A method for predicting a protein distance map, a storage medium, and an electronic device

Abstract

In the method for predicting a protein distance map, the storage medium and the electronic equipment provided in the specification, the protein residue sequence is confirmed, and a first attention map is generated. By a triangular attention module, residues satisfying the triangular inequality are weighted to determine a second attention map. And by a residue mixing module, the local features of residues with specified structural relationships are weighted. Finally, using a reduction module, the output of the residue mixing module is superimposed with the first attention map to highlight the contrast effect, and the size transformation output distance map of the regression prediction map is output. By determining the residues satisfying the triangular constraint and highlighting the distance map of the specified two-dimensional structure and the hyper two-dimensional structure, the problem of invalid prediction results in the separate prediction of protein residue distance is solved, thereby improving the efficiency of protein synthesis.

Description

Technical Field

[0001] This specification relates to the field of bioinformatics, and in particular to a method, storage medium, and electronic device for predicting protein distance maps. Background Technology

[0002] In recent years, with the development of bioinformatics, the technology for artificially synthesizing proteins has also made rapid progress. Proteins are the main carriers of life activities, performing many important biological functions such as catalytic regulation, immune response, and cell signal transduction. Since the biological functions of proteins are determined by their specific three-dimensional spatial structures, accurately understanding protein structure is of great significance for understanding protein function. This is crucial for drug development, vaccine therapy, and the study of biological functions.

[0003] In existing technologies, the three-dimensional structure of proteins is generally obtained by first predicting the distances between residues in the protein sequence. However, the protein structure predicted by this method may not satisfy the inherent characteristics of protein structure, thus preventing the predicted protein from being synthesized. For example, simply obtaining the distances between multiple sets of residues may not satisfy the requirement that at least one set of residues in the distances must meet the triangle inequality.

[0004] It is evident that existing methods for predicting protein structure neglect the structural characteristics of the protein itself, thus failing to provide accurate predictions. In contrast, protein distance maps, which not only display the protein residues that satisfy the triangle inequality but also present a two-dimensional view of the protein's local structure, allow for a more accurate and intuitive observation of the protein's three-dimensional structure. Therefore, this specification provides a method for predicting protein distance maps. Summary of the Invention

[0005] This specification provides a method for predicting protein distance maps, a storage medium, and an electronic device to partially solve the aforementioned problems existing in the prior art.

[0006] The following technical solution is adopted in this specification:

[0007] A method for predicting protein distance maps includes:

[0008] Determine the residue sequence of the target protein;

[0009] The residue sequence is input into the trained distance map prediction model, and the residue sequence is weighted by the attention subnetwork of the distance map prediction model to determine the first attention map corresponding to the residue sequence.

[0010] The first attention map is preprocessed, and the result of the first preprocessing is input into the triangular attention module of the distance map prediction model to predict the distance between each residue, determine each residue whose distance between residues conforms to the triangular distance constraint relationship, and perform attention weighting on the determined residues that satisfy the triangular distance constraint relationship to obtain the second attention map corresponding to the first attention map.

[0011] The second attention map is subjected to a second preprocessing, and the second preprocessing result is input into the residual mixing module of the distance map prediction model. The residual mixing module filters the data, and the size of the filtered result is adjusted according to the second attention map. The residues with specified structural relationships in the adjusted filtered result are weighted by local features to obtain the third attention map corresponding to the second attention map.

[0012] The first attention map and the third attention map are input into the residual convolution module of the distance map prediction model. The third attention map is transposed and convolved by the residual convolution module. The size of the filtering result is adjusted according to the first attention map. The adjusted result is superimposed with the first attention map to obtain the fourth attention map corresponding to the third attention map.

[0013] The fourth attention map is input into the regression prediction module of the distance map prediction model. The regression prediction module normalizes the fourth attention map to obtain the protein distance map corresponding to the fourth attention map.

[0014] Optionally, the first attention map undergoes a first preprocessing step, specifically including:

[0015] The first attention map is input into the feedforward convolutional module of the distance map prediction model, and the size of the first attention map is reduced through the pooling layer of the feedforward convolutional module.

[0016] The convolutional layer of the feedforward convolutional module adjusts and reduces the dimension of the attention map input to the triangular attention module of the distance map prediction model, and determines the adjusted result as the first preprocessing result.

[0017] Optionally, the second attention map undergoes a second preprocessing, specifically including:

[0018] The third attention map is input into the feedforward mixing module of the distance map prediction model. The third attention map is compressed through the lightweight layer of the feedforward mixing module so that the residual mixing module of the distance map prediction model can perform local feature weighting.

[0019] The size of the compressed third attention map is reduced through the pooling layer of the feedforward mixing module;

[0020] The convolutional layer of the feedforward convolutional module adjusts the dimension of the attention map input to the residual mixing module of the distance map prediction model, based on the dimension of the input attention map, and determines the adjusted result as the preprocessing result.

[0021] Optionally, filtering via the residual mixing module and adjusting the size of the filtering result based on the second attention map specifically includes:

[0022] The transpose layer of the residual mixing module reduces the dimension of the second preprocessing result and increases its size.

[0023] The size of the transposed layer processing result is adjusted according to the difference between the transposed layer processing result and the second attention map through the filling layer of the residual mixing module until it is the same as the size of the second attention map.

[0024] The adjustment result is superimposed on the third attention map through the superposition layer of the residual mixing module to obtain the adjusted filtering result.

[0025] Optionally, the first attention map and the third attention map are input into the residual convolution module of the distance map prediction model. The third attention map is transposed and convolved by the residual convolution module. The filtering result is then resized according to the first attention map, and the resized result is superimposed on the first attention map to obtain a fourth attention map. Specifically, this includes:

[0026] The third attention map is input into the residual convolution module of the distance map prediction model. The dimension of the third attention map is adjusted according to the dimension of the first attention map through the transpose layer of the residual convolution module.

[0027] The dimension adjustment result is adjusted based on the size of the first attention map by using the filling layer of the residual convolution module, according to the difference between the dimension adjustment result and the size of the first attention map.

[0028] The size adjustment result is superimposed on the first attention map through the superposition layer of the residual convolution module;

[0029] The convolutional layer of the residual convolution module is used to convolve the superimposed result, and the convolution result is determined to be the fourth attention map corresponding to the third attention map.

[0030] Optionally, before performing the first preprocessing on the first attention map, the method further includes:

[0031] The first attention map is input into the group pooling module of the distance map prediction model. The group pooling module divides the first attention map into a preset number of sub-attention maps according to a preset number.

[0032] Then, the maximum value of each sub-attention map is extracted to determine the sub-feature map corresponding to each sub-attention map;

[0033] By concatenating the sub-feature maps, a first attention map with a predetermined number of dimensions is obtained.

[0034] Optionally, the method further includes obtaining the protein distance map corresponding to the fourth attention map:

[0035] The regression prediction module results are input into the size adjustment module of the distance map prediction model. The size adjustment module scales the attention map values ​​of the regression prediction module results as a whole to obtain a protein residue distance map that is easy to observe.

[0036] This specification provides a model training method for predicting protein distance maps, including:

[0037] Determine the residue sequence of the target protein;

[0038] The residue sequence is input into the trained distance map prediction model, and the residue sequence is weighted by the attention subnetwork of the distance map prediction model to determine the first attention map corresponding to the residue sequence.

[0039] The first attention map is preprocessed, and the result of the first preprocessing is input into the triangular attention module of the distance map prediction model to predict the distance between each residue, determine each residue whose distance between residues conforms to the triangular distance constraint relationship, and perform attention weighting on the determined residues that satisfy the triangular distance constraint relationship to obtain the second attention map corresponding to the first attention map.

[0040] The second attention map is subjected to a second preprocessing, and the second preprocessing result is input into the residual mixing module of the distance map prediction model. The residual mixing module filters the data, and the size of the filtered result is adjusted according to the second attention map. The residues with specified structural relationships in the adjusted filtered result are weighted by local features to obtain the third attention map corresponding to the second attention map.

[0041] The first attention map and the third attention map are input into the residual convolution module of the distance map prediction model. The third attention map is transposed and convolved by the residual convolution module. The size of the filtering result is adjusted according to the first attention map. The adjusted result is superimposed with the first attention map to obtain the fourth attention map corresponding to the third attention map.

[0042] The fourth attention map is input into the regression prediction module of the distance map prediction model. The regression prediction module normalizes the fourth attention map to obtain the protein distance map corresponding to the fourth attention map.

[0043] The regression prediction module of the distance map prediction model is supervised learning using the labels of the protein distance map training set, and the model parameters are updated using a preset optimization algorithm based on the calculated loss value.

[0044] The above training method is iterated and the current iteration number is determined; if the current iteration number reaches the preset number, the iteration is stopped and the target model is obtained.

[0045] Optionally, before performing the first preprocessing on the first attention map, the training method further includes:

[0046] The first attention map is input into the group pooling module of the distance map prediction model. The group pooling module divides the first attention map into a preset number of sub-attention maps according to a preset number.

[0047] Then, the maximum value of each sub-attention map is extracted to determine the sub-feature map corresponding to each sub-attention map;

[0048] By concatenating the sub-feature maps, a first attention map with a predetermined number of dimensions is obtained.

[0049] This specification provides an apparatus for predicting protein distance maps, comprising:

[0050] The sequence acquisition module determines the residue sequence of the target protein;

[0051] The sequence confirmation module inputs the residue sequence into the trained distance map prediction model, and performs attention weighting on the residue sequence through the attention sub-network of the distance map prediction model to determine the first attention map corresponding to the residue sequence.

[0052] The triangular attention module performs a first preprocessing on the first attention map, inputs the first preprocessing result into the triangular attention module of the distance map prediction model, predicts the distance between each residue, determines each residue whose distance between residues conforms to the triangular distance constraint relationship, and performs attention weighting on the determined residues that satisfy the triangular distance constraint relationship to obtain the second attention map corresponding to the first attention map.

[0053] The local structure attention module performs a second preprocessing on the second attention map, inputs the second preprocessing result into the residual mixing module of the distance map prediction model, filters the result through the residual mixing module, adjusts the size of the filtered result according to the second attention map, and performs local feature weighting on the residues with specified structural relationships in the adjusted filtered result to obtain the third attention map corresponding to the second attention map.

[0054] The attention map restoration module inputs the first attention map and the third attention map into the residual convolution module of the distance map prediction model. The residual convolution module transposes and convolves the third attention map, adjusts the size of the filtering result according to the first attention map, and superimposes the adjustment result with the first attention map to obtain the fourth attention map corresponding to the third attention map.

[0055] The distance map output module inputs the fourth attention map into the regression prediction module of the distance map prediction model. The regression prediction module normalizes the fourth attention map to obtain the protein distance map corresponding to the fourth attention map.

[0056] This specification provides an apparatus for training a protein distance map prediction module, comprising:

[0057] The sequence acquisition training module determines the residue sequence of the target protein;

[0058] The sequence confirmation training module inputs the residue sequence into the trained distance map prediction model, and performs attention weighting on the residue sequence through the attention sub-network of the distance map prediction model to determine the first attention map corresponding to the residue sequence.

[0059] The triangular attention training module performs a first preprocessing on the first attention map, inputs the first preprocessing result into the triangular attention module of the distance map prediction model, predicts the distance between each residue, determines each residue whose distance between residues conforms to the triangular distance constraint relationship, and performs attention weighting on the determined residues that satisfy the triangular distance constraint relationship to obtain the second attention map corresponding to the first attention map.

[0060] The local structure attention training module performs a second preprocessing on the second attention map, inputs the second preprocessing result into the residual mixing module of the distance map prediction model, filters the result through the residual mixing module, adjusts the size of the filtered result according to the second attention map, and performs local feature weighting on the residues with specified structural relationships in the adjusted filtered result to obtain the third attention map corresponding to the second attention map.

[0061] The attention map restoration training module inputs the first attention map and the third attention map into the residual convolution module of the distance map prediction model. The residual convolution module transposes and convolves the third attention map, adjusts the size of the filtering result according to the first attention map, and superimposes the adjustment result with the first attention map to obtain the fourth attention map corresponding to the third attention map.

[0062] The distance map output training module inputs the fourth attention map into the regression prediction module of the distance map prediction model. Through the regression prediction module, the fourth attention map is normalized to obtain the protein distance map corresponding to the fourth attention map.

[0063] The parameter update training module uses the labels of the protein distance map training set to supervise the regression prediction module of the distance map prediction model, and updates the model parameters using a preset optimization algorithm based on the calculated loss value.

[0064] The iterative training module iterates the above training method and determines the current iteration number; if the current iteration number reaches the preset number, the iteration stops and the target model is obtained.

[0065] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method.

[0066] This specification provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for predicting protein distance maps.

[0067] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects:

[0068] In a method for predicting protein distance maps provided in this specification, the protein residue sequence is identified, and a first attention map is generated. A second attention map is determined by weighting residues that satisfy the triangle inequality using a triangular attention module. A residue mixing module is then used to weight residues with specified structural relationships based on their local features. Finally, a reduction module is used to overlay the output of the residue mixing module with the first attention map to highlight the contrast, and the distance map is output by regressing the size of the predicted map.

[0069] As can be seen from the above methods, emphasizing residues satisfying the triangle inequality in the protein distance map helps professionals determine whether a predicted protein can be synthesized. Emphasizing residues satisfying specific structures in the protein distance map helps professionals understand the helical structure of proteins from their two-dimensional structure and infers the specific functions of proteins from their ultra-two-dimensional structure. This solves the problem of prediction failure in protein residue distance prediction alone, thereby improving the efficiency of protein synthesis. Attached Figure Description

[0070] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings:

[0071] Figure 1 This is a schematic flowchart of a method for predicting protein distance maps provided in this specification;

[0072] Figure 2 This is a schematic diagram of a feedforward convolution module structure for predicting protein distance maps, as provided in this specification.

[0073] Figure 3 A flowchart illustrating an example of a method for predicting protein distance maps provided in this specification;

[0074] Figure 4 This is a schematic diagram of a feedforward hybrid module structure for predicting protein distance maps provided in this specification;

[0075] Figure 5 This is a schematic diagram of a residual mixing module structure for predicting protein distance maps provided in this specification;

[0076] Figure 6 This is a schematic diagram of a residual convolution module structure for predicting protein distance maps provided in this specification;

[0077] Figure 7 This is a schematic diagram of the training process for a model that predicts protein distance maps, as provided in this specification.

[0078] Figure 8 This is a schematic diagram of a device for predicting protein distance maps provided in this specification;

[0079] Figure 9 This is a schematic diagram of the apparatus for training a protein distance map prediction model provided in this specification.

[0080] Figure 10 This specification provides a corresponding Figure 1 A schematic diagram of an electronic device. Detailed Implementation

[0081] Currently, methods for predicting the three-dimensional structure of proteins sometimes fail to meet the inherent characteristics of protein structure, leading to situations where the predicted protein cannot be synthesized. For example, simply obtaining the distances between multiple sets of residues may not satisfy the requirement that at least one set of residues must meet the triangle inequality. While all the distances between these residues can be predicted, the predicted protein should at least possess the characteristics of being synthesizable and having a local structure. Since synthetic proteins have a three-dimensional structure, at least one set of residues in the protein must satisfy the triangle inequality for the protein to be synthesized normally. Furthermore, only when the triangle inequality is satisfied and the synthesized protein possesses a local structure can the predicted protein's residue function be better reflected.

[0082] It is evident that existing methods for predicting protein structure ignore the structural characteristics of the protein itself, resulting in a significant problem: they cannot accurately predict protein structure.

[0083] The protein distance map predicted in this application can analyze and solve the aforementioned problems. A protein distance map is a two-dimensional matrix where each value represents the residue-residue distance. It contains rich structural information, including secondary structure, inter-residue distances, and interaction types related to specific residue types. The distance map helps assess the feasibility of generating tertiary structures for a given protein sequence in protein design. High-quality distance maps can make the accuracy of two-stage protein structure prediction algorithms approach that of single-stage algorithms, while requiring far fewer computational resources. In target protein mining, candidate active sites or candidate proteins are screened by identifying whether they are located on specific secondary structures and whether they satisfy corresponding distance relationships. Protein distance maps are the data material used in protein domain segmentation algorithms, which follow the basic principle of "maximizing intra-domain contact and minimizing inter-domain contact."

[0084] Therefore, this specification provides a method for predicting protein distance maps to solve the above-mentioned problems.

[0085] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0086] The technical solutions provided in the various embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0087] Figure 1This specification provides a flowchart illustrating a method for predicting protein distance maps, which includes the following steps:

[0088] S101: Determine the residue sequence of the target protein.

[0089] In the embodiments of this specification, since the arrangement of protein residue sequences has an important influence on protein function, protein structure prediction is usually based on protein residue sequences as the basic data.

[0090] Specifically, this specification describes the prediction of protein distance maps using a computer or a computational group capable of performing convolutional calculations, without limitation herein. For ease of description, this specification collectively refers to them as servers.

[0091] In this specification, the server can identify the protein for which distance map prediction is required as the target protein and obtain the residue sequence of the target protein. The server can determine the target protein residue sequence through methods such as manual compilation input or acquisition from existing protein residue sequence libraries (such as UniRef50, UniRef90, and UniRef100). This specification does not limit the specific acquisition method.

[0092] For example, taking datasets_15051 as an example, datasets_15051 consists of 15051 protein sequences. In the embodiments of this specification, the server uses one protein sequence from datasets_15051 for prediction, and its length is 510.

[0093] S103: Input the residue sequence into the trained distance map prediction model, and use the attention subnetwork of the distance map prediction model to perform attention weighting on the residue sequence to determine the first attention map corresponding to the residue sequence.

[0094] In the embodiments of this specification, the selected protein residue sequences need to be extracted and analyzed into attention maps in order to analyze the proteins composed of the selected sequences.

[0095] Specifically, there are many methods for how the server extracts selected protein sequences from the dataset and processes them into attention maps, including the ESM (Self-supervised protein language model) series. This application does not limit the methods in this regard.

[0096] For example, the server uses ESM-2 to extract attention maps of selected protein sequences from the datasets_15051. The ESM-2 consists of several encoder-decoder layers, where the encoder-decoder layer is a common transformer model structure, and it outputs a first attention map with dimensions {1440, 510, 510}.

[0097] S105: Perform a first preprocessing on the first attention map, input the first preprocessing result into the triangular attention module of the distance map prediction model, predict the distance between each residue, determine each residue whose distance between residues conforms to the triangular distance constraint relationship, and perform attention weighting on the determined residues that satisfy the triangular distance constraint relationship to obtain the second attention map corresponding to the first attention map.

[0098] In the embodiments of this specification, to facilitate the determination of whether the predicted protein can be synthesized in the distance graph, an existing triangular attention model for protein synthesis is used to strengthen the residues that satisfy the triangular distance constraint relationship, which are then reflected in the second attention graph. However, the existing triangular attention model for protein synthesis has a requirement for the dimensionality of the input attention graph, so the first attention graph is preprocessed to meet the requirement.

[0099] Specifically, the server inputs the first attention map into the feedforward convolution module of the distance map prediction model. Through the feedforward convolution module, the size of the first attention map is reduced. Based on the dimension of the attention map input by the triangular attention module of the distance map prediction model, the dimension of the reduced result is adjusted, and the adjusted result is determined as the first preprocessing result.

[0100] For example, the server runs a feedforward convolution module such as Figure 2 As shown: It consists of a max-pooling layer and a two-layer convolutional neural network connected in series. The window size of the max-pooling layer is 2×2, as shown... Figure 3 As shown, its input is a first attention map with dimensions {36, 510, 510}, and its output is an attention map with dimensions {36, 255, 255}. The two-layer convolutional neural network consists of two concatenated convolutional layers. Each convolutional layer consists of a filter with a 3×3 kernel, an instance normalization layer, and a ReLU activation function concatenated. The output of the first convolutional layer is an attention map with dimensions {36, 255, 255}. The output of the second convolutional layer is an attention map with dimensions {48, 255, 255}. Inputting the attention map output from the feedforward convolutional module into the triangular attention module outputs a second attention map with dimensions {48, 255, 255}.

[0101] S107: Perform a second preprocessing on the second attention map, input the second preprocessing result into the residual mixing module of the distance map prediction model, filter through the residual mixing module, adjust the size of the filtering result according to the second attention map, and perform local feature weighting on the residues with specified structural relationships in the adjusted filtering result to obtain the third attention map corresponding to the second attention map.

[0102] In the embodiments of this specification, to facilitate the analysis and prediction of protein functionality in the distance graph, an existing method is used to enhance the proportion of protein residues satisfying a specified structure, which is reflected in the third attention graph. However, existing local lightweight models have dimensionality requirements for the input attention graph; therefore, the second attention graph is preprocessed to meet these requirements.

[0103] Specifically, the server inputs the third attention map into the feedforward mixing module of the distance map prediction model. The feedforward mixing module compresses the third attention map to facilitate local feature weighting by the residual mixing module of the distance map prediction model, thus reducing the size of the compressed third attention map. Based on the dimension of the attention map input to the residual mixing module of the distance map prediction model, the dimension of the reduced image is adjusted, and the adjusted result is determined as the preprocessing result.

[0104] Specifically, the server inputs the third attention map into the residual convolution module of the distance map prediction model. The residual convolution module adjusts the dimension of the third attention map based on the dimension of the first attention map. Based on the difference between the dimension adjustment result and the size of the first attention map, the dimension adjustment result is adjusted according to the size of the first attention map. The size adjustment result is then superimposed on the first attention map. The superimposed result is convolved to determine the fourth attention map corresponding to the third attention map.

[0105] For example, the server runs a feedforward hybrid module such as Figure 4 As shown, it consists of a locally lightweight module, a max-pooling layer, and a two-layer convolutional neural network connected in series. The output of the locally lightweight module is an attention map with dimensions {48, 255, 255}. The window of the max-pooling layer is 2×2, as shown... Figure 3 As shown, its output is an attention map with dimensions {48, 127, 127}. The two-layer convolutional network consists of two concatenated convolutional layers. Each convolutional layer comprises a filter with a 3×3 kernel, an instance normalization layer, and a ReLU activation function. The output of the first convolutional layer is an attention map with dimensions {48, 127, 127}, and the output of the second convolutional layer is an attention map with dimensions {128, 127, 127}.

[0106] For example, the server runs a residual hybrid module such as Figure 5 As shown, it consists of a transposed convolutional layer, a padding module, a stacking module, a local lightweighting module, and a two-layer convolutional neural network connected in series. It performs the following process:

[0107] The server runs a transposed convolutional layer consisting of a 2×2 transposed filter, whose input is the output of the feedforward mixing module, such as... Figure 3 As shown, the output of the transposed filter of the transposed convolutional layer is an attention map with dimensions {48, 254, 254}. The padding module calculates the size difference between the output of the transposed filter and the output of the triangular attention module, and increases the dimensions of the second and third dimensions of the transposed filter output by 1 by padding with zeros, ultimately outputting an attention map with dimensions {48, 255, 255}. The overlay module adds the output of the padding module to the output of the triangular attention module to obtain an attention map with dimensions {48, 255, 255}. The local lightweighting module outputs an attention map with dimensions {48, 255, 255}, which will be used as the input to the two-layer convolutional neural network. The two-layer convolutional neural network consists of two concatenated convolutional layers, each consisting of a 3×3 kernel filter, an instance normalization layer, and a ReLU activation function concatenated. The output of the first convolutional layer is an attention map with dimensions {48,255,255}, and the output of the second convolutional layer is also an attention map with dimensions {48,255,255}, which will be used as the input to the residual convolutional module.

[0108] S109: Input the first attention map and the third attention map into the residual convolution module of the distance map prediction model, transpose and convolve the third attention map through the residual convolution module, adjust the size of the filtering result according to the first attention map, and superimpose the adjustment result with the first attention map to obtain the fourth attention map corresponding to the third attention map.

[0109] In the embodiments of this specification, since the attention map is transformed to the required dimension using the triangular attention module and the local lightweight module, the third attention map form is directly adopted and the transformation is easy to calculate, but not easy to analyze and compare. Therefore, the third attention map is restored, and the first attention map is compared and adjusted with the third attention map and superimposed to enhance the comparison effect of the fourth attention map for easier observation and analysis.

[0110] Specifically, the server inputs the third attention map into the residual convolution module of the distance map prediction model. The residual convolution module adjusts the dimension of the third attention map based on the dimension of the first attention map. Based on the difference between the dimension adjustment result and the size of the first attention map, the dimension adjustment result is adjusted according to the size of the first attention map. The size adjustment result is then superimposed on the first attention map. The superimposed result is convolved to determine the fourth attention map corresponding to the third attention map.

[0111] For example, the server runs a residual convolution module such as Figure 6 As shown, it consists of a transposed convolutional layer, a padding module, a stacking module, and a two-layer convolutional neural network connected in series. It performs the following process:

[0112] The server runs transposed convolutional layers consisting of transposed convolutions with a 2×2 kernel, such as... Figure 3 As shown, its output is an attention map with dimensions {36, 510, 510}, which serves as the input to the padding module. The padding module calculates the differences between the third and first attention maps in the second and third dimensions, and then directly outputs the dimension-adjusted third attention map. The overlay module adds the output of the padding module to the first attention map to obtain an attention map with dimensions {36, 510, 510}, which serves as the input to the two-layer convolutional neural network. The two-layer convolutional neural network consists of two concatenated convolutional layers. Each convolutional layer comprises a 3×3 kernel filter, an instance normalization layer, and a ReLU activation function. The output of the first convolutional layer is an attention map with dimensions {36, 510, 510}, and the output of the second convolutional layer is a fourth attention map with dimensions {36, 510, 510}.

[0113] S111: Input the fourth attention map into the regression prediction module of the distance map prediction model, and normalize the fourth attention map through the regression prediction module to obtain the protein distance map corresponding to the fourth attention map.

[0114] In the embodiments described in this specification, a single attention map is used for easier observation in the analysis; therefore, the fourth attention map is normalized. Specifically, the server normalizes the fourth attention map.

[0115] For example, the server's regression prediction layer consists of a filter with a 1×1 convolutional kernel, such as... Figure 3 As shown, its output is a distance map of protein residues with dimensions {1,510,510}.

[0116] based on Figure 1The proposed protein distance mapping scheme achieves the following: the server represents residues satisfying the triangle inequality in the protein distance map to help professionals determine whether the predicted protein can be synthesized. It emphasizes residues in the protein distance map that satisfy specific structures, helping professionals understand the helical structure of the protein from a two-dimensional perspective, thus aiding in the inference of the protein's specific function. This scheme solves the problem of prediction results failing due to solely predicting protein residue distances, thereby significantly improving the efficiency of protein synthesis.

[0117] Optionally, in step S105, the server performs a first preprocessing on the first attention map, inputting the first attention map into the feedforward convolutional module of the distance map prediction model, and reducing the size of the first attention map through the pooling layer of the feedforward convolutional module. Then, through the convolutional layer of the feedforward convolutional module, the dimension of the reduced attention map is adjusted according to the dimension of the attention map input to the triangular attention module of the distance map prediction model, and the adjusted result is determined as the first preprocessing result.

[0118] Optionally, in step S107, the server performs a second preprocessing on the second attention map, inputs the third attention map into the feedforward mixing module of the distance map prediction model, and compresses the third attention map through the lightweight layer of the feedforward mixing module to facilitate local feature weighting by the residual mixing module of the distance map prediction model. The pooling layer of the feedforward mixing module reduces the size of the compressed third attention map. The convolutional layer of the feedforward convolution module adjusts the dimension of the reduced attention map according to the dimension of the attention map input to the residual mixing module of the distance map prediction model, and the adjusted result is determined as the preprocessing result.

[0119] Optionally, the server execution step S109 includes: inputting the third attention map into the residual convolution module of the distance map prediction model; adjusting the dimension of the third attention map according to the dimension of the first attention map through the transpose layer of the residual convolution module; adjusting the dimension adjustment result according to the size of the first attention map through the padding layer of the residual convolution module based on the difference between the dimension adjustment result and the size of the first attention map; and convolving the superimposed result through the convolution layer of the residual convolution module to determine the convolution result as the fourth attention map corresponding to the third attention map.

[0120] Optionally, before performing the first preprocessing on the first attention map in step S105, the server inputs the first attention map into the group pooling module of the distance map prediction model. Through the group pooling module, the first attention map is divided into a preset number of sub-attention maps according to a preset number. The maximum value of each sub-attention map is extracted to determine the sub-feature map corresponding to each sub-attention map. The sub-feature maps are then concatenated to obtain a first attention map of a preset number of dimensions.

[0121] Optionally, in step S111, the server inputs the regression prediction module result into the size adjustment module of the distance map prediction model. Through the size adjustment module, the attention map values ​​of the regression prediction module result are scaled as a whole to obtain a protein residue distance map that is easy to observe.

[0122] It should be noted that protein distance maps are not the same as protein contact maps. Furthermore, protein distance maps are inherently superior to protein contact maps in applications such as "estimating the feasibility of generating tertiary structures for a given protein sequence in protein design" and "screening candidate active sites or candidate proteins." Secondly, existing contact map prediction algorithms do not simultaneously predict both the triangular inequality distance constraint and the two-dimensional protein structure, thus affecting the accuracy of protein structure predictions based on contact maps.

[0123] This manual also provides information related to Figure 1 The model training method corresponding to the flowchart for predicting protein distance graphs, such as... Figure 7 As shown:

[0124] S201: Determine the residue sequence of the target protein;

[0125] S203: Input the residue sequence into the trained distance map prediction model, and use the attention subnetwork of the distance map prediction model to perform attention weighting on the residue sequence to determine the first attention map corresponding to the residue sequence;

[0126] S205: Perform a first preprocessing on the first attention map, input the first preprocessing result into the triangular attention module of the distance map prediction model, predict the distance between each residue, determine each residue whose distance between residues conforms to the triangular distance constraint relationship, and perform attention weighting on the determined residues that satisfy the triangular distance constraint relationship to obtain the second attention map corresponding to the first attention map;

[0127] S207: Perform a second preprocessing on the second attention map, input the second preprocessing result into the residual mixing module of the distance map prediction model, filter through the residual mixing module, adjust the size of the filtering result according to the second attention map, and perform local feature weighting on the residues with specified structural relationships in the adjusted filtering result to obtain the third attention map corresponding to the second attention map;

[0128] S209: Input the first attention map and the third attention map into the residual convolution module of the distance map prediction model, transpose and convolve the third attention map through the residual convolution module, adjust the size of the filtering result according to the first attention map, and superimpose the adjustment result with the first attention map to obtain the fourth attention map corresponding to the third attention map;

[0129] S211: Input the fourth attention map into the regression prediction module of the distance map prediction model, and normalize the fourth attention map through the regression prediction module to obtain the protein distance map corresponding to the fourth attention map;

[0130] S213: Supervised learning is performed on the regression prediction module of the distance map prediction model using the labels of the protein distance map training set, and the model parameters are updated using a preset optimization algorithm based on the calculated loss value.

[0131] S215: Iterate the above training method and determine the current iteration number; if the current iteration number reaches the preset number, stop the iteration and obtain the target model.

[0132] Optionally, in step S205, before performing the first preprocessing on the first attention map, the first attention map is input into the group pooling module of the distance map prediction model. The group pooling module divides the first attention map into a preset number of sub-attention maps according to a preset number. The maximum value of each sub-attention map is extracted to determine the sub-feature map corresponding to each sub-attention map. The sub-feature maps are then concatenated to obtain a first attention map of a preset number of dimensions.

[0133] Optionally, before performing the first preprocessing on the first attention map in step S205, the first attention map is input into the group pooling module of the distance map prediction model. In any step before performing supervised learning on the regression prediction module of the distance map prediction model using the labels of the protein distance map training set in S213, supervised learning is performed on the group pooling module of the distance map prediction model using the labels of the protein distance map training set: the first attention map is input into the group pooling module of the distance map prediction model; the group pooling module divides the first attention map into a preset number of sub-attention maps according to a preset number; the maximum value of each sub-attention map is extracted to determine the sub-feature map corresponding to each sub-attention map; and the sub-feature maps are concatenated to obtain a first attention map of a preset number of dimensions.

[0134] This manual also provides information related to Figure 1 The device corresponding to the flowchart for predicting protein distance maps, such as Figure 8 As shown:

[0135] Sequence acquisition module 301 determines the residue sequence of the target protein;

[0136] The sequence confirmation module 303 inputs the residue sequence into the trained distance map prediction model, and performs attention weighting on the residue sequence through the attention sub-network of the distance map prediction model to determine the first attention map corresponding to the residue sequence;

[0137] The triangular attention module 305 performs a first preprocessing on the first attention map, inputs the first preprocessing result into the triangular attention module of the distance map prediction model, predicts the distance between each residue, determines each residue whose distance between residues conforms to the triangular distance constraint relationship, and performs attention weighting on the determined residues that satisfy the triangular distance constraint relationship to obtain the second attention map corresponding to the first attention map.

[0138] The local structure attention module 307 performs a second preprocessing on the second attention map, inputs the second preprocessing result into the residual mixing module of the distance map prediction model, filters the result through the residual mixing module, adjusts the size of the filtered result according to the second attention map, and performs local feature weighting on the residues with specified structural relationships in the adjusted filtered result to obtain the third attention map corresponding to the second attention map.

[0139] The restoration module 309 inputs the first attention map and the third attention map into the residual convolution module of the distance map prediction model, transposes and convolves the third attention map through the residual convolution module, adjusts the size of the filtering result according to the first attention map, and superimposes the adjustment result with the first attention map to obtain the fourth attention map corresponding to the third attention map.

[0140] The output module 311 inputs the fourth attention map into the regression prediction module of the distance map prediction model, and normalizes the fourth attention map through the regression prediction module to obtain the protein distance map corresponding to the fourth attention map.

[0141] Optionally, the triangular attention module 305 is used to input the first attention map into the feedforward convolutional module of the distance map prediction model, and reduce the size of the first attention map through the pooling layer of the feedforward convolutional module. Through the convolutional layer of the feedforward convolutional module, the dimension of the reduced result is adjusted according to the dimension of the attention map input to the triangular attention module of the distance map prediction model, and the adjusted result is determined as the first preprocessing result.

[0142] Optionally, the local structure attention module 307 is used to input the third attention map into the feedforward mixing module of the distance map prediction model. The third attention map is compressed through a lightweight layer of the feedforward mixing module to facilitate local feature weighting by the residual mixing module of the distance map prediction model. The size of the compressed third attention map is reduced through a pooling layer of the feedforward mixing module. The dimension of the reduced attention map is adjusted according to the dimension of the attention map input to the residual mixing module of the distance map prediction model through a convolutional layer of the feedforward convolution module, and the adjusted result is determined as the preprocessing result.

[0143] Optionally, the local structure attention module 307 is used to reduce the dimensionality of the second preprocessing result and increase its size through the transposed layer of the residual mixing module. Through the fill layer of the residual mixing module, the size of the transposed layer processing result is adjusted until it is the same as the size of the second attention map, based on the difference between the transposed layer processing result and the second attention map. Through the superposition layer of the residual mixing module, the adjusted result is superimposed on the third attention map to obtain the adjusted filtering result.

[0144] Optionally, the restoration module 309 is used to input the third attention map into the residual convolution module of the distance map prediction model, and adjust the dimension of the third attention map according to the dimension of the first attention map through the transpose layer of the residual convolution module. Through the padding layer of the residual convolution module, the dimension adjustment result is adjusted according to the size of the first attention map based on the difference between the dimension adjustment result and the size of the first attention map. Through the stacking layer of the residual convolution module, the size adjustment result is stacked with the first attention map. Through the convolutional layer of the residual convolution module, the stacked result is convolved to determine the convolution result as the fourth attention map corresponding to the third attention map.

[0145] Optionally, the sequence confirmation module 303 is used to input the first attention map into the group pooling module of the distance map prediction model, and through the group pooling module, divide the first attention map into a preset number of sub-attention maps according to a preset number. The maximum value of each sub-attention map is extracted to determine the sub-feature map corresponding to each sub-attention map. The sub-feature maps are then concatenated to obtain a first attention map of a preset number of dimensions.

[0146] Optionally, the output module 311 is used to input the regression prediction module result into the size adjustment module of the distance map prediction model. Through the size adjustment module, the attention map values ​​of the regression prediction module result are scaled as a whole to obtain a protein residue distance map that is easy to observe.

[0147] This manual also provides information related to Figure 7 The device corresponding to the training flowchart of the protein distance graph prediction model, such as... Figure 9 As shown:

[0148] Sequence acquisition training module 401 determines the residue sequence of the target protein;

[0149] The sequence confirmation training module 403 inputs the residue sequence into the trained distance map prediction model, and performs attention weighting on the residue sequence through the attention sub-network of the distance map prediction model to determine the first attention map corresponding to the residue sequence.

[0150] The triangular attention training module 405 performs a first preprocessing on the first attention map, inputs the first preprocessing result into the triangular attention module of the distance map prediction model, predicts the distance between each residue, determines each residue whose distance between residues conforms to the triangular distance constraint relationship, and performs attention weighting on the determined residues that satisfy the triangular distance constraint relationship to obtain the second attention map corresponding to the first attention map.

[0151] The local structure attention training module 407 performs a second preprocessing on the second attention map, inputs the second preprocessing result into the residual mixing module of the distance map prediction model, filters the result through the residual mixing module, adjusts the size of the filtered result according to the second attention map, and performs local feature weighting on the residues with specified structural relationships in the adjusted filtered result to obtain the third attention map corresponding to the second attention map.

[0152] The attention map restoration training module 409 inputs the first attention map and the third attention map into the residual convolution module of the distance map prediction model, transposes and convolves the third attention map through the residual convolution module, adjusts the size of the filtering result according to the first attention map, and superimposes the adjustment result with the first attention map to obtain the fourth attention map corresponding to the third attention map.

[0153] The distance map output training module 411 inputs the fourth attention map into the regression prediction module of the distance map prediction model, and normalizes the fourth attention map through the regression prediction module to obtain the protein distance map corresponding to the fourth attention map.

[0154] The parameter update training module 413 uses the labels of the protein distance map training set to supervise the regression prediction module of the distance map prediction model, and updates the model parameters using a preset optimization algorithm based on the calculated loss value.

[0155] The iterative training module 415 iterates the above training method and determines the current iteration number; if the current iteration number reaches the preset number, the iteration stops and the target model is obtained.

[0156] Optionally, the triangular attention training module 405 is used to input the first attention map into the group pooling module of the distance map prediction model. Through the group pooling module, the first attention map is divided into a preset number of sub-attention maps according to a preset number. The maximum value of each sub-attention map is extracted to determine the sub-feature map corresponding to each sub-attention map. The sub-feature maps are then concatenated to obtain a first attention map of a preset number of dimensions.

[0157] Optionally, the triangular attention training module 405 is used to input the first attention map into the feedforward convolution module of the distance map prediction model, reduce the size of the first attention map through the pooling layer of the feedforward convolution module, and adjust the dimension of the reduction result through the convolution layer of the feedforward convolution module according to the dimension of the attention map input by the triangular attention module of the distance map prediction model, and determine the adjusted result as the first preprocessing result.

[0158] Optionally, the local structure attention training module 407 is used to input the third attention map into the feedforward mixing module of the distance map prediction model, compress the third attention map through the lightweight layer of the feedforward mixing module so that the residual mixing module of the distance map prediction model can perform local feature weighting, reduce the size of the compressed third attention map through the pooling layer of the feedforward mixing module, and adjust the dimension of the reduction result through the convolutional layer of the feedforward convolution module according to the dimension of the attention map input by the residual mixing module of the distance map prediction model, and determine the adjusted result as the preprocessing result.

[0159] Optionally, the local structure attention training module 407 is used to reduce the dimension of the second preprocessing result and increase the size of the second preprocessing result through the transposed layer of the residual mixing module, adjust the size of the transposed layer processing result according to the difference between the transposed layer processing result and the second attention map through the filling layer of the residual mixing module until it is the same as the size of the second attention map, and superimpose the adjusted result with the third attention map through the superposition layer of the residual mixing module as the adjusted filtering result.

[0160] Optionally, the attention map restoration training module 409 is used to input the third attention map into the residual convolution module of the distance map prediction model. Through the transpose layer of the residual convolution module, the dimension of the third attention map is adjusted according to the dimension of the first attention map. Through the fill layer of the residual convolution module, the dimension adjustment result is adjusted according to the size of the first attention map based on the difference between the dimension adjustment result and the size of the first attention map. Through the stacking layer of the residual convolution module, the size adjustment result is stacked with the first attention map. Through the convolution layer of the residual convolution module, the stacked result is convolved to determine the convolution result as the fourth attention map corresponding to the third attention map.

[0161] Optionally, the distance map output training module 411 is used to input the regression prediction module results into the size adjustment module of the distance map prediction model. Through the size adjustment module, the attention map values ​​of the regression prediction module results are scaled as a whole to obtain a protein residue distance map that is easy to observe.

[0162] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described method for predicting protein distance maps.

[0163] This instruction manual also provides Figure 10 One of the corresponding Figure 1 A schematic diagram of the structure of an electronic device. (e.g.) Figure 10At the hardware level, the electronic device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for the business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above-mentioned functions. Figure 1 The described method for detecting graphics processor anomalies. Of course, besides software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution entity of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0164] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0165] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0166] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0167] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components.

[0168] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0169] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0170] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0171] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0172] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0173] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0174] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0175] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0176] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0177] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0178] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0179] The above description is merely an embodiment of this specification and is not intended to limit this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this application.

Claims

1. A method for predicting protein distance maps, characterized in that, include: Determine the residue sequence of the target protein; The residue sequence is input into the trained distance map prediction model, and the residue sequence is weighted by the attention subnetwork of the distance map prediction model to determine the first attention map corresponding to the residue sequence. The first attention map is preprocessed, and the result of the first preprocessing is input into the triangular attention module of the distance map prediction model to predict the distance between each residue, determine each residue whose distance between residues conforms to the triangular distance constraint relationship, and perform attention weighting on the determined residues that satisfy the triangular distance constraint relationship to obtain the second attention map corresponding to the first attention map. The second attention map is subjected to a second preprocessing, and the second preprocessing result is input into the residual mixing module of the distance map prediction model. The residual mixing module filters the data, and the size of the filtered result is adjusted according to the second attention map. Local feature weighting is performed on the residues with specified structural relationships in the adjusted filtered result to obtain a third attention map corresponding to the second attention map. The specified structural relationship includes specified two-dimensional structures and ultra-two-dimensional structures. The first attention map and the third attention map are input into the residual convolution module of the distance map prediction model. The third attention map is transposed and convolved by the residual convolution module. The size of the filtering result is adjusted according to the first attention map. The adjusted result is superimposed with the first attention map to obtain the fourth attention map corresponding to the third attention map. The fourth attention map is input into the regression prediction module of the distance map prediction model. The regression prediction module normalizes the fourth attention map to obtain the protein distance map corresponding to the fourth attention map. The first preprocessing of the first attention map includes: The first attention map is input into the feedforward convolutional module of the distance map prediction model, and the size of the first attention map is reduced through the pooling layer of the feedforward convolutional module. The convolutional layer of the feedforward convolutional module adjusts the dimension of the reduced result by adjusting the dimension of the reduced result based on the dimension of the triangular attention module of the distance map prediction model, and determines the adjusted result as the first preprocessing result. The second preprocessing of the second attention map includes: The third attention map is input into the feedforward mixing module of the distance map prediction model. The third attention map is compressed through the lightweight layer of the feedforward mixing module so that the residual mixing module of the distance map prediction model can perform local feature weighting. The size of the compressed third attention map is reduced through the pooling layer of the feedforward mixing module; The convolutional layer of the feedforward convolutional module adjusts the dimension of the attention map input to the residual mixing module of the distance map prediction model, based on the dimension of the input attention map, and determines the adjusted result as the preprocessing result.

2. The method as described in claim 1, characterized in that, The filtering process, performed using the residual mixing module, involves adjusting the size of the filtered result based on the second attention map. Specifically, this includes: The transpose layer of the residual mixing module reduces the dimension of the second preprocessing result and increases its size. The size of the transposed layer processing result is adjusted according to the difference between the transposed layer processing result and the second attention map through the filling layer of the residual mixing module until it is the same as the size of the second attention map. The adjustment result is superimposed on the third attention map through the superposition layer of the residual mixing module to obtain the adjusted filtering result.

3. The method as described in claim 1, characterized in that, The first attention map and the third attention map are input into the residual convolution module of the distance map prediction model. The third attention map is transposed and convolved by the residual convolution module. The filtering result is then resized according to the first attention map, and the resized result is superimposed on the first attention map to obtain the fourth attention map. Specifically, this includes: The third attention map is input into the residual convolution module of the distance map prediction model. The dimension of the third attention map is adjusted according to the dimension of the first attention map through the transpose layer of the residual convolution module. The dimension adjustment result is adjusted based on the size of the first attention map by using the filling layer of the residual convolution module, according to the difference between the dimension adjustment result and the size of the first attention map. The size adjustment result is superimposed on the first attention map through the superposition layer of the residual convolution module; The convolutional layer of the residual convolution module is used to convolve the superimposed result, and the convolution result is determined to be the fourth attention map corresponding to the third attention map.

4. The method as described in claim 1, characterized in that, Before performing the first preprocessing on the first attention map, the method further includes: The first attention map is input into the group pooling module of the distance map prediction model. The group pooling module divides the first attention map into a preset number of sub-attention maps according to a preset number. Then, the maximum value of each sub-attention map is extracted to determine the sub-feature map corresponding to each sub-attention map; By concatenating the sub-feature maps, a first attention map with a predetermined number of dimensions is obtained.

5. The method as described in claim 1, characterized in that, The method further includes obtaining the protein distance map corresponding to the fourth attention map: The regression prediction module results are input into the size adjustment module of the distance map prediction model. The size adjustment module scales the attention map values ​​of the regression prediction module results as a whole to obtain a protein residue distance map that is easy to observe.

6. A model training method for predicting protein distance maps, characterized in that, include: Determine the residue sequence of the target protein; The residue sequence is input into the trained distance map prediction model, and the residue sequence is weighted by the attention subnetwork of the distance map prediction model to determine the first attention map corresponding to the residue sequence. The first attention map is preprocessed, and the result of the first preprocessing is input into the triangular attention module of the distance map prediction model to predict the distance between each residue, determine each residue whose distance between residues conforms to the triangular distance constraint relationship, and perform attention weighting on the determined residues that satisfy the triangular distance constraint relationship to obtain the second attention map corresponding to the first attention map. The second attention map is subjected to a second preprocessing, and the second preprocessing result is input into the residual mixing module of the distance map prediction model. The residual mixing module filters the data, and the size of the filtered result is adjusted according to the second attention map. Local feature weighting is performed on the residues with specified structural relationships in the adjusted filtered result to obtain a third attention map corresponding to the second attention map. The specified structural relationship includes specified two-dimensional structures and ultra-two-dimensional structures. The first attention map and the third attention map are input into the residual convolution module of the distance map prediction model. The third attention map is transposed and convolved by the residual convolution module. The size of the filtering result is adjusted according to the first attention map. The adjusted result is superimposed with the first attention map to obtain the fourth attention map corresponding to the third attention map. The fourth attention map is input into the regression prediction module of the distance map prediction model. The regression prediction module normalizes the fourth attention map to obtain the protein distance map corresponding to the fourth attention map. The regression prediction module of the distance map prediction model is supervised learning using the labels of the protein distance map training set, and the model parameters are updated using a preset optimization algorithm based on the calculated loss value. Iterate through all the above steps and determine the current iteration number; if the current iteration number reaches the preset number, stop the iteration and obtain the target model; The first preprocessing of the first attention map includes: The first attention map is input into the feedforward convolutional module of the distance map prediction model, and the size of the first attention map is reduced through the pooling layer of the feedforward convolutional module. The convolutional layer of the feedforward convolutional module adjusts the dimension of the reduced result by adjusting the dimension of the reduced result based on the dimension of the triangular attention module of the distance map prediction model, and determines the adjusted result as the first preprocessing result. The second preprocessing of the second attention map includes: The third attention map is input into the feedforward mixing module of the distance map prediction model. The third attention map is compressed through the lightweight layer of the feedforward mixing module so that the residual mixing module of the distance map prediction model can perform local feature weighting. The size of the compressed third attention map is reduced through the pooling layer of the feedforward mixing module; The convolutional layer of the feedforward convolutional module adjusts the dimension of the attention map input to the residual mixing module of the distance map prediction model, based on the dimension of the input attention map, and determines the adjusted result as the preprocessing result.

7. The method as described in claim 6, characterized in that, Before performing the first preprocessing on the first attention map, the training method further includes: The first attention map is input into the group pooling module of the distance map prediction model. The group pooling module divides the first attention map into a preset number of sub-attention maps according to a preset number. Then, the maximum value of each sub-attention map is extracted to determine the sub-feature map corresponding to each sub-attention map; By concatenating the sub-feature maps, a first attention map with a predetermined number of dimensions is obtained.

8. An apparatus for predicting protein distance maps, characterized in that, include: The sequence acquisition module determines the residue sequence of the target protein; The sequence confirmation module inputs the residue sequence into the trained distance map prediction model, and performs attention weighting on the residue sequence through the attention sub-network of the distance map prediction model to determine the first attention map corresponding to the residue sequence. The triangular attention module performs a first preprocessing on the first attention map, inputs the first preprocessing result into the triangular attention module of the distance map prediction model, predicts the distance between each residue, determines each residue whose distance between residues conforms to the triangular distance constraint relationship, and performs attention weighting on the determined residues that satisfy the triangular distance constraint relationship to obtain the second attention map corresponding to the first attention map. The first preprocessing of the first attention map includes: The first attention map is input into the feedforward convolutional module of the distance map prediction model, and the size of the first attention map is reduced through the pooling layer of the feedforward convolutional module. The convolutional layer of the feedforward convolutional module adjusts the dimension of the reduced result by adjusting the dimension of the reduced result based on the dimension of the triangular attention module of the distance map prediction model, and determines the adjusted result as the first preprocessing result. The local structure attention module performs a second preprocessing on the second attention map, inputs the second preprocessing result into the residual mixing module of the distance map prediction model, filters the result through the residual mixing module, adjusts the size of the filtered result according to the second attention map, and performs local feature weighting on the residues with specified structural relationships in the adjusted filtered result to obtain a third attention map corresponding to the second attention map. The specified structural relationship includes specified two-dimensional structures and ultra-two-dimensional structures. The second preprocessing of the second attention map includes: The third attention map is input into the feedforward mixing module of the distance map prediction model. The third attention map is compressed through the lightweight layer of the feedforward mixing module so that the residual mixing module of the distance map prediction model can perform local feature weighting. The size of the compressed third attention map is reduced through the pooling layer of the feedforward mixing module; The convolutional layer of the feedforward convolutional module adjusts the dimension of the attention map input to the residual mixing module of the distance map prediction model by adjusting the convolutional layer of the feedforward convolutional module, and determines the adjusted result as the preprocessing result. The attention map restoration module inputs the first attention map and the third attention map into the residual convolution module of the distance map prediction model. The residual convolution module transposes and convolves the third attention map, adjusts the size of the filtering result according to the first attention map, and superimposes the adjustment result with the first attention map to obtain the fourth attention map corresponding to the third attention map. The distance map output module inputs the fourth attention map into the regression prediction module of the distance map prediction model. The regression prediction module normalizes the fourth attention map to obtain the protein distance map corresponding to the fourth attention map.

9. An apparatus for training a protein distance graph prediction model, characterized in that, include: The sequence acquisition training module determines the residue sequence of the target protein; The sequence confirmation training module inputs the residue sequence into the trained distance map prediction model, and performs attention weighting on the residue sequence through the attention sub-network of the distance map prediction model to determine the first attention map corresponding to the residue sequence. The triangular attention training module performs a first preprocessing on the first attention map, inputs the first preprocessing result into the triangular attention module of the distance map prediction model, predicts the distance between each residue, determines each residue whose distance between residues conforms to the triangular distance constraint relationship, and performs attention weighting on the determined residues that satisfy the triangular distance constraint relationship to obtain the second attention map corresponding to the first attention map. The first preprocessing of the first attention map includes: The first attention map is input into the feedforward convolutional module of the distance map prediction model, and the size of the first attention map is reduced through the pooling layer of the feedforward convolutional module. The convolutional layer of the feedforward convolutional module adjusts the dimension of the reduced result by adjusting the dimension of the reduced result based on the dimension of the triangular attention module of the distance map prediction model, and determines the adjusted result as the first preprocessing result. The local structure attention training module performs a second preprocessing on the second attention map, inputs the second preprocessing result into the residual mixing module of the distance map prediction model, filters the result through the residual mixing module, adjusts the size of the filtered result according to the second attention map, and performs local feature weighting on the residues with specified structural relationships in the adjusted filtered result to obtain a third attention map corresponding to the second attention map. The specified structural relationship includes specified two-dimensional structures and ultra-two-dimensional structures. The second preprocessing of the second attention map includes: The third attention map is input into the feedforward mixing module of the distance map prediction model. The third attention map is compressed through the lightweight layer of the feedforward mixing module so that the residual mixing module of the distance map prediction model can perform local feature weighting. The size of the compressed third attention map is reduced through the pooling layer of the feedforward mixing module; The convolutional layer of the feedforward convolutional module adjusts the dimension of the attention map input to the residual mixing module of the distance map prediction model by adjusting the convolutional layer of the feedforward convolutional module, and determines the adjusted result as the preprocessing result. The attention map restoration training module inputs the first attention map and the third attention map into the residual convolution module of the distance map prediction model. The residual convolution module transposes and convolves the third attention map, adjusts the size of the filtering result according to the first attention map, and superimposes the adjustment result with the first attention map to obtain the fourth attention map corresponding to the third attention map. The distance map output training module inputs the fourth attention map into the regression prediction module of the distance map prediction model. Through the regression prediction module, the fourth attention map is normalized to obtain the protein distance map corresponding to the fourth attention map. The parameter update training module uses the labels of the protein distance map training set to supervise the regression prediction module of the distance map prediction model, and updates the model parameters using a preset optimization algorithm based on the calculated loss value. The iterative training module iterates through all the above steps and determines the current iteration number; if the current iteration number reaches the preset number, the iteration stops and the target model is obtained.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 5.

11. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 5.

Images