Protein ligand affinity prediction method and system based on voxel compound and CNN-GAM

Through voxel compounds and CNN-GAM methods, the spatial characteristics of proteins and ligands are extracted using three-dimensional convolutional neural networks and global attention mechanisms, and the problem of plane representation limitations in protein sequence affinity prediction is solved, which significantly improves the prediction accuracy. Suitable for computational drug development and virtual screening processes.

CN118969123APending Publication Date: 2024-11-15SHANGHAI SHEN LIAN BIOMEDICAL CORP
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411228305.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-03
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art is limited to plane representation in the affinity prediction of protein sequences, resulting in low prediction accuracy and inability to effectively utilize three-dimensional models.

Method used

Using voxel compounds and CNN-GAM methods, the spatial characteristics of proteins and ligands are extracted through three-dimensional convolutional neural networks and global attention mechanisms, single-hot encoding vectors are generated, and feature extraction and affinity prediction are performed, using PDBbind-2019 The data set evaluates prediction accuracy.

Benefits of technology

It significantly improves the accuracy of protein and ligand affinity prediction, which is better than traditional deep learning and machine learning methods, and does not require prior functional form assumptions, and has good practicality and application prospects.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118969123A_ABST
    Figure CN118969123A_ABST
Patent Text Reader

Abstract

The invention provides a voxel compound and CNN-GAM-based protein ligand affinity prediction method and system, and the method comprises the steps: generating a single thermal coding vector: extracting the spatial features of a ligand and a protein through a trained model, and generating a corresponding single thermal coding vector after voxelization; a feature extraction step: performing feature extraction on the single thermal coding vectors of the ligands and the proteins based on a three-dimensional convolutional neural network to obtain extracted feature vectors; and an affinity prediction step: predicting the binding affinity related to the pKi value based on a global attention mechanism and the extracted feature vector, and outputting a prediction result. The method provided by the invention not only solves the problem of predicting the affinity of the protein and the ligand in the processes of drug development, drug virtual screening and reutilization, but also remarkably improves the precision of predicting the space structure of the compound, and has a relatively good application prospect.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for predicting protein ligand affinity based on voxel compounds and CNN-GAM. Background Art

[0002] Accurately predicting the binding affinity of compounds to proteins is a major challenge in the virtual screening of candidate drugs. Using computational methods to conduct effective high-throughput virtual screening of lead compounds can greatly accelerate the process of drug development by reducing R&D time and experimental workload.

[0003] In recent years, deep learning has been widely used in the field of protein structure, but the constructed protein sequence graph structure is limited to planar representation, and three-dimensional models are rarely used.

[0004] In the Chinese patent document with the publication number CN116504310A, a protein affinity prediction method, device, electronic device and storage medium are disclosed, wherein the method encodes the protein sequence to obtain a protein string representation, and constructs a graph structure of the protein sequence based on the protein string representation. After initializing the multi-task prediction model, the multi-task prediction model is iteratively trained based on the protein string representation, the graph structure of the protein sequence and the corresponding protein's true affinity. At the end of the iterative training, a target multi-task prediction model is obtained, and the target multi-task prediction model is used to predict the affinity of the protein sequence to be tested. However, the patent document only encodes the protein sequence to obtain the protein string representation, and constructs the graph structure of the protein sequence based on this, which results in the constructed protein sequence graph structure being limited to a plane representation; proteins are all spatial three-dimensional structures, and the affinity prediction of the protein sequence to be tested is only performed after iterative training with a plane model, and its accuracy is significantly different from that of analysis and prediction using a three-dimensional model. Summary of the invention

[0005] In view of the defects in the prior art, the object of the present invention is to provide a protein ligand affinity prediction method and system based on voxel compounds and CNN-GAM.

[0006] According to the present invention, a protein ligand affinity prediction method based on voxel compounds and CNN-GAM is provided, comprising:

[0007] One-hot encoding vector generation step: Use the trained model to extract the spatial features of ligands and proteins, and generate the corresponding one-hot encoding vectors after voxelization;

[0008] Feature extraction step: feature extraction of the one-hot encoding vectors of ligands and proteins based on a three-dimensional convolutional neural network to obtain extracted feature vectors;

[0009] Affinity prediction step: Based on the global attention mechanism and the extracted feature vector, the binding affinity related to the pKi value is predicted and the prediction results are output.

[0010] Preferably, the one-hot encoding vector generation step comprises:

[0011] Step S1.0: constructing an experimental model framework; the experimental model framework includes a three-dimensional convolutional neural network and a global attention mechanism;

[0012] Step S1.1: Configure a structure-based data-driven approach within the experimental model framework; use an end-to-end deep learning framework and train on hypothetical protein-ligand complexes;

[0013] Step S1.2: Use the designed data-driven method to obtain the required data from the database, establish the interaction model between protein and ligand, and extract the spatial representation of the ligand chemical features generated by the model;

[0014] Step S1.3: Based on the spatial representation of the ligand generated by the combined model, the properties of the ligand and protein are voxelized separately.

[0015] Preferably, step S1.2 includes extracting a refined subset for establishing an interaction model between protein and ligand based on the protein-ligand complexes in the database and the corresponding experimentally determined binding affinity data obtained from published studies and protein databases through a data-driven method, according to the quality standards of structural resolution and experimental accuracy of the binding measurement.

[0016] Preferably, step S1.3 includes combining the spatial representation of the ligand generated by the model, voxelizing the properties of the ligand and the protein respectively; for each feature, creating a cubic grid centered on the geometric center of the ligand; creating a voxel from the grid position of each atom and assigning it a corresponding attribute value so that each voxel generates a one-hot encoding vector.

[0017] Preferably, the interaction model includes a 3D-CNN module and a GAM module; the single-hot encoding vectors of the ligand and the protein are passed to the 3D-CNN module for feature extraction, and the extracted feature vectors are provided to the GAM module to predict the binding affinity associated with the pKi value.

[0018] According to the present invention, a protein ligand affinity prediction system based on voxel compounds and CNN-GAM is provided, comprising:

[0019] One-hot encoding vector generation module: Use the trained model to extract the spatial features of ligands and proteins, and generate the corresponding one-hot encoding vectors after voxelization;

[0020] Feature extraction module: extract features from the one-hot encoding vectors of ligands and proteins based on a three-dimensional convolutional neural network to obtain extracted feature vectors;

[0021] Affinity prediction module: predicts the binding affinity related to the pKi value based on the global attention mechanism and the extracted feature vector, and outputs the prediction results.

[0022] Preferably, the one-hot encoding vector generation module comprises:

[0023] Module M1.0: construct an experimental model framework; the experimental model framework includes a three-dimensional convolutional neural network and a global attention mechanism;

[0024] Module M1.1: Configure a structure-based data-driven approach within an experimental model framework; use an end-to-end deep learning framework trained on hypothetical protein-ligand complexes;

[0025] Module M1.2: Use the designed data-driven method to obtain the required data from the database, establish the interaction model between protein and ligand, and extract the spatial representation of the ligand chemical features generated by the model;

[0026] Module M1.3: Spatial representation of ligands generated by the binding model, voxelizing the properties of ligands and proteins separately.

[0027] Preferably, the module M1.2 comprises extracting a refined subset for establishing an interaction model between protein and ligand based on the protein-ligand complexes in the database through a data-driven approach, as well as the corresponding experimentally determined binding affinity data obtained from published studies and protein databases, according to the quality standards of structural resolution and experimental accuracy of the binding measurements.

[0028] Preferably, the module M1.3 includes a spatial representation of the ligand generated by the binding model, voxelizing the properties of the ligand and the protein respectively; for each feature, creating a cubic grid centered on the geometric center of the ligand; creating a voxel from the grid position of each atom and assigning it a corresponding attribute value so that each voxel generates a one-hot encoding vector.

[0029] Preferably, the interaction model includes a 3D-CNN module and a GAM module; the single-hot encoding vectors of the ligand and the protein are passed to the 3D-CNN module for feature extraction, and the extracted feature vectors are provided to the GAM module to predict the binding affinity associated with the pKi value.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. The present invention studies the application of three-dimensional voxel compounds based on CNN and GAM in the prediction of protein-ligand affinity. By adopting the three-dimensional convolutional neural network (3D-CNN) and global attention mechanism (GAM) architecture, the model's ability to capture complex interactions is enhanced.

[0032] 2. The present invention uses the PDBbind-2019 dataset to evaluate the prediction accuracy of different models, and verifies that this method is significantly better than existing deep learning, machine learning and traditional scoring algorithms under a fully data-driven situation, without the need for prior functional form assumptions, and has good practicality.

[0033] 3. The method provided by the present invention not only solves the problem of protein-ligand affinity prediction in the process of computational drug development, drug virtual screening and reuse, but also significantly improves the accuracy of compound spatial structure prediction and has good application prospects.

[0034] Other beneficial effects of the present invention will be explained in the specific implementation manner through the introduction of specific technical features and technical solutions. Through the introduction of these technical features and technical solutions, those skilled in the art should be able to understand the beneficial technical effects brought about by the technical features and technical solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0036] Figure 1 Schematic diagram of the prediction process of the present invention.

[0037] Figure 2 The present invention is a flow chart of the method. DETAILED DESCRIPTION

[0038] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0039] Reference Figure 1 and Figure 2As shown, the present invention provides a protein ligand affinity prediction method based on voxel compounds and CNN-GAM, which extracts the spatial features of ligands and proteins through a model, and generates a single-hot encoding vector after voxelization. Then, the single-hot encoding vectors of ligands and proteins are passed to the corresponding 3D-CNN module for feature extraction. The extracted feature vectors are then provided to the GAM module to predict the binding affinity associated with the pKi value. In this way, the method not only solves the problem of protein-ligand affinity prediction in computational drug development, drug virtual screening and reuse, but also significantly improves the accuracy of compound spatial structure prediction.

[0040] The specific steps are as follows:

[0041] Step 1: Build an experimental model framework including a three-dimensional convolutional neural network (3DCNN) and a global attention mechanism (GAM). The experimental model construction includes the following steps: ① Design the network structure of 3DCNN, including input layer, multiple 3D convolution layers, activation layer (such as ReLU), pooling layer and fully connected layer. ② Select the appropriate convolution kernel size and stride, which is usually used to extract spatial and temporal features. ③ Introduce a global attention mechanism in the middle of 3DCNN to enhance the representation of important features. Use a global average pooling layer or a weighted mechanism to calculate the importance weight of each feature. ④ Create a special Loss loss function to replace the traditional mean square error (MSE) as the loss function in the model training process. ⑤ Conduct model training and monitor the loss and accuracy during training. ⑥ Evaluate the model performance on the validation set and test set, and calculate indicators such as accuracy and F1-score. And analyze and visualize the results. ⑦ Adjust the model structure and parameters according to the evaluation results, and conduct multiple experiments to optimize the performance. Used to accurately predict the affinity between proteins and ligands.

[0042] Step 2: Within the experimental model framework, a structure-based data-driven method is designed. The data-driven process includes the following steps: ① Select 19,443 protein-ligand complex data from the PDBbind (v.2019) database ② Clean the original data, remove duplicate data, and standardize or normalize to ensure scale consistency between different features. At the same time, the data is divided into training set, validation set, and test set. ③ Use the pre-constructed experimental model to train the model on the training set, optimize the model parameters, evaluate the model performance on the validation set, and adjust the hyperparameters to prevent overfitting. ④ Evaluate the generalization ability of the model on the test set, and use the Pearson correlation coefficient to measure the model performance. Adopt an end-to-end deep learning framework to train on hypothetical protein-ligand complexes to achieve accurate prediction of affinity.

[0043] Step 3: A data-driven approach was used to extract a smaller refined subset (4852 complexes) from 19443 protein-ligand complexes in the PDBbind (v.2019) database, as well as the corresponding experimentally determined binding affinity data (including dissociation constant (Kd), inhibition constant (Ki) or half-valence (IC50)) obtained from published studies and the Protein Data Bank (PDB). The complexes were used to establish the interaction model between protein and ligand, and to extract the spatial representation of the ligand chemical features generated by the model, based on the quality criteria of structural resolution and experimental accuracy of the binding measurement.

[0044] Step 4: Combine the ligand spatial representation generated by the model and voxelize the properties of the ligand and protein separately. For each feature, create a A cubic grid with spacing of Centered on the geometric center of the ligand, a voxel is created from the grid position of each atom and assigned its corresponding attribute value, resulting in a one-hot encoded vector for each voxel.

[0045] Step 5: The one-hot encoded vectors of ligands and proteins are passed to the corresponding 3D-CNN modules in the model for feature extraction. The extracted feature vectors are then provided to the GAM module in the model to predict the binding affinity associated with the pKi value. The global attention mechanism (GAM) module achieves accurate and reliable prediction of binding affinity mainly through the following process: ① The GAM module weights the extracted feature vectors. The weight of each feature is calculated using the global attention mechanism. ② The input feature vector is passed through a feedforward neural network to calculate the attention weight. The weight reflects the importance of each feature in the prediction process. ③ Each feature vector is multiplied by its corresponding attention weight, and then the weighted features are aggregated to obtain a global context vector. ④ The aggregated global context vector is mapped to the output node through a fully connected layer, which represents the predicted pKi value or binding affinity. ⑤ Compare the predicted pKi value with the true value, calculate the loss, and adjust the model parameters through backpropagation to continuously optimize the prediction performance.

[0046] The present invention also provides a protein ligand affinity prediction system based on voxel compounds and CNN-GAM. The protein ligand affinity prediction system based on voxel compounds and CNN-GAM can be realized by executing the process steps of the protein ligand affinity prediction method based on voxel compounds and CNN-GAM, that is, those skilled in the art can understand the protein ligand affinity prediction method based on voxel compounds and CNN-GAM as a preferred embodiment of the protein ligand affinity prediction system based on voxel compounds and CNN-GAM.

[0047] Specifically, a protein ligand affinity prediction system based on voxel compounds and CNN-GAM includes:

[0048] One-hot encoding vector generation module: Use the trained model to extract the spatial features of ligands and proteins, and generate the corresponding one-hot encoding vectors after voxelization;

[0049] Feature extraction module: extract features from the one-hot encoding vectors of ligands and proteins based on a three-dimensional convolutional neural network to obtain extracted feature vectors;

[0050] Affinity prediction module: predicts the binding affinity related to the pKi value based on the global attention mechanism and the extracted feature vector, and outputs the prediction results.

[0051] The one-hot encoding vector generation module includes:

[0052] Module M1.0: construct an experimental model framework; the experimental model framework includes a three-dimensional convolutional neural network and a global attention mechanism;

[0053] Module M1.1: Configure a structure-based data-driven approach within an experimental model framework; use an end-to-end deep learning framework trained on hypothetical protein-ligand complexes;

[0054] Module M1.2: Use the designed data-driven method to obtain the required data from the database, establish the interaction model between protein and ligand, and extract the spatial representation of the ligand chemical features generated by the model;

[0055] Module M1.3: Spatial representation of ligands generated by the binding model, voxelizing the properties of ligands and proteins separately.

[0056] The module M1.2 consists of extracting a refined subset for building interaction models between proteins and ligands based on the quality criteria of structural resolution and experimental accuracy of binding measurements based on a data-driven approach from protein-ligand complexes in the database and corresponding experimentally determined binding affinity data obtained from published studies and protein databases.

[0057] The module M1.3 includes a spatial representation of the ligand generated by the binding model, voxelizing the properties of the ligand and protein respectively; for each feature, creating a cubic grid centered on the geometric center of the ligand; creating a voxel from the grid position of each atom and assigning it the corresponding attribute value so that each voxel generates a one-hot encoded vector.

[0058] The interaction model includes a 3D-CNN module and a GAM module; the single-hot encoding vectors of the ligand and the protein are passed to the 3D-CNN module for feature extraction, and the extracted feature vectors are provided to the GAM module for predicting the binding affinity associated with the pKi value.

[0059] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0060] In the description of the present application, it should be understood that the terms "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0061] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A protein ligand affinity prediction method based on voxel compounds and CNN-GAM, characterized in that: include: One-hot encoding vector generation step: Use the trained model to extract the spatial features of ligands and proteins, and generate the corresponding one-hot encoding vectors after voxelization; Feature extraction step: feature extraction of the one-hot encoding vectors of ligands and proteins based on a three-dimensional convolutional neural network to obtain extracted feature vectors; Affinity prediction step: Based on the global attention mechanism and the extracted feature vector, the binding affinity related to the pKi value is predicted and the prediction results are output.

2. The protein ligand affinity prediction method based on voxel compounds and CNN-GAM according to claim 1, characterized in that: The one-hot encoding vector generation step includes: Step S1.0: constructing an experimental model framework; the experimental model framework includes a three-dimensional convolutional neural network and a global attention mechanism; Step S1.1: Configure a structure-based data-driven approach within the experimental model framework; use an end-to-end deep learning framework and train on hypothetical protein-ligand complexes; Step S1.2: Use the designed data-driven method to obtain the required data from the database, establish the interaction model between protein and ligand, and extract the spatial representation of the ligand chemical features generated by the model; Step S1.3: Based on the spatial representation of the ligand generated by the combined model, the properties of the ligand and protein are voxelized separately.

3. The protein ligand affinity prediction method based on voxel compounds and CNN-GAM according to claim 2, characterized in that: The step S1.2 includes extracting a refined subset for establishing an interaction model between protein and ligand based on the quality standards of structural resolution and experimental accuracy of binding measurements based on the protein-ligand complexes in the database and the corresponding experimentally determined binding affinity data obtained from published studies and protein databases through a data-driven approach.

4. The protein ligand affinity prediction method based on voxel compounds and CNN-GAM according to claim 2, characterized in that: The step S1.3 includes combining the spatial representation of the ligand generated by the model, voxelizing the properties of the ligand and the protein respectively; for each feature, creating a cubic grid centered on the geometric center of the ligand; creating a voxel from the grid position of each atom and assigning it a corresponding attribute value so that each voxel generates a one-hot encoding vector.

5. The protein ligand affinity prediction method based on voxel compounds and CNN-GAM according to claim 3, characterized in that: The interaction model includes a 3D-CNN module and a GAM module; the single-hot encoding vectors of the ligand and the protein are passed to the 3D-CNN module for feature extraction, and the extracted feature vectors are provided to the GAM module for predicting the binding affinity associated with the pKi value.

6. A protein ligand affinity prediction system based on voxel compounds and CNN-GAM, characterized in that: include: One-hot encoding vector generation module: Use the trained model to extract the spatial features of ligands and proteins, and generate the corresponding one-hot encoding vectors after voxelization; Feature extraction module: extract features from the one-hot encoding vectors of ligands and proteins based on a three-dimensional convolutional neural network to obtain extracted feature vectors; Affinity prediction module: predicts the binding affinity related to the pKi value based on the global attention mechanism and the extracted feature vector, and outputs the prediction results.

7. The protein ligand affinity prediction system based on voxel compounds and CNN-GAM according to claim 6, characterized in that: The one-hot encoding vector generation module includes: Module M1.0: construct an experimental model framework; the experimental model framework includes a three-dimensional convolutional neural network and a global attention mechanism; Module M1.1: Configure a structure-based data-driven approach within an experimental model framework; use an end-to-end deep learning framework trained on hypothetical protein-ligand complexes; Module M1.2: Use the designed data-driven method to obtain the required data from the database, establish the interaction model between protein and ligand, and extract the spatial representation of the ligand chemical features generated by the model; Module M1.3: Spatial representation of ligands generated by the binding model, voxelizing the properties of ligands and proteins separately.

8. The protein ligand affinity prediction system based on voxel compounds and CNN-GAM according to claim 7, characterized in that: The module M1.2 consists of extracting a refined subset for building interaction models between proteins and ligands based on the quality criteria of structural resolution and experimental accuracy of binding measurements based on a data-driven approach from protein-ligand complexes in the database and corresponding experimentally determined binding affinity data obtained from published studies and protein databases.

9. The protein ligand affinity prediction system based on voxel compounds and CNN-GAM according to claim 7, characterized in that: The module M1.3 includes a spatial representation of the ligand generated by the binding model, voxelizing the properties of the ligand and protein respectively; for each feature, creating a cubic grid centered on the geometric center of the ligand; creating a voxel from the grid position of each atom and assigning it the corresponding attribute value so that each voxel generates a one-hot encoded vector.

10. The protein ligand affinity prediction method based on voxel compounds and CNN-GAM according to claim 8, characterized in that: The interaction model includes a 3D-CNN module and a GAM module; the single-hot encoding vectors of the ligand and the protein are passed to the 3D-CNN module for feature extraction, and the extracted feature vectors are provided to the GAM module for predicting the binding affinity associated with the pKi value.

Citation Information

Patent Citations

  • Binding affinity prediction system and method

    CN106575320A

  • Attention mechanism-based psychotropic drug combination side effect prediction method and system

    CN116153527A