Deep learning-based methods for predicting protein-protein interaction sites
By employing a deep learning-based approach and utilizing a fine-tuned protBERT model and feature extraction module, the problem of high time cost in predicting protein-protein interaction sites was solved, achieving high-precision and efficient prediction results.
Patent Information
- Application Number
- CN202410829250.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-25
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-06-25
AI Technical Summary
Current technologies for identifying protein-protein interaction sites in life science experiments are time-consuming and costly, and make it difficult to make predictions efficiently.
A deep learning-based approach is adopted, using a fine-tuned protBERT model to generate protein sequence embeddings, combining a feature extraction module and the PageRank mechanism for feature classification, and finally using a fully connected layer for binary classification to predict protein interaction sites.
It achieves high-precision prediction of protein-protein interaction sites, significantly reducing time costs and improving prediction efficiency.
Smart Images

Figure CN118692553B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of protein interaction site prediction, and more particularly to a method for predicting protein interaction sites based on deep learning. Background Technology
[0002] Protein-protein interactions refer to the process by which two or more protein molecules combine through non-covalent bonds to exert specific biochemical effects, playing a crucial role in many life activities. Studying protein interaction sites is of great significance for predicting protein biochemical functions and developing novel drugs.
[0003] However, conventional life science experiments for identifying protein-protein interaction sites are very time-consuming and costly. Therefore, how to use deep learning models to predict protein-protein interaction sites is a technical problem that urgently needs to be solved.
[0004] To this end, this invention proposes a method for predicting protein interaction sites based on deep learning. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a deep learning-based method for predicting protein interaction sites.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A deep learning-based method for predicting protein-protein interaction sites first takes a protein sequence as input and generates embeddings using a fine-tuned protBERT model. Then, the embeddings are passed to a feature extraction module, which consists of L layers of co-evolutionary graph attention (CGAT) blocks with initial residuals and identity mappings. The PageRank mechanism is used for thorough feature classification across all layers. Finally, the prediction module performs binary classification on each residue to determine whether it is an interaction site using fully connected layers, specifically including the following steps:
[0008] S1: Graph construction, calculating the original DCA matrix for each protein sequence based on MSAs generated by Blastp;
[0009] S2: Feature input generates the initial feature matrix. Given a protein with n amino acids, the embedding is obtained by fine-tuning the prot-BERT model. Then inject position encoding Provides absolute or relative position information of parameters in the sequence;
[0010] S3: L-layer co-evolutionary graph attention block;
[0011] S4: Prediction, passing the output vector of the features to a fully connected layer, where ReLU and dropout are used in every layer except the last one; after the last layer, the Sigmoid function is used to generate the final prediction;
[0012] S5: Evaluation, which uses residual-based true and predicted classes to comprehensively evaluate prediction performance through multiple evaluation metrics.
[0013] Preferably, the fine-tuned protBERT model is an adjusted model based on the protBERT model, and its adjustment method includes the following steps:
[0014] A1: Initialize using the original protBERT model parameters;
[0015] A2: During training, the parameters of all self-attention layers except the last two are frozen;
[0016] A3: Two convolutional kernels and a sig-moid function were added on top of the protBERT model to transform the output vector of each residual into a prediction about whether it is a PPI site;
[0017] A4: Use the Adam optimizer and adjust the parameters to a learning rate of 1e. -6 The batch size is set to 1.
[0018] Preferably: In step S1, graph construction is performed by converting any value less than a selected cutoff value to 0 and keeping any value greater than or equal to the cutoff value unchanged. The cutoff value is different for each protein and is set to the highest 10% quantile of the original DCA matrix for each protein sequence. The formula for generating graph P is as follows:
[0019]
[0020] Preferably: In step S2, the formula for feature calculation is as follows: Where d pe This refers to the size of the position in the encoding; in step S2, the dimensionality is reduced by using a fully connected layer without an activation function: H (0) =XW+b, to obtain the initial feature matrix of the protein.
[0021] Preferably, step S3 includes the following steps:
[0022] S31: Graph attention mechanism enhanced by co-evolutionary information: Given a protein with n amino acids, compute the i-th co-evolutionary graph attention block;
[0023] S32: Further, the initial residuals and identity mappings are added to the feature matrix of proteins across different layers;
[0024] S33: Use the PageRank mechanism to aggregate all output feature matrices.
[0025] Preferably: In step S31, the calculation formula is as follows: in It represents the intermediate feature matrix of the (l-1)th layer; This represents the intermediate feature matrix of the l-th layer; This represents the DCA value between the remainders i and j; It is an attention matrix. This is a layer matrix, where W1, W2, W3, and q are trainable parameters. σ is the activation function. Preferably, in step S32, the formula used is: Here H (l) This represents the final feature matrix generated by the l-th layer of the CGAT block. α and β l It's a hyperparameter, β l =log(λ / 1+1) is determined by another hyperparameter λ.
[0026] Preferably, in step S33, the formula used is:
[0027]
[0028] Preferably: In step S4, the formula for passing the output vector of the feature to the fully connected layer is x. (t) =Relu(x (t-1) W (t-1) +b (t-1) ), t∈[1,T], x (t-1) and x (t) Let x represent the input vector and output vector of the t-th fully connected layer, respectively. (0) =ZW (t) Let b represent the weight matrix. (t) This indicates a bias.
[0029] Preferably, in step S5, the evaluation metrics include the area under the accuracy-recall curve (AUPRC), accuracy (ACC), recall, precision, F-measure (F1), and Matthews correlation coefficient (MCC), and the evaluation formula is as follows: TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false negative; TP and TN refer to correctly predicted interaction sites and non-interaction sites, respectively; FP and FN are applicable to incorrectly predicted interaction sites and non-interaction sites, respectively.
[0030] The beneficial effects of this invention are as follows:
[0031] 1. This invention uses a deep learning model to predict protein interaction sites, which significantly reduces the time cost compared to conventional life science experiments, and also provides relatively high prediction accuracy. Attached Figure Description
[0032] Figure 1 This is a model architecture diagram of the deep learning-based protein interaction site prediction method proposed in this invention. Detailed Implementation
[0033] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.
[0034] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "setting" should be interpreted broadly. For example, they can refer to a fixed connection or setting, a detachable connection or setting, or an integral connection or setting. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0035] Example 1:
[0036] The deep learning-based protein interaction site prediction method first generates embeddings using a fine-tuned protBERT model, taking the protein sequence as input. Then, the embeddings are passed to a feature extraction module, which consists of L layers of co-evolutionary graph attention (CGAT) blocks with initial residuals and identity mappings. The PageRank mechanism is used for thorough feature classification across all layers. Finally, the prediction module performs binary classification on each residue to determine whether it is an interaction site using fully connected layers, specifically including the following steps:
[0037] S1: Graph construction, calculating the original DCA matrix for each protein sequence based on MSAs generated by Blastp;
[0038] S2: Feature input generates the initial feature matrix. Given a protein with n amino acids, the embedding is obtained by fine-tuning the prot-BERT model. Then inject position encoding Provides absolute or relative position information of parameters in the sequence;
[0039] S3: L-layer co-evolutionary graph attention block;
[0040] S4: Prediction, passing the output vector of the features to a fully connected layer, where ReLU and dropout are used in every layer except the last one; after the last layer, the Sigmoid function is used to generate the final prediction;
[0041] S5: Evaluation, which uses residual-based true and predicted classes to comprehensively evaluate prediction performance through multiple evaluation metrics.
[0042] The fine-tuned protBERT model is an adjusted model based on the protBERT model, and its adjustment method includes the following steps:
[0043] A1: Initialize using the original protBERT model parameters;
[0044] A2: During training, the parameters of all self-attention layers except the last two are frozen;
[0045] A3: Two convolutional kernels and a sig-moid function were added on top of the protBERT model to transform the output vector of each residual into a prediction about whether it is a PPI site;
[0046] A4: Use the Adam optimizer and adjust the parameters to a learning rate of 1e. -6 The batch size is set to 1.
[0047] Example 2:
[0048] The deep learning-based protein interaction site prediction method first generates embeddings using a fine-tuned protBERT model, taking the protein sequence as input. Then, the embeddings are passed to a feature extraction module, which consists of L layers of co-evolutionary graph attention (CGAT) blocks with initial residuals and identity mappings. The PageRank mechanism is used for thorough feature classification across all layers. Finally, the prediction module performs binary classification on each residue to determine whether it is an interaction site using fully connected layers, specifically including the following steps:
[0049] S1: Graph construction, calculating the original DCA matrix for each protein sequence based on MSAs generated by Blastp;
[0050] S2: Feature input generates the initial feature matrix. Given a protein with n amino acids, the embedding is obtained by fine-tuning the prot-BERT model. Then inject position encoding Provides absolute or relative position information of parameters in the sequence;
[0051] S3: L-layer co-evolutionary graph attention block;
[0052] S4: Prediction, passing the output vector of the features to a fully connected layer, where ReLU and dropout are used in every layer except the last one; after the last layer, the Sigmoid function is used to generate the final prediction;
[0053] S5: Evaluation, which uses residual-based true and predicted classes to comprehensively evaluate prediction performance through multiple evaluation metrics.
[0054] The fine-tuned protBERT model is an adjusted model based on the protBERT model, and its adjustment method includes the following steps:
[0055] A1: Initialize using the original protBERT model parameters;
[0056] A2: During training, the parameters of all self-attention layers except the last two are frozen;
[0057] A3: Two convolutional kernels and a sig-moid function were added on top of the protBERT model to transform the output vector of each residual into a prediction about whether it is a PPI site;
[0058] A4: Use the Adam optimizer and adjust the parameters to a learning rate of 1e. -6 The batch size is set to 1.
[0059] In step S1, graph construction is performed by converting any value less than a selected cutoff value to 0 and keeping any value greater than or equal to the cutoff value unchanged. The cutoff value is different for each protein and is set to the highest 10% quantile of the original DCA matrix for each protein sequence. The formula for generating graph P is as follows:
[0060] In step S2, the formula for feature calculation is: Where d pe It refers to the size of the position in the encoding.
[0061] In step S2, to reduce computational burden, a fully connected layer without an activation function is used to reduce dimensionality: H (0) =XW+b, to obtain the initial feature matrix of the protein.
[0062]
[0063] Step S3 includes the following steps:
[0064] S31: Graph attention mechanism enhanced by co-evolutionary information: Given a protein with n amino acids, compute the i-th co-evolutionary graph attention block;
[0065] S32: Further, the initial residuals and identity mappings are added to the feature matrix of proteins across different layers;
[0066] S33: Use the PageRank mechanism to aggregate all output feature matrices.
[0067] In step S31, the calculation formula is as follows: in It represents the intermediate feature matrix of the (l-1)th layer; This represents the intermediate feature matrix of the l-th layer; This represents the DCA value between the remainders i and j; It is an attention matrix. It is a layer matrix, where W1, W2, W3, and q are trainable parameters. σ is the activation function.
[0068] In step S32, the formula used is: H (l) =[(1-α)H~ (l) +αH (0) ][(1-β l )I+β l ], here H (l) This represents the final feature matrix generated by the l-th layer of the CGAT block. α and β l It's a hyperparameter, β l =log(λ / 1+1) is determined by another hyperparameter λ.
[0069] In step S33, the formula used is: In step S4, the formula for passing the feature output vector to the fully connected layer is x. (t) =Relu(x (t-1) W (t-1) +b (t-1) ), t∈[1,T], x (t-1) and x (t) Let x represent the input vector and output vector of the t-th fully connected layer, respectively. (0) =ZW (t) Let b represent the weight matrix. (t) This indicates a bias.
[0070] In step S4, the Sigmoid function is: This represents the predicted probability that the residue is an interaction site. It is the predicted probability that the residue is a non-interacting site.
[0071] In step S5, the evaluation metrics include the area under the accuracy-recall curve (AUPRC), accuracy (ACC), recall, precision, F-measure (F1), and Matthews correlation coefficient (MCC), and the evaluation formula is as follows: TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false negative; TP and TN refer to correctly predicted interaction sites and non-interaction sites, respectively; FP and FN are applicable to incorrectly predicted interaction sites and non-interaction sites, respectively.
[0072] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A deep learning-based method for predicting protein-protein interaction sites: First, the protein sequence is used as input, and an embedding is generated through a fine-tuned protBERT model. Then, the embedding is passed to the feature extraction module, which consists of L layers of co-evolutionary graph attention blocks with initial residuals and identity mappings, and the PageRank mechanism is used for thorough classification of features across all layers; finally, the prediction module performs binary classification of whether each residue is an interaction site through fully connected layers, characterized by including the following steps: S1: Graph construction, calculating the original DCA matrix for each protein sequence based on MSAs generated by Blastp; S2: Feature input generates the initial feature matrix. Given a protein with n amino acids, the embedding is obtained by fine-tuning the prot-BERT model. Then inject position encoding Provides absolute or relative position information of parameters in the sequence; S3: L-layer co-evolutionary graph attention block; S4: Prediction, passing the output vector of the features to a fully connected layer, where ReLU and dropout are used in every layer except the last one; after the last layer, the Sigmoid function is used to generate the final prediction; S5: Evaluation, which uses residual-based true and predicted classes to comprehensively evaluate prediction performance through multiple evaluation metrics; Step S3 includes the following steps: S31: Graph attention mechanism enhanced by co-evolutionary information: Given a protein with n amino acids, compute the i-th co-evolutionary graph attention block; S32: Further, the initial residuals and identity mappings are added to the feature matrix of proteins across different layers; S33: Use the PageRank mechanism to aggregate all output feature matrices; In step S31, the calculation formula is as follows: , in , which indicates the first Layer intermediate feature matrix; This represents the intermediate feature matrix of the l-th layer; , representing the DCA value between remainders i and j; It is an attention matrix. It is a layer matrix, where W1, W2, W3, and q are trainable parameters, and σ is the activation function; The formula used in step S32 is: H here (l) Let α and β represent the final feature matrix generated by the l-th layer of the CGAT block. l It's a hyperparameter. It is determined by another hyperparameter λ.
2. The method for predicting protein interaction sites based on deep learning according to claim 1, characterized in that, The fine-tuned protBERT model is an adjusted model based on the protBERT model, and its adjustment method includes the following steps: A1: Initialize using the original protBERT model parameters; A2: During training, the parameters of all self-attention layers except the last two are frozen; A3: Two convolutional kernels and a sig-moid function were added on top of the protBERT model to transform the output vector of each residual into a prediction about whether it is a PPI site; A4: Use the Adam optimizer and adjust the parameters to a learning rate of 1e. -6 The batch size is set to 1.
3. The method for predicting protein interaction sites based on deep learning according to claim 1, characterized in that, In step S1, graph construction is performed by converting any value less than a selected cutoff value to 0 and keeping any value greater than or equal to the cutoff value unchanged. The cutoff value is different for each protein and is set to the highest 10% quantile of the original DCA matrix for each protein sequence. The formula for generating graph P is as follows: 。 4. The method for predicting protein interaction sites based on deep learning according to claim 1, characterized in that, In step S2, the formula for feature calculation is: , in This refers to the dimension of the position in the encoding; in step S2, the dimensionality is reduced by using a fully connected layer without an activation function. The initial feature matrix of the protein is obtained. .
5. The method for predicting protein interaction sites based on deep learning according to claim 1, characterized in that, In step S33, the formula used is: 。 6. The method for predicting protein interaction sites based on deep learning according to claim 1, characterized in that, In step S4, the formula for passing the feature output vector to the fully connected layer is as follows: x (t-1) and x (t) Let x represent the input vector and output vector of the t-th fully connected layer, respectively. (0) =ZW (t) Let b represent the weight matrix. (t) This indicates a bias.
7. The method for predicting protein interaction sites based on deep learning according to claim 1, characterized in that, In step S5, the evaluation metrics include the area under the accuracy-recall curve, accuracy, recall, F-measure, and Matthews correlation coefficient. The evaluation formula is: ; TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false negative; TP and TN refer to correctly predicted interaction sites and non-interaction sites, respectively; FP and FN are applicable to incorrectly predicted interaction sites and non-interaction sites, respectively.
Citation Information
Patent Citations
Protein interaction site prediction method based on residue position characteristics
CN113257342A
Protein interaction point location prediction method based on cross-graph attention
CN115620803A
