, the present application provides an optional implementation of a method for generating a protein three-dimensional structure distribution, which comprises:

[0067] S1, sequence data processing, specifically comprising:

[0068] For all unique protein sequences in the collected dataset, we first searched for corresponding multiple sequence alignment data for each protein sequence in the Uniref90, MGnify, and BFD databases. The obtained protein sequences and their corresponding multiple sequence alignment data were then input into a pre-trained Alphafold model. The embedding vectors (including residue embedding vectors and residue pair embedding vectors) output by the EVOformer module in the Alphafold model were used as input to the initial neural network model.

[0069] S2, Determination of structural information, specifically including:

[0070] S21, Determination of protein backbone structural parameters. Specifically, the three-dimensional structure of the protein backbone can be represented by the α-carbon atoms (C atoms) of each residue. α The (x, y, z) coordinates p, and (N, C) α C) A rotation matrix R (3x3 matrix) of three atoms forming a triangle, where N represents a nitrogen atom and C represents a carbon atom.

[0071] S22, Setting the initial neural network model structure. Specifically, the kernel of the initial neural network model set in this embodiment is a Transformer network with three-dimensional rotational and translational covariance. The invariant point attention (IPA) module in the Alphafold model can be used to ensure the network has three-dimensional rotational and translational covariance. The main chain positions (p...) of residues mixed with noise... t ,R t In this process, the noise main chain structure parameters, along with the corresponding initialized residue embeddings and residue pair embeddings, are input into an initial neural network model. This initial neural network model is then trained using machine learning to obtain the diffusion model. The specific implementation is as follows:

[0072] In the initial neural network model, each input data consists of the noisy main chain structure parameters, residue embedding parameters, and residue pair embedding parameters of a protein sequence. First, the residue embedding parameters of the input protein sequence are fed into the first linear layer of the initial neural network model. Linear transformation and feature extraction are then performed on the residue embedding parameters to obtain the fifth embedding vector. The main chain positions (p...) of the protein sequence's residues are then... t ,R t ) and residue pairs are embedded into the second linear layer of the initial neural network model, and the residue main chain position (p) is used as the input parameter. t ,R t) and linearly transforms and extracts features of the residue pair embedding parameters to obtain a sixth embedding vector; the obtained fifth embedding vector and the sixth embedding vector are input into a converter network (i.e., a multi-layer Transformer module) in the initial neural network model, and finally the network outputs a predicted residue main chain position (p0, R0) after removing noise. In the multi-layer Transformer module, the input of the first layer is the input fifth embedding vector h and the sixth embedding vector z corresponding to the input protein sequence, and the noise main chain structure parameters (p t ,R t ); the input of the remaining layers of the Transformer module is the output of the previous layer, including two embedding vectors (one of which corresponds to the input fifth embedding vector h, and the other of which corresponds to the input sixth embedding vector z) processed by the previous layer of the Transformer module, and the main chain structure parameters output by the previous layer of the Transformer module. Each layer of the Transformer module includes an Invariant Point Attention layer, an edge update layer, and a backbone update layer.

[0073] S3, training phase, specifically comprising: in the training phase of the model, each data sample contains a protein sequence, an embedding vector (including a residue embedding vector and a residue pair embedding vector) of the corresponding protein sequence extracted from the Alphafold model, main chain structure parameters (p, R) of the protein sequence, and a time step t randomly selected from [1, T]. In the training data, one sequence may correspond to multiple structures, and the main chain structure parameters (p, R) of the protein sequence are randomly selected from the multiple structures. In the training process, a certain order of random noise is mixed into (p t ,R t ) to obtain (p0, R0) to be predicted by the kernel of the initial neural network model.

[0074] In the model training phase, a part of the amino acid residues can be randomly selected, and the embedding vectors corresponding to the amino acid residues extracted from the Alphafold model are set to 0.

[0075] The main loss function of the model training is the difference between the real protein main chain structure parameters (p0, R0) and the predicted main chain structure parameters , which is specifically represented as the root mean square error (rmsd) between the predicted coordinates and the real coordinates p0, and The L2 norm. When training a model based on multiple protein sequence-related data, the initial neural network model can be fully trained using protein sequences obtained from the PDB database and the Swiss-Model repository, along with the residue embedding vectors and residue pair embedding vectors corresponding to each protein sequence. Then, the trained network model can be fine-tuned using only the first protein sequence obtained from the PDB database and structural data determined by NMR to obtain the diffusion model.

[0076] S4, the reasoning stage, specifically includes:

[0077] During the inference phase of the model, initial random main chain structure parameters are randomly generated from the prior distribution and used as the input (p) at the T-th time step. T ,R T The predicted main chain structure parameters are obtained through the kernel of the diffusion model. The p-value was calculated using a general diffusion model. T-1 ,R T-1 This is used as the input for step T-1 (the inference phase of the diffusion model starts at step T, followed by steps T-1, T-2, ..., step 1). This process continues until the final prediction is obtained. These parameters serve as the final predicted main chain structure parameters. Finally, based on the ideal structure of the residues, from... The three-dimensional coordinates of the main chain atoms are calculated. After the main chain atom coordinates are determined, the coordinates of other side chain atoms can be filled in using tools such as OpenMM to obtain the complete three-dimensional protein structure. Repeated generation can yield multiple different three-dimensional protein structures; based on these multiple different three-dimensional protein structures, the distribution of the protein's three-dimensional structure is obtained.

[0078] It should be noted that although molecular dynamics simulation data was not used in the model training phase of this embodiment, this embodiment can generate dynamic structures that are similar to the results of molecular dynamics simulations from the protein amino acid sequence. Figure 4 This is an optional amino acid fluctuation comparison diagram according to an embodiment of the present invention, used to indicate the fluctuation of each amino acid in the Kras protein. The blue line represents the result of the protein three-dimensional structure distribution prediction based on the molecular dynamics (MD) method (denoted by MD), and the orange line represents the result of the protein three-dimensional structure distribution prediction based on this embodiment (denoted by prediction). It can be seen that the change trends of the results obtained based on the MD method and the results obtained based on this embodiment are basically consistent. Figure 5 This is an optional protein three-dimensional structure comparison diagram according to an embodiment of the present invention, such as... Figure 5As shown, the green part is the predicted protein three-dimensional structure based on the present embodiment, and the blue part is the predicted protein three-dimensional structure based on the MD method. It can be seen that the predicted protein three-dimensional structure based on the MD method and the predicted protein three-dimensional structure based on the present embodiment are basically coincident.

[0079] Figure 6 is a comparison diagram of an optional Kras protein amino acid displacement correlation matrix according to an embodiment of the present application, wherein Figure 6 The left graph in is a schematic diagram of the Kras protein amino acid displacement correlation matrix obtained based on the present embodiment, and the right graph is a schematic diagram of the Kras protein amino acid displacement correlation matrix obtained based on the MD method. It can be seen that the Kras protein amino acid displacement correlation matrix distributions of the two are basically consistent. Figure 7 is a comparison diagram of the Kras protein amino acid displacement correlation curve according to an embodiment of the present application, Figure 7 The left graph in is the correlation curve of the 50th amino acid in the Kras protein with other amino acids, and the right graph is the correlation curve of the 90th amino acid in the Kras protein with other amino acids. The orange line is the result obtained based on the present embodiment, and the blue line is the result obtained based on the MD method. It can be seen that the results obtained based on the present embodiment and the results obtained based on the MD method are basically consistent.

[0080] As can be seen from the above, the protein three-dimensional structure distribution prediction based on the present embodiment and the protein three-dimensional structure distribution prediction based on the MD method have basically consistent prediction effects. Moreover, the protein sampling a structure based on the present embodiment with a length of 180 amino acids only takes about 8 seconds, which has a relatively high sampling efficiency.

[0081] It should be noted that the present embodiment can efficiently utilize the information in the Alphafold model embedding vector, so that the neural network described in the present application needs a small amount of parameters, and the speed of batch generation sequence is fast, which is suitable for the scene of high-throughput prediction required in protein engineering. The present embodiment only trains the model on the public protein structure database, avoiding the expensive calculation cost of large-scale molecular dynamics simulation. The model training does not need to rely on molecular dynamics simulation data, but only uses the public protein structure database. The model is first trained on the sequence and PDB database overlapping part in the PDB database and Swiss-Model repository, and then the model is fine-tuned on the NMR structure data in the PDB database. The three-dimensional structure distribution generated by the model is similar to the molecular dynamics simulation data.

[0082] In the embodiment, a protein three-dimensional structure distribution generation device is also provided, which is used to implement the above-mentioned embodiments and preferred embodiments, and will not be described again. As used below, the term "module" "device" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware, or a combination of software and hardware implementation is also possible and contemplated.

[0083] According to the embodiment of the present application, a device for implementing the above-mentioned protein three-dimensional structure distribution generation method is also provided, Figure 8 is a structural diagram of a protein three-dimensional structure distribution generation device according to an embodiment of the present application, as Figure 8 shown, the above-mentioned protein three-dimensional structure distribution generation device comprises a data acquisition module 700, a vector generation module 702, an initial parameter determination module 704, a parameter prediction module 706, and a three-dimensional structure distribution generation module 708, wherein:

[0084] The data acquisition module 700 is configured to acquire a target protein sequence and a plurality of sequence comparison data corresponding to the target protein sequence, wherein the plurality of sequence comparison data is used to indicate a comparison result between the protein sequence and a plurality of related proteins corresponding thereto.

[0085] The vector generation module 702 is connected to the data acquisition module 700 and is configured to obtain a target residue embedding parameter corresponding to the target protein sequence and a target residue pair embedding parameter based on the target protein sequence and the plurality of sequence comparison data corresponding to the target protein sequence, wherein the target residue embedding parameter is used to describe a feature of a plurality of amino acid residues included in the target protein sequence, and the target residue pair embedding parameter is used to describe an interaction and distance relationship between the plurality of amino acid residues under the influence of the plurality of sequence comparison data corresponding to the target protein sequence.

[0086] The initial parameter determination module 704 is connected to the vector generation module 702 and is configured to determine a plurality of groups of initial random main chain structure parameters corresponding to the target protein sequence, wherein the initial random main chain structure parameter is used to describe a spatial arrangement of each residue in a main chain structure of the target protein sequence, and the initial random main chain structure parameter is randomly generated.

[0087] The parameter prediction module 706 is connected to the initial parameter determination module 704, and is configured to perform multiple predictions based on the multiple groups of initial random main chain structure parameters, the target residue embedding parameters, and the target residue pair embedding parameters, to obtain multiple groups of predicted main chain structure parameters corresponding to the target protein sequence, wherein the multiple groups of initial random main chain structure parameters correspond to the multiple groups of predicted main chain structure parameters in a one-to-one manner, and the diffusion model is pre-learned to associate the residue embedding parameters and the residue pair embedding parameters corresponding to the protein sequence with the changes in the three-dimensional structure of the protein.

[0088] The three-dimensional structure distribution generation module 708 is connected to the parameter prediction module 706, and is configured to generate a three-dimensional structure distribution of the target protein sequence based on the multiple groups of predicted main chain structure parameters, wherein the three-dimensional structure distribution is used to indicate multiple different three-dimensional structures of the target protein sequence.

[0089] In the embodiment of the present application, by setting the data acquisition module 700, the vector generation module 702, the initial parameter determination module 704, the parameter prediction module 706, and the three-dimensional structure distribution generation module 708, the multiple three-dimensional structures of the protein under different conditions are obtained by using the diffusion model to perform multiple predictions based on the residue embedding parameters, the residue pair embedding parameters, and the randomly generated main chain structure parameters of the target protein, and the three-dimensional structure distribution of the protein is formed, so that the purpose of efficiently predicting the three-dimensional structure distribution of the protein is achieved without using the complex molecular dynamics simulation method, thereby achieving the technical effects of simplifying the protein three-dimensional structure distribution generation process and improving the protein three-dimensional structure distribution generation efficiency, and further solving the technical problem of low protein three-dimensional structure distribution generation efficiency in the related art method of generating the three-dimensional structure distribution of the protein based on the molecular dynamics simulation method.

[0090] It should be noted that each of the above modules can be implemented by software or hardware. For example, for the latter, each of the above modules can be located in the same processor, or any combination of the above modules can be located in different processors.

[0091] It should be noted that the data acquisition module 700, the vector generation module 702, the initial parameter determination module 704, the parameter prediction module 706, and the three-dimensional structure distribution generation module 708 correspond to steps S102 to S110 in the embodiment, and the above modules have the same instances and application scenarios as the corresponding steps, but are not limited to the contents disclosed in the above embodiment. It should be noted that the above modules can run in a computer terminal as part of the device.

[0092] It should be noted that the optional or preferred embodiments of the present embodiment can refer to the related description in the embodiment, which will not be repeated here.

[0093] The protein three-dimensional structure distribution generation apparatus can further include a processor and a memory, the data acquisition module 700, the vector generation module 702, the initial parameter determination module 704, the parameter prediction module 706, the three-dimensional structure distribution generation module 708, and the like are stored in the memory as program modules, and the processor executes the above-mentioned program modules stored in the memory to realize the corresponding functions.

[0094] The processor includes a core, and the core calls the corresponding program modules from the memory, and the core can be one or more. The memory can include a non-permanent memory in a computer readable medium, a random access memory (RAM) and / or a non-volatile memory such as a read-only memory (ROM) or a flash memory (flash RAM), and the memory includes at least one memory chip.

[0095] According to the embodiments of the present application, an embodiment of a non-volatile storage medium is also provided. Optionally, in the embodiment, the non-volatile storage medium includes a stored program, and when the program runs, controls a device in which the non-volatile storage medium is located to execute any one of the protein three-dimensional structure distribution generation methods.

[0096] Optionally, in the embodiment, the non-volatile storage medium can be located in any one of a computer terminal group in a computer network or any one of a mobile terminal group, and the non-volatile storage medium includes a stored program.

[0097] Optionally, the device where the non-volatile storage medium is located performs the following functions under the control of the program running at runtime: obtaining a target protein sequence and multi-sequence comparison data corresponding to the target protein sequence, wherein the multi-sequence comparison data is used to indicate comparison results between the protein sequence and a plurality of related proteins corresponding to the protein sequence; obtaining target residue embedding parameters corresponding to the target protein sequence and target residue pair embedding parameters based on the target protein sequence and the multi-sequence comparison data corresponding to the target protein sequence, wherein the target residue embedding parameters are used to describe characteristics of a plurality of amino acid residues included in the target protein sequence, and the target residue pair embedding parameters are used to describe interactions and distance relationships between the plurality of amino acid residues under the influence of the multi-sequence comparison data corresponding to the target protein sequence; determining a plurality of groups of initial random main chain structure parameters corresponding to the target protein sequence, wherein the initial random main chain structure parameters are used to describe spatial arrangements of each residue in a main chain structure of the target protein sequence; performing a plurality of predictions based on the plurality of groups of initial random main chain structure parameters, the target residue embedding parameters, and the target residue pair embedding parameters using a diffusion model to obtain a plurality of groups of predicted main chain structure parameters corresponding to the target protein sequence, wherein the plurality of groups of initial random main chain structure parameters and the plurality of groups of predicted main chain structure parameters correspond to each other, and the diffusion model is pre-learned to associate residue embedding parameters and residue pair embedding parameters of a corresponding protein sequence with changes in a three-dimensional structure of the protein; and generating a three-dimensional structure distribution of the target protein sequence based on the plurality of groups of predicted main chain structure parameters, wherein the three-dimensional structure distribution is used to indicate a plurality of different three-dimensional structures in which the target protein sequence exists.

[0098] According to an embodiment of the present application, a processor is also provided. Optionally, in the embodiment, the processor is used to run a program, and the program performs any of the above methods for generating a three-dimensional structure distribution of a protein when running.

[0099] According to an embodiment of the present application, a computer program product is also provided, which, when executed on a data processing device, is adapted to execute a program that initializes steps of any of the above methods for generating a three-dimensional structure distribution of a protein.

[0100] Optionally, the computer program product described above, when executed on the data processing device, is adapted to execute the program steps of: obtaining a target protein sequence, and multi-sequence comparison data corresponding to the target protein sequence, wherein the multi-sequence comparison data is used to indicate comparison results between the protein sequence and a plurality of related proteins corresponding thereto; obtaining target residue embedding parameters corresponding to the target protein sequence and target residue pair embedding parameters based on the target protein sequence and the multi-sequence comparison data corresponding to the target protein sequence, wherein the target residue embedding parameters are used to describe characteristics of a plurality of amino acid residues included in the target protein sequence, and the target residue pair embedding parameters are used to describe interactions and distance relationships between the plurality of amino acid residues under the influence of the multi-sequence comparison data corresponding to the target protein sequence; determining a plurality of groups of initial random main chain structure parameters corresponding to the target protein sequence, wherein the initial random main chain structure parameters are used to describe spatial arrangements of each residue in a main chain structure of the target protein sequence; performing a plurality of predictions based on the plurality of groups of initial random main chain structure parameters, the target residue embedding parameters, and the target residue pair embedding parameters using a diffusion model to obtain a plurality of groups of predicted main chain structure parameters corresponding to the target protein sequence, wherein the plurality of groups of initial random main chain structure parameters and the plurality of groups of predicted main chain structure parameters correspond one-to-one, and the diffusion model is pre-learned to associate residue embedding parameters and residue pair embedding parameters of corresponding protein sequences with changes in protein three-dimensional structures; and generating a three-dimensional structure distribution of the target protein sequence based on the plurality of groups of predicted main chain structure parameters, wherein the three-dimensional structure distribution is used to indicate a plurality of different three-dimensional structures in which the target protein sequence exists.

[0101] The electronic device provided by the embodiment of the present application comprises a processor, a memory and a program stored on the memory and executable on the processor. When the processor executes the program, the following steps are implemented: obtaining a target protein sequence and multi-sequence comparison data corresponding to the target protein sequence, wherein the multi-sequence comparison data is used to indicate comparison results between the protein sequence and a plurality of related proteins corresponding to the protein sequence; obtaining target residue embedding parameters corresponding to the target protein sequence and target residue pair embedding parameters based on the target protein sequence and the multi-sequence comparison data corresponding to the target protein sequence, wherein the target residue embedding parameters are used to describe characteristics of a plurality of amino acid residues included in the target protein sequence, and the target residue pair embedding parameters are used to describe interactions and distance relationships between the plurality of amino acid residues under the influence of the multi-sequence comparison data corresponding to the target protein sequence; determining a plurality of groups of initial random main chain structure parameters corresponding to the target protein sequence, wherein the initial random main chain structure parameters are used to describe spatial arrangements of each residue in a main chain structure of the target protein sequence; performing a plurality of predictions by using a diffusion model based on the plurality of groups of initial random main chain structure parameters, the target residue embedding parameters and the target residue pair embedding parameters to obtain a plurality of groups of predicted main chain structure parameters corresponding to the target protein sequence, wherein the plurality of groups of initial random main chain structure parameters and the plurality of groups of predicted main chain structure parameters correspond to each other, and the diffusion model is pre-learned to associate the residue embedding parameters and the residue pair embedding parameters of the corresponding protein sequence with changes in the three-dimensional structure of the protein; and generating a three-dimensional structure distribution of the target protein sequence based on the plurality of groups of predicted main chain structure parameters, wherein the three-dimensional structure distribution is used to indicate a plurality of different three-dimensional structures existing in the target protein sequence.

[0102] The sequence of the above embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments.

[0103] In the above embodiments of the present application, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0104] In several embodiments provided in the present application, it should be understood that the disclosed technical content can be implemented by other ways. Among them, the device embodiments described above are only schematic, for example, the division of the above modules can be a logical function division, and actual implementation can have another division way, for example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed each other can be through some interface, indirect coupling or communication connection between modules or modules, which can be electrical or other forms.

[0105] The modules described above as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, i.e., may be located in one place, or may be distributed to multiple modules. Some or all of the modules can be selected as needed to achieve the purpose of the embodiment.

[0106] In addition, each functional module in each embodiment of the present application can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module. The integrated module can be realized in the form of hardware or in the form of a software functional module.

[0107] The integrated module, if realized in the form of a software functional module and sold or used as an independent product, can be stored in a computer-readable nonvolatile storage medium. Based on this understanding, the technical solutions of the present application, essentially or the part that contributes to the prior art, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a nonvolatile storage medium, including a number of instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the embodiments of the present application. The aforementioned nonvolatile storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0108] The above is only the preferred embodiment of the present application, and it should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A method for generating a distribution of protein three-dimensional structures, characterized by, The method comprises the following steps: obtaining a target protein sequence and multi-sequence comparison data corresponding to the target protein sequence, wherein the multi-sequence comparison data is used to indicate comparison results between the protein sequence and a plurality of related proteins corresponding to the protein sequence; based on the target protein sequence and the multi-sequence comparison data corresponding to the target protein sequence, obtaining target residue embedding parameters and target residue pair embedding parameters corresponding to the target protein sequence, wherein the target residue embedding parameters are used to describe characteristics of a plurality of amino acid residues included in the target protein sequence, and the target residue pair embedding parameters are used to describe interactions and distance relationships between the plurality of amino acid residues under the influence of the multi-sequence comparison data corresponding to the target protein sequence; determining a plurality of groups of initial random main chain structure parameters corresponding to the target protein sequence, wherein the initial random main chain structure parameters are used to describe spatial arrangements of respective residues in a main chain structure of the target protein sequence, and the initial random main chain structure parameters are randomly generated; based on the plurality of groups of initial random main chain structure parameters, the target residue embedding parameters, and the target residue pair embedding parameters, performing a plurality of times of prediction by using a diffusion model to obtain a plurality of groups of predicted main chain structure parameters corresponding to the target protein sequence, wherein the plurality of groups of initial random main chain structure parameters and the plurality of groups of predicted main chain structure parameters correspond to each other, and the diffusion model is pre-learned to associate residue embedding parameters and residue pair embedding parameters of a corresponding protein sequence with changes in a three-dimensional structure of the protein sequence; based on the plurality of groups of predicted main chain structure parameters, generating a three-dimensional structure distribution of the target protein sequence, wherein the three-dimensional structure distribution is used to indicate a plurality of different three-dimensional structures in which the target protein sequence exists.

2. The method of claim 1, wherein, The method comprises the following steps: encoding the target protein sequence and the multi-sequence comparison data corresponding to the target protein sequence by using a pre-trained AlphaFold model to obtain the target residue embedding parameters and the target residue pair embedding parameters.

3. The method of claim 1, wherein, The method comprises the following steps: masking embedding vectors corresponding to predetermined amino acid residues in the target residue embedding parameters and the target residue pair embedding parameters to obtain processed residue embedding parameters and processed residue pair embedding parameters; based on the plurality of groups of initial random main chain structure parameters, the processed residue embedding parameters, and the processed residue pair embedding parameters, performing a plurality of times of prediction by using the diffusion model to obtain the plurality of groups of predicted main chain structure parameters.

4. The method of claim 3, wherein, The diffusion model comprises a first linear layer, a second linear layer and a transformer network, and the diffusion model is used for multiple times of prediction based on the multiple groups of initial random main chain structure parameters, the processed residue embedding parameters and the processed residue pair embedding parameters, to obtain the multiple groups of predicted main chain structure parameters, comprising: The multiple groups of initial random main chain structure parameters are sequentially taken as first main chain structure parameters, and the multiple groups of predicted main chain structure parameters are obtained by the following manner: Starting from the last time step in the multiple time steps, the multiple time steps are sequentially taken as target time steps, and the second main chain structure parameters are obtained by the following manner: the processed residue embedding parameters are processed by the first linear layer to obtain a third embedding vector; the processed residue pair embedding parameters and the first main chain structure parameters are processed by the second linear layer to obtain a fourth embedding vector; the third embedding vector and the fourth embedding vector are processed by the transformer network to obtain the second main chain structure parameters, wherein the transformer network comprises multiple transformer layers; the new first main chain structure parameters are calculated according to the second main chain structure parameters, and the above operations are repeatedly performed until the first time step in the multiple time steps is reached; The second main chain structure parameters corresponding to the output of the first time step are taken as a group of predicted main chain structure parameters.

5. The method according to any one of claims 1 to 4, characterized in that, Before the multiple times of prediction based on the multiple groups of initial random main chain structure parameters, the target residue embedding parameters and the target residue pair embedding parameters are performed by the diffusion model to obtain the multiple groups of predicted main chain structure parameters corresponding to the target protein sequence, the method further comprises: Obtaining multiple protein sequences, multiple sequence comparison data corresponding to the multiple protein sequences respectively, and actual main chain structure parameters corresponding to the multiple protein sequences respectively; Based on the multiple protein sequences and the multiple sequence comparison data corresponding to the multiple protein sequences respectively, obtaining residue embedding parameters and residue pair embedding parameters corresponding to the multiple protein sequences respectively; Performing noise adding processing on the actual main chain structure parameters corresponding to the multiple protein sequences respectively to obtain noise main chain structure parameters corresponding to the multiple protein sequences respectively; Taking the noise main chain structure parameters corresponding to the multiple protein sequences respectively, the residue embedding parameters and the residue pair embedding parameters corresponding to the multiple protein sequences respectively as inputs of an initial neural network model, and taking the actual main chain structure parameters corresponding to the multiple protein sequences respectively as target outputs of the initial neural network model, training the initial neural network model to obtain the diffusion model.

6. The method of claim 5, wherein, Taking the noise main chain structure parameters corresponding to the multiple protein sequences respectively, the residue embedding parameters and the residue pair embedding parameters corresponding to the multiple protein sequences respectively as inputs of an initial neural network model, and taking the actual main chain structure parameters corresponding to the multiple protein sequences respectively as target outputs of the initial neural network model, training the initial neural network model to obtain the diffusion model, comprises: corresponding to the plurality of protein sequences as input of the initial neural network model, and actual main chain structure parameters corresponding to the plurality of protein sequences as target output of the initial neural network model, training the initial neural network model to obtain a first neural network model; corresponding to the first quantity of protein sequences as input of the first neural network model, and actual main chain structure parameters corresponding to the first quantity of protein sequences as target output of the first neural network model, continuing to train and adjust parameters of the first neural network model to obtain the diffusion model; wherein the first quantity of protein sequences are protein sequences in the plurality of protein sequences, and corresponding structure data of the first quantity of protein sequences are determined based on a nuclear magnetic resonance spectroscopy method.

7. The method according to any one of claims 1 to 4, characterized in that, In a case where the multi-sequence comparison data corresponding to the target protein sequence is multiple, the obtaining, based on the target protein sequence and the multi-sequence comparison data corresponding to the target protein sequence, of target residue embedding parameters and target residue pair embedding parameters corresponding to the target protein sequence includes: clustering and grouping the multiple pieces of multi-sequence comparison data corresponding to the target protein sequence to obtain multiple groups of multi-sequence comparison data; obtaining, based on the target protein sequence and the multiple groups of multi-sequence comparison data, multiple residue embedding parameters corresponding to the target protein sequence and corresponding multiple residue pair embedding parameters, wherein a quantity of the multiple residue embedding parameters corresponds to a number of groups of the multiple groups of multi-sequence comparison data; performing average calculation on the multiple residue embedding parameters to obtain the target residue embedding parameters; performing average calculation on the multiple residue pair embedding parameters to obtain the target residue pair embedding parameters.

8. A device for generating a distribution of three-dimensional structures of a protein, characterized by includes: a data acquisition module configured to acquire a target protein sequence and multi-sequence comparison data corresponding to the target protein sequence, wherein the multi-sequence comparison data is used to indicate comparison results between the protein sequence and a plurality of related proteins corresponding to the protein sequence; a vector generation module configured to obtain target residue embedding parameters and target residue pair embedding parameters corresponding to the target protein sequence based on the target protein sequence and the multi-sequence comparison data corresponding to the target protein sequence, wherein the target residue embedding parameters are used to describe characteristics of a plurality of amino acid residues included in the target protein sequence, and the target residue pair embedding parameters are used to describe mutual interaction and distance relationship between the plurality of amino acid residues under influence of the multi-sequence comparison data corresponding to the target protein sequence; and An initial parameter determination module is configured to determine a plurality of groups of initial random main-chain structure parameters corresponding to the target protein sequence, wherein the initial random main-chain structure parameters are used to describe the spatial arrangement of each residue in the main-chain structure of the target protein sequence, and the initial random main-chain structure parameters are randomly generated. A parameter prediction module is configured to perform multiple predictions based on the plurality of groups of initial random main-chain structure parameters, the target residue embedding parameters, and the target residue pair embedding parameters, by using a diffusion model, to obtain a plurality of groups of predicted main-chain structure parameters corresponding to the target protein sequence, wherein the plurality of groups of initial random main-chain structure parameters and the plurality of groups of predicted main-chain structure parameters correspond to each other, and the diffusion model is pre-learned to associate the residue embedding parameters and the residue pair embedding parameters of the corresponding protein sequence with the changes in the three-dimensional structure of the protein. A three-dimensional structure distribution generation module is configured to generate a three-dimensional structure distribution of the target protein sequence based on the plurality of groups of predicted main-chain structure parameters, wherein the three-dimensional structure distribution is used to indicate a plurality of different three-dimensional structures of the target protein sequence.

9. A non-volatile storage medium, comprising: The non-volatile storage medium stores a plurality of instructions, and the instructions are adapted to be loaded and executed by a processor to implement the method for generating a protein three-dimensional structure distribution according to any one of claims 1 to 7.

10. An electronic device, comprising: The device includes one or more processors and a memory configured to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the method for generating a protein three-dimensional structure distribution according to any one of claims 1 to 7.