Method for predicting the pose of the docking between a protein and a ligand based on a graph neural network

Abstract

The application discloses a kind of based on graph neural network's protein and ligand between the prediction method of pose of docking pose, first, obtain the biological information sample set of protein-ligand complex, sample set includes sample data and sample annotation data;Second, construct the pose generation model of docking pose based on graph neural network and the pose evaluation model of docking pose based on multi-view, further adjust the parameter of model, the sample data is handled by the structure generation model obtained by training, obtain the pose of protein ligand and actually output;Finally, the stability of output result is evaluated using mainstream pose docking structure evaluation index.This application directly uses the biological structure information of ligand protein to generate the optimal docking pose structure, and evaluates the generation result through multi-angle comprehensive evaluation model, so as to improve the accuracy of ligand-protein pose structure docking prediction, and improve the effectiveness of ligand-protein pose structure docking prediction result evaluation.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided drug design, specifically relating to a method for predicting the docking posture between proteins and ligands based on graph neural networks. Background Technology

[0002] In computer-aided drug design, identifying docking postures with high affinity between specific proteins and ligands has always been a challenge. Traditional methods consider combining existing techniques to generate an infinite number of docking posture combinations, then selecting the most suitable set. This approach ignores global information about the relationship between drug molecules and target proteins. Furthermore, existing docking posture information is limited, and assessing binding affinity requires extensive experimentation. Good screening results necessitate substantial dataset annotation. Insufficient annotation often leads to inaccurate predictions of ligand-target protein binding affinity, and the financial and labor costs of such experiments are prohibitive for most research projects. However, emerging methods utilizing three-dimensional molecular structures to predict docking postures show great promise in predicting protein-ligand binding affinity, improving the efficiency and accuracy of computer-aided drug design.

[0003] Existing methods for predicting docking posture and affinity using the three-dimensional structure of molecules employ various sampling methods to improve prediction efficiency. For example, Glide uses Monte Carlo sampling for global posture information, thus improving the accuracy and speed of docking posture prediction when dealing with a large number of drug molecule ligands. However, methods like Glide often neglect biological optimization of the docking posture. Without optimized posture information, inaccurate predictions of docking posture and affinity are inevitable. More intelligent and accurate prediction models are urgently needed in drug manufacturing. Furthermore, these methods often only consider one type of interaction—intramolecular or intermolecular forces—without comprehensively considering both. This leads to the loss of crucial information during affinity prediction, resulting in inaccurate predictions.

[0004] In general, current methods for predicting the docking posture between proteins and ligands have many limitations, such as long prediction time, limited existing docking posture information, insufficient utilization of molecular spatial structure information, and inaccurate prediction of binding affinity. This is mainly due to the large amount of data and complex spatial structures of candidate drug molecules. This invention addresses the limitations of existing protein-ligand docking posture prediction methods by proposing a prediction model based on a combination of graph neural networks and generative adversarial networks to solve the problem of low prediction efficiency of ligand docking posture structure and improve the effectiveness of ligand-protein binding affinity assessment. Summary of the Invention

[0005] Objective of the invention: This invention provides a method for predicting the docking posture between proteins and ligands based on graph neural networks. This method fully utilizes global information between protein and ligand molecules, avoids the loss of important intermolecular information in traditional prediction methods, and greatly improves the accuracy and efficiency of predicting the docking posture between target proteins and ligand molecules.

[0006] Technical Solution: This invention aims to provide a method for predicting the docking posture between proteins and ligands based on graph neural networks, specifically including the following steps:

[0007] (1) Obtain a bioinformatics sample set of protein-ligand complexes and preprocess it; the sample set includes sample data and sample annotations of the sample data. Encode the sample data to obtain feature vectors.

[0008] (2) Construct a docking posture generation model based on graph neural network and train it using generative adversarial network; fix the target protein, use multi-step posture prediction to simulate the docking posture structure of the atoms of the ligand molecules, evaluate the docking posture by a docking posture discriminator based on multiple perspectives, calculate the loss difference, and adjust the spatial position of the atoms in the docking posture; iterate repeatedly until the discrimination result meets the threshold requirement, and the output of all atoms will be considered as the final predicted posture.

[0009] (3) Construct a docking posture evaluation model based on graph neural network to evaluate the generated ideal docking posture; output the actual ideal docking posture of protein ligand, and evaluate the predicted docking posture results based on existing mainstream evaluation indicators.

[0010] Furthermore, the preprocessing of the bioinformatics sample set described in step (1) is as follows:

[0011] Remove protein-ligand complexes with fewer than two rotatable bonds and proteins with more than one ligand, remove proteins with missing or repeated residues, and form a label containing N protein-ligand complexes and their native binding affinity.

[0012] By converting the molecular graph of three-dimensional biomolecular data into a two-dimensional adjacency matrix based on structural properties such as chemical bonds and atomic arrangement, the resulting adjacency matrix conforms to the input format of graph neural networks.

[0013] This yields the protein molecular structure adjacency matrix, the ligand molecular structure adjacency matrix, and the natural binding affinity sample annotations.

[0014] Furthermore, in step (2), a training method for generating docking pose structures based on generative adversarial networks specifically includes the following steps:

[0015] (21) Obtain the initial simulated docking posture structure of the protein ligand; use the AutoDock docking posture structure simulation model to fix the protein molecules and change the spatial distribution of the ligand molecules to simulate the random docking posture structure.

[0016] (22) Construct a docking posture generation model based on graph neural network to generate candidate docking postures; initialize the neural network model generator according to the simulated docking posture structure; predict the docking posture structure of simulated atoms of ligand molecules using multi-step posture prediction by the sample data extracted from the training sample set and the generator using the defined noise distribution, calculate the motion of each ligand atom and output the movement vector, and obtain the actual output of the docking posture structure of the generator.

[0017] (23) Construct a multi-view docking posture validity discrimination model and evaluate the actual output of the generated model; take the natural binding affinity labeling in the original dataset PDBbind2016 as the benchmark result, perform multi-view affinity prediction between the docking posture result of the generated ligand and the target protein, and judge the validity of the generated docking posture result; feed back the deviation between the calculated binding affinity of the docking posture and the benchmark result to the generated model for parameter adjustment and optimization.

[0018] (24) Repeatedly train and iterate the generated docking posture so that the docking posture reaches the ideal expectation in the affinity discrimination index, obtain the trained model and output the ideal docking posture result.

[0019] Furthermore, the implementation process of step (3) is as follows:

[0020] The root mean square deviation (RMSD) index is measured using an evaluation function, defined as follows:

[0021]

[0022] Where δ represents the position of an atom in a certain frame minus its position in the reference frame, i.e., the position offset, 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. The larger the value, the larger the spatial range of the atom's motion, and the smaller the spatial steric hindrance of the atom; in the evaluation model, the smaller the RMSD, the more accurate and effective the generated docking posture structure is.

[0023] Furthermore, step (23) of determining the validity of the generated docking posture result specifically includes the following steps:

[0024] (231) Data feature representation; Obtain input data, namely ligand molecular diagram structure, protein molecular diagram structure, whole protein sequence and generated protein ligand docking posture, wherein the protein ligand docking posture is represented by a bipartite graph;

[0025] (232) Data encoding; The input data is encoded and converted into feature vectors using the Attentive FP model, BIGGNN model and ProBert model of graph neural networks respectively;

[0026] (233) Affinity prediction: The encoded feature vector is used to predict the affinity result using a multilayer perceptron. The obtained binding affinity value is compared with the natural binding affinity label to determine the validity of the generated pose result.

[0027] (234) The deviation between the predicted results and the actual posture was calculated; four indicators were used to measure it: mean absolute error (MAE), root mean square error (RMSE), root mean square error standard deviation (SD), and Pearson correlation coefficient (R). The definitions of these four indicators are as follows:

[0028]

[0029] Where D is the number of samples in the dataset, and y and These are the experimentally determined and model-predicted protein-ligand binding affinity values, respectively, and a and b are the intercept and slope of the regression line, respectively. The loss function is determined using the values ​​calculated from these four indicators and fed back to the docking posture generation model for parameter adjustment and optimization.

[0030] Beneficial Effects: Compared with the prior art, the beneficial effects of this invention are as follows: This invention uses a graph neural network to fully utilize the global information between proteins and ligand molecules, avoiding the loss of important intermolecular information in traditional prediction methods, and greatly improving the accuracy and efficiency of predicting the docking posture between target proteins and ligand molecules; This invention employs two optimized models: a docking posture structure generation model based on generative adversarial networks and a docking posture evaluation model based on graph neural networks, to fully utilize the global information of biomolecules, generate a model of protein-ligand docking posture based on the obtained information, and perform prediction and evaluation. Attached Figure Description

[0031] Figure 1 This is a flowchart of the present invention;

[0032] Figure 2 This is a schematic diagram of the preprocessing flow for bioinformatics data of protein-ligand complexes according to the present invention.

[0033] Figure 3 This is a schematic diagram of the protein-ligand generation posture structure evaluation method of the present invention. Detailed Implementation

[0034] The present invention will now be described in further detail with reference to the accompanying drawings.

[0035] like Figure 1 As shown, this invention proposes a method for predicting the docking posture between proteins and ligands based on graph neural networks, specifically including the following steps:

[0036] Step 1: Obtain a bioinformatics sample set of protein-ligand complexes and preprocess it; the sample set includes sample data and sample annotations of the sample data. Encode the sample data to obtain feature vectors.

[0037] Data was extracted from the original PDBbind2016 dataset to obtain the biomolecular structure information of ligands and proteins. Protein-ligand complexes with fewer than two rotatable bonds and proteins with more than one ligand were removed. Proteins with missing or repeated residues were also removed to form a label containing N protein-ligand complexes and their natural binding affinity.

[0038] Then, based on structural properties such as chemical bonds and atomic arrangement, the three-dimensional biomolecular data is converted into a two-dimensional adjacency matrix (this adjacency matrix conforms to the input format of the PyTorch_Geometric graph neural network framework). The obtained multidimensional information of the protein-ligand combination is then used to generate docking posture structures, randomly acquiring a large number of initial docking posture structures. Currently, there are mature models for docking posture structure generation; we use the AutoDock docking posture structure generation model. The final training and testing sample sets are randomly partitioned based on the affinity prediction and annotation of the protein-ligand docking posture structures. Each set of sample data includes the protein molecular structure adjacency matrix, the ligand molecular structure adjacency matrix, the protein-ligand docking posture structure, and the corresponding binding affinity sample annotations.

[0039] Step 2: Construct a docking posture generation model based on a graph neural network and train it using a generative adversarial network; fix the target protein, use multi-step posture prediction to simulate the docking posture structure of atoms for ligand molecules, evaluate the docking posture using a multi-view docking posture discriminator, calculate the loss difference, and adjust the spatial position of atoms in the docking posture; iterate repeatedly until the discrimination result meets the threshold requirement, and the output of all atoms will be considered as the final predicted posture. Figure 2 As shown, the specific steps include:

[0040] (2.1) Obtain the initial simulated docking posture structure of the protein ligand; use the docking posture structure generation model AutoDock to fix the protein molecule and change the spatial distribution of the ligand molecule to generate a random docking posture structure.

[0041] (2.2) Construct a docking posture generation model based on graph neural network to generate candidate docking postures; initialize the neural network model generator, and use the sample data extracted from the training sample set and the generator to use the defined noise distribution to predict the docking posture structure of the simulated atoms of the ligand molecules using multi-step posture prediction, calculate the motion of each ligand atom and output the movement vector to obtain the actual output of the docking posture structure of the generator.

[0042] (2.3) Construct a multi-view docking posture validity discrimination model and evaluate the actual output of the generated model; take the natural binding affinity label in the original dataset PDBbind2016 as the benchmark result, perform multi-view affinity prediction between the docking posture result of the generated ligand and the target protein, and judge the validity of the generated docking posture result; feed back the deviation between the calculated binding affinity of the docking posture and the benchmark result to the generated model for parameter adjustment and optimization.

[0043] To determine the validity of the generated docking posture results, such as... Figure 3 As shown, the specific process is as follows:

[0044] 1) Data feature representation; Obtain input data, namely ligand molecular diagram structure, protein molecular diagram structure, whole protein sequence and generated protein ligand docking posture, wherein the protein ligand docking posture is represented by a bipartite graph.

[0045] 2) Data encoding: The input data is encoded and converted into feature vectors using Attentive FP, BIGGNN and ProBert models of graph neural networks.

[0046] 3) Affinity prediction: The encoded feature vector is used to predict the affinity result using a multilayer perceptron. The obtained binding affinity value is compared with the natural binding affinity label to determine the validity of the generated pose result.

[0047] 4) Calculate the deviation between the predicted results and the actual posture; four indicators are used for measurement: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Root Mean Square Error Standard Deviation (SD), and Pearson Correlation Coefficient (R). The definitions of these four indicators are as follows:

[0048]

[0049] Where D is the number of samples in the dataset, and y and These are the experimentally determined and model-predicted protein-ligand binding affinity values, respectively, and a and b are the intercept and slope of the regression line, respectively. The values ​​calculated using these four indicators determine the loss function, which is then fed back to the docking posture generation model for parameter adjustment and optimization.

[0050] (2.4) Repeatedly train and iterate the generated docking posture until the docking posture reaches the ideal expectation in the affinity discrimination index, obtain the trained model and output the ideal docking posture result.

[0051] Step 3: Construct a docking posture evaluation model based on graph neural networks to evaluate the generated ideal docking posture; output the actual ideal docking posture of the protein ligand, and evaluate the predicted docking posture results based on existing mainstream evaluation indicators.

[0052] The root mean square deviation (RMSD) index is measured using a traditional evaluation function, and its definition is as follows:

[0053]

[0054] Where δ represents the position of an atom in a given frame minus its position in the reference frame, i.e., the position offset; x represents the position of the atom at a certain moment; the RMSD value represents the magnitude of the motion amplitude of each atom. The larger the value, the larger the spatial range of the atom's motion, and the smaller the spatial steric hindrance of the atom. In the evaluation model, the smaller the RMSD, the more accurate and effective the generated docking posture structure.

Claims

1. A method for predicting the docking posture between proteins and ligands based on graph neural networks, characterized in that, Includes the following steps: (1) Obtain a bioinformatics sample set of protein-ligand complexes and preprocess it; the sample set includes sample data and sample annotations of the sample data. Encode the sample data to obtain feature vectors. (2) Construct a docking posture generation model based on graph neural network and train it using generative adversarial network; fix the target protein, use multi-step posture prediction to simulate the docking posture structure of the atoms of the ligand molecules, evaluate the docking posture by a docking posture discriminator based on multiple perspectives, calculate the loss difference, and adjust the spatial position of the atoms in the docking posture. The process is iterated repeatedly until the discrimination result meets the threshold requirement. The output of all atoms will then be considered as the final predicted posture. (3) Construct a docking posture evaluation model based on graph neural network to evaluate the generated ideal docking posture; The actual output of the ideal docking posture for protein ligands is evaluated based on existing mainstream evaluation metrics.

2. The method for predicting the docking posture between proteins and ligands based on graph neural networks according to claim 1, characterized in that, The preprocessing of the bioinformatics 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 proteins with missing or repeated residues, and form a label containing N protein-ligand complexes and their native binding affinity. By converting the molecular graph of three-dimensional biomolecular data into a two-dimensional adjacency matrix based on structural properties such as chemical bonds and atomic arrangement, the resulting adjacency matrix conforms to the input format of graph neural networks. This yields the protein molecular structure adjacency matrix, the ligand molecular structure adjacency matrix, and the natural binding affinity sample annotations.

3. The method for predicting the docking posture between proteins and ligands based on graph neural networks according to claim 1, characterized in that, In step (2), a training method for generating docking pose structures based on generative adversarial networks 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 molecules and change the spatial distribution of the ligand molecules to simulate the random docking posture structure. (22) Construct a docking posture generation model based on graph neural network to generate candidate docking postures; initialize the neural network model generator according to the simulated docking posture structure; By using sample data extracted from the training sample set and the generator using the defined noise distribution, the docking posture structure of the simulated atoms is predicted using multi-step posture prediction for the ligand molecules. The motion of each ligand atom is calculated and the movement vector is output to obtain the actual output of the generator's docking posture structure. (23) Construct a docking posture validity discrimination model based on multiple perspectives and evaluate the actual output of the generated model; Using the natural binding affinity annotations in the original dataset PDBbind2016 as the baseline, the docking posture results of the generated ligands are compared with the target protein from multiple perspectives to predict the affinity and determine the validity of the generated docking posture results. The deviation between the calculated docking posture and the baseline result is fed back to the generative model for parameter adjustment and optimization. (24) Repeatedly train and iterate the generated docking posture so that the docking posture reaches the ideal expectation in the affinity discrimination index, obtain the trained model and output the ideal docking posture result.

4. The method for predicting the docking posture between proteins and ligands based on graph neural networks according to claim 1, characterized in that, The implementation process of step (3) is as follows: The root mean square deviation (RMSD) index is measured using an evaluation function, defined as follows: Where δ represents the position of an atom in a certain frame minus its position in the reference frame, i.e., the position offset, 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. The larger the value, the larger the spatial range of the atom's motion, and the smaller the spatial steric hindrance of the atom; in the evaluation model, the smaller the RMSD, the more accurate and effective the generated docking posture structure is.

5. The method for predicting the docking posture between proteins and ligands based on graph neural networks according to claim 3, characterized in that, Step (23) of determining the validity of the generated docking posture result specifically includes the following steps: (231) Data feature representation; Obtain input data, namely ligand molecular diagram structure, protein molecular diagram structure, whole protein sequence and generated protein ligand docking posture, wherein the protein ligand docking posture is represented by a bipartite graph; (232) Data encoding; The input data is encoded and converted into feature vectors using Attentive FP, BIGGNN and ProBert models of graph neural networks, respectively. (233) Affinity prediction: The encoded feature vector is used to predict the affinity result using a multilayer perceptron. The obtained binding affinity value is compared with the natural binding affinity label to determine the validity of the generated pose result. (234) The deviation between the predicted results and the actual posture was calculated; four indicators were used to measure it: mean absolute error (MAE), root mean square error (RMSE), root mean square error standard deviation (SD), and Pearson correlation coefficient (R). The definitions of these four indicators are as follows: Where D is the number of samples in the dataset, and y and These are the experimentally determined and model-predicted protein-ligand binding affinity values, respectively, and a and b are the intercept and slope of the regression line, respectively. The loss function is determined using the values ​​calculated from these four indicators and fed back to the docking posture generation model for parameter adjustment and optimization.

Images