A protein interaction prediction method based on graph neural network

Abstract

This invention discloses a protein interaction prediction method based on a graph neural network, belonging to the technical field of protein interaction prediction. The method comprises the following steps: S1: amino acid sequence preprocessing; S2: extraction of protein structure and sequence information; S3: fusion of structure information and feature information; S4: construction of a graph neural network; S5: model network training; and S6: protein-protein interaction prediction. This method uses a fusion of two features as a protein feature expression; employs the ESM‑2 module to extract protein structural features from the amino acid sequence of the protein, eliminating the need for separate structural features; and employs an asymmetric loss function to effectively address the sample label imbalance problem in protein-protein interactions.

Description

Technical Field

[0001] The present invention relates to the technical field of protein interaction prediction, and in particular to a protein interaction prediction method based on graph neural network. Background Art

[0002] Proteins play an important role in organisms and play a vital role in life activities. They can exist alone or interact with other proteins in the organism. However, they usually do not act as a single substance, but as members of a biological network. Protein interactions are particularly important in biological processes, such as signal transduction, promoting or inhibiting biological functions, etc. Therefore, the study of protein interactions helps us understand the biological processes in which proteins participate in organisms, which is beneficial to the treatment of diseases, drug development, medical development, and understanding of molecular mechanisms. However, the current prediction of protein-protein interactions mostly relies on laboratory methods, which are time-consuming, material-consuming, and labor-intensive, and can only be used for specific proteins and are difficult to apply to all proteins. Obviously, for protein-protein interactions, a high-precision and reliable method is expected to predict unknown protein reaction pairs.

[0003] With the development of neural networks, deep learning methods and graph neural networks are increasingly being used in this research area. Neural networks for predicting protein-protein interactions can be categorized into those that rely solely on amino acid sequences and those that combine amino acid sequences with protein structure. Methods that rely solely on amino acid sequences extract sequence features as protein signatures, but these methods ignore protein structural features, which can negatively impact prediction results. Methods that combine amino acid sequence and protein structural features consider both sequence information and structural features. While this complements the protein's structural features and enriches its signature, the process of extracting protein structural features is extremely complex.

[0004] In summary, laboratory methods for predicting protein-protein interactions are time-consuming and labor-intensive. Methods that rely solely on amino acid sequences ignore protein structural features, while methods that combine amino acid sequences with protein structural features are complex in extracting structural information. Therefore, a protein interaction prediction method based on graph neural networks is proposed. Summary of the Invention

[0005] The technical problem to be solved by the present invention is: how to extract the key features representing proteins, and when the representative features are obtained, how to accurately and efficiently predict the interactions between proteins, realize automatic prediction through machines, reduce laboratory research costs, and provide a protein interaction prediction method based on graph neural network.

[0006] The present invention solves the above technical problems through the following technical solutions, which include the following steps:

[0007] S1: Amino acid sequence preprocessing

[0008] Obtain protein amino acid sequences and public interaction datasets, and preprocess the amino acid sequences according to task requirements;

[0009] S2: Extract protein structure and sequence information

[0010] The ESM-2 model is used to extract structural information, and the recurrent gating unit is used to extract sequence information.

[0011] S3: Fusion of structural information and feature information

[0012] Fuse the structural information and sequence information to obtain fusion features;

[0013] S4: Building a Graph Neural Network

[0014] Build a graph neural network based on protein interaction and fusion features to obtain a protein-protein interaction graph neural network;

[0015] S5: Model network training

[0016] The training set is used to train the protein-protein interaction graph neural network to obtain a protein-protein interaction prediction model;

[0017] S6: Protein-protein interaction prediction

[0018] The unknown protein is input into the protein-protein interaction prediction model for prediction to obtain the prediction results.

[0019] Furthermore, in step S1, the specific processing process is as follows:

[0020] S11: Obtain the amino acid sequence of proteins and a public interaction dataset. Fix the length of each protein's amino acid sequence to N amino acids. Then use the Skip-Gram model to train the amino acid sequence. Concatenate the resulting embedding vector and the one-hot encoding defined based on the electrical stability and hydrophobicity similarity between amino acids to obtain a preliminary expression of the protein.

[0021] S12: The amino acid sequence of the protein with a fixed length of N amino acids is divided into a training set and a test set according to a set ratio. The three division modes are random mode, depth-first search mode and breadth-first search mode.

[0022] Furthermore, in step S11, for amino acid sequences with a sequence length greater than N, the first N amino acids are trimmed, and for amino acid sequences with a sequence length less than N, 0 is added to the end of the sequence to obtain an amino acid sequence with a length of N.

[0023] Furthermore, in step S2, the specific process is as follows:

[0024] S21: Input the amino acid sequence of the protein into the ESM-2 module to obtain the protein structural characteristics;

[0025] S22: The initial expression of the protein is passed through a one-dimensional convolutional layer, a pooling layer, a recurrent gating unit, and a fully connected layer to obtain the amino acid sequence features of the protein.

[0026] Furthermore, in step S3, feature fusion is to splice the vector of the protein's amino acid sequence features and the vector of the protein's structural features to obtain a new protein feature expression.

[0027] Furthermore, in step S4, a graph neural network is constructed, and a protein-protein interaction graph neural network is constructed with protein expression as nodes and protein-protein interactions as edges.

[0028] Furthermore, in the protein-protein interaction graph neural network, the update of protein feature expression depends on its neighboring nodes. At the same time, a PPI graph with protein feature expression as nodes and protein-protein interactions as edges is constructed, which transforms the original independent learning task into a graph-related learning task. The protein-protein interaction graph neural network is a graph isomorphic network, which uses a multilayer perceptron to combine neighboring nodes and its own nodes to update node features. The node update formula is as follows:

[0029]

[0030] in, is the feature of node v at the kth iteration, ∈ is a learnable parameter or a fixed scalar.

[0031] Furthermore, in the protein-protein interaction graph neural network, the characteristic expression of each protein is called a node, and the protein characteristic expression directly connected to the protein characteristic expression to be updated is called the neighborhood node of the protein.

[0032] Furthermore, in step S5, during the model training process, the loss function is an asymmetric loss function, and the formula is as follows:

[0033]

[0034] Among them, p is the network output probability, γ is the focusing parameter, and p m is the transition probability, m is the probability margin which is an adjustable hyperparameter, L + and L - It is the positive and negative loss part.

[0035] Furthermore, in step S6, protein-protein interactions include reaction, binding, post-translational modification, activation, inhibition, catalysis, and expression. When testing whether two proteins interact, each type of reaction between the proteins is predicted. If the predicted value of this type of reaction is greater than 0.5, it is considered that the protein pair has this type of reaction. There is a possibility of multiple reactions between the same pair of proteins.

[0036] Compared with the existing technology, the present invention has the following advantages: the protein interaction prediction method based on graph neural network adopts the fusion of two features as the protein feature expression; the ESM-2 module is used to extract protein structural features using the amino acid sequence of the protein, without the need to provide separate structural features; the asymmetric loss function is used to effectively solve the sample label imbalance problem existing in the interaction between proteins. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a schematic flow chart of a protein interaction prediction method based on a graph neural network in an embodiment of the present invention;

[0038] Figure 2 Schematic diagrams of three partitioning methods in an embodiment of the present invention, where (a) is a random mode (Random), (b) is a random mode (BFS), and (c) is a breadth-first search mode (DFS);

[0039] Figure 3 Schematic diagram of the overall structure of the protein-protein interaction prediction model in an embodiment of the present invention;

[0040] Figure 4 Schematic diagram of the structure of the cyclic gating unit in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. However, the protection scope of the present invention is not limited to the following embodiment.

[0042] like Figure 1 As shown, this embodiment provides a technical solution: a protein interaction prediction method based on graph neural network, comprising the following steps:

[0043] S1: Amino acid sequence preprocessing.

[0044] In step S1, the following contents are specifically included:

[0045] S11: Obtain public datasets of protein amino acid sequences and interactions, and preprocess the amino acid sequences as needed. The length of each protein's amino acid sequence is fixed at 2000 amino acids. Sequences longer than 2000 are pruned to the first 2000 amino acids, while sequences shorter than 2000 are padded with zeros at the end to obtain a 2000 amino acid sequence. These amino acid sequences are trained using the Skip-Gram model, and the resulting embedding vectors are concatenated with a one-hot encoding defined based on the electrical stability and hydrophobicity similarity between amino acids to obtain a preliminary representation of the protein.

[0046] S12: The data was divided into training and test sets in a ratio of 8:2. To simulate the distribution of proteins in the real world, three partitioning modes were used: random mode, depth-first search mode, and breadth-first search mode. The random mode randomly selects proteins as training and test sets according to the ratio. The depth-first search mode searches for proteins as deep as possible along the starting protein node, thereby dividing the training and test sets according to the ratio. The breadth-first search mode starts from a given starting protein node and diffuses outward layer by layer, thereby dividing the training and test sets according to the ratio. The data distribution of the test set after preprocessing is shown in Table 1. BS, ES, and NS indicate that both proteins were seen, only one protein was seen, and neither protein was seen during the training process, respectively. Protein-protein interactions include reaction, binding, post-translational modification, activation, inhibition, catalysis, and expression.

[0047] Table 1 Distribution of test set after preprocessing

[0048]

[0049] S2: Extract protein structure and sequence information;

[0050] In step S2, the following sub-steps are included:

[0051] S21: Select appropriate ESM-2 module pre-training weights and input the protein amino acid sequence into the ESM-2 module to obtain structural information;

[0052] In step S21, on an evolutionary scale, the amino acid sequence of a protein reflects its biological structure and function, so the structure and function of a protein can be inferred from the sequence. The ESM-2 module is trained based on the Transformer model and has 15 billion parameters. This large number of parameters improves the model's fidelity in modeling protein sequences, enhances understanding of protein sequences, and thus better predicts structural information.

[0053] S22: The initial expression of the protein is sequentially passed through a one-dimensional convolutional layer, a pooling layer, a recurrent gating unit, and a fully connected layer to obtain the amino acid sequence information of the protein;

[0054] In step S22, the recurrent gating unit effectively retains the information in the sequence using the reset gate and the update gate. In this way, the network discards irrelevant information later, allowing the information to be expressed more compactly.

[0055] S3: Fusion of structural information and feature information

[0056] In step S3, the extracted structural information and sequence information are fused to obtain fused features.

[0057] S4: Building a Graph Neural Network

[0058] In step S4, a protein-protein interaction graph neural network is constructed, with protein feature expressions as nodes and protein-protein interactions as edges. In a graph neural network, protein expression updates depend on their neighboring nodes. Simultaneously, a PPI graph is constructed, with protein feature expressions as nodes and interactions as edges. This transforms the original independent learning task into a graph-dependent learning task. A graph isomorphism network is used here, employing a multilayer perceptron to combine neighboring nodes with the node itself to update node features.

[0059] S5: Model network training

[0060] Use the training set to train the graph neural network to obtain a protein-protein interaction prediction model;

[0061] Specifically, in step S5, the specific training scheme is as follows:

[0062] S51: Set the initial learning rate to 0.01 and train for 320 rounds.

[0063] S52: During the training process, the optimal F1 value of the model is recorded, and the AUPR value of each type of reaction in each round of training is also recorded. Table 2 shows the performance of different graph neural networks in three partitioning modes for the STRING dataset and its subsets SHS27K and SHS148K. As can be seen from Table 2, the predicted F1 value decreases when the graph neural network is replaced with a graph convolutional neural network (GCN) and a graph attention network (GAT), reflecting the effectiveness of the graph isomorphism network.

[0064] Table 2 Performance of different graph neural networks

[0065]

[0066]

[0067] S53: Optimizing the loss function for unbalanced samples. Although there are many types of protein-protein interactions, the types are relatively unbalanced, and the reactions between proteins are also unbalanced. Some proteins have only one reaction, while others have multiple reactions. This imbalance in sample labels will affect the training process and ultimately the prediction results. To solve this problem, an asymmetric loss function is used here. The specific form has been explained above. Table 3 shows the performance of different loss functions under three partitioning modes for the STRING dataset and its subsets SHS27K and SHS148K. As can be seen from Table 3, the two complementary asymmetric mechanisms contained in the asymmetric loss function can dynamically adjust the asymmetry level during training due to their different effects on positive and negative samples, effectively balancing positive and negative samples.

[0068] Table 3 Performance of different loss functions

[0069]

[0070] S6: Protein-protein interaction prediction

[0071] The unknown protein was input into the protein-protein interaction prediction model for prediction, and the prediction results were obtained. The results are shown in Table 4. As can be seen from Table 4, the performance of the present invention is superior to that of other methods. The performance of each partitioning method on the three datasets is better than that of other methods, reflecting the effectiveness of the ESM-2 structural model extracted and the asymmetric loss function that can effectively balance the samples.

[0072] Table 4 Performance of different methods

[0073]

[0074]

[0075] In summary, the protein interaction prediction method based on graph neural network in the above embodiment first performs necessary preprocessing on the STRING dataset on the public protein STRING database, including amino acid sequence clipping and supplementation, uses ESM-2 to extract structural information, and uses a recurrent gating unit to extract sequence information, and then fuses the extracted structural information with the sequence information. A protein-protein interaction graph neural network is constructed with protein expression as nodes and protein-protein interactions as edges. The network model is then trained. Here, an asymmetric loss function is used to balance the sample imbalance problem, thereby improving the model performance. Finally, the unknown protein is input into the protein-protein interaction prediction model for prediction to obtain the prediction result. By using the ESM-2 module to extract structural information and the asymmetric loss function to balance samples, the performance of different models was verified and compared on the test set. The experimental results show that these techniques can accelerate the convergence and robustness of the network; the model proposed by the present invention has high prediction accuracy. In the future, it is expected to be deployed on edge devices for production and application.

[0076] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A protein interaction prediction method based on graph neural network, characterized in that: The following steps are involved: S1: Amino acid sequence preprocessing Obtain protein amino acid sequences and public interaction datasets, and preprocess the amino acid sequences according to task requirements; S2: Extract protein structure and sequence information The ESM-2 model is used to extract structural information, and the recurrent gating unit is used to extract sequence information. S3: Fusion of structural information and feature information Fuse the structural information and sequence information to obtain fusion features; S4: Building a Graph Neural Network Build a graph neural network based on protein interaction and fusion features to obtain a protein-protein interaction graph neural network; S5: Model network training The training set is used to train the protein-protein interaction graph neural network to obtain a protein-protein interaction prediction model; S6: Protein-protein interaction prediction Inputting unknown proteins into the protein-protein interaction prediction model for prediction and obtaining prediction results; In the protein-protein interaction graph neural network, the update of protein feature expression depends on its neighboring nodes. At the same time, a PPI graph with protein feature expression as nodes and protein interactions as edges is constructed, which transforms the original independent learning task into a graph-related learning task. The protein-protein interaction graph neural network is a graph isomorphic network. It uses a multilayer perceptron to combine neighboring nodes and its own nodes to update node features. The node update formula is as follows: in, is the feature of node v at the kth iteration, ∈ is a learnable parameter or a fixed scalar; In the protein-protein interaction graph neural network, the characteristic expression of each protein is called a node, and the protein characteristic expression directly connected to the protein characteristic expression to be updated is called the neighborhood node of the protein.

2. A protein interaction prediction method based on graph neural network according to claim 1, characterized in that: In step S1, the specific processing process is as follows: S11: Obtain the amino acid sequence of proteins and a public interaction dataset. Fix the length of each protein's amino acid sequence to N amino acids. Then use the Skip-Gram model to train the amino acid sequence. Concatenate the resulting embedding vector and the one-hot encoding defined based on the electrical stability and hydrophobicity similarity between amino acids to obtain a preliminary expression of the protein. S12: The amino acid sequence of the protein with a fixed length of N amino acids is divided into a training set and a test set according to a set ratio. The three division modes are random mode, depth-first search mode and breadth-first search mode.

3. The protein interaction prediction method based on graph neural network according to claim 2, characterized in that: In step S11, for amino acid sequences with a length greater than N, the first N amino acids are trimmed, while for amino acid sequences with a length less than N, 0 is added to the end of the sequence to obtain an amino acid sequence with a length of N.

4. The protein interaction prediction method based on graph neural network according to claim 2, characterized in that: In step S2, the specific process is as follows: S21: Input the amino acid sequence of the protein into the ESM-2 module to obtain the protein structural characteristics; S22: The initial expression of the protein is passed through a one-dimensional convolutional layer, a pooling layer, a recurrent gating unit, and a fully connected layer to obtain the amino acid sequence features of the protein.

5. The protein interaction prediction method based on graph neural network according to claim 4, characterized in that: In step S3, feature fusion is to splice the vector of the protein's amino acid sequence features and the vector of the protein's structural features to obtain a new protein feature expression.

6. The protein interaction prediction method based on graph neural network according to claim 1, characterized in that: In step S4, a graph neural network is constructed, and a protein-protein interaction graph neural network is constructed with protein expressions as nodes and protein-protein interactions as edges.

7. The protein interaction prediction method based on graph neural network according to claim 1, characterized in that: In step S5, during the model training process, the loss function is an asymmetric loss function, and the formula is as follows: Among them, p is the network output probability, γ is the focusing parameter, and p m is the transition probability, m is the probability margin which is an adjustable hyperparameter, L + and L - It is the positive and negative loss part.

8. The protein interaction prediction method based on graph neural network according to claim 1, characterized in that: In step S6, protein-protein interactions include reaction, binding, post-translational modification, activation, inhibition, catalysis, and expression. When testing whether two proteins interact with each other, each type of reaction between the proteins is predicted. If the predicted value of this type of reaction is greater than 0.5, it is considered that the protein pair has this type of reaction.

Images