CN116343910A - Prediction method of docking posture between protein and ligand based on graphic neural network - Google Patents

Abstract

The invention discloses a method for predicting the docking posture between a protein and a ligand based on a graph neural network. First, a biological information sample set of a protein-ligand complex is obtained, and the sample set includes sample data and sample label data; secondly, Build a docking pose generation model based on a graph neural network and a docking pose evaluation model based on multiple perspectives, further adjust the parameters of the model, process the sample data through the structure generation model obtained through training, and obtain the actual output of the pose docking of the protein ligand; finally , using the mainstream pose docking structure evaluation index to evaluate the stability of the output results. The present invention directly uses the biological structure information of the ligand protein to generate the optimal docking posture structure, and evaluates the generated result through a multi-angle comprehensive evaluation model, thereby improving the accuracy of the docking prediction of the ligand-protein posture structure, and improving Evaluating the validity of ligand-protein pose-structure docking predictions.

Description

Prediction method of docking pose between protein and ligand based on graph neural network

technical field

The invention belongs to the field of computer-aided drug design, in particular to a method for predicting the docking posture between a protein and a ligand based on a graph neural network.

Background technique

In the process of computer-aided drug design, it has always been a difficult problem to screen out the docking pose with high binding affinity between a specific protein and its ligand. The traditional method considers combining existing methods to generate an infinite number of combinations of docking poses, and on this basis to screen a set of poses that best meet the conditions, ignoring the global information between the drug molecule and the target protein. In addition, the existing The docking pose information is relatively small, and the evaluation of binding affinity also requires a large number of experiments. Good screening results require a large number of annotations on the data set. Insufficient labeling often leads to inaccurate predictions of the binding affinity between ligand molecules and target proteins, and the capital costs and labor costs required for experiments in such methods are unaffordable for most scientific research projects. The emerging method of using three-dimensional molecular structure to predict the docking posture is very promising in predicting the binding affinity between proteins and ligands, which improves the efficiency and accuracy of computer-aided drug design.

In the existing methods for predicting docking posture and affinity using molecular three-dimensional structure, in order to improve the efficiency of prediction, various sampling methods are used. For example, Glide adopts the Monte Carlo sampling method for the global information of posture, thereby improving the Accuracy and speed of docking pose prediction in the face of large numbers of drug molecule ligands. However, various methods similar to Glide often ignore the biological optimization of the docking posture, and without optimized posture information, it will cause inaccurate prediction of the docking posture and affinity. There is an urgent need for more intelligent and Accurate predictive models. In addition, in this type of method, only one of the intramolecular force and the intermolecular force is often considered, and the two are not discussed comprehensively. In the process of predicting the affinity, some important information is lost, resulting in the prediction inaccurate.

Generally speaking, there are still many limitations in the current prediction of the docking pose between proteins and ligands, such as: the time-consuming prediction, the lack of existing docking pose information, the insufficient use of molecular spatial structure information, and the The prediction of binding affinity is not accurate enough. This is mainly due to the large amount of data and complex spatial structure of candidate drug molecules. Aiming at the limitations of existing protein ligand docking posture prediction methods, the present invention proposes a prediction model based on the combination of graph neural network and generative adversarial network to solve the low efficiency of ligand docking posture structure prediction and improve the binding of ligand to protein Validity of affinity assessment.

Contents of the invention

Purpose of the invention: The present invention provides a method for predicting the docking pose between a protein and a ligand based on a graph neural network, which realizes the full utilization of the global information between the protein and the ligand molecule, and avoids the loss in the traditional prediction method The defect of important information between molecules greatly improves the accuracy and efficiency of predicting the docking posture between the target protein and the ligand molecule.

Technical solution: The present invention aims at a method for predicting the docking posture between a protein and a ligand based on a graph neural network, which specifically includes the following steps:

(1) Acquire the biological information sample set of the protein-ligand complex, and preprocess it; the sample set includes sample data and sample annotation of the sample data, encode the sample data, and obtain the feature vector;

(2) Construct a docking pose generation model based on a graph neural network, and use a generative adversarial network to train it; fix the target protein, use multi-step pose prediction for the ligand molecule to simulate the docking pose structure of the atom, and use the multi-view based The docking pose discriminator evaluates the docking pose, calculates the loss difference, and adjusts the spatial position of the atoms in the docking pose; iterates repeatedly until the discriminant result meets the threshold requirement, and the output for all atoms will be considered as the final predicted pose;

(3) Construct a docking pose evaluation model based on graph neural network, and evaluate the generated ideal docking pose; based on the actual output of the ideal docking pose of protein ligands, evaluate the predicted docking pose results according to the existing mainstream evaluation indicators.

Further, the preprocessing process of the biological information sample set described in step (1) is as follows:

Remove protein-ligand complexes with fewer than two rotatable bonds and proteins with more than one ligand, remove missing or repeated residues, form protein-ligand complexes containing N and their native associations Affinity labeling;

By converting the molecular graph of three-dimensional biomolecular data into a two-dimensional adjacency matrix according to structural properties such as chemical bonds and atomic arrangements, the resulting adjacency matrix conforms to the input format of the graph neural network;

In this way, the protein molecular structure adjacency matrix, the ligand molecular structure adjacency matrix, and the natural binding affinity sample annotations are obtained.

Further, in step (2), a training method for generating a docking pose structure based on a generative confrontation network specifically includes the following steps:

(21) Obtain the initial simulated docking posture structure of the protein ligand; use the AutoDock docking posture structure simulation model to fix the protein molecule, change the spatial position distribution of the ligand molecule, and perform random docking posture structure simulation;

(22) Build a docking pose generation model based on graph neural network to generate candidate docking poses; initialize the neural network model generator according to the simulated docking pose structure; use the sample data extracted from the training sample set and the generator to use the defined noise distribution , use multi-step attitude prediction to simulate the docking pose structure of the atom, calculate the motion of each ligand atom and output the movement vector, and obtain the actual output of the generator's docking pose structure;

(23) Construct a docking posture validity discrimination model based on multiple perspectives, and evaluate the actual output of the generated model; take the natural binding affinity annotation in the original data set PDBbind2016 as the benchmark result, and compare the docking posture results of the generated ligands with Predict the affinity of the target protein from multiple perspectives, and judge the validity of the generated docking posture results; the deviation between the calculated binding affinity of the docking posture and the benchmark result is fed back to the generated model for parameter adjustment and optimization;

(24) Perform repeated training iterations on the generated docking poses, so that the docking poses can reach the ideal expectation on the affinity discrimination index, obtain the trained model and output the ideal docking pose results.

Further, the implementation process of the step (3) is as follows:

The evaluation function is used to measure the root mean square deviation RMSD index, which is defined as follows:

Among them, δ represents the position of the atom in a certain frame minus its position in the reference frame, that is, the position offset, T represents the time, and x represents the position of the atom at a certain moment; the RMSD value represents the magnitude of the motion amplitude of each atom, and the value The larger the , the larger the spatial range of motion of the atom is, and the smaller the steric hindrance of the atom is; in the evaluation model, the smaller the RMSD, the more accurate and effective the generated docking pose structure is.

Further, the validity of the docking posture result generated by judging in step (23) specifically includes the following steps:

(231) Data feature representation; obtain input data, namely ligand molecular graph structure, protein molecular graph structure, the entire protein sequence and the generated protein-ligand docking pose, wherein the protein-ligand docking pose is represented by a bipartite graph;

(232) Data coding; use the Attentive FP model, the BIGGNN model and the ProBert model of the graph neural network to encode the input data respectively, and convert them into feature vectors;

(233) Affinity prediction; the eigenvector obtained by encoding is used to predict the affinity result by a multi-layer perceptron, and the obtained binding affinity value is compared with the natural binding affinity label to determine the validity of the generated posture result;

(234) Calculate the deviation between the predicted result and the actual posture; four indicators are used to measure, the average absolute error MAE, the root mean square error RMSE, the root mean square error standard deviation SD, the Pearson correlation coefficient R, these four indicators is defined as follows:

分别是实验确定的和模型预测的蛋白质-配体结合亲和力的值，a和b分别为回归线的截距和斜率；利用这四项指标计算的值确定损失函数，并反馈给对接姿势生成模型，进行参数的调节和优化。where D is the number of samples in the dataset, y and

are the experimentally determined and model-predicted protein-ligand binding affinity values, a and b are the intercept and slope of the regression line, respectively; the values calculated using these four indicators determine the loss function and feed back to the docking pose generation model, Adjust and optimize parameters.

Beneficial effect: compared with the prior art, the beneficial effect of the present invention: the present invention adopts the graph neural network, realizes the full use of the global information between the protein and the ligand molecule, and avoids the loss of intermolecular information in the traditional prediction method. The defect of important information greatly improves the accuracy and efficiency of predicting the docking posture between the target protein and the ligand molecule; the present invention adopts the docking posture structure generation model based on the generative confrontation network and the docking posture based on the graph neural network Evaluation model Two optimization models make full use of the global information of biomolecules, generate a model of protein-ligand docking pose based on the obtained information, and perform prediction evaluation.

Description of drawings

Fig. 1 is a flowchart of the present invention;

Fig. 2 is a schematic diagram of the preprocessing flow chart of the biological information data of the protein-ligand complex of the present invention;

Fig. 3 is a schematic flowchart of a method for evaluating a protein-ligand generated pose structure of the present invention.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings.

As shown in Figure 1, the present invention proposes a method for predicting the docking posture between a protein and a ligand based on a graph neural network, which specifically includes the following steps:

Step 1: Obtain the biological information sample set of the protein-ligand complex, and preprocess it; the sample set includes sample data and sample annotation of the sample data, encode the sample data, and obtain the feature vector.

Extract the data in the original PDBbind2016 dataset to obtain the biomolecular structure information of ligands and proteins, remove protein-ligand complexes with less than two rotatable bonds and proteins with more than one ligand, and remove missing residues or repeating residues in a protein, forming a protein-ligand complex comprising N and their native binding affinity annotations.

Then, based on structural properties such as chemical bonds and atomic arrangements, the 3D biomolecular data is converted into a 2D adjacency matrix (the adjacency matrix conforms to the input format of the Pytorch_Geometric graph neural network framework). Then the obtained multi-dimensional information of protein-ligand combination is used for docking pose structure generation to randomly obtain a large number of initial docking pose structures. At present, there are mature models for docking pose structure generation, and we use the AutoDock docking pose structure generation model. The final training sample set and test sample set are randomly divided according to the affinity prediction and labeling of the protein-ligand docking pose structure. Among them, each set of sample data includes protein molecular structure adjacency matrix, ligand molecular structure adjacency matrix, protein ligand docking pose structure and corresponding binding affinity sample annotation.

Step 2: Build a docking pose generation model based on a graph neural network, and use a generative confrontation network to train it; fix the target protein, use multi-step pose prediction for the ligand molecule to simulate the docking pose structure of the atom, and use the multi-view based The docking pose discriminator evaluates the docking pose, calculates the loss difference, and adjusts the spatial position of the atoms in the docking pose; iterates repeatedly until the discriminant results meet the threshold requirements, and the output for all atoms will be considered as the final predicted pose. As shown in Figure 2, it specifically includes the following steps:

(2.1) Obtain the initial simulated docking pose structure of the protein ligand; use the docking pose structure generation model AutoDock to fix the protein molecule, change the spatial position distribution of the ligand molecule, and generate a random docking pose structure;

(2.2) Build a docking pose generation model based on a graph neural network to generate candidate docking poses; initialize the neural network model generator, use the sample data extracted from the training sample set and the generator to use the defined noise distribution, and use multiple The step pose prediction simulates the docking pose structure of the atom, calculates the motion of each ligand atom and outputs the movement vector, and obtains the actual output of the generator's docking pose structure;

(2.3) Construct a docking posture validity discrimination model based on multiple perspectives, and evaluate the actual output of the generated model; take the natural binding affinity annotation in the original data set PDBbind2016 as the benchmark result, and compare the docking posture results of the generated ligands with The target protein performs multi-view affinity prediction to judge the validity of the generated docking pose results; the deviation between the calculated docking pose binding affinity and the benchmark result is fed back to the generated model for parameter adjustment and optimization.

Judging the validity of the generated docking pose results, as shown in Figure 3, the specific process is as follows:

1) Data feature representation; the input data are obtained, namely ligand molecular graph structure, protein molecular graph structure, the entire protein sequence and the generated protein-ligand docking pose, where the protein-ligand docking pose is represented by a bipartite graph.

2) Data encoding: Use the Attentive FP model, BIGGNN model and ProBert model of the graph neural network to encode the input data and convert them into feature vectors.

3) Affinity prediction: use the encoded eigenvectors to predict the affinity results using a multi-layer perceptron, and compare the obtained binding affinity values with the natural binding affinity labels to determine the validity of the generated pose results.

4) Calculate the deviation between the predicted result and the actual posture; four indicators are used to measure, the average absolute error MAE, the root mean square error RMSE, the root mean square error standard deviation SD, the Pearson correlation coefficient R, the four indicators It is defined as follows:

分别是实验确定的和模型预测的蛋白质-配体结合亲和力的值，a和b分别为回归线的截距和斜率。利用这四项指标计算的值确定损失函数，并反馈给对接姿势生成模型，进行参数的调节和优化。where D is the number of samples in the dataset, y and

are the experimentally determined and model-predicted values of protein-ligand binding affinity, respectively, and a and b are the intercept and slope of the regression line, respectively. Use the values calculated by these four indicators to determine the loss function, and feed it back to the docking pose generation model for parameter adjustment and optimization.

(2.4) Perform repeated training iterations on the generated docking poses, so that the docking poses can reach the ideal expectation on the affinity discrimination index, obtain the trained model and output the ideal docking pose results.

Step 3: Build a docking pose evaluation model based on graph neural network, and evaluate the generated ideal docking pose; for the actual output of the ideal docking pose of the protein ligand, evaluate the predicted docking pose results based on the existing mainstream evaluation indicators.

Use the traditional evaluation function to measure the root mean square deviation RMSD index, which is defined as follows:

Among them, δ represents the position of the atom in a certain frame minus its position in the reference system, that is, the position offset, T represents the time, x represents the position of the atom at a certain moment; the RMSD value represents the magnitude of the motion range of each atom, and the value The larger the , the larger the spatial range of the movement of the atom is, and the smaller the steric hindrance of the atom is. In the evaluation model, the smaller the RMSD, the more accurate and efficient the generated docking pose structure.

Claims (5)

1. A method for predicting the docking posture between a graph neural network-based protein and a ligand, characterized in that it comprises the following steps: (1) Acquire the biological information sample set of the protein-ligand complex, and preprocess it; the sample set includes sample data and sample annotation of the sample data, encode the sample data, and obtain the feature vector; (2) Construct a docking pose generation model based on a graph neural network, and use a generative adversarial network to train it; fix the target protein, use multi-step pose prediction for the ligand molecule to simulate the docking pose structure of the atom, and use the multi-view based The docking pose discriminator evaluates the docking pose, calculates the loss difference, and adjusts the spatial position of the atoms in the docking pose; iterates repeatedly until the discriminant result meets the threshold requirement, and the output for all atoms will be considered as the final predicted pose; (3) Construct a docking pose evaluation model based on graph neural network, and evaluate the generated ideal docking pose; based on the actual output of the ideal docking pose of protein ligands, evaluate the predicted docking pose results according to the existing mainstream evaluation indicators. 2. the method for predicting the docking posture between the protein and the ligand based on the graph neural network according to claim 1, characterized in that, the preprocessing process of the biological information sample set described in step (1) is as follows: Remove protein-ligand complexes with fewer than two rotatable bonds and proteins with more than one ligand, remove missing or repeated residues, form protein-ligand complexes containing N and their native associations Affinity labeling; By converting the molecular graph of three-dimensional biomolecular data into a two-dimensional adjacency matrix according to structural properties such as chemical bonds and atomic arrangements, the resulting adjacency matrix conforms to the input format of the graph neural network; In this way, the protein molecular structure adjacency matrix, the ligand molecular structure adjacency matrix, and the natural binding affinity sample annotations are obtained. 3. The prediction method of the docking posture between the protein and the ligand based on the graph neural network according to claim 1, characterized in that, in step (2), a training based on the docking posture structure generation of the generative confrontation network The method specifically includes the following steps: (21) Obtain the initial simulated docking posture structure of the protein ligand; use the AutoDock docking posture structure simulation model to fix the protein molecule, change the spatial position distribution of the ligand molecule, and perform random docking posture structure simulation; (22) Build a docking pose generation model based on graph neural network to generate candidate docking poses; initialize the neural network model generator according to the simulated docking pose structure; use the sample data extracted from the training sample set and the generator to use the defined noise distribution , use multi-step attitude prediction to simulate the docking pose structure of the atom, calculate the motion of each ligand atom and output the movement vector, and obtain the actual output of the generator's docking pose structure; (23) Construct a docking posture validity discrimination model based on multiple perspectives, and evaluate the actual output of the generated model; take the natural binding affinity annotation in the original data set PDBbind2016 as the benchmark result, and compare the docking posture results of the generated ligands with Predict the affinity of the target protein from multiple perspectives, and judge the validity of the generated docking posture results; the deviation between the calculated binding affinity of the docking posture and the benchmark result is fed back to the generated model for parameter adjustment and optimization; (24) Perform repeated training iterations on the generated docking poses, so that the docking poses can reach the ideal expectation on the affinity discrimination index, obtain the trained model and output the ideal docking pose results. 4. the method for predicting the docking posture between the protein and the ligand based on the graph neural network according to claim 1, characterized in that, the step (3) implementation process is as follows: The evaluation function is used to measure the root mean square deviation RMSD index, which is defined as follows:

Among them, δ represents the position of the atom in a certain frame minus its position in the reference frame, that is, the position offset, T represents the time, and x represents the position of the atom at a certain moment; the RMSD value represents the magnitude of the motion amplitude of each atom, and the value The larger the , the larger the spatial range of motion of the atom is, and the smaller the steric hindrance of the atom is; in the evaluation model, the smaller the RMSD, the more accurate and effective the generated docking pose structure is. 5. The method for predicting the docking posture between a protein and a ligand based on a graph neural network according to claim 3, wherein the validity of the docking posture result generated by judging in step (23) specifically includes the following steps : (231) Data feature representation; obtain input data, namely ligand molecular graph structure, protein molecular graph structure, the entire protein sequence and the generated protein-ligand docking pose, wherein the protein-ligand docking pose is represented by a bipartite graph; (232) data encoding; Utilize the AttentiveFP model, the BIGGNN model and the ProBert model of the graph neural network to encode the input data respectively, and convert them into feature vectors; (233) Affinity prediction; the eigenvector obtained by encoding is used to predict the affinity result by a multi-layer perceptron, and the obtained binding affinity value is compared with the natural binding affinity label to determine the validity of the generated posture result; (234) Calculate the deviation between the predicted result and the actual posture; four indicators are used to measure, the average absolute error MAE, the root mean square error RMSE, the root mean square error standard deviation SD, the Pearson correlation coefficient R, these four indicators is defined as follows:

where D is the number of samples in the dataset, y and

are the experimentally determined and model-predicted protein-ligand binding affinity values, a and b are the intercept and slope of the regression line, respectively; the values calculated using these four indicators determine the loss function and feed back to the docking pose generation model, Adjust and optimize parameters.

Images