Target protein drug binding prediction method based on meta-learning and subgraph matching

Abstract

A target protein drug binding prediction method based on meta learning and subgraph matching, by constructing protein structure, drug small molecule structure and its binding energy value, the meta learning training task is established according to protein grouping, that is, as the main model of the meta model; the sub model and its loss function for the protein prediction task are obtained after fine tuning of the meta model; then the task self-adaptive self-attention model is established to balance the optimization contribution of the sub model of each protein prediction task to the meta model, the weighted average is used for the loss function of the sub model of each protein prediction task to obtain the meta model loss function; then the meta model loss function is used to calculate the gradient, the meta model parameters are updated and optimized based on the preset learning rate, and after the training is completed, the trained meta model is used for fine tuning of the test set of new protein samples, the prediction sub model of the new protein is obtained, the sub model is used for prediction of the protein test set, and the evaluation result is obtained. The training method combining meta learning and subgraph matching can effectively avoid shortcut learning, enhance the generalization effect, and also solve the problem that the previous prediction model is difficult to predict the newly discovered protein.

Description

Technical Field

[0001] This invention relates to a technology in the field of protein and drug small molecules, specifically a target protein drug binding prediction method based on meta-learning and subgraph matching. Background Technology

[0002] Identifying the interactions between drugs and their targets (proteins) plays a crucial role in drug discovery. However, traditional experimental methods for identifying drug-target interactions by resolving the crystal structure of drug-protein complexes are both expensive and time-consuming. To reduce costs, computational methods for identifying drug-target interactions are gaining increasing attention. Instead of incorporating a large number of candidate drugs into experimental validation, computational methods can screen out most candidates before experimental validation, which is more efficient and less costly. Generally, computational methods fall into two main categories: docking simulation methods and data-driven methods. Docking simulations utilize the 3D structures of drug molecules and target proteins to identify their potential binding sites, but this remains extremely time-consuming. In contrast, due to the rapid development of machine learning, identifying interactions using the characteristics of proteins and drugs can achieve both high accuracy and low cost.

[0003] Recently, deep learning has achieved exciting results in predictions, primarily learning from known drug-target interactions. However, these methods do not generalize well to newly discovered proteins and drugs. Similarity / distance-based and network-based methods mainly utilize protein-protein similarity, drug-drug similarity, and known drug-target interactions, but their performance is severely overestimated during testing. Furthermore, for newly discovered target proteins, there are often few known drugs that bind to them, making it particularly difficult to predict binding effects. Currently, most existing methods focus on how to effectively learn molecular and protein representations and then feed these representations into classification or regression models. However, these methods suffer from shortcut learning, memorizing the ratio of positive to negative samples binding annotations in the training set instead of learning the molecular features of the interactions, thus failing in tests with unseen new proteins and drugs. Moreover, proteins may have different binding patterns with different drugs, and these patterns can be learned by training protein-specific models for individual proteins. Summary of the Invention

[0004] To address the aforementioned shortcomings of existing technologies, this invention proposes a target protein drug binding prediction method based on meta-learning and subgraph matching. By combining meta-learning with subgraph matching in the training method, shortcut learning can be effectively avoided, the generalization effect can be enhanced, and the problem of previous prediction models being unable to predict newly discovered proteins can also be solved.

[0005] This invention is achieved through the following technical solution:

[0006] This invention relates to a method for predicting target protein drug binding based on meta-learning and subgraph matching, comprising:

[0007] S1. Construct protein structures, drug small molecule structures and their binding energy values, grouped by protein to establish a meta-learning training task, which serves as the main model of the meta-model; after fine-tuning the meta-model, obtain the sub-model and its loss function for the protein prediction task.

[0008] The sub-models for the protein prediction task include graph neural networks and multilayer perceptrons.

[0009] S2. Establish a task-adaptive self-attention model to balance the optimization contribution of each protein prediction task sub-model to the meta-model. Use a weighted average of the loss functions of each protein prediction task sub-model obtained in S1 to obtain the meta-model loss function.

[0010] S3. Calculate the gradient using the meta-model loss function. After obtaining the meta-model gradient, update and optimize the meta-model parameters based on the preset learning rate.

[0011] S4. Repeat steps S1-S3 multiple times to train the meta-model. After training, use the trained meta-model to fine-tune the new protein samples in the test set to obtain a new protein prediction sub-model. Use this sub-model to predict the protein in the test set and obtain the evaluation results.

[0012] Step S1 specifically includes:

[0013] S1.1 Group the protein structure, drug small molecule structure and their binding energy values ​​according to the protein, and construct the initial characterization of the protein map structure and drug small molecule map structure according to the protein structure data and drug small molecule structure data respectively.

[0014] S1.2 For each group, the initial representations of the protein graph structure and the drug small molecule graph structure are learned based on the graph neural network to obtain the protein graph node representation and the drug small molecule graph node representation after graph neural network processing.

[0015] S1.3. Based on the characterization of each node in the protein graph, under the condition of maximizing the mutual information between the protein key subgraph and the target protein-drug binding interaction, a multilayer perceptron network is used to select protein nodes to form a protein key subgraph, thereby obtaining the protein key subgraph characterization.

[0016] S1.4. Combine protein key subgraph characterization with drug small molecule characterization to construct a multilayer perceptron network for prediction, thus obtaining a sub-model for the protein prediction task.

[0017] Step S1.1 specifically includes:

[0018] S1.1.1. Group the protein structure, drug small molecule structure, and their binding data according to protein. Each task is responsible for predicting the binding effect of the protein and the drug small molecule. For each task's data, randomly sample k positive samples and k negative samples to form the training set for that task, and then randomly sample m positive samples and m negative samples to form the test set for that task.

[0019] S1.1.2 Read the protein PDB file in the training set of this task to obtain the chemical information and three-dimensional coordinates of each amino acid residue of the protein.

[0020] S1.1.3. Set each amino acid residue as a node in a graph network, and calculate the Euclidean distance between them using the three-dimensional coordinates of each amino acid. At this point, there is an edge between the nodes corresponding to the two amino acids, where (x1, y1, z1) and (x2, y2, z2) are the three-dimensional coordinates of the two amino acid residues, respectively. A pre-trained model is then used to encode the protein residue information, ultimately yielding the protein graph structure and the initial representation of the graph nodes.

[0021] S1.1.4. The RDKit software package is used to convert the drug small molecule SMILES string into a graph network structure, where small molecule atoms are graph network nodes and intermolecular chemical bonds are graph network edges. The Open Graph Benchmark encoding method is used to encode the node features and edge features of the drug small molecule to obtain the initial representation of the small molecule graph nodes.

[0022] Step S1.2, which involves characterizing the protein graph structure based on a graph neural network, specifically includes:

[0023] S1.2.1. The initial representation of the protein graph nodes obtained in step S1.1.3 is processed using a graph convolutional neural network. The node representation is updated by aggregating features from neighboring nodes: Where: ReLU is the activation function, A is the adjacency matrix, D is the degree matrix of the adjacency matrix, H( l ) represents the features of the nodes in layer l, and W represents the parameters of the neural network.

[0024] S1.2.2 Combining protein graph node representations into protein graph structure representations:

[0025] Step S1.2, which involves characterizing the graph structure of drug small molecules based on a graph neural network, specifically includes:

[0026] S1.2.3. The initial representation of the drug small molecule graph nodes obtained in step S1.1.4 is processed using a graph convolutional neural network. Node representations are updated by aggregating features from neighboring nodes. Where: ReLU is the activation function, A is the adjacency matrix, D is the degree matrix of the adjacency matrix, and H... (l) Let W represent the features of the nodes in layer l, and W represent the parameters of the neural network.

[0027] S1.2.4. Combine the node representations of the drug small molecule graph into a drug small molecule graph structure representation:

[0028] The protein keystone characterization in step S1.3 is obtained through the following method:

[0029] S1.3.1. A protein subgraph node selection model is established using a multilayer perceptron. The protein graph representation from S1.2.2 and the drug small molecule graph representation from S1.2.4 are concatenated, and then input into the protein subgraph node selection model to obtain the node selection matrix Z = Softmax(MLP(concatenate(G)). P G m ))), where: MLP is the protein subgraph node selection model. G P For protein mapping, G m This is used for small molecule characterization. Softmax is the activation function. A loss function that promotes node selection is set. Where: I2 is the identity second-order matrix, and A is the protein graph adjacency matrix. Minimize This allows the node selection matrix Z to be close to [0,1] / [1,0].

[0030] S1.3.2 Obtain the protein key subgraph representation G based on the node selection matrix Z. sub =Z T G p [0].

[0031] The condition in step S1.3 that maximizes the mutual information between the protein keymap and the target protein-drug binding interaction refers to: Where: Y is the target protein-drug binding tag, β is the Lagrange multiplier for optimizing the target, and I(Y, G) sub The lower bound can be viewed as the cross-entropy loss between Y and the key subgraph; I(G, G) sub The upper bound of ) is represented by the Donsker-Varadhan divergence of the KL divergence, specifically as follows: Among them: G i For protein mapping, For protein sub-graph characterization, This is a multilayer perceptron network used to extract mutual information between protein maps and subgraphs; by maximizing... The loss can be maximized by I(G, G). subThe upper bound of ) is found, and minimizing the cross-entropy loss between Y and the key subgraph can minimize I(Y, G). sub The lower bound is determined by updating the network parameters of both parts together to optimize the determination of the protein key subgraph.

[0032] The aforementioned construction of a multilayer perceptron network for prediction refers to combining protein key graph characterization with drug small molecule characterization, and then using a multilayer perceptron for prediction to obtain model prediction values. Among them: G m For small molecule characterization of drugs, G sub For protein keystone characterization, θ cls Here are the parameters of the multilayer perceptron network, and the classification loss function of this multilayer perceptron. y represents the true value; this leads to the generation of the loss function for the protein prediction task.

[0033] Step S2 specifically includes:

[0034] S2.1 For each sub-model of the protein prediction task, the protein key sub-graph representation is combined with the small molecule drug representations of all drugs for that protein task to form the task representation, specifically: Among them: G sub For protein keystone map characterization, G m For the characterization of small molecules of the drug, the following were obtained This is a characterization of the task.

[0035] S2.2 Establish a task-adaptive self-attention model and output the optimization contribution weights of each task to the meta-model, specifically: Where: Q, K, V are the task representations, d k η is the number of dimensions representing this task. Tb This represents the contribution weight of the task to the meta-model. Add d k Normalizable QK T Variance, stable training gradient.

[0036] S2.3, Use The meta-model loss function is obtained by weighted averaging the sub-model losses of each protein prediction task. The specific loss function formula is as follows: in: The loss function for each task's test set.

[0037] This invention relates to a system for implementing the above-mentioned method, comprising: a data preprocessing unit, a model building unit, a model training unit, and a model testing unit, wherein: the data preprocessing unit reads the three-dimensional coordinates of the protein to build a graph structure based on the protein structure information, and encodes the protein graph structure using a pre-trained model to obtain the protein graph structure and node representation; based on the drug small molecule structure information, it uses rdkit to convert the drug small molecule structure into a graph structure, and uses the open graph benchmark encoding method to obtain the drug small molecule graph structure and node representation; based on the protein and drug small molecule binding information, it performs threshold screening to obtain binding and non-binding binary labels; the model building unit builds each graph neural network and multilayer perceptron based on the number of network layers and dimensionality information; the model training unit trains the model based on the protein and drug small molecule graph structure and node representation information obtained by the data preprocessing unit to obtain a trained meta-model; the model testing unit fine-tunes the meta-model based on the protein and drug small molecule information in the test set to obtain the protein prediction model, and then makes predictions on the test set to obtain the evaluation results.

[0038] Technical effect

[0039] This invention fully utilizes the characteristic information of proteins and drug small molecules by constructing protein graph structures and drug small molecule graph structures, more accurately capturing potential important binding features compared to existing technologies. Based on the characterization of each node in the protein graph, by maximizing the mutual information between key protein subgraphs and target protein-drug binding interactions, key protein subgraphs are selected to find the protein subgraphs most relevant to the binding interaction. Compared to existing technologies, redundant node information is removed. A task-adaptive self-attention model is established to balance the optimization contributions of sub-models of each protein prediction task to the meta-model, resulting in initialization parameters that can be quickly fine-tuned from the meta-model to a specific protein model. This invention more accurately captures potential important binding features compared to existing technologies; by finding the protein subgraphs most relevant to the binding interaction, and by removing redundant node information compared to existing technologies, a task-adaptive self-attention model is used to obtain initialization parameters that can be quickly fine-tuned from the meta-model to a specific protein model. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the process of the present invention;

[0041] Figure 2 This is a flowchart of the training process of the present invention. Detailed Implementation

[0042] like Figure 1 As shown, this embodiment illustrates a target protein drug binding prediction method based on meta-learning and subgraph matching, specifically including:

[0043] Step 1: Group the protein structure, drug small molecule structure, and their binding data according to protein. Each task is responsible for predicting the binding effect between the protein and the drug small molecule. For each task's data, randomly sample k positive samples and k negative samples to form the training set for that task, and then randomly sample m positive samples and m negative samples to form the test set for that task.

[0044] Step 2: Construct initial characterizations of protein graph structures and drug small molecule graph structures based on protein structure data and drug small molecule structure data, respectively.

[0045] Step 3: For each sub-model, characterize and learn the protein graph structure and drug small molecule graph structure based on graph neural networks.

[0046] Step 4: Use a multilayer perceptron to build a protein subgraph node selection model, concatenate the protein graph representation with the small molecule graph representation, and then input the protein subgraph node selection model to obtain the node selection matrix.

[0047] Step 5: Use the node selection matrix to obtain the protein key graph and maximize the mutual information between the protein key graph and the target protein-drug binding interaction.

[0048] Step 6: Combine protein subgraph characterization with drug small molecule characterization to construct a multilayer perceptron network for prediction, and obtain the loss function for the protein prediction task.

[0049] Step 7: Establish a task-adaptive self-attention model. Combine the protein subgraph representation with the drug small molecule representation of all test sets of the protein task as the input model for the task representation, and output the optimization contribution weight of each task to the meta-model.

[0050] Step 8: Use the meta-model loss function to train the meta-model, including gradient descent and network parameter updates.

[0051] Step 9: Repeat steps 2-8 multiple times to train the meta-model. After training, use the trained meta-model to fine-tune the new protein samples in the test set to obtain a new protein prediction sub-model. Use the sub-model to predict the binding effect of the protein with each small drug molecule.

[0052] Through specific practical experiments, the above method was run in a Linux environment with GPU training settings and random initialization parameters. The number of proteins, drug molecules, and protein-drug molecule binding samples in the training and test sets are shown in Table 1. The accuracy of the experimental data reached 90%. The target protein drug prediction effect is shown in Table 2, which includes the predicted binding energy, drug molecule InChI key, protein ID, and predicted binding energy of 10 drug molecules binding to proteins.

[0053] Table 1

[0054]

[0055] Table 2

[0056]

[0057] Compared with existing technologies, this invention captures semantically richer node representations by performing representation learning on protein graph structures and drug small molecule graph structures separately based on graph neural networks; it removes redundant information and improves prediction performance by using a node selection matrix to obtain the protein key subgraph; and it improves the performance of the meta-model by establishing a task-adaptive self-attention model to balance the optimization contributions of each task to the meta-model.

[0058] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for predicting target protein drug binding based on meta-learning and subgraph matching, characterized in that, include: S1. Group protein structures, drug small molecule structures and their binding energy values ​​according to proteins to establish meta-learning training tasks, which serve as the main model of the meta-model. After fine-tuning the meta-model, a sub-model and its loss function for the protein prediction task are obtained. The sub-models for the protein prediction task include: graph neural networks and multilayer perceptrons; S2. Establish a task-adaptive self-attention model to balance the optimization contributions of each protein prediction task's sub-model to the meta-model. The meta-model loss function is obtained by weighted averaging the loss functions of each protein prediction task's sub-model obtained in S1, specifically including: S2.1 For each sub-model of the protein prediction task, the protein key subgraph representation is combined with the small molecule drug representations of all drugs for that protein prediction task as the task representation, specifically: ,in: For protein keystone characterization. For the characterization of small molecules of the drug, the following were obtained This is a characterization of the task; S2.2 Establish a task-adaptive self-attention model and output the optimization contribution weights of each task to the meta-model, specifically: , ,in: This is a characterization of the task. The number of dimensions representing this task is obtained. Assign weights to the task's contribution to the meta-model; add... Normalizable Variance, stable training gradient; S2.3, Use The meta-model loss function is obtained by weighted averaging the sub-model losses of each protein prediction task; the specific loss function formula is as follows: ,in: The loss function for the test set of each task; S3. Calculate the gradient using the meta-model loss function. After obtaining the meta-model gradient, update and optimize the meta-model parameters based on the preset learning rate. S4. Repeat steps S1-S3 multiple times to train the meta-model. After training, use the trained meta-model to fine-tune the new protein samples in the test set to obtain a new protein prediction sub-model. Use this sub-model to predict the protein test set and obtain the evaluation results.

2. The target protein drug binding prediction method based on meta-learning and subgraph matching according to claim 1, characterized in that, Step S1 specifically includes: S1.1 Group the protein structure, drug small molecule structure and their binding energy values ​​according to the protein, and construct the initial characterization of the protein map structure and drug small molecule map structure according to the protein structure data and drug small molecule structure data respectively. S1.2 For each group, the initial representations of the protein graph structure and the drug small molecule graph structure are learned based on the graph neural network to obtain the protein graph node representation and the drug small molecule graph node representation after graph neural network processing. S1.3 Based on the characterization of each node in the protein graph, under the condition of maximizing the mutual information between the protein key subgraph and the target protein-drug binding interaction, a multilayer perceptron network is used to select protein nodes to form a protein key subgraph, thereby obtaining the protein key subgraph characterization. S1.

4. Combine protein key subgraph characterization with drug small molecule characterization to construct a multilayer perceptron network for prediction, thus obtaining a sub-model for the protein prediction task.

3. The target protein drug binding prediction method based on meta-learning and subgraph matching according to claim 2, characterized in that, Step S1.1 specifically includes: S1.1.1 Group the protein structure, drug small molecule structure and their binding data according to protein, and each task is responsible for predicting the binding effect of the protein and the drug small molecule; for the data of each task, randomly sample k positive samples and k negative samples to form the training set of the task, and then randomly sample m positive samples and m negative samples to form the test set of the task. S1.1.2 Read the protein PDB file in the training set of this task to obtain the chemical information and three-dimensional coordinates of each amino acid residue of the protein; S1.1.

3. Set each amino acid residue as a node in a graph network, and calculate the Euclidean distance from the three-dimensional coordinates of each amino acid. At that time, there is an edge between the nodes corresponding to the two amino acids, where: ( , , ), ( , , The three-dimensional coordinates of the two amino acid residues are shown below; then, a pre-trained model is used to encode the protein residue information, and finally the protein graph structure and the initial characterization of the graph nodes are obtained. S1.1.

4. Use the RDKit software package to convert the drug small molecule SMILES string into a graph network structure, where: small molecule atoms are graph network nodes, and intermolecular chemical bonds are graph network edges; use the Open Graph Benchmark encoding method to encode the drug small molecule node features and edge features to obtain the initial representation of the small molecule graph nodes.

4. The target protein drug binding prediction method based on meta-learning and subgraph matching according to claim 3, characterized in that, Step S1.2, which involves characterizing the protein graph structure based on a graph neural network, specifically includes: S1.2.

1. Process the initial representation of the protein graph nodes obtained in step S1.1.3 using a graph convolutional neural network; update the representation of nodes by aggregating the features of neighboring nodes: ,in: For activation function, It is an adjacency matrix. Let be the degree matrix of the adjacency matrix. for Layer node features For neural network parameters; S1.2.2 Combining protein graph node representations into protein graph structure representations: .

5. The target protein drug binding prediction method based on meta-learning and subgraph matching according to claim 4, characterized in that, Step S1.2, which involves characterizing the graph structure of drug small molecules based on a graph neural network, specifically includes: S1.2.

3. Process the initial representation of the drug small molecule graph nodes obtained in step S1.1.4 using a graph convolutional neural network; update the representation of nodes by aggregating the features of neighboring nodes: ,in: For activation function, It is an adjacency matrix. Let be the degree matrix of the adjacency matrix. for Layer node features For neural network parameters; S1.2.

4. Combine the node representations of the drug small molecule graph into a drug small molecule graph structure representation: .

6. The target protein drug binding prediction method based on meta-learning and subgraph matching according to claim 4, characterized in that, The protein keystone characterization in step S1.3 is obtained through the following method: S1.3.

1. A protein subgraph node selection model is established using a multilayer perceptron. The protein graph representation from S1.2.2 and the drug small molecule graph representation from S1.2.4 are concatenated and then input into the protein subgraph node selection model to obtain the node selection matrix. ,in: Select a model for protein subgraph nodes; For protein mapping, Characterization of small molecules; Set the activation function; set the loss function to promote node selection. ,in: It is a unit second-order matrix. Given the protein graph adjacency matrix; minimize Allows for the selection of nodes using a matrix. Close to [0,1] or [1,0]; S1.3.2 Selecting a matrix based on nodes Protein keystone characterization .

7. The target protein drug binding prediction method based on meta-learning and subgraph matching according to claim 6, characterized in that, The condition in step S1.3 that maximizes the mutual information between the protein keymap and the target protein-drug binding interaction refers to: ,in: For target protein-drug binding tag, To optimize the Lagrange multipliers of the objective, where: The lower bound can be regarded as Cross-entropy loss with key subgraphs; The upper bound is represented by the Donsker-Varadhan divergence of the KL divergence, specifically: ,in: For protein mapping, For protein sub-graph characterization, This is a multilayer perceptron network used to extract mutual information between protein maps and subgraphs; by maximizing... Loss can be maximized upper bound, minimize The cross-entropy loss with the key subgraph can be minimized. The lower bound is used to jointly update the network parameters of both parts for optimization, thereby determining the protein key subgraph.

8. The target protein drug binding prediction method based on meta-learning and subgraph matching according to claim 7, characterized in that, The construction of a multilayer perceptron network for prediction mentioned in step S1 refers to combining protein key graph representation with drug small molecule representation, and then using a multilayer perceptron for prediction to obtain the model's predicted value. ,in: For drug small molecule characterization, For protein keystone characterization. Here are the parameters of the multilayer perceptron network, and the classification loss function of this multilayer perceptron. , The true value is used to generate the loss function for the protein prediction task. .

9. A system for implementing the target protein drug binding prediction method based on meta-learning and subgraph matching as described in any one of claims 1-8, characterized in that, include: The system comprises a data preprocessing unit, a model building unit, a model training unit, and a model testing unit. Specifically: The data preprocessing unit reads the three-dimensional coordinates of proteins to build a graph structure based on protein structure information, and encodes the protein graph structure using a pre-trained model to obtain the protein graph structure and node representations. Based on drug small molecule structure information, it uses rdkit to convert the drug small molecule structure into a graph structure, and uses the open graph benchmark encoding method to obtain the drug small molecule graph structure and node representations. Based on the protein-drug small molecule binding information, it performs threshold screening to obtain binding and non-binding binary labels. The model building unit builds various graph neural networks and multilayer perceptrons based on the number of network layers and dimensionality information. The model training unit trains the model based on the protein and drug small molecule graph structure and node representation information obtained from the data preprocessing unit to obtain a trained meta-model. The model testing unit fine-tunes the meta-model based on the protein and drug small molecule information in the test set to obtain a protein prediction model, and then performs predictions on the test set to obtain evaluation results.

Images