Protein side chain conformation prediction method and system and electronic equipment

Abstract

The invention provides a protein side chain conformation prediction method and system and electronic equipment, and relates to the technical field of bioengineering.The method comprises the steps that feature information of each original amino acid is extracted based on a target position and a protein structure file; the feature information comprises self features of the original amino acids and associated features between the original amino acids; masking the characteristics of the amino acid needing to be designed; updating all the feature information through a protein structure prediction model; based on the updating result, predicting the category of amino acid needing to be designed, and obtaining a structure prediction result of the target protein; by combining the characteristics of adjacent amino acids, the accuracy of protein structure prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of bioengineering technology, and in particular to a protein side chain conformation prediction method, system and electronic equipment. Background Art

[0002] Proteins are long-chain polymer compounds composed of amino acids. At present, proteins that can form stable three-dimensional structures are mainly natural proteins formed by long-term natural evolution. Their amino acid sequences are basically formed by natural amino acids arranged in a certain combination order, and the sequence length is uncertain. There are 20 kinds of natural amino acids, with different chemical compositions and properties. The interaction between them in the sequence determines the shape, structure, and function of the folded protein.

[0003] When the structure and function of natural proteins cannot meet the needs of industrial or medical applications, in order to obtain specific functional proteins, protein sequence design is necessary. By arranging the amino acid sequence of the protein, it can fold spontaneously to form the required three-dimensional structure.

[0004] However, traditional protein sequence design methods and protein structure analysis methods mainly rely on laboratory experiments, which are time-consuming and costly. Machine learning and deep learning-based algorithms also have obvious limitations when using known structural information to predict complex protein structures.

[0005] Therefore, it is necessary to propose a protein side chain conformation prediction method, system and electronic device. Summary of the invention

[0006] The present specification provides a protein side chain conformation prediction method, system and electronic device, which extracts characteristic information of each original amino acid based on the target position and protein structure file; masks the self-characteristics of the amino acid to be designed in combination with the side chain attribute information; updates all the characteristic information through a protein structure prediction model; predicts the category of the amino acid to be designed based on the update result, and obtains the structure prediction result of the target protein, and increases the interpretability of the prediction and improves the accuracy of the prediction by combining the characteristic information of adjacent amino acids.

[0007] The present application provides a protein side chain conformation prediction method using the following technical solutions, including:

[0008] Obtain protein structure files and obtain original amino acid sequences;

[0009] Obtaining user design requirements, wherein the user design requirements include: at least one target position; the target position includes: the original position of the amino acid to be designed in the original amino acid sequence;

[0010] Based on the target position and the protein structure file, extract the characteristic information of each original amino acid; the characteristic information includes: the original amino acid's own characteristics and the correlation characteristics between the original amino acids; specifically, extract the basic information of the target atom from the protein structure file; construct the original amino acid's own characteristics according to the basic information of the target atom; determine the interaction information between the adjacent original amino acids according to the protein structure file; determine the correlation characteristics between the adjacent original amino acids based on the interaction information; the target atoms include main chain target atoms and side chain target atoms;

[0011] The self-features of the amino acids to be designed are masked; all the feature information is updated through a protein structure prediction model; based on the update result, the category of the amino acids to be designed is predicted to obtain a structure prediction result of the target protein.

[0012] Optionally, the step of obtaining a protein structure file to obtain an original amino acid sequence includes:

[0013] Obtaining the protein structure file uploaded by the user;

[0014] The protein structure file is first parsed to obtain an original amino acid sequence, wherein the protein structure file is verified according to a verification condition; if the verification passes, a linear arrangement order of each of the original amino acids is extracted from the protein structure file as the original amino acid sequence.

[0015] Optionally, constructing the original amino acid's own characteristics according to the basic information of the target atom includes:

[0016] Constructing the main chain attribute information of the original amino acid according to the basic information of the main chain target atom;

[0017] Constructing the side chain attribute information of the original amino acid according to the basic information of the side chain target atom;

[0018] The main chain attribute information and the side chain attribute information are spliced ​​to obtain the original amino acid's own characteristics.

[0019] Optionally, the extracting characteristic information of each original amino acid based on the target position and the protein structure file further includes:

[0020] Traversing each of the original amino acids, and calculating the distance between the current original amino acid and other original amino acids;

[0021] Based on the prediction range condition, determining the adjacent amino acid corresponding to the current original amino acid;

[0022] An association relationship is established between the current original amino acid and the corresponding adjacent amino acid, and the two original amino acids with an association relationship are each other's adjacent original amino acids.

[0023] Optionally, the method updates all the feature information by using a protein structure prediction model; based on the update result, predicts the category of the amino acid to be designed, and obtains the structure prediction result of the target protein, including:

[0024] Preprocessing the self-features and the associated features to obtain preprocessed self-features and associated features;

[0025] According to the arrangement order of the original amino acids in the original amino acid sequence, the preprocessed self-features and associated features are updated in sequence to obtain updated node features and updated associated features;

[0026] Decode the updated self-features and predict the target amino acid at the target position;

[0027] Combined with the inherent characteristics of the target amino acid, the structure prediction results of the target protein are obtained.

[0028] Optionally, the preprocessed self-features and associated features are updated in sequence according to the arrangement order of the original amino acids in the original amino acid sequence to obtain updated node features and updated associated features, including:

[0029] Perform feature splicing on the features of the adjacent amino acids of the current original amino acid and the associated features between the current original amino acid and the adjacent amino acids to obtain a splicing result;

[0030] Weighting the splicing results of the adjacent amino acids by using learnable weights to obtain a weighted result for each adjacent amino acid;

[0031] Summing the weighted results of all adjacent amino acids corresponding to the current original amino acid to obtain a weighted sum result;

[0032] The product of the weighted summation result and the activation function is used as an updated intermediate value; and according to the updated intermediate value, the own features and associated features are updated.

[0033] Optionally, the combining of the characteristics of the target amino acid to obtain the structure prediction result of the target protein includes:

[0034] Determining a target amino acid at a target position according to the probability distribution of the category of the target amino acid and / or the feedback information of the user;

[0035] Combining the characteristics of the target amino acid, the predicted properties of the predicted atom are obtained;

[0036] According to the original amino acid corresponding to the target position, searching for the main chain target atom corresponding to the original amino acid, and using the coordinate information of the main chain target atom as the coordinate information of the backbone atom;

[0037] Determining the positioning order of the predicted atoms according to the coordinate information of the predicted atoms and the central carbon atom of the target amino acid;

[0038] According to the positioning order, the coordinate information of the predicted atoms is determined in sequence;

[0039] The coordinate information of all predicted atoms and the coordinate information of all backbone atoms are summarized to obtain the structural prediction result of the target protein.

[0040] The protein side chain conformation prediction system provided in this application adopts the following technical solutions, including:

[0041] The first acquisition module is used to acquire the protein structure file and obtain the original amino acid sequence;

[0042] The second acquisition module is used to acquire user design requirements, wherein the user design requirements include: at least one target position; the target position includes: the original position of the amino acid to be designed in the original amino acid sequence;

[0043] A feature extraction module, for extracting feature information of each original amino acid based on the target position and the protein structure file; the feature information includes: the original amino acid's own features and the correlation features between the original amino acids;

[0044] A masking module, used for masking the characteristics of the amino acid to be designed;

[0045] The multi-task prediction module updates all the feature information through a protein structure prediction model; based on the update result, the category of the amino acid to be designed is predicted to obtain the structure prediction result of the target protein.

[0046] Optionally, the feature extraction module includes:

[0047] A basic information extraction submodule, used to extract basic information of target atoms from the protein structure file;

[0048] A self-feature construction submodule, used to construct the self-feature of the original amino acid according to the basic information of the target atom;

[0049] An interaction confirmation submodule, used for determining the interaction information between adjacent original amino acids according to the protein structure file;

[0050] A correlation feature construction submodule, used for determining the correlation features between adjacent original amino acids based on the interaction information;

[0051] The target atoms include main chain target atoms and side chain target atoms;

[0052] Optionally, the first acquisition module includes:

[0053] A file acquisition submodule, used to acquire the protein structure file uploaded by the user;

[0054] The file parsing submodule is used to perform a first analysis on the protein structure file to obtain an original amino acid sequence, wherein the protein structure file is verified according to a verification condition; if the verification passes, a linear arrangement order of each of the original amino acids is extracted from the protein structure file as the original amino acid sequence.

[0055] Optionally, the self-feature construction submodule includes:

[0056] A main chain attribute information construction unit, used to construct the main chain attribute information of the original amino acid according to the basic information of the main chain target atom;

[0057] A side chain attribute information construction unit, used to construct the side chain attribute information of the original amino acid according to the basic information of the side chain target atom;

[0058] The attribute information splicing unit is used to splice the main chain attribute information and the side chain attribute information to obtain the original amino acid's own characteristics.

[0059] Optionally, the feature extraction module further includes: an association relationship confirmation submodule;

[0060] The association confirmation submodule includes:

[0061] A traversal unit, used for traversing each of the original amino acids and calculating the distance between the current original amino acid and other original amino acids;

[0062] A screening unit, used for determining the adjacent amino acid corresponding to the current original amino acid based on the prediction range condition;

[0063] The association confirmation unit is used to establish an association relationship between the current original amino acid and the corresponding adjacent amino acid, and the two original amino acids with an association relationship are each other's adjacent original amino acids.

[0064] Optionally, the multi-task prediction module includes:

[0065] A preprocessing submodule, used for preprocessing the self-features and the associated features to obtain the preprocessed self-features and associated features;

[0066] An updating submodule, used to update the preprocessed self-features and associated features in sequence according to the arrangement order of the original amino acids in the original amino acid sequence, to obtain updated node features and updated associated features;

[0067] The category prediction submodule is used to decode the updated self-features and predict the target amino acid at the target position;

[0068] The structure prediction submodule is used to combine the characteristics of the target amino acids to obtain the structure prediction results of the target protein.

[0069] Optionally, the updating submodule includes:

[0070] A feature splicing unit, used for performing feature splicing on the features of the adjacent amino acids of the current original amino acid and the associated features between the current original amino acid and the adjacent amino acids to obtain a splicing result;

[0071] A weighting unit, used for weighting the splicing result of the adjacent amino acids by a learnable weight to obtain a weighted result of each adjacent amino acid;

[0072] A summing unit, used for summing the weighted results of all adjacent amino acids corresponding to the current original amino acid to obtain a weighted summation result;

[0073] An updating unit is used to take the product of the weighted sum result and the activation function as an updated intermediate value; and update the own features and associated features according to the updated intermediate value.

[0074] Optionally, the structure prediction submodule includes:

[0075] a target amino acid determination unit, configured to determine the target amino acid at the target position according to the probability distribution of the category of the target amino acid and / or the feedback information of the user;

[0076] A prediction unit, used to obtain predicted properties of predicted atoms by combining the characteristics of the target amino acid;

[0077] A first coordinate determination unit is used to search for a main chain target atom corresponding to the original amino acid according to the original amino acid corresponding to the target position, and use the coordinate information of the main chain target atom as the coordinate information of the backbone atom;

[0078] A positioning unit, used to determine the positioning order of the predicted atoms according to the coordinate information of the predicted atoms and the central carbon atom of the target amino acid;

[0079] A second coordinate determination unit, used to determine the coordinate information of the predicted atoms in sequence according to the positioning order;

[0080] The summary unit is used to summarize the coordinate information of all predicted atoms and the coordinate information of all skeleton atoms to obtain the structure prediction result of the target protein.

[0081] This specification also provides an electronic device, wherein the electronic device includes:

[0082] processor; and,

[0083] A memory storing computer executable instructions, which when executed cause the processor to perform any of the above methods.

[0084] The present specification also provides a computer-readable storage medium, wherein the computer-readable storage medium stores one or more programs, and when the one or more programs are executed by a processor, any of the above methods is implemented.

[0085] In the present application, the basic information of the target atom is extracted from the protein structure file; the self-characteristics of the original amino acid are constructed according to the basic information of the target atom; the interaction information between the adjacent original amino acids is determined according to the protein structure file; the association characteristics between the adjacent original amino acids are determined based on the interaction information; the target atoms include main chain target atoms and side chain target atoms; the self-characteristics of the amino acid to be designed are masked; all the characteristic information is updated through a protein structure prediction model; based on the update result, the category of the amino acid to be designed is predicted to obtain the structure prediction result of the target protein, and the prediction accuracy is improved by combining the characteristics of adjacent amino acids. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 A schematic diagram of the principle of a protein side chain conformation prediction method provided in the embodiments of this specification;

[0087] Figure 2 A schematic diagram of a protein structure prediction model for a protein side chain conformation prediction method provided in an embodiment of this specification;

[0088] Figure 3 A schematic diagram of the positioning sequence of a protein side chain conformation prediction method provided in the embodiments of this specification;

[0089] Figure 4 A schematic diagram of a model prediction test result of a protein side chain conformation prediction method provided in an embodiment of this specification;

[0090] Figure 5A schematic diagram of amino acid connections provided in the embodiments of this specification;

[0091] Figure 6 A schematic diagram of the structure of a protein side chain conformation prediction system provided in the embodiments of this specification;

[0092] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of this specification;

[0093] Figure 8 A schematic diagram of a computer-readable storage medium provided in accordance with an embodiment of the present specification. DETAILED DESCRIPTION

[0094] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art can think of other obvious variations. The basic principles of the present invention defined in the following description can be applied to other embodiments, variations, improvements, equivalents, and other technical solutions that do not deviate from the spirit and scope of the present invention.

[0095] The exemplary embodiments of the present invention will now be described more fully with reference to the accompanying drawings. Under the premise of being consistent with the technical concept of the present invention, the features, structures, characteristics or other details described in a particular embodiment do not exclude that they can be combined in one or more other embodiments in a suitable manner.

[0096] The term "and / or" or "and / or" includes all combinations of any one or more of the associated listed items.

[0097] Figure 1 A schematic diagram of the principle of a protein side chain conformation prediction method provided in the embodiments of this specification, the method comprising:

[0098] S2 obtains the protein structure file and obtains the original amino acid sequence;

[0099] S3 obtains user design requirements, wherein the user design requirements include: at least one target position; the target position includes: the original position of the amino acid to be designed in the original amino acid sequence;

[0100] S4 extracts characteristic information of each original amino acid based on the target position and the protein structure file;

[0101] S5 masks the self-features of the amino acids to be designed; updates all the feature information through a protein structure prediction model; and predicts the category of the amino acids to be designed based on the update result to obtain a structure prediction result of the target protein.

[0102] Proteins perform a variety of key functions in organisms, including catalyzing chemical reactions, signal transduction, and molecular transport. Designing protein sequences with specific functions has broad application prospects in drug development, industrial enzyme design, and biomaterial engineering.

[0103] Traditional protein sequence design methods and protein structure analysis methods mainly rely on laboratory experiments, which are time-consuming and costly. In recent years, algorithms based on machine learning and deep learning have shown great potential in the design of protein sequences, such as AlphaFold. However, these methods and tools still have limitations when using known structural information to predict complex protein structures. For example, when designing protein sequences through existing models, the sequence results given have low interpretability; the existing models have low accuracy in designing sequences for complex protein structures; when existing models predict protein structures, there are problems such as unreasonable structures and structural conflicts.

[0104] Based on this, the present invention provides a method for predicting protein side chain conformation, specifically:

[0105] S1 builds, trains, and tests protein structure prediction models.

[0106] S11 builds a protein structure prediction model.

[0107] The protein structure prediction model is preferably a graph neural network. The protein structure prediction model includes: an input layer, an encoding layer, a decoding layer and an output layer in sequence. The present invention introduces graph neural networks into protein sequence design and structure prediction. Graph neural networks can effectively capture the three-dimensional structural information of protein molecules, and update node features and edge features through a message passing mechanism, thereby improving the accuracy of prediction and the rationality of design.

[0108] S12 obtains the original sample.

[0109] S121 obtains structural information of historical proteins;

[0110] The sources of protein structural information include: non-redundant data from the Protein Structure Database (PDB, website: http: / / www.rcsb.org / ) and enhanced structural data predicted by Alphafold.

[0111] The structural information includes, but is not limited to, the total sequence length and basic information of all atoms. The basic information includes, but is not limited to, coordinate information. In one embodiment of the present specification, the structural information of the protein is a pdb file. The pdb file is a standard storage format for the three-dimensional structural data of the protein.

[0112] S122 verifies the structural information of historical proteins and determines the original sample;

[0113] If the total sequence length of the historical protein is less than 2000 and the main chain atoms of the historical protein are complete, the structural information of the historical protein is deemed to have passed the verification, and the historical protein corresponding to the structural information that has passed the verification is used as a sample protein, and the structural information of each sample protein is used as an original sample. In one embodiment of the present specification, it is allowed that the protein structure has missing positions, but the missing positions are not the positions where the main chain atoms are located.

[0114] S13 calculates the feature information corresponding to each original sample.

[0115] Existing graph neural network technology still has some defects and shortcomings in protein sequence design and protein structure prediction tasks, such as: (1) Insufficient feature selection and representation. For example, the lack of utilization of partially known three-dimensional structural information and physicochemical properties of proteins results in the model being unable to fully capture the complexity of proteins. This may affect the prediction accuracy and design effect of the model. (2) High computational complexity. Graph neural networks need to process a large number of nodes and edges, especially large protein structures. The high computational complexity not only increases the calculation time, but also puts higher requirements on hardware resources, limiting the practical application scope of the algorithm. (3) The rationality of the predicted structure is insufficient. Most protein structure prediction work is an independent task based on the sequence of the determined protein. It does not utilize the features and representation in the predicted protein sequence, resulting in insufficient rationality of the predicted structure and prone to spatial overlap or conflict.

[0116] Based on this, the present invention is based on a graph neural network and makes full use of information such as atomic dihedral angles, atomic distances, and amino acid side chain structures in protein structures to train models and improve the accuracy of sequence prediction. Therefore, in order to help the model capture the complexity and details of proteins, it is necessary to calculate feature information. Among them, the feature information includes the original amino acid's own features and the correlation features between the original amino acids.

[0117] S131 determines the self-signatures corresponding to the amino acids in the original sample.

[0118] S131-1 obtains the coordinate information of all target atoms in the original sample;

[0119] The coordinate information of the target atoms is extracted from the original sample. The target atoms include main chain target atoms and side chain target atoms.

[0120] In one embodiment of the present specification, each ATOM record and HETATM record in the original sample is traversed to identify the target atom of each amino acid and find the coordinate information of the target atom. Since the x-coordinate, y-coordinate, and z-coordinate of each atom are directly recorded in the original sample, the coordinate information of the target atom can be directly retrieved from the original sample after the target atom is determined.

[0121] S131-2 determines the intrinsic characteristics of the amino acid based on the coordinate information of the target atom.

[0122] The amino acid's own characteristics are obtained based on the chemical characteristics and structural characteristics of the amino acid. In one embodiment of the present specification, information such as dihedral angles, distances, angles, etc. between main chain atoms and side chain atoms are used as initialized node characteristics.

[0123] Specifically, the main chain characteristics and side chain characteristics of amino acids are calculated. The main chain characteristics include main chain dihedral angle information, main chain distance information, and main chain angle information. The side chain characteristics include side chain dihedral angle information, side chain distance information, and side chain angle information.

[0124] The main chain dihedral angle information, main chain distance information, and main chain angle information are summed to obtain the main chain attribute information of the amino acid. The side chain dihedral angle information, side chain distance information, and side chain angle information are summed to obtain the side chain attribute information of the amino acid. The main chain attribute information and side chain attribute information of the amino acid are spliced ​​to obtain the amino acid's own characteristics.

[0125] S132 determines the correlation features between amino acids in the original sample.

[0126] S132-1 constructs the relationship between amino acids.

[0127] Traverse each amino acid and calculate the distance between the current amino acid and other amino acids. In one embodiment of the present specification, the central carbon atom (C α ) and the central carbon atom (C α ) as the Euclidean distance between the current amino acid and other amino acids.

[0128] Select the K amino acids closest to the current starting amino acid as the adjacent amino acids corresponding to the current amino acid. Summarize all the adjacent amino acids corresponding to the current amino acid to construct the adjacent amino acid set corresponding to the current amino acid.

[0129] An association relationship is established between the current amino acid and the corresponding adjacent amino acid, and the two amino acids with the association relationship are adjacent original amino acids.

[0130] S132-2 determines the correlation characteristics between adjacent amino acids based on the chemical bonds and interaction relationships between the amino acids.

[0131] The correlation features are obtained based on the chemical bonds and interactions between amino acids. Chemical bonds include: peptide bonds, hydrogen bonds, hydrophobic interactions, ionic bonds (salt bridges) and disulfide bonds. Interaction relationships include: the bond length of chemical bonds.

[0132] In one embodiment of the present specification, information such as dihedral angles, distances, and angles between atomic chemical bonds is used as initialized edge features.

[0133] S132-2-1 Use molecular visualization software or computational chemistry tools to display chemical bonds and corresponding interaction relationships.

[0134] Molecular visualization software includes but is not limited to: PyMOL software, VMD (Visual Molecular Dynamics) software, Maestro software ( Suite). Computational chemistry tools include but are not limited to: AutoDock Vina molecular docking software, Gaussian and other quantum chemistry software.

[0135] S132-2-2 Determine the associated dihedral angle information between adjacent original amino acids based on the chemical bonds, determine the associated distance information based on the bond length of the chemical bonds, and determine the associated angle information based on the bond angle.

[0136] S132-2-3 sums the associated dihedral angle information, the associated distance information, and the associated angle information to obtain the association characteristics between adjacent amino acids.

[0137] The present invention introduces protein main chain structure and protein side chain structure into graph neural network for multi-task prediction, making the prediction level more efficient. Prediction is performed through comprehensive protein structure information, which is suitable for various application scenarios such as protein function research and drug design.

[0138] S14 divides the original samples into training set, test set and validation set.

[0139] The original samples are randomly selected as one of the training samples, test samples and validation samples according to the preset division ratio. The set of training samples is used as the training set, the set of test samples is used as the validation sample, and the set of validation samples is used as the test set. In order to ensure the training effect of the model and the accuracy of the evaluation, the division ratio is determined according to the specific data set size and actual needs.

[0140] S15 randomly selects amino acids from the training samples according to a preset mask ratio, performs masking, and updates the training samples.

[0141] Through masking, the model can learn context information based on enough unmasked data during the training process, and also provide enough masked data to train the model's ability to predict missing parts. During the training process, the preset mask ratio is preferably 15%. That is, the protein involved in the training sample includes N amino acids, and B amino acids are randomly selected from the N amino acids as masked amino acids, where B = int(0.15·N).

[0142] A mask value is set for the self-features in the masked amino acid. In one embodiment of the present specification, the mask value is a vector of all zeros. By setting the mask, the ability of the model to process incomplete data is improved.

[0143] S16 uses the training set to train the protein structure prediction model.

[0144] An undirected graph is constructed based on the association relationship of amino acids, in which, for N amino acids (nodes), each amino acid has adjacent amino acids (adjacent nodes), each amino acid has its own characteristics (node ​​characteristics), and there are association characteristics (edge ​​characteristics) between two adjacent amino acids.

[0145] Among them, in the protein structure prediction model, the decoding layer is used to decode the information output by the encoding layer. The decoding layer includes a fully connected layer. The output layer is used to output the amino acid category probability and side chain torsion information, and the coordinates of each atom on the side chain are determined according to the side chain torsion information, thereby obtaining the three-dimensional structure information of the protein. First, the present invention will try to predict the characteristics of the masked amino acids and the three-dimensional structure of the protein based on the unmasked amino acids through the protein structure prediction model. Specifically, as Figure 2 As shown:

[0146] S161 obtains the input result through the input layer and passes the input result to the encoding layer.

[0147] The input layer is used to receive the characteristic information of amino acids and perform dimensionality unification processing on the characteristic information.

[0148] Specifically, all the own features and associated features are initialized and standardized, which helps to maintain data consistency and facilitates efficient feature processing later. Considering that the own features and associated features need to be spliced ​​later, the feature dimensions of the own features are consistent with the feature dimensions of the associated features.

[0149] S162 updates its own features and associated features through the coding layer.

[0150] The encoding layer is used to update feature information. The encoding layer includes multiple graph network layers connected in sequence. Each graph network layer captures local structural information through graph convolution operations, and then updates its own features and related features. The graph network layer includes: attention module. The attention module includes: attention layer, normalization layer, fully connected layer, normalization layer in sequence. The output of the current graph network layer is used as the input of the next graph network layer.

[0151] In one embodiment of the present specification, a local graph adjacency matrix is ​​constructed for each node and its related adjacent nodes, and the node features and edge features in the graph are initialized, wherein the node features represent the amino acid features, and the edge features represent the association features between amino acids. Then, the node features are updated by the adjacent node features, the adjacent edge features, and the original node features; the edge features are updated by the adjacent node features and the original edge features. The feature update method is to form a matrix by splicing the node features and edge features, weight them through multiple linear layers, obtain multi-layer weighted results, and obtain the updated features after normalization.

[0152] Specific:

[0153] S162-1 uses the attention layer to concatenate the self-features and associated features of the adjacent amino acids to obtain a concatenation result; and performs weighted summation on the concatenation results of all adjacent amino acids to obtain a weighted summation result.

[0154] According to the order of each amino acid in the amino acid sequence, for each current amino acid, all adjacent amino acids corresponding to the current amino acid are traversed. For each adjacent amino acid, the self-features of the same adjacent amino acid are spliced ​​with the associated features between the current amino acid and the adjacent amino acid to obtain the splicing result of the adjacent amino acids. The splicing results of all adjacent amino acids are weighted summed to obtain the weighted summation result.

[0155] S162-2 performs a nonlinear transformation based on the weighted sum result and the activation function to obtain an updated intermediate value.

[0156] Based on the weighted summation result and the activation function, a nonlinear transformation is performed, and the product of the weighted summation result and the activation function is used as the updated intermediate value. In one embodiment of the present specification, the calculation method of the updated intermediate value is:

[0157]

[0158] Among them, node i is the current amino acid h i The characteristics of node j h is the adjacent amino acid j The characteristics of edge ij h is the adjacent amino acid jThe associated features, that is, the current amino acid h i and adjacent amino acids h j α is the learnable weight in the network module, which is optimized by the back propagation algorithm during the training process to learn the optimal weighting method. j Corresponding to a learnable weight α j . K is the adjacent amino acid h j The number of. σ represents the activation function. Update(node i ,edge i ) is the updated intermediate value.

[0159] S162-3 is based on updating the current amino acid's own features and associated features.

[0160] The current amino acid's own characteristics and the updated intermediate value are added to determine the current amino acid's new own characteristics. According to the current amino acid's new own characteristics, each associated characteristic in the associated characteristic set is updated to obtain a new associated characteristic.

[0161] S162-4 In order to balance the influence of different feature scales, the original features are regularized by randomly freezing neurons during the training process through the dropout layer. In one embodiment of the present specification, Adam is selected as the optimizer, the initial learning rate is 0.01, and the weight decay is 1e-6.

[0162] S162-5 performs feature transformation through multiple linear layers and finally performs feature scaling through a normalization function.

[0163] Take our own characteristics as an example:

[0164] Input its own features into the dropout layer, regularize its own features, and obtain the regularized result node′ of its own features i _dropout, where node′ i _dropout = dropout(node′ i ).

[0165] In one embodiment of the present specification, the outputs of a portion of neurons in the neural network are randomly discarded based on a preset probability (dropout rate) to reduce the model's dependence on specific neurons, thereby preventing overfitting.

[0166] Then, the optimized self-features are transformed by multi-layer linear layers to obtain the target transformation result node′ of the self-features i _dense. Among them,

[0167] In one embodiment of the present specification, the multi-layer linear layer includes a first dense layer and a second dense layer. The first dense layer performs a linear transformation on the optimized self-features to obtain a first transformation result. The second dense layer performs a linear transformation on the first transformation result to obtain a target transformation result of the self-features.

[0168] Finally, the target conversion result of its own features is normalized to obtain the normalized result of its own features.

[0169] The amino acid correlation features were transformed and normalized.

[0170] in,

[0171] The specific steps are similar to the above steps and will not be repeated here.

[0172] Based on the above steps, the output result of the current graph network layer is obtained. The output result of the current graph network layer includes: optimized own features and optimized associated features.

[0173] S162-6 uses the output result of each graph network layer as the input result of the next graph network layer, and adjusts its own features and related features.

[0174] Wherein, h represents the depth of the graph network layer, that is, the number of graph network layers. Preferably, h=10.

[0175] The output of the last graph network layer is used as the output of the entire encoding layer.

[0176] The present invention can optimize multiple objective functions in parallel and is suitable for large-scale protein sequence design and side chain structure prediction tasks, so that the designed protein sequence can be more accurately fitted to the structure of the target protein.

[0177] S163 determines the prediction result through the decoding layer.

[0178] The decoding layer includes a fully connected layer, which converts the high-dimensional output of the encoding layer into a specific prediction result.

[0179] Specifically, the prediction results include the predicted category and predicted attributes of the amino acid. In one embodiment of the present specification, a 20-dimensional vector representation is obtained, the probabilities of 20 standard amino acids at different residue positions are predicted, the probability distribution is given and a residue-site scoring matrix is ​​obtained as the amino acid category probability matrix, and the category with the highest probability is used as the predicted category of the current amino acid.

[0180] S163-1 obtains the predicted category of each amino acid, updates the amino acid sequence, and obtains a new amino acid sequence.

[0181] The predicted attributes describe the geometric conformation of the side chain rotation by predicting the torsion angles chi1, chi2, etc. of the side chain atoms at different residue positions. By constraining the spatial conformation between the side chain atoms and avoiding position conflicts, the formed side chain conformation is more stable and the structure is more reasonable.

[0182] S163-2 uses the coordinate information of the main chain target atoms as the atomic coordinates of the backbone atoms.

[0183] S163-3 predicts and calculates the spatial structure of the side chains of new amino acid sequences based on predicted attributes.

[0184] The predicted attributes include side chain torsion information, which includes: torsion angles of side chain target atoms and bond angles of chemical bonds. The torsion angles of side chain target atoms are determined by the side chain dihedral angle information.

[0185] First, in the same amino acid, the side chain target atoms are numbered from near to far according to the distance from the central carbon atom in the amino acid to which they belong, and the positioning order is obtained. Figure 3 As shown, those skilled in the art will know that the convention for side chains is that when leaving C α When the main chain is used, it is in accordance with the main chain C α The equilibrium distance is used to name the atoms in the heavy atom side chain (C β , C γ , C δ , C ε ……), for multiple C at the same equilibrium distance δ For example, it will match the subscript, such as δ 1 ,δ 2 ,δ 3 .

[0186] In one embodiment of the present specification, when the right atom is observed from the left atom of the bond, the sign of the side chain torsion angle corresponds to a clockwise rotation. The position of each side chain target atom is predicted in sequence according to the numbering order. The position of the side chain target atom is determined based on the side chain torsion angle and the bond angle.

[0187] According to the central carbon atom C of the amino acid α With the first side chain atom C β The fixed bond length of the first side chain atom C β The first spherical range of the atomic coordinates of the first side chain atom C β The first atomic coordinate set of .

[0188] S163-4 summarizes the atomic coordinates of all side chain target atoms and the atomic coordinates of all backbone atoms to obtain the structural prediction results of the protein.

[0189] The present invention innovatively solves the limitations of traditional methods in designing complex protein sequences and predicting side chain structures, and has important scientific research and practical application value.

[0190] S164 calculates the model loss through the output layer.

[0191] S164-1 selects masked amino acids to calculate the cross entropy loss.

[0192] The class with the highest probability in the amino acid class probability matrix of the masked amino acid is selected as the predicted class of the masked original amino acid. The true class of the masked amino acid is compared with the predicted class, and the cross entropy loss is calculated.

[0193] S164-2 selects masked amino acids to calculate structural loss (FAPE loss).

[0194] The predicted properties of each side chain target atom in the masked amino acid are selected, and the predicted coordinates of each side chain target atom are calculated.

[0195] The FAPE loss (Frame Aligned Point Error) was calculated based on the predicted coordinates, true coordinates, and the total number of side chain target atoms. Among them, p i is the predicted coordinate of the side chain target atom i, r i is the real coordinate of the side chain target atom i, and N is the total number of side chain target atoms.

[0196] The FAPE loss is used to measure the difference between the predicted atomic coordinates and the reference coordinates, and the error is calculated by aligning the predicted and reference coordinate systems.

[0197] S164-3 Select masked amino acids to calculate torsion angle loss.

[0198] The torsion angle loss is used to measure the difference between the predicted torsion angle and the true torsion angle. Specifically, the predicted attributes of each side chain target atom in the masked amino acid are selected to calculate the predicted torsion angle of each side chain target atom. The predicted torsion angle includes the predicted dihedral angle and the predicted bond angle. The difference between the predicted torsion angle and the true torsion angle is averaged to obtain the torsion angle loss AngleLoss. in, is the predicted dihedral torsion angle of the side chain target atom i, is the predicted bond angle of the side chain target atom i; is the torsion angle of the real side chain target atom i, is the bond angle of the actual side chain target atom i, and M is the total number of torsion angles. The torsion angle (χ angle) of the side chain target atom i, that is, the side chain dihedral angle information corresponding to the side chain target atom i.

[0199] S164-4 selects masked amino acids to calculate the atomic conflict penalty of atom pairs.

[0200] The two atoms i and j with a bond are regarded as an atom pair, and the Euclidean distance between each atom pair is calculated. For each pair of atoms, if the distance is less than a threshold (set as the sum of the atomic van der Waals radii), a conflict is considered to have occurred. All conflicts are counted, and the total conflict loss ClashLoss is calculated based on the severity of the conflict (for example, the shorter the distance, the greater the loss). Among them, R vdw Refers to the sum of the van der Waals radii of an atom pair. Atomic conflict loss is used to measure the steric conflict between atoms in the predicted structure.

[0201] In one embodiment of the present specification, specifically, the Euclidean distance between two atoms i and j in the atom pair is calculated as the first distance d of the atom pair. ij ,

[0202] By comparing the first distance and a preset comparison threshold, it is determined whether there is a conflict between the atom pairs; wherein the preset comparison threshold T ij is the sum of the van der Waals radii of the two atoms in the atomic pair; T ij =R i +R j If the first distance d between the atom pairs ij Less than the preset comparison threshold T ij , then it is considered that there is a conflict between the pair of atoms.

[0203] For each pair of colliding atoms, a collision penalty needs to be calculated.

[0204] The conflict loss can be defined based on the severity of the conflict, using the inverse of the distance or some nonlinear function of the distance to weight the conflict. For example, the conflict loss L ij With distance d ij Inversely proportional, that is, the shorter the distance, the greater the loss: ∈ is a small positive number used when d ij Very close to T ij Avoid division by zero when . The total atomic conflict loss L total =∑ (i，j)∈conflicts L ij ; where conflicts is the set of all conflicting atom pairs.

[0205] In the model training process of the present invention, the sum of the squares of the differences between all predicted angles and true angles is minimized, so that the model has the value of the predicted angle, and thus after a specific amino acid is given, the model has the ability to predict the angle.

[0206] S17 tests protein structure prediction models.

[0207] This method was tested on the samples of CASP14 (https: / / predictioncenter.org / casp14 / ). Table 1 compares the sequence design and side chain prediction results of this method with similar methods (CASP14 samples).

[0208] Note: The higher the sequence design accuracy (recovery), the better; the lower the side chain predicted atomic coordinate loss (fape), the better.

[0209]

[0210] (Table 1)

[0211] The prediction difficulties of CASP14 samples are as follows: T1073: TMB-easy (easy). T1080: FM / TBM (no fixed template prediction, free template prediction). T1067: TMB-hard (difficult). T1100: TMB-hard (difficult).

[0212] Table 1 Similar methods from:

[0213] AttnPacker refers to a new AI protein side chain prediction algorithm launched by Xu Jinbo's team at Molecular Heart (from Proceedings of the National Academy of Sciences, 2023Jun 6).

[0214] ZetaDesign refers to an end-to-end deep learning method for protein sequence design and side chain packing (from Briefings in Bioinformatics, Volume 24, Issue 4, 2023Jul 20).

[0215] DLPacker refers to a deep learning method for predicting the conformation of amino acid side chains in proteins (from Proteins, Volume 90, Issue 6, Page 1278-1290, 2022Jun).

[0216] This method has been tested on multiple samples. The test results are as follows: Figure 4 As shown:

[0217] Among them, Our model (on uniprot) indicates the prediction effect of the model of the present invention based on the UniProt protein database;

[0218] Our model muti-Task (on uniprot) represents the prediction effect of the multi-task processing of the model of the present invention based on the UniProt protein database;

[0219] Our model (on CATH 4-3) The model of the present invention is based on the prediction effect of CATH 4-3 classification;

[0220] Pifold (on CATH 4-3) indicates the prediction effect of Pifold based on CATH 4-3 classification;

[0221] ProteinMPNN(on CATH 4-3) represents the prediction effect of protein message passing neural network based on CATH 4-3 classification.

[0222] Figure 4 Similar methods come from:

[0223] ProteinMPNN refers to a robust protein sequence design method based on deep learning (from Science, Volume 378, Issue 6615, Page 49-56, 2022).

[0224] PiFold refers to a method for achieving effective and efficient protein reverse folding (from Proceedings of ICLR, 2023).

[0225] Compared with the Pifold model and the protein message passing neural network model, the prediction effect of the present invention has higher sequence design accuracy and lower model uncertainty.

[0226] The present invention optimizes the model structure, does not rely on high-demand operating hardware when used, and is convenient for deployment and application. Moreover, the present invention can simultaneously handle the multi-task work of protein sequence design and protein side chain structure prediction, and simultaneously form a reasonable structural conformation and a high-accuracy sequence.

[0227] Based on Table 1 and Figure 4 It can be seen that the present invention, combined with subsequent biological experimental verification, can achieve a level that is more accurate and more efficient in screening than existing protein design methods, providing guidance value for subsequent bioengineering development.

[0228] S2 obtains the protein structure file and obtains the original amino acid sequence;

[0229] S21 obtains the protein structure file uploaded by the user.

[0230] Protein structure files are used to describe the three-dimensional structure information of single-chain or multi-chain proteins. The file format of protein structure files includes but is not limited to: PDB (Protein Data Bank) format. Protein structure files include several parts, including: header part, primary structure, miscellaneous factors, secondary structure, connection annotation, unit cell characteristics and coordinate transformation, coordinate part, connection information part, bookkeeping, etc.

[0231] Each part contains multiple records, and different records store different contents, such as:

[0232] The HEADER record in the header section includes global information of the protein structure file, such as deposition date, source, etc.

[0233] The SEQRES record of the primary structure includes: residue sequence.

[0234] The connection annotation describes the non-covalent interactions between different residues in the protein. Specifically, the SSBOND records of the connection annotation include: disulfide bonds, defining the disulfide bonds between cysteine ​​CYS residues; the LINK records of the connection annotation include: chemical bonds between residues; the HYDBND records of the connection annotation include: hydrogen bonds; the SLTBRG records of the connection annotation include: salt bridges; and the CISPEP records of the connection annotation include: cis residues.

[0235] The ATOM record in the coordinate part is used to describe the atomic coordinates of standard amino acids and nucleic acids, which specifically includes: atoms of standard residues, atomic names of each atom in standard residues (amino acids and nucleic acids), residue names, rectangular coordinates (unit angstroms), occupancy, temperature factors and other information.

[0236] The HETATM record in the coordinate section is similar to the ATOM record, but is used to describe the atomic coordinates of non-standard residues (such as inhibitors, cofactors, ions, solvents, etc.). HETATM residues are not connected to other residues by default, and water molecules are also placed in this record.

[0237] The HETATM record of the coordinate part specifically includes: atoms of non-standard residues, atomic names of each atom in non-standard residues (compounds other than standard amino acids and nucleic acids, including inhibitors, cofactors, ions, solvents), residue names, rectangular coordinates (unit angstroms), occupancy, temperature factors and other information.

[0238] S22 performs a first analysis on the protein structure file to obtain an original amino acid sequence.

[0239] S221 verifies the protein structure file according to the verification condition;

[0240] The verification condition includes several check sub-conditions. In one embodiment of the present specification, the verification condition includes: a first check sub-condition and a second check sub-condition. The first check sub-condition includes: the total sequence length of the protein is less than 2000. The second check sub-condition includes: the main chain atoms of the protein are complete.

[0241] In one embodiment of the present specification, missing positions are allowed to exist in the protein structure, but the missing positions are not positions where main chain atoms are located.

[0242] If the protein structure file satisfies all the verification sub-conditions in the verification condition at the same time, it is considered that the protein structure file has passed the verification.

[0243] If the verification is passed, then extract the linear arrangement order of each original amino acid from the protein structure file as the original amino acid sequence. The original amino acid sequence includes the original position of each original amino acid.

[0244] In one embodiment of the present specification, the residue sequence is obtained by the SEQRES record in the protein structure file, and each original amino acid is numbered one by one according to the sequence arrangement order. Specifically, the two ends of the protein are the amino end (N-terminus) and the hydroxyl end (C-terminus), and the amino acids in a protein chain are indexed one by one from the amino end (N-terminus).

[0245] In another embodiment of the present specification, the ATOM and HETATM records in the protein structure file are traversed, the category of the corresponding original amino acid is determined according to the residue name, the order of appearance of each original amino acid is determined according to the order of appearance of the residue names in the protein structure file, and an index is established for each original amino acid based on the order of appearance of each original amino acid.

[0246] The index represents the original position of each original amino acid in the original amino acid sequence. Each original position corresponds to a class of original amino acids.

[0247] S3 obtains user design requirements, wherein the user design requirements include: at least one target position; the target position includes: the original position of the amino acid to be designed in the original amino acid sequence;

[0248] The basic information of the original amino acid sequence is displayed to the user through the first display page.

[0249] In one embodiment of the present specification, the basic information includes the original position of the original amino acid. In another embodiment of the present specification, the basic information also includes the category of the original amino acid corresponding to the original position, so as to improve the convenience of user selection and viewing.

[0250] The user selects the amino acid to be designed on the first display page, and based on the user's selection, the original position of the amino acid to be designed in the original amino acid sequence is determined as the target position.

[0251] S4 extracts characteristic information of each original amino acid based on the target position and the protein structure file;

[0252] The characteristic information includes the original amino acid's own characteristics and the correlation characteristics between the original amino acids.

[0253] S41 extracts the basic information of target atoms from protein structure files.

[0254] S411 determines the target atom.

[0255] Protein is a substance with a certain spatial structure formed by the coiled folding of a polypeptide chain composed of multiple amino acids in the form of "dehydration condensation". After dehydration condensation, each amino acid residue contains a central carbon atom (C α ), the central carbon atom is connected to four different groups / atoms, namely: amino group (-NH2), hydroxyl group (-COOH), hydrogen atom (-H) and side chain group (R group).

[0256] The target atoms include main chain target atoms and side chain target atoms. The main chain target atoms include the main atoms on the main chain of all original amino acids. The main atoms on the main chain of the original amino acids include: the central carbon atom (C α ), the nitrogen atom (N) in the amino group connected to the central carbon atom, the non-central carbon atom (C) in the hydroxyl group connected to the central carbon atom, and the oxygen atom (O) in the hydroxyl group connected to the central carbon atom. The side chain target atoms include: all atoms on all side chain groups.

[0257] S412 performs a second analysis on the protein structure file to extract basic information of the target atoms.

[0258] Traverse the ATOM and HETATM records in the protein structure file to find each atom in each original amino acid;

[0259] For atoms on the main chain, whether they are main chain target atoms is determined based on the element type of the atom. Specifically, if the element type of the atom is one of amino nitrogen atom (N), carbon atom (C), and hydroxyl oxygen atom (O), the atom is identified as the main chain target atom, and the basic information of the main chain target atom is extracted.

[0260] All atoms on the side chain are identified as side chain target atoms, and the basic information of the side chain target atoms is extracted.

[0261] The basic information includes coordinate information. The coordinate information includes the coordinates (x coordinate, y coordinate and z coordinate) of the atom.

[0262] S42 constructs the original amino acid's own characteristics according to the basic information of the target atom;

[0263] Find the basic information of the main chain target atom and the basic information of the side chain target atom from the atomic information, specifically:

[0264] S421 constructs the main chain attribute information of the original amino acid according to the basic information of the main chain target atom.

[0265] S421-1 calculates the main chain dihedral angle information according to the basic information of the main chain target atoms.

[0266] Main chain dihedral angle information includes: first torsion angle ( angle), a second torsion angle (ψ angle) and a third torsion angle (ω angle).

[0267] like Figure 5 As shown in the figure, there is a peptide bond (-CO-NH-) between two adjacent amino acid residues, and the group composed of the peptide bond and the two adjacent central carbon atoms (-C α -CO-NH-C α -) as a peptide unit; multiple sequentially connected peptide units are connected to form the main chain (skeleton) of the protein.

[0268] The peptide unit is a planar structure, in which the six atoms are located in the same plane, which is the peptide bond plane. The six atoms in a peptide unit are: C i , O i 、N i+1 , H i+1 , In one embodiment of the present specification, the main chain (skeleton) of a protein is considered to be a series of planes surrounding the central carbon atom C α Relatively free to rotate. That is, around the NC α Key, C α -C bonds and peptide bonds can undergo torsional rotation.

[0269] Around NC α The angle of key twist is the first twist angle ( horn), N in the peptide unit i , The rotation state is directly related to N i The peptide plane The rotation result.

[0270] Around C α -C bond torsion angle is the second torsion angle (ψ angle), ψ i With peptide unit C i The rotation state is directly related to C i The peptide plane The rotation result.

[0271] The angle of twist around the CN bond is the third torsion angle (ω angle). The CN bond here refers to the CN bond in the peptide bond connecting two amino acid residues. The ω angle has two stable states, namely trans (ω = 180°) and cis (ω = 0°). Both states appear in protein structures, but the trans state is more common. Specifically, ω i C in the peptide unit i 、N i+1 is directly related to the rotation state of .

[0272] In one embodiment of the present specification, the main chain dihedral angle information is calculated by a special bioinformatics software or programming tool (such as Python combined with NumPy and SciPy libraries). For example, the main chain dihedral angle information di_angle in the original amino acid is calculated by the DSSP tool main .

[0273] S421-2 calculates the main chain distance information according to the basic information of the main chain target atoms.

[0274] The distance between two main chain target atoms is calculated by the Euclidean distance formula. Preferably, the coordinates of the first main chain target atom are (x1, y1, z1), and the coordinates of the second main chain target atom are (x2, y2, z2), then the distance between the first main chain target atom and the second main chain target atom is

[0275] In one embodiment of the present specification, the distance between two adjacent main chain target atoms in the original amino acid is calculated to obtain the first distance information, and all the first distance information is summed to obtain the main chain distance information distance main .

[0276] S421-3 calculates the main chain angle information according to the basic information of the main chain target atoms.

[0277] In one embodiment of the present specification, the coordinates of the first main chain target atom are (x1, y1, z1), the coordinates of the second main chain target atom are (x2, y2, z2), and the coordinates of the third main chain target atom are (x3, y3, z3).

[0278] Calculate the first vector from the first main chain target atom to the second main chain target atom

[0279] Calculate the second vector from the first main chain target atom to the third main chain target atom

[0280] Calculates the dot product of the first vector and the second vector

[0281] Calculate the length of the first vector

[0282] Calculate the length of the second vector

[0283] The angles between the first main chain target atom, the second main chain target atom, and the third main chain target atom are calculated based on the dot product and the length. Specifically,

[0284] In one embodiment of the present specification, the angles between the three main chain target atoms in the original amino acid are calculated to obtain the angle information between the three main chain target atoms, and the angle information between all three main chain target atoms is summed to obtain the main chain angle information dirtect main .

[0285] S421-4 sums the main chain dihedral angle information, main chain distance information, and main chain angle information to obtain the main chain attribute information V of the original amino acid. main , where V main =di_angle main +distance main +dirtect main .

[0286] S422 constructs the side chain attribute information of the original amino acid according to the basic information of the side chain target atom.

[0287] The side chain attribute information includes: side chain dihedral angle information, side chain distance information and side chain angle information.

[0288] S422-1 calculates the side chain dihedral angle information according to the basic information of the side chain target atom.

[0289] In one embodiment of the present specification, the side chain dihedral angle information di_angle in the original amino acid is calculated by a special bioinformatics tool or software. side . Side chain dihedral angle information includes: χ angle (chi angle).

[0290] S422-2 calculates the side chain distance information according to the basic information of the side chain target atom.

[0291] The distance between two side chain target atoms is calculated using the Euclidean distance formula. The distance between two adjacent side chain target atoms in the original amino acid is calculated to obtain the distance information between the side chain target atoms. The distance information between all side chain target atoms is summed to obtain the side chain distance information distance. side .

[0292] S422-3 calculates the side chain angle information based on the basic information of the side chain target atom.

[0293] Calculate the angles between the three side chain target atoms in the original amino acid to obtain the angle information between the three side chain target atoms, and sum the angle information between all three side chain target atoms to obtain the side chain angle information dirtect side .

[0294] S422-4 sums the side chain dihedral angle information, side chain distance information, and side chain angle information to obtain the side chain attribute information V of the original amino acid side , where V side =di_angle side +distance side +dirtect side .

[0295] S423 concatenates the main chain attribute information and the side chain attribute information to obtain the original amino acid's own characteristic node. Wherein node = Concat (V main , V side ).

[0296] S43 determines the association relationship of each original amino acid according to the basic information of each central carbon atom and the adjacency condition;

[0297] S431 traverses each of the original amino acids and calculates the distance between the current original amino acid and other original amino acids;

[0298] In one embodiment of the present specification, the current original amino acid h is calculated based on the coordinate information.i The central carbon atom (C α ) and the central carbon atom (C α ) as the current original amino acid h i The distance from other original amino acids.

[0299] S432: determining the adjacent amino acid corresponding to the current original amino acid based on the prediction range condition;

[0300] Specifically, based on the original amino acid h i Sort the distances between the original amino acids and the current original amino acid h i The most recent K original amino acids are used as the current original amino acid h i The corresponding adjacent amino acid h j , where i≠j.

[0301] Summarize the current raw amino acids h i All the corresponding adjacent amino acids h j , construct the current original amino acid h i The corresponding adjacent amino acid set neighbor(h i ). K is a hyperparameter determined artificially, and preferably, K=30.

[0302] S433 converts the current original amino acid h i and the corresponding adjacent amino acid h j An association relationship is established between them, and the two original amino acids h with an association relationship are adjacent original amino acids to each other.

[0303] Summarizes the current raw amino acid h i Related associations, construct the current original amino acid h i The corresponding association relationship set co(h i ).

[0304] S44 determines the interaction information between the adjacent original amino acids according to the protein structure file;

[0305] The interaction information includes: chemical bonds and corresponding interaction relationships. The types of chemical bonds include: peptide bonds, hydrogen bonds, hydrophobic interactions, ionic bonds, and disulfide bonds. The interaction relationships include: the types of chemical bonds, the bond lengths of chemical bonds, and the bond angles of chemical bonds.

[0306] S441 determines whether there are peptide bonds between adjacent original amino acids;

[0307] In one embodiment of the present specification, it is determined whether there is a peptide bond between adjacent original amino acids by using coordinate information; specifically, all carbon atoms and nitrogen atoms between adjacent original amino acids are searched, and it is determined whether the distance between the carbon atom in one original amino acid and the nitrogen atom in another original amino acid between the adjacent original amino acids is a first preset distance, and a first determination result is obtained; wherein the first preset distance is

[0308] It is determined whether there is a peptide bond between adjacent original amino acids according to the first determination result.

[0309] Specifically, the adjacent original amino acids include a first original amino acid and a second original amino acid, and it is determined whether the distance between the carbon atom in the first original amino acid and the nitrogen atom in the second original amino acid is a first preset distance; if so, it is determined that there is a peptide bond between the carbon atom in the first original amino acid and the nitrogen atom in the second original amino acid.

[0310] S442 determines whether there are hydrogen bonds between adjacent original amino acids.

[0311] Hydrogen bonds are a common non-covalent interaction between amino acids, which are mainly formed between amide hydrogen and carbonyl oxygen on the peptide chain backbone, and between side chain groups.

[0312] In one embodiment of the present specification, the ATOM and HETATM records in the protein structure file are traversed to find out whether there are potential hydrogen bond acceptors (such as carbonyl oxygen, amide nitrogen, etc.) and potential hydrogen bond donors (such as amide hydrogen, hydroxyl hydrogen, etc.) on the side chains or peptide chain backbones of adjacent original amino acids. Between adjacent original amino acids, if one of the original amino acids has a potential hydrogen bond acceptor and the other original amino acid has a potential hydrogen bond donor, it is determined that there is a possible hydrogen bond pair between the adjacent original amino acids. For possible hydrogen bond pairs, the distance between the potential hydrogen bond donor and the potential hydrogen bond acceptor and the angle between the potential hydrogen bond donor-hydrogen-potential hydrogen bond acceptor are calculated. If the distance between the potential hydrogen bond donor and the potential hydrogen bond acceptor is less than If the angle between the potential hydrogen bond donor-hydrogen-potential hydrogen bond acceptor is close to 180 degrees, it is determined that there is a hydrogen bond between the potential hydrogen bond acceptor and the potential hydrogen bond donor.

[0313] S443 determines whether there is a hydrophobic interaction between adjacent original amino acids.

[0314] Hydrophobic interaction refers to the interaction between hydrophobic groups (such as non-polar side chains) in proteins due to avoiding contact with water.

[0315] In one embodiment of the present specification, it is determined whether the amino acid residue contains a hydrophobic group according to the type of the adjacent original amino acid. Common hydrophobic amino acids include alanine (Ala), valine (Val), leucine (Leu), isoleucine (Ile), proline (Pro), phenylalanine (Phe), tryptophan (Trp) and methionine (Met), etc. The side chains of these amino acids usually contain non-polar hydrocarbon groups or aromatic rings.

[0316] According to the protein structure file, it is determined whether the hydrophobic groups between adjacent original amino acids are close to each other to form an aggregation area inside the protein; if so, the arrangement of water molecules around the aggregation area is analyzed. If the distribution of water molecules around the aggregation area is restricted or arranged in an orderly manner, it is determined that a hydrophobic effect exists.

[0317] S444 determines whether there are ionic bonds between adjacent original amino acids.

[0318] Ionic bonds (salt bridges) Under certain circumstances, ionic bonds (salt bridges) can form between the side chains of acidic amino acids (such as aspartic acid and glutamic acid) and basic amino acids (such as lysine and arginine).

[0319] In one embodiment of the present specification, based on the element type of each atom in the protein structure file, the elements that may form ions are searched; based on the electronegativity and ionization tendency of the elements, the elements that may form ion pairs are searched to determine the charged groups. Charged groups include: positively charged groups (such as the ε-amino group of lysine and the guanidine group of arginine) and negatively charged groups (such as the β-carboxyl group of aspartic acid and the α-carboxyl group of glutamic acid). The distance between the charged groups is calculated. If the distance between the charged groups is arrive If there are two charged groups with opposite charges, it is considered that there is an ionic bond between the adjacent amino acids.

[0320] S445 determines whether there are disulfide bonds between adjacent native amino acids.

[0321] A disulfide bond is a covalent bond formed by the oxidation of the sulfhydryl groups of two cysteine ​​residues. It has strong stability and can form a cross-linked structure in proteins.

[0322] Determine whether there are cysteine ​​residues in adjacent original amino acids. If so, determine the distance between the sulfhydryl groups of the cysteine ​​residues based on the coordinate information. If the distance between the sulfhydryl groups of two cysteine ​​residues is less than If there is a covalent bond, it is considered that there is a disulfide bond between adjacent amino acids.

[0323] Of course, you can also search for SSBOND records in the connection annotation section of the protein structure file. Each SSBOND record will contain the sequence number (resSeq) and chain identifier (chainID) of two or more cysteine ​​residues.

[0324] Verified based on SSBOND records.

[0325] S446 uses molecular visualization software or computational chemistry tools to display chemical bonds and corresponding interaction relationships.

[0326] Molecular visualization software includes but is not limited to: PyMOL software, VMD (Visual Molecular Dynamics) software, Maestro software ( Suite).

[0327] Computational chemistry tools include but are not limited to: AutoDock Vina molecular docking software, Gaussian and other quantum chemistry software.

[0328] S45 determines the association features between the adjacent original amino acids based on the interaction information.

[0329] S451 determines the associated dihedral angle information between adjacent original amino acids according to the interaction relationship.

[0330] In one embodiment of the present specification, the chemical bonds between adjacent original amino acids are searched, and two atoms to be detected are determined based on the two ends of the chemical bonds. The two atoms to be detected are: a first atom to be detected and a second atom to be detected.

[0331] In one embodiment of the present specification, the first auxiliary atom closest to the first atom to be detected is searched, the first auxiliary atom is connected to the first atom to be detected, and the first auxiliary atom and the first atom to be detected are located in the same original amino acid. The second auxiliary atom closest to the second atom to be detected is searched, the second auxiliary atom is connected to the second atom to be detected, and the second auxiliary atom and the second atom to be detected are located in the same original amino acid.

[0332] The dihedral angle information corresponding to the chemical bond is determined according to the first atom to be measured, the second atom to be measured, the first auxiliary atom, and the second auxiliary atom. Specifically, the fourth vector from the first atom to be measured to the first auxiliary atom is calculated, the fifth vector from the first atom to be measured to the second atom to be measured is calculated, and the sixth vector from the second atom to be measured to the second auxiliary atom is calculated.

[0333] Use the cross product of the vectors to calculate the normal vector of the first plane and the normal vector of the second plane. Multiply the fourth vector by the fifth vector to get the normal vector of the first plane. Multiply the fifth vector by the sixth vector to get the normal vector of the second plane. The angle between the normal vector of the first plane and the normal vector of the second plane is used as the dihedral angle information corresponding to the chemical bond.

[0334] Sum the dihedral angle information corresponding to all chemical bonds to obtain the associated dihedral angle information di_a ij .

[0335] S452 determines the associated distance information according to the bond length of the chemical bond.

[0336] In one embodiment of the present specification, the chemical bonds between adjacent original amino acids are searched, and the bond lengths of the chemical bonds are used as the distance information corresponding to the chemical bonds.

[0337] Sum the distance information corresponding to all chemical bonds to obtain the associated distance information dis ij .

[0338] S453 determines the associated angle information according to the bond angle.

[0339] In one embodiment of the present specification, the chemical bonds between adjacent original amino acids are searched, and the bond angles of the chemical bonds are used as the angle information corresponding to the chemical bonds. The angle information corresponding to all chemical bonds is summed to obtain the associated angle information dir ij .

[0340] Of course, the present invention can also calculate the associated dihedral angle information di_a between adjacent original amino acids by using the DSSP tool. ij , associated distance information dis ij and associated angle information dir ij .

[0341] S454 sums the associated dihedral angle information, associated distance information, and associated angle information to obtain the associated feature edge between adjacent original amino acids. ij ,edge ij =di_a ij +dis ij +dir ij .

[0342] S5 masks the intrinsic characteristics of the amino acid to be designed;

[0343] The self-features of the original amino acid at the target position are masked, and a mask value is set for the self-features in the masked amino acid. In one embodiment of the specification, the mask value is an all-zero vector.

[0344] S6 updates all the characteristic information through a protein structure prediction model; based on the update result, predicts the category of the amino acid to be designed, and obtains the structure prediction result of the target protein.

[0345] S61 preprocesses the self-features and the associated features to obtain preprocessed self-features and associated features;

[0346] In one embodiment of the present specification, the own features and the associated features are initialized and standardized. Specifically, the own features and the feature dimensions and the feature dimensions of the associated features are all 128 dimensions.

[0347] S62 updates the preprocessed self-features and associated features in sequence according to the arrangement order of the original amino acids in the original amino acid sequence to obtain updated node features and updated associated features;

[0348] For each original amino acid, do the following:

[0349] S621 extracts and updates the intermediate value by combining the preprocessed self-features and the associated features;

[0350] S621-1 The current original amino acid h i The characteristics of the adjacent amino acids and the current original amino acid h i Perform feature splicing with the correlation features between adjacent amino acids to obtain a splicing result;

[0351] Determine the current original amino acid h i , traverse the central amino acid h i For each adjacent amino acid h j , the characteristic node of the same adjacent amino acid j With the current original amino acid h i The edge feature of the adjacent amino acid i，j Splice to obtain adjacent amino acids h j The concatenation result of Concat(node j ,edge ij ).

[0352] S621-2 through learnable weight α j For the adjacent amino acid h j The splicing results are weighted to obtain each adjacent amino acid h j The weighted result α j Concat(node j ,edge ij ).

[0353] S621-3 for the current original amino acid h i All the corresponding adjacent amino acids h j The weighted results are summed to obtain a weighted summation result.

[0354] S621-4 uses the product of the weighted sum result and the activation function σ as an updated intermediate value;

[0355] The activation function σ is applied to the weighted summation result, and by introducing nonlinearity, the network can learn more complex feature representations. The activation function is preferably ReLU. The update intermediate value is the Update function, which also considers the central amino acid h i Its own characteristics and the information of adjacent nodes (through edges).

[0356] S622 updates the own features and associated features according to the updated intermediate value.

[0357] In each graph network layer, according to the central amino acid h i The node's own characteristics i and update the intermediate value Update(node i ,edge i ) are added to obtain the central amino acid h i The new self-feature node′ i The residual connection helps prevent the gradient vanishing / exploding problem and makes it easier for the network to learn the identity mapping. i =node i +Update(node i ,edge i ).

[0358] According to the central amino acid h i The new self-feature node′ i , for each associated feature edge in the associated feature set i，j Update to get the new associated feature edge′ ij Among them, edge′ i，j =edge ij +Update(node i ,edge ij ).

[0359] After performing the above operations on all central amino acids, all self-features and associated features are updated.

[0360] After updating the self-features and associated features corresponding to all original amino acids, S52 is executed repeatedly according to the preset update times, and the output result of the last update is used as the input result of the next update to obtain the final update result and determine the updated self-features and associated features.

[0361] The preset number of updates is the depth of the graph network layer. That is, Among them, h represents the depth of the graph network layer, that is, the number of graph network layers.

[0362] S63 decodes the updated self-features and predicts the target amino acid at the target position;

[0363] S631 converts the self-features corresponding to the target position to obtain the category probability distribution of the target amino acid corresponding to the target position;

[0364] The high-dimensional self-features corresponding to the target position are converted into a vector of fixed length. Specifically, the self-features of each original amino acid are converted into a 20-dimensional vector, and each dimensional vector represents an amino acid category. Each element of the 20-dimensional vector can be regarded as the score or probability of the corresponding amino acid category, which can be further converted into a probability distribution through activation functions such as the softmax function, thereby predicting the type of amino acid at each position.

[0365] S632 determines the prediction result based on the category probability distribution.

[0366] In one embodiment of the present specification, the amino acid category with the highest probability is selected from the category probability distribution as the prediction result.

[0367] Considering that the above method is simple and direct, but may ignore low-probability but reasonable options. Therefore, in another embodiment of the present specification, the probability sampling of the sampling temperature is set: by setting the sampling temperature (temperature), sampling is performed according to the probability distribution.

[0368] Specifically, set the sampling temperature T to adjust the probability distribution of the softmax output. The higher the temperature, the more random the sampling results; the lower the temperature, the more concentrated the sampling results are in the high probability category.

[0369] The amino acid classes were randomly sampled using the adjusted probability distribution. This approach can introduce diversity to a certain extent and prevent the model from being too deterministic.

[0370] S64 combines the inherent characteristics of the target amino acids to obtain the structural prediction results of the target protein.

[0371] S641, determining a target amino acid at a target position according to the probability distribution of the category of the target amino acid and / or the feedback information of the user;

[0372] By default, the category with the highest probability in the probability distribution is the final category of the target amino acid. Of course, the final category of the target amino acid can also be determined by the user. Specifically, the prediction result of the target amino acid is displayed to the user on the second display page, and a feedback component is provided.

[0373] Based on the user's interaction on the feedback component, feedback information from the user is obtained, and based on the feedback information, a final category of the target amino acid is determined. Based on the final category, a target amino acid at a target position is determined.

[0374] S642 combines the characteristics of the target amino acid to obtain predicted properties of the predicted atom;

[0375] Combine the target amino acid and the self characteristics of the target amino acid, map the self characteristics of the target amino acid, and obtain the predicted attributes of the predicted atom. The predicted atom is the side chain target atom of the target amino acid. The predicted attributes include: side chain torsion information of the side chain target atom. In the embodiment of the present specification, the side chain torsion information includes: torsion angle (side chain dihedral angle information, chi angle) and bond angle (θ).

[0376] S643, according to the original amino acid corresponding to the target position, searching for the main chain target atom corresponding to the original amino acid, and using the coordinate information of the main chain target atom as the coordinate information of the backbone atom;

[0377] S644 determines the positioning order of the predicted atoms according to the coordinate information of the predicted atoms and the central carbon atom of the target amino acid;

[0378] Among them, the fewer the connections with the central carbon atom of the target amino acid, the higher the predicted atom's positioning order.

[0379] According to the original amino acid corresponding to the target position, the main chain target atom corresponding to the original amino acid is searched, and the coordinate information of the central carbon atom therein is used as the coordinate information of the central carbon atom of the target amino acid.

[0380] S645, determining the coordinate information of the predicted atoms in sequence according to the positioning order;

[0381] Specifically, the following operations are performed on each predicted atom one by one in the positioning order:

[0382] Determine the current bond length based on the type of chemical bond between the current predicted atom and the previous atom connected to it. Determine the spherical range in the three-dimensional space where the current predicted atom is located, with the previous atom as the center and the current bond length as the radius. Determine the unique coordinate information of the current predicted atom in the spherical range based on the torsion angle and included angle of the current predicted atom.

[0383] S646 summarizes the coordinate information of all predicted atoms and the coordinate information of all backbone atoms to obtain the structural prediction result of the target protein.

[0384] According to the coordinate information of all predicted atoms and the coordinate information of all skeleton atoms, the three-dimensional structure of the target protein is constructed as the structure prediction result of the target protein.

[0385] In the present invention, the protein structure file is used as data input, a single amino acid is used as a node, and the connection between amino acids is used as an edge; structural information such as dihedral angles and distances are extracted as initial node features and initial edge features, and adjacent nodes and edges between adjacent nodes are selected to construct a graph network. The target node features and edge features are updated by adjacent node features, adjacent edge features and self-nodes, and the amino acid sequence probability and corresponding side chain structure coordinates of the position to be predicted are obtained as results. Compared with the prior art, the present invention has a higher sequence recovery rate, a higher design sequence diversity, and a more reasonable side chain structure conformation.

[0386] Compared with the existing technology, the present invention combines the characteristics of adjacent amino acids for prediction, and its sequence recovery rate is higher and the interpretability is stronger. By screening the generated new protein sequences, the corresponding predicted structural information is provided, which facilitates further verification of the rationality of the sequence design, improves the efficiency of downstream bioengineering, and provides a more effective tool for protein space exploration tasks.

[0387] Figure 6 A schematic diagram of the structure of a protein side chain conformation prediction system provided in an embodiment of this specification, the system comprising:

[0388] The first acquisition module 620 is used to acquire a protein structure file to obtain an original amino acid sequence;

[0389] The second acquisition module 630 is used to acquire user design requirements, wherein the user design requirements include: at least one target position; the target position includes: the original position of the amino acid to be designed in the original amino acid sequence;

[0390] A feature extraction module 640 is used to extract feature information of each original amino acid based on the target position and the protein structure file; the feature information includes: the original amino acid's own features and the correlation features between the original amino acids;

[0391] A masking module 650 is used to mask the characteristics of the amino acid to be designed;

[0392] The multi-task prediction module 660 is used to update all the feature information through a protein structure prediction model; based on the update result, the category of the amino acid to be designed is predicted to obtain the structure prediction result of the target protein.

[0393] Optionally, the feature extraction module 640 includes:

[0394] A basic information extraction submodule, used to extract basic information of target atoms from the protein structure file;

[0395] A self-feature construction submodule, used to construct the self-feature of the original amino acid according to the basic information of the target atom;

[0396] An interaction confirmation submodule, used for determining the interaction information between adjacent original amino acids according to the protein structure file;

[0397] A correlation feature construction submodule, used for determining the correlation features between adjacent original amino acids based on the interaction information;

[0398] The target atoms include main chain target atoms and side chain target atoms;

[0399] Optionally, the first acquisition module 620 includes:

[0400] A file acquisition submodule, used to acquire the protein structure file uploaded by the user;

[0401] The file parsing submodule is used to perform a first analysis on the protein structure file to obtain an original amino acid sequence, wherein the protein structure file is verified according to a verification condition; if the verification passes, a linear arrangement order of each of the original amino acids is extracted from the protein structure file as the original amino acid sequence.

[0402] Optionally, the self-feature construction submodule includes:

[0403] A main chain attribute information construction unit, used to construct the main chain attribute information of the original amino acid according to the basic information of the main chain target atom;

[0404] A side chain attribute information construction unit, used to construct the side chain attribute information of the original amino acid according to the basic information of the side chain target atom;

[0405] The attribute information splicing unit is used to splice the main chain attribute information and the side chain attribute information to obtain the original amino acid's own characteristics.

[0406] Optionally, the feature extraction module 640 further includes: an association relationship confirmation submodule;

[0407] The association confirmation submodule includes:

[0408] A traversal unit, used for traversing each of the original amino acids and calculating the distance between the current original amino acid and other original amino acids;

[0409] A screening unit, used for determining the adjacent amino acid corresponding to the current original amino acid based on the prediction range condition;

[0410] The association confirmation unit is used to establish an association relationship between the current original amino acid and the corresponding adjacent amino acid, and the two original amino acids with an association relationship are each other's adjacent original amino acids.

[0411] Optionally, the multi-task prediction module 660 includes:

[0412] A preprocessing submodule, used for preprocessing the self-features and the associated features to obtain the preprocessed self-features and associated features;

[0413] An updating submodule, used to update the preprocessed self-features and associated features in sequence according to the arrangement order of the original amino acids in the original amino acid sequence, to obtain updated node features and updated associated features;

[0414] The category prediction submodule is used to decode the updated self-features and predict the target amino acid at the target position;

[0415] The structure prediction submodule is used to combine the characteristics of the target amino acids to obtain the structure prediction results of the target protein.

[0416] Optionally, the updating submodule includes:

[0417] A feature splicing unit, used for performing feature splicing on the features of the adjacent amino acids of the current original amino acid and the associated features between the current original amino acid and the adjacent amino acids to obtain a splicing result;

[0418] A weighting unit, used for weighting the splicing result of the adjacent amino acids by a learnable weight to obtain a weighted result of each adjacent amino acid;

[0419] A summing unit, used for summing the weighted results of all adjacent amino acids corresponding to the current original amino acid to obtain a weighted summation result;

[0420] An updating unit is used to take the product of the weighted sum result and the activation function as an updated intermediate value; and update the own features and associated features according to the updated intermediate value.

[0421] Optionally, the structure prediction submodule includes:

[0422] a target amino acid determination unit, configured to determine the target amino acid at the target position according to the probability distribution of the category of the target amino acid and / or the feedback information of the user;

[0423] A prediction unit, used to obtain predicted properties of predicted atoms by combining the characteristics of the target amino acid;

[0424] A first coordinate determination unit is used to search for a main chain target atom corresponding to the original amino acid according to the original amino acid corresponding to the target position, and use the coordinate information of the main chain target atom as the coordinate information of the backbone atom;

[0425] A positioning unit, used to determine the positioning order of the predicted atoms according to the coordinate information of the predicted atoms and the central carbon atom of the target amino acid;

[0426] A second coordinate determination unit, used to determine the coordinate information of the predicted atoms in sequence according to the positioning order;

[0427] The summary unit is used to summarize the coordinate information of all predicted atoms and the coordinate information of all skeleton atoms to obtain the structure prediction result of the target protein.

[0428] The functions of the device in the embodiment of the present invention have been described in the above method embodiment, so for details not provided in the description of this embodiment, please refer to the relevant description in the above embodiment, and no further description will be given here.

[0429] The embodiments of the present invention may be provided as a method, an apparatus or a computer program product. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects.

[0430] Figure 7 The present invention provides a schematic diagram of the structure of an electronic device provided in an embodiment of the present specification. Specifically, the present invention provides an electronic device, which can implement each process and / or box in the aforementioned flowchart and / or block diagram, and the combination of the processes and / or boxes in the flowchart and / or block diagram, by the executable instructions of the computer program 703. The electronic device of the present invention can execute the executable instructions of the computer program 703 through the processor 702 of a computer or other programmable data processing device to implement one or more processes and / or boxes. Figure 1 A function specified in one or more boxes.

[0431] The executable instructions of the computer program 703 may also be stored in a computer-readable memory 701 that can guide a computer or other programmable data processing device to work in a specific manner, and executed by the processor 702 to implement corresponding functions.

[0432] Figure 8 A schematic diagram of a computer-readable storage medium provided for an embodiment of the present specification. The present invention may be in the form of a computer program product implemented on one or more computer-readable storage media 800 (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing a computer program 703.

[0433] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for predicting protein side chain conformation, characterized in that: include; Obtain protein structure files and obtain original amino acid sequences; Obtaining user design requirements, wherein the user design requirements include: at least one target position; the target position includes: the original position of the amino acid to be designed in the original amino acid sequence; Based on the target position and the protein structure file, extract the characteristic information of each original amino acid; the characteristic information includes: the original amino acid's own characteristics and the correlation characteristics between the original amino acids; specifically, extract the basic information of the target atom from the protein structure file; construct the original amino acid's own characteristics according to the basic information of the target atom; determine the interaction information between the adjacent original amino acids according to the protein structure file; determine the correlation characteristics between the adjacent original amino acids based on the interaction information; the target atoms include main chain target atoms and side chain target atoms; The self-features of the amino acids to be designed are masked; all the feature information is updated through a protein structure prediction model; based on the update result, the category of the amino acids to be designed is predicted to obtain a structure prediction result of the target protein.

2. A protein side chain conformation prediction method according to claim 1, characterized in that: The step of obtaining the protein structure file to obtain the original amino acid sequence includes: Obtaining the protein structure file uploaded by the user; The protein structure file is first parsed to obtain an original amino acid sequence, wherein the protein structure file is verified according to a verification condition; if the verification passes, a linear arrangement order of each of the original amino acids is extracted from the protein structure file as the original amino acid sequence.

3. A protein side chain conformation prediction method according to claim 1, characterized in that: The step of constructing the original amino acid's own characteristics according to the basic information of the target atom includes: Constructing the main chain attribute information of the original amino acid according to the basic information of the main chain target atom; Constructing the side chain attribute information of the original amino acid according to the basic information of the side chain target atom; The main chain attribute information and the side chain attribute information are spliced ​​to obtain the original amino acid's own characteristics.

4. A protein side chain conformation prediction method according to claim 1, characterized in that: The extracting characteristic information of each original amino acid based on the target position and the protein structure file also includes: Traversing each of the original amino acids, and calculating the distance between the current original amino acid and other original amino acids; Based on the prediction range condition, determining the adjacent amino acid corresponding to the current original amino acid; An association relationship is established between the current original amino acid and the corresponding adjacent amino acid, and the two original amino acids with an association relationship are each other's adjacent original amino acids.

5. A protein side chain conformation prediction method according to claim 1, characterized in that: The method comprises updating all the feature information by using a protein structure prediction model; Based on the updated results, the types of amino acids to be designed are predicted, and the structural prediction results of the target protein are obtained, including: Preprocessing the self-features and the associated features to obtain preprocessed self-features and associated features; According to the arrangement order of the original amino acids in the original amino acid sequence, the preprocessed self-features and associated features are updated in sequence to obtain updated node features and updated associated features; Decode the updated self-features and predict the target amino acid at the target position; Combined with the inherent characteristics of the target amino acid, the structure prediction results of the target protein are obtained.

6. A protein side chain conformation prediction method according to claim 5, characterized in that: The preprocessed self-features and associated features are sequentially updated according to the arrangement order of the original amino acids in the original amino acid sequence to obtain updated node features and updated associated features, including: Perform feature splicing on the features of the adjacent amino acids of the current original amino acid and the associated features between the current original amino acid and the adjacent amino acids to obtain a splicing result; Weighting the splicing results of the adjacent amino acids by using learnable weights to obtain a weighted result for each adjacent amino acid; Summing the weighted results of all adjacent amino acids corresponding to the current original amino acid to obtain a weighted sum result; The product of the weighted summation result and the activation function is used as an updated intermediate value; and according to the updated intermediate value, the own features and associated features are updated.

7. A protein side chain conformation prediction method according to claim 5, characterized in that: The method of combining the characteristics of the target amino acid to obtain the structure prediction result of the target protein includes: Determining a target amino acid at a target position according to the probability distribution of the category of the target amino acid and / or the feedback information of the user; Combining the characteristics of the target amino acid, the predicted properties of the predicted atom are obtained; According to the original amino acid corresponding to the target position, searching for the main chain target atom corresponding to the original amino acid, and using the coordinate information of the main chain target atom as the coordinate information of the backbone atom; Determining the positioning order of the predicted atoms according to the coordinate information of the predicted atoms and the central carbon atom of the target amino acid; According to the positioning order, the coordinate information of the predicted atoms is determined in sequence; The coordinate information of all predicted atoms and the coordinate information of all backbone atoms are summarized to obtain the structural prediction result of the target protein.

8. A protein side chain conformation prediction system, characterized in that: include: The first acquisition module is used to acquire the protein structure file and obtain the original amino acid sequence; The second acquisition module is used to acquire user design requirements, wherein the user design requirements include: at least one target position; the target position includes: the original position of the amino acid to be designed in the original amino acid sequence; A feature extraction module is used to extract feature information of each original amino acid based on the target position and the protein structure file; the feature information includes: the original amino acid's own features and the correlation features between the original amino acids; specifically, the basic information of the target atom is extracted from the protein structure file; the original amino acid's own features are constructed according to the basic information of the target atom; the interaction information between adjacent original amino acids is determined according to the protein structure file; the correlation features between adjacent original amino acids are determined based on the interaction information; the target atoms include main chain target atoms and side chain target atoms; A masking module, used for masking the characteristics of the amino acid to be designed; The multi-task prediction module is used to update all the feature information through a protein structure prediction model; based on the update result, the category of the amino acid to be designed is predicted to obtain the structure prediction result of the target protein.

9. An electronic device, wherein: The electronic device includes: processor; and, A memory storing computer executable instructions which, when executed, cause the processor to perform a method according to any one of claims 1-7.

10. A computer-readable storage medium, wherein: The computer-readable storage medium stores one or more programs, and when the one or more programs are executed by a processor, the method of any one of claims 1 to 7 is implemented.

Images