A single-sequence protein contact map prediction method based on a graph neural network
By constructing a protein contact matrix prediction model based on graph neural networks, the problem of low accuracy in predicting protein contacts with low homology single sequences was solved, and higher accuracy in predicting amino acid contact relationships was achieved. This model can be applied to the analysis of protein three-dimensional structure and interactions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-09
- Publication Date
- 2026-04-14
AI Technical Summary
Existing protein contact prediction methods have low accuracy for low homology single-sequence proteins and cannot effectively predict the contact relationships between amino acids.
A graph neural network-based approach is used to predict protein residue contact matrices by constructing residue contact matrices, one-hot encoding, pre-trained model feature extraction, and graph convolutional encoding/decoding models, combined with graph convolutional models and convolutional neural networks.
It improves the accuracy of protein contact matrix prediction, enabling more accurate prediction of contact relationships between amino acids, and is suitable for protein three-dimensional structure prediction and proteome interaction analysis.
Smart Images
Figure CN115312118B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of protein map prediction, and specifically relates to a method for predicting single-sequence protein contact maps based on graph neural networks. Background Technology
[0002] Proteins are the main carriers of life activities, performing numerous 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 crucial for understanding protein function. This is important for drug development, vaccine therapy, and research on biological functions. Currently, the acquisition of protein three-dimensional structures mainly falls into two categories: experimental determination and computational methods. Experimental determination methods can accurately obtain protein structures, but they suffer from drawbacks such as long processing time, high cost, and high difficulty. Methods such as nuclear magnetic resonance (NMR) and cryo-electron microscopy are time-consuming, labor-intensive, require demanding experimental equipment and targets, and are expensive, hindering widespread adoption. Computational methods, on the other hand, predict protein structures based on protein sequences, significantly saving time and offering a simpler and lower-cost process.
[0003] In the process of predicting protein structure using computational methods, protein contact map prediction is a crucial intermediate step. The protein contact map is a simplified representation of protein structure, containing rich structural information. Analyzing the prediction results of the protein contact map can serve as constraints on the structure prediction model to improve its accuracy, and the results can also be used to construct a scoring function to evaluate and select the structure prediction model. In short, analyzing the protein contact map reveals protein structure-related features.
[0004] Protein contact is defined as the contact between two amino acids in a protein when the spatial distance between them is less than 1 / 3. When protein contact is considered to be in contact with each other, it is understood from the definition that protein contact provides the spatial distance between each amino acid in a protein, which plays a crucial role in subsequent prediction of the protein's three-dimensional structure. Furthermore, since the contact matrix can provide spatial information about proteins, it provides structural information in tasks such as protein function prediction and proteome interaction. Existing protein contact prediction methods infer the contact probability between amino acids by predicting the co-evolutionary relationship between each amino acid based on homology information such as the target protein and its homologous sequences. Co-evolutionary relationships can be simply understood as follows: in a protein, two or more amino acids are spatially close to each other, forming a stable local structure. When one or more amino acids mutate, causing a change in the local structure, other amino acids in this local structure will also mutate to maintain the protein's structural stability. This "mutual mutation" relationship between two or more amino acids is the co-evolutionary relationship. Homologous sequences are protein sequences in known protein structures that are structurally similar to the target protein but whose sequence originates from a common ancestor. However, a considerable number of known protein sequences lack homologous sequences. For these protein sequences, existing methods are not very effective and the accuracy remains low. Therefore, there is an urgent need for a new single-sequence-based protein contact prediction method. Summary of the Invention
[0005] The problem to be solved by this invention is that existing protein contact prediction methods have low accuracy in predicting contacts of proteins with low homology single sequences.
[0006] To address the above problems, this invention provides a method for predicting single-sequence protein contact maps based on graph neural networks, comprising the following steps:
[0007] S1 collects protein sequences and collects corresponding three-dimensional protein structure data from the PDB database to construct an inter-residue contact matrix;
[0008] S2 performs one-hot encoding on each protein sequence to obtain one-hot one-dimensional features;
[0009] S3 uses a pre-trained model to obtain one-dimensional protein coding features and two-dimensional attention matrix features for each protein sequence;
[0010] S4 constructs graph convolutional coding and convolutional decoding models;
[0011] S5 inputs the one-hot one-dimensional features and the one-dimensional protein coding features obtained through pre-training into the graph convolutional coding model to obtain the protein structure features; the protein structure features and the two-dimensional attention matrix obtained through pre-training are then input into the convolutional decoding model to predict the protein residue contact matrix.
[0012] Furthermore, step S1 is implemented as follows:
[0013] S1.1 Collect a dataset of protein sequences with 95% sequence homology from the ProteinNet dataset, totaling 93,054 proteins;
[0014] S1.2 100 proteins were randomly separated from the ProteinNet dataset. All protein sequences were searched with HHsearch on these 100 protein sequences. Any protein with a cutoff value less than 0.1 was deleted. Proteins with a sequence length of more than 500 were also deleted. Finally, a total of 34,578 proteins were selected.
[0015] S1.3 Download the three-dimensional structure data of the protein from the PDB database, and extract the residue pair contact matrix of the target protein from the three-dimensional structure data of the protein.
[0016] S1.4 Constructing the inter-residue contact matrix: This involves extracting the three-dimensional coordinates of each residue from the target protein's three-dimensional structural data and calculating the three-dimensional distance between each residue pair. The distance between residue pairs is less than [value missing]. If they are in contact, they are represented by 1; otherwise, they are not in contact and are represented by 0. Assuming the protein sequence length is L, a contact matrix with dimension L*L can be calculated.
[0017] Furthermore, the pre-trained model mentioned in step S3 is Esm-1b; each protein sequence is processed by Esm-1b to obtain a one-dimensional protein coding feature with a dimension of L*1280 and a two-dimensional attention matrix with a dimension of L*L*20, where L refers to the number of protein amino acids contained in the target protein.
[0018] Furthermore, the graph convolutional coding model and convolutional decoding model described in step S4 are both composed of a series of graph convolutional models.
[0019] Furthermore, the graph convolution model consists of two graph convolutional layers, a batch normalization layer, and a ReLU activation layer; the graph convolutional layer can be represented as:
[0020]
[0021] Where A represents the adjacency matrix, W represents the network weights, and H...l This represents the node information of the l-th layer. This represents the ReLU activation function.
[0022] Furthermore, the ReLU activation function can be expressed as:
[0023]
[0024] Furthermore, the convolutional decoding model described in step S4 consists of a ResNet module and a multilayer perceptron.
[0025] Furthermore, the ResNet module consists of 5 convolutional layers, with a batch normalization layer and a ReLU activation layer between each convolutional layer.
[0026] Furthermore, the multilayer perceptron consists of multiple fully connected layers, with a Drout layer added to each convolutional layer, and residual operations are also incorporated.
[0027] Furthermore, step S5 is implemented as follows:
[0028] S5.1 The one-hot one-dimensional feature is concatenated with the one-dimensional protein coding feature to form a one-dimensional feature with a dimension of L*1300.
[0029] As described in S5.2, a one-dimensional feature with dimension L*1300 is input into the graph convolutional coding model, and a graph structure feature with dimension L*420 is output.
[0030] S5.3 The graph structure features of dimension L*420 are concatenated to obtain a two-dimensional matrix of dimension L*L*420;
[0031] S5.4 The two-dimensional matrix with dimension L*L*420 is concatenated with the pre-trained attention matrix to obtain a two-dimensional matrix with dimension L*L*440.
[0032] S5.5 Input the two-dimensional matrix with dimension L*L*440 into the convolutional decoding model to obtain three different residue pair contact prediction method matrices.
[0033] The beneficial effects of this invention are as follows: The network model consists of a graph convolution model and a convolutional neural network. The graph convolution model greatly improves the receptive field, focuses more on protein graph structure information, obtains more protein features, transforms the prediction task from single-objective prediction to multi-objective prediction, sets corresponding parameters for different prediction indicators and reduces training parameters by sharing parameters, thereby reducing the difficulty of training and improving the prediction accuracy of protein contact matrix. Attached Figure Description
[0034] Figure 1 This is an example flowchart illustrating the method of implementing the present invention.
[0035] Figure 2 This is a schematic diagram of the workflow of a deep learning network model.
[0036] Figure 3 This is a diagram showing the actual residue contact of the 5N5P protein as an example.
[0037] Figure 4 This is a predicted residue contact diagram of the example protein 5NP obtained in this invention. Detailed Implementation
[0038] To more clearly illustrate the purpose, technical solutions, and advantages of the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.
[0039] like Figure 1 As shown, a method for predicting single-sequence protein contact maps based on graph neural networks includes the following steps:
[0040] S1 collects protein sequences and collects corresponding three-dimensional protein structure data from the PDB database to construct an inter-residue contact matrix;
[0041] S2 performs one-hot encoding on each protein sequence to obtain one-hot one-dimensional features;
[0042] S3 uses a pre-trained model to obtain one-dimensional protein coding features and two-dimensional attention matrix features for each protein sequence;
[0043] S4 constructs graph convolutional coding and convolutional decoding models;
[0044] S5 inputs the one-hot one-dimensional features and the one-dimensional protein coding features obtained through pre-training into the graph convolutional coding model to obtain the protein structure features; the protein structure features and the two-dimensional attention matrix obtained through pre-training are then input into the convolutional decoding model to predict the protein residue contact matrix.
[0045] Step S1 involves collecting protein sequences and their corresponding 3D structure data from the PDB database, and constructing a residue contact matrix. Specifically, the collected protein sequences are primarily from the ProteinNet dataset, with a sequence homology of 95% selected. To provide protein sequences suitable for single-sequence residue contact prediction, 100 proteins are randomly isolated from the ProteinNet dataset. All protein sequences are then compared with these 100 sequences using HHsearch. Any protein with a cutoff value less than 0.1 is deleted, followed by proteins with a sequence length exceeding 500. Finally, a total of 34,578 proteins are selected. The 3D structure data is downloaded from the PDB database. The residue contact matrix of the target protein is extracted from the 3D structure data. The construction process mainly involves extracting the 3D coordinates of each residue from the target protein's 3D structure data and calculating the 3D distance between each residue. The distance between residue pairs is less than... If the protein sequence is in contact, it is considered to be in contact and represented by 1; otherwise, it is not in contact and represented by 0. Assuming the protein sequence length is L, a contact matrix with dimension L*L can be calculated.
[0046] Step S2 performs one-hot encoding on each protein sequence to obtain one-hot one-dimensional features. Specifically, since there are only 20 types of amino acids in a protein, each type of protein is encoded, and each amino acid is encoded with a natural number, as shown in Table 1, which encodes amino acid sequences 1-20 in this example.
[0047] Table 1 Amino acid coding
[0048] Amino acid name Ala Cys Asp Glu Phe Gly His Ile Natural number encoding 0 1 2 3 4 5 6 7 Amino acid name Lys Leu Met Asn Pro Gln Arg Ser Natural number encoding 8 9 10 11 12 13 14 15 Amino acid name Thr Val Trp Tyr Natural number encoding 16 17 18 19
[0049] If the amino acid type of the test case is Met, then the one-hot code is (0,0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0).
[0050] Step S3 involves using a pre-trained model to obtain one-dimensional protein coding features and two-dimensional attention matrix features for each protein sequence. Specifically, each protein sequence is input into the ESM-1b pre-trained model to obtain a one-dimensional protein sequence code and a two-dimensional attention matrix, respectively.
[0051] Step S4 constructs a graph convolutional network and a decoding model; specifically, the graph convolutional network consists of a series of graph convolutional models, preferably, the graph convolutional model consists of two graph convolutional layers, a batch normalization layer, and a ReLU activation layer; wherein the graph convolutional layer can be represented as:
[0052]
[0053] Specifically, where A represents the adjacency matrix, W represents the network weights, and H... l This represents the node information of the l-th layer. This represents the ReLU activation function. Specifically, the ReLU activation function can be represented as:
[0054]
[0055] Specifically, the decoding model consists of a ResNet module and feedforward layers. Each convolutional layer is connected by a batch normalization layer and a ReLU activation layer. Preferably, the feedforward layer consists of multiple fully connected layers. Since the model's output is for different prediction targets, the feedforward layer outputs prediction matrices for different targets by setting different parameters. To improve the model's generalization ability and prevent overfitting, a Droout layer is added to each convolutional layer, along with residual operations.
[0056] Step S5 inputs the one-hot one-dimensional features and the one-dimensional protein coding features obtained through pre-training into the graph convolutional coding model to obtain the protein structure features; then inputs the protein structure features and the two-dimensional attention matrix obtained through pre-training into the convolutional decoding model to predict the protein residue contact matrix.
[0057] Specifically, the one-dimensional one-hot feature obtained from claim 2 and the one-dimensional protein-coding feature obtained from claim 3 are first concatenated together to form a one-dimensional feature with a dimension of L*1300. This one-dimensional feature is then input into the graph convolution model, which outputs a graph sequence feature with a dimension of L*420. This graph sequence feature is then concatenated to obtain a two-dimensional matrix with a dimension of L*L*420. This matrix, along with the attention matrix obtained from claim 3, is then input into the decoding module to obtain the residue pair contact prediction matrix.
[0058] In protein contact prediction tasks, different methods predict residue pairs with varying probabilities. To better compare the accuracy of different methods, CASP (Critical Assessment of Structure Prediction) proposes comparing the top K residue pairs with the highest probabilities from different prediction methods with the actual protein contacts. Here, K is typically set to L / 1, L / 2, or L / 5, where L is the protein sequence length. Based on the number of residues between residue pairs, contacts are further categorized as short-range contacts (6-12 residues apart), medium-range contacts (12-24 residues apart), and long-range contacts (more than 24 residues apart). We transform the prediction task into a multi-objective task based on these three comparison metrics. Specifically, most parameters in the convolutional decoding model are shared; the biggest difference lies in the feedforward layer at the final output. This approach reduces the number of training parameters and lowers the training difficulty. During training, independent loss functions are set for different targets. The total loss value is the weighted sum of the losses corresponding to each sub-target. The weights of different losses are calculated by taking the average of the proportions of short-range, medium-range, and long-range protein sequence lengths in the entire training set as the loss weights for different sub-targets.
[0059] This invention provides a single-sequence protein contact prediction method based on graph deep learning. It predicts contact matrices for protein sequences, which is of great significance for predicting protein 3D structures, proteome interactions, and antigen-antibody experiments. The network model consists of a graph convolutional model and a convolutional neural network. The graph convolutional model significantly increases the receptive field, focusing more on protein graph structural information and obtaining more protein features. Appropriate parameters are set for different prediction metrics, and a parameter-sharing method is used to reduce training parameters, lowering the training difficulty and thus improving the accuracy of protein contact matrix prediction.
[0060] The above provides a specific implementation method, but the present invention is not limited to the described implementation method. The basic idea of the present invention lies in the above solution. For those skilled in the art, various modified models, formulas, and parameters can still be designed based on the teachings of the present invention without requiring creative effort. Changes, modifications, substitutions, and variations made to the implementation method without departing from the principles and spirit of the present invention still fall within the protection scope of the present invention.
Claims
1. A method for predicting single-sequence protein contact maps based on graph neural networks, characterized in that, Includes the following steps: S1 collects protein sequences and collects corresponding three-dimensional protein structure data from the PDB database to construct an inter-residue contact matrix; S2 performs one-hot encoding on each protein sequence to obtain one-hot one-dimensional features; S3 obtains one-dimensional protein coding features and two-dimensional attention matrix features for each protein sequence through a pre-trained model; the pre-trained model is Esm-1b; each protein sequence obtains one-dimensional protein coding features with dimension L*1280 and two-dimensional attention matrix with dimension L*L*20 through Esm-1b, where L refers to the number of protein amino acids contained in the target protein. S4 constructs a graph convolutional coding model and a convolutional decoding model; the graph convolutional coding model consists of two graph convolutional layers, a batch normalization layer, and a ReLU activation layer; the graph convolutional layer is represented as follows: Where A represents the adjacency matrix and W represents the network weights. This represents the node information of the l-th layer. This represents the ReLU activation function; the convolutional decoding model described is composed of a ResNet module and a multilayer perceptron. S5 inputs the one-hot one-dimensional features and the one-dimensional protein coding features obtained through the pre-trained model into the graph convolutional coding model to obtain the protein structure features; then inputs the protein structure features and the two-dimensional attention matrix obtained through the pre-trained model into the convolutional decoding model to predict the protein residue contact matrix.
2. The method for predicting single-sequence protein contact maps based on graph neural networks according to claim 1, characterized in that, Step S1 is implemented as follows: S1.1 Collect a dataset of protein sequences with 95% sequence homology from the ProteinNet dataset, totaling 93,054 proteins; S1.2 100 proteins were randomly separated from the ProteinNet dataset. All protein sequences were searched with HHsearch on these 100 protein sequences. Any protein with a cutoff value less than 0.1 was deleted. Proteins with a sequence length of more than 500 were also deleted. Finally, a total of 34,578 proteins were selected. S1.3 Download the three-dimensional structure data of the protein from the PDB database, and extract the residue pair contact matrix of the target protein from the three-dimensional structure data of the protein. S1.4 Constructing the inter-residue contact matrix: By extracting the three-dimensional coordinates of each residue from the three-dimensional structural data of the target protein, the three-dimensional distance between each residue is calculated. Residue pairs with a distance of less than 8 Å are considered to be in contact and are represented by 1, while those with a distance of more than 8 Å are not in contact and are represented by 0. Assuming the protein sequence length is L, an L*L contact matrix is obtained by calculation.
3. The method for predicting single-sequence protein contact maps based on graph neural networks according to claim 1, characterized in that, The ReLU activation function is expressed as follows: 。 4. The method for predicting single-sequence protein contact maps based on graph neural networks according to claim 1, characterized in that, The ResNet module consists of 5 convolutional layers, with batch normalization layers and ReLU activation layers between each convolutional layer.
5. The method for predicting single-sequence protein contact maps based on graph neural networks according to claim 1, characterized in that, The multilayer perceptron consists of multiple fully connected layers, with a Droout layer added to each convolutional layer, and residual operations are also incorporated.
6. The method for predicting single-sequence protein contact maps based on graph neural networks according to any one of claims 1-5, characterized in that, Step S5 is implemented as follows: S5.1 The one-hot one-dimensional feature is concatenated with the one-dimensional protein coding feature to form a one-dimensional feature with a dimension of L*1300; The one-dimensional feature with dimension L*1300 described in S5.2 is input into the graph convolutional coding model, and outputs a graph structure feature with dimension L*420. S5.3 The graph structure features of dimension L*420 are concatenated to obtain a two-dimensional matrix of dimension L*L*420; S5.4 The two-dimensional matrix with dimension L*L*420 is concatenated with the two-dimensional attention matrix obtained by the pre-trained model to obtain a two-dimensional matrix with dimension L*L*440. S5.5 Input the two-dimensional matrix with dimension L*L*440 into the convolutional decoding model to obtain three different residue pair contact prediction method matrices.