Prediction method for mining protein interaction types based on deep learning
By constructing the SE3NET-PPI model, fusing protein structure information and network topology information, and using deep learning technology, the prediction accuracy of protein interaction types is improved.
Patent Information
- Application Number
- CN202211320402.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-26
- Publication Date
- 2025-08-08
- Estimated Expiration
- 2042-10-26
AI Technical Summary
The existing evaluation framework for predicting protein interaction types is unreasonable, with low accuracy, and failing to effectively integrate protein structure information and network topology information.
The SE3NET-PPI model was constructed, and end-to-end protein interaction type prediction was carried out by fusing protein structure information and interaction network topology information, using twin architecture, convolutional neural network CNN, pyramid pooling technology SPP and isomorphic neural network GIN.
It improves the prediction accuracy of protein interaction types and can effectively utilize protein sequence information, structural information and network topology information.
Smart Images
Figure CN115588463B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of computer-aided drug design and relates to a prediction method for mining protein interaction types based on deep learning. Background Art
[0002] As common molecules in cells, proteins play a crucial role in regulating various biological processes. The activity of biological cells is controlled by interacting proteins through metabolic pathways and non-complexes. Protein-protein interaction (PPI) networks and transcriptional regulatory networks play a crucial role in regulating cells and their signaling. Furthermore, from a practical perspective, monitoring and studying PPIs can provide interesting and important candidates for medically applicable diagnostic and therapeutic targets, thereby facilitating the design of new drugs. Therefore, predicting PPIs is a fundamental research topic in systems biology and has garnered increasing attention in recent years.
[0003] There are generally two approaches to predicting PPIs: experimental and computational. However, laboratory-based methods are often time-consuming and labor-intensive, and the results obtained are often unreliable and subject to significant uncertainty. Furthermore, the recent development of high-throughput technologies has generated a vast amount of proteomics data, creating an urgent need to learn reliable computational methods from this accumulated PPI data to accurately predict unknown PPIs. The development of computational methods can be roughly divided into two phases. Early research based on machine learning (ML) provided viable solutions, but their performance was limited by the expressiveness of PPI feature representations and models. With the tremendous success of deep learning in computer vision and natural language processing, deep learning (DL) has garnered widespread attention as a powerful tool that can automatically learn data representations at multiple levels of abstraction without requiring prior knowledge. The application of deep learning to PPI prediction has recently been discussed in detail, with numerous studies demonstrating that deep learning has become a powerful tool for PPI prediction, leading to scientific discoveries and practical solutions. For example, Chen et al. proposed an end-to-end framework called PIPR that uses only protein sequence information. Using a twin architecture and a deep bidirectional recurrent convolutional neural network with residual connections, they achieved good results on three protein prediction problems: 1) protein classification, 2) protein binary classification, and 3) protein affinity. However, the model did not integrate protein structural information and network topology information. Based on this, the present invention designs a deep learning-based SE3NET-PPI model that integrates the topological structure of the PPI network and protein structural features to improve the accuracy of the model in predicting protein interaction types. Summary of the Invention
[0004] The purpose of this invention is to address the shortcomings of existing protein interaction type prediction methods, such as unreasonable evaluation framework and low accuracy, and to propose a method based on deep learning that integrates protein structure information and protein interaction network topology information to improve the accuracy of protein interaction type prediction.
[0005] The present invention specifically comprises the following steps:
[0006] Step 1: Build a data set, specifically:
[0007] First, protein sequence information and corresponding protein structure information are obtained from existing public databases, and proteins with a protein sequence length of less than 2700 are screened out to form a protein sequence structure database.
[0008] Then, the protein ID of the protein interaction database STRING is matched in the Uniport website according to the protein number in the protein sequence structure database. The corresponding protein interaction pairs and the corresponding interaction types of the interaction pairs are collected in the STRING database according to the STRING ID to construct the protein interaction database.
[0009] Finally, 3000 and 9000 proteins were randomly selected from the protein interaction database, and all proteins in the database were used to generate three datasets with different data amounts: String_3000, String_9000, and String_all;
[0010] The three data sets were clustered for homology, and the homology was set to be less than or equal to 40%.
[0011] Step 2: Construct a protein interaction network and divide it into test and training sets:
[0012] A protein-protein interaction network was constructed based on all protein-protein interaction pairs in the String_3000, String_9000, and String_all datasets. The three datasets were then partitioned into training and test sets using random walks, depth-first sparse search (DFS), and breadth-first sparse search (BFS). The training set comprised 80% of the dataset, while the test set comprised the remaining 20%. This resulted in nine datasets (partitioned using three different partitioning methods: String_3000-BFS, String_3000-DFS, String_3000-Random, String_9000-BFS, String_9000-DFS, String_9000-Random, String_all-BFS, String_all-DFS, and String_all-Random).
[0013] Step 3: Construct the protein interaction type prediction model SE3NET-PPI and train the model;
[0014] The protein interaction type prediction model SE3NET-PPI includes a structure module, a protein network module, and a prediction module;
[0015] The structural module adopts a twin architecture, which is composed of two neural networks with the same structure and shared weights.
[0016] In the structure module, the protein structure information in the protein information structure database is converted into a SE (3) invariant matrix diagram, and then the SE (3) invariant matrix diagram is input into the convolutional neural network CNN to obtain the convolved feature vector; the convolved feature vector is average pooled by the pyramid pooling layer SPP to obtain the pooled feature vector; finally, all the feature vectors after the pooling operation are spliced together to obtain the protein structure feature vector, and the multi-layer perceptron (MLP) is used to convert the protein structure feature vector into a feature vector of a specific length.
[0017] Specifically: Get the protein structure file for each amino acid residue The atomic coordinates are expressed as ,in is the protein sequence length, It is residues The three-dimensional coordinates of the atoms are obtained by The formula calculates the Euclidean distance between them to construct the SE (3) invariant matrix diagram ; The SE (3) invariant matrix diagram Input into CNN and get the convolved feature vector ; Use three pooling windows of different sizes The feature vector after convolution Perform average pooling operation on features in different fields of view to obtain the pooled feature vector , and finally all the feature vectors after the pooling operation are Stitched together ;Will Input to the multi-layer perceptron MLP to get .
[0018] Based on the twin architecture, a pair of protein structure information files are input, and after the above steps, a pair of protein structure features are obtained. 、 .
[0019] In the protein network module, the ProtT5 protein pre-training model is used to convert protein sequence information into corresponding protein sequence feature vectors; the protein node features of the protein interaction network are obtained through the protein sequence feature vectors, and the protein node features of the protein interaction network and the interaction relationship features (edge features) of the protein interaction network are input into the homomorphic graph neural network GIN layer; the interaction relationship features (edge features) of the protein interaction network are obtained through the protein interaction network.
[0020] Specifically, the node features of a protein are summed with the node features of its neighboring proteins with a degree of 1, that is, ; Use MLP to update the current protein node, that is ; Input node features and edge features to obtain the network topology information features of all proteins in the dataset , and then screen out the predicted protein pair features 、 .
[0021] In the prediction module, feature fusion is used to concatenate a pair of protein structure features output by the structure module and a pair of protein features output by the protein network module to obtain a feature fusion vector. The feature fusion vector is input into the MLP, and the output is a 1*7 dimensional 0, 1 matrix representing the prediction of the protein's corresponding action category.
[0022] Specifically: feature fusion is used to transform a pair of protein structure features output by the structure module as well as Protein pair features output by the protein network module as well as Splicing together to get a pair of information-rich latent vectors is represented as , ,Right now , , and then calculate the characteristic information of the protein pair by the following formula : ; ; ;
[0023] in, Represents the difference calculation for each element, Represents the result of difference calculation; On behalf of Adamaj, Represents the result of the Hadamard product calculation.
[0024] Fusion of feature vectors Input into a layer of MLP layer, the output is a 1*7 dimensional 0, 1 matrix, which represents the prediction of the protein to its corresponding action category, that is, .
[0025] The SE3NET-PPI model was trained using data from the training set to obtain the protein interaction type prediction model SE3NET-PPI, which can predict the type of unknown interaction relationships.
[0026] Compared with existing methods, this method constructs a protein interaction type prediction model SE3NET-PPI, which only requires protein sequence information to complete end-to-end training. It converts protein 3D structural information into SE (3) invariant matrix graphs and uses convolutional neural network CNN and pyramid pooling technology SPP to extract structural features. A protein interaction network is constructed based on protein pairs in the protein interaction database. The protein sequence is embedded in the pre-trained model layer to obtain protein node features and the aforementioned protein interaction network is input into the homogeneous graph neural network GIN to extract the topological information structure of the protein network. The above features are input into the MLP after feature fusion, and the prediction results of the corresponding interaction category of the protein pair are output. The protein sequence information, structural information and topological information of the PPI network are integrated to predict the protein interaction type, thereby improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 is a flow chart of the present invention;
[0028] Figure 2 Structural diagram of a model for predicting protein interaction types. DETAILED DESCRIPTION
[0029] The present invention will be further described below with reference to the accompanying drawings and specific examples. It should be understood that the examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention.
[0030] like Figure 1 As shown in FIG, a prediction method for mining protein interaction types based on deep learning includes the following steps:
[0031] Step 1: Build a data set, specifically:
[0032] First, obtain the protein structure file (which contains protein sequence information and protein structure information) and build a protein sequence structure database;
[0033] In this example, all human subset proteins were retrieved from the Uniport website and the sequence information of the corresponding proteins was downloaded. The aforementioned sequence information was input into the AlphaFold2 model to obtain all protein structure files of the human subset. Then, proteins with a protein sequence length of less than 2700 were screened to form a protein sequence structure database.
[0034] Then, according to the protein number (Uniport ID) in the protein sequence structure database, the STRING ID of the protein interaction database STRING is matched in the Uniport website. According to the STRING ID, the corresponding protein interaction pairs and the corresponding interaction types of the interaction pairs are collected in the STRING database to construct the protein interaction database.
[0035] Finally, 3000 and 9000 proteins were randomly selected from the protein interaction database, and all proteins in the database were used to generate three datasets with different data amounts, String_3000, String_9000, and String_all. The three datasets were then clustered by homology (eliminating some similarities in protein sequences) and set to be less than or equal to 40%.
[0036] Through the above steps, a protein sequence structure database and three data sets for model training and evaluation were obtained (the data sets contained different numbers of protein interaction pairs and the corresponding interaction types of the interaction pairs).
[0037] Step 2: Divide the three datasets String_3000, String_9000, and String_all obtained in Step 1 into 9 datasets using the depth-first search algorithm BFS, breadth-first search algorithm DFS, and random partitioning algorithm Random for scientific evaluation of subsequent model effects;
[0038] In each of the nine datasets, 80% of the protein interaction pairs were used for training, with the remaining 20% used for testing. Each sample in the training set consists of a single protein interaction pair. Each protein interaction pair is assigned a label in seven categories, corresponding to seven types of protein interactions (reaction, binding, post-translational modification, activation, inhibition, catalysis, and expression). If the pair has the corresponding type of interaction, it is labeled 1; otherwise, it is labeled 0. It is worth noting that a protein pair may have multiple interaction types. The test set consists of a single protein interaction pair, whose interaction type is unknown.
[0039] Step 3: Build Figure 2 The protein interaction type prediction model SE3NET-PPI is used and the model is trained; the specific steps are as follows:
[0040] (1) Protein structure feature extraction: The protein sequence is predicted by AlphaFold2 to output 3D structural information; the 3D structural information is converted into a SE (3) invariant matrix and the structural features are extracted using convolutional neural network (CNN) and pyramid pooling technology (SPP).
[0041] (2) Extraction of topological information features of protein interaction network: A protein interaction network is constructed based on the protein pairs in the protein interaction database. The protein sequence is then embedded in the pre-trained model layer to obtain protein node features and the aforementioned protein interaction network is input into the homomorphic graph neural network (GIN) to extract the topological information structure of the protein network.
[0042] (3) Feature fusion: by combining a pair of protein structural features output by the structural module as well as A pair of protein features output by the protein network module as well as Splicing together to get a pair of information-rich latent vectors is represented as , ,Right now , , and then calculate the characteristic information of the protein pair by the following formula .
[0043]
[0044]
[0045]
[0046] in, Represents the difference calculation for each element, Represents the result of difference calculation; On behalf of Adamaj, Represents the result of the Hadamard product calculation.
[0047] The prediction layer is used to give the prediction result of the interaction type; specifically, the feature fusion vector Input into a layer of MLP layer, the output is a 1*7 dimensional 0, 1 matrix, which represents the prediction of the protein to its corresponding action category, that is, .
[0048] The SE3NET-PPI model is trained using the data in the training set with the following training parameters: Learning rate: 0.001, Optimizer: AdamW, Batch_size: 256, Dropout 0.5.
[0049] For a given training set and protein interaction types , trained using multi-task binary cross entropy as the loss function:
[0050] Loss function:
[0051] in, represents the real protein interaction type; Represents the protein interaction type predicted by the model; n represents the number of protein pairs involved in the training.
[0052] Step 4: Use different test sets to predict the trained protein interaction type prediction model SE3NET-PPI, obtain the test results of the model in different test sets and give an evaluation.
[0053] This invention expands the work of AlphaFold2, introduces SE (3) invariant matrix graph and pyramid pooling layer technology to extract protein structural information for protein prediction; and integrates protein structure and interaction network information between proteins to improve the prediction accuracy of protein interaction type prediction model.
Claims
1. A method for predicting protein interaction types based on deep learning, characterized by: The specific steps include: Step 1: Build a data set, specifically: First, we obtained protein sequence information and corresponding protein structure information from existing public databases, and screened out proteins with a protein sequence length of less than 2700 to form a protein sequence structure database; Then, the protein ID of the protein interaction database STRING is matched on the Uniport website according to the protein number in the protein sequence structure database. The corresponding protein interaction pairs and the corresponding interaction types of the interaction pairs are collected in the STRING database according to the STRING ID to construct the protein interaction database; Finally, 3000 and 9000 proteins were randomly selected from the protein interaction database, and all proteins in the database were used to generate three datasets with different data amounts: String_3000, String_9000, and String_all; Step 2: Construct a protein interaction network and divide it into test and training sets: Construct a protein interaction network based on all protein interaction pairs in the datasets String_3000, String_9000, and String_all in step 1; use random walk Random, depth-first algorithm DFS, and breadth-first algorithm BFS to evolve the datasets String_3000, String_9000, and String_all into nine datasets, namely String_3000-BFS, String_3000-DFS, String_3000- Random, String_9000-BFS, String_9000-DFS, String_9000- Random, String_all-BFS, String_all-DFS, and String_all- Random; and divide the data in all datasets into training and test sets; Step 3: Construct the protein interaction type prediction model SE3NET-PPI and train the model; The protein interaction type prediction model SE3NET-PPI includes a structure module, a protein network module, and a prediction module; The structural module adopts a twin architecture, which is composed of two neural networks with the same structure and shared weights; In the structure module, the protein structure information in the protein information structure database is converted into a SE (3) invariant matrix diagram, and then the SE (3) invariant matrix diagram is input into the convolutional neural network CNN to obtain the convolutional feature vector; the convolutional feature vector is average pooled by the pyramid pooling layer SPP to obtain the pooled feature vector; finally, all the features after the pooling operation are spliced together to obtain the protein structure feature vector, and the multi-layer perceptron MLP is used to convert the protein structure feature vector into a feature vector of a specific length; Based on the twin architecture, a pair of protein structure information files are input. After the above steps, a pair of protein structure feature vectors are obtained; In the protein network module, the ProtT5 protein pre-training model is used to convert protein sequence information into corresponding protein sequence feature vectors; The protein node features of the protein interaction network are obtained through the protein sequence feature vector, and the protein node features of the protein interaction network and the interaction relationship features of the protein interaction network are input into the homomorphic graph neural network (GIN) layer; the interaction relationship features of the protein interaction network are obtained through the protein interaction network; In the prediction module, feature fusion is used to concatenate a pair of protein structure features output by the structure module and a pair of protein features output by the protein network module to obtain a feature fusion vector. The feature fusion vector is input into the MLP, and the output is a 1*7 dimensional 0, 1 matrix representing the prediction of the protein's corresponding action category. The SE3NET-PPI model was trained using data from the training set to obtain the protein interaction type prediction model SE3NET-PPI, which can predict the type of unknown interaction relationships.
2. The method for predicting protein interaction types based on deep learning according to claim 1, characterized in that: The homology clustering of the three data sets described in step 1 is less than or equal to 40%.
3. The method for predicting protein interaction types based on deep learning according to claim 1, characterized in that: In the structure module, the structure of each amino acid residue is obtained from the protein structure file. The atomic coordinates are expressed as ,in is the protein sequence length, It is residues The three-dimensional coordinates of the atoms and The formula calculates the Euclidean distance between them to construct the SE (3) invariant matrix diagram ; The SE (3) invariant matrix diagram Input into CNN and get the convolved feature vector ; Use three pooling windows of different sizes The feature vector after convolution Perform average pooling operation on features in different fields of view to obtain the pooled feature vector , and finally all the feature vectors after the pooling operation are Stitched together ;Will Input to the multi-layer perceptron MLP to get ; Based on the twin architecture, a pair of protein structure information files are input, and after the above steps, a pair of protein structure features are obtained. as well as .
4. The method for predicting protein interaction types based on deep learning according to claim 1, characterized in that: In the protein network module, the node features of a protein are summed with the node features of its neighboring proteins with a degree of 1, that is, ; Use MLP to update the current protein node, that is ; Input node features and edge features to obtain the network topology information features of all proteins in the dataset , and then screen out the predicted protein pair features 、 .
5. The method for predicting protein interaction types based on deep learning according to claim 1, characterized in that: In the prediction module, feature fusion is used to transform a pair of protein structure features output by the structure module as well as Protein pair features output by the protein network module as well as Splicing together to get a pair of information-rich latent vectors is represented as , ,Right now , , and then calculate the characteristic information of the protein pair by the following formula : ; ; ; in, Represents the difference calculation for each element, Represents the result of difference calculation; On behalf of Adamaj, Represents the result of Hadamard product calculation; Fusion of feature vectors Input into MLP, the output is a 1*7 dimensional 0, 1 matrix, which represents the prediction of the protein to its corresponding action category.