Accelerated method and system for virtual screening of drug molecules targeting
By using graph neural networks and a hierarchical progressive screening process, combined with molecular complexity scores and experimental feedback optimization, the problem of the imbalance between accuracy and efficiency in virtual screening methods was solved, achieving efficient and reliable drug molecule screening.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 河南省儿童医院郑州儿童医院
- Filing Date
- 2026-05-15
- Publication Date
- 2026-07-14
Smart Images
Figure CN122392708A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer technology, and more specifically to a method and system for accelerating virtual screening of targeted drug molecules. Background Technology
[0002] Drug development is a long and costly process, with lead compound discovery being a key bottleneck. As compound libraries grow to the millions or tens of millions, traditional experimental screening methods can no longer balance efficiency and comprehensiveness. Therefore, computer-based virtual screening technology has become an important preliminary step in drug discovery, used to rapidly narrow down the pool of candidate molecules.
[0003] However, existing virtual screening methods generally suffer from an imbalance between accuracy and efficiency. Rapid screening based on pharmacophore or conformational matching has a high false positive rate, while high-precision molecular docking and molecular dynamics simulations are computationally too time-consuming and difficult to handle ultra-large-scale compound libraries. Meanwhile, deep learning-based prediction methods often use uniform batch processing, failing to consider differences in molecular complexity, resulting in uneven GPU utilization, unstable screening accuracy, and a lack of ability to self-optimize the model using experimental feedback.
[0004] Therefore, there is an urgent need for a virtual screening method that can dynamically allocate computing resources based on molecular characteristics, improve screening accuracy in a hierarchical and progressive manner, and achieve self-optimization through closed-loop experimental data, so as to realize efficient and high-precision screening of large-scale compound libraries under limited computing resources. Summary of the Invention
[0005] In view of the above-mentioned shortcomings of the prior art, the present invention provides a method and system for accelerating virtual screening of targeted drug molecules, which can effectively solve the problems mentioned in the prior art.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] This invention provides a method for accelerating virtual screening of targeted drug molecules, the specific steps of which include:
[0008] S100. Data Preprocessing: Receive the three-dimensional structure data of the target protein and the compound library data; perform structural preprocessing on the target protein and identify active sites; perform chemical structure normalization on each molecule in the compound library; calculate the complexity score of each molecule, which is obtained by weighting the number of atoms, the number of rotatable bonds, and the number of rings; allocate molecules to different processing queues according to the complexity score, and set a batch size for each queue that matches the complexity level;
[0009] S200, First coarse screening: The active sites of the target protein and candidate molecules are converted into graph representations, where nodes represent atoms and edges represent chemical bonds; GPU batch inference is performed on molecules in each processing queue using a graph neural network model to predict protein-ligand binding affinity; the molecules with the highest predicted affinity are selected to enter the next layer.
[0010] S300, Second-layer fine screening: GPU parallel molecular docking calculations are performed on candidate molecules after coarse screening, and ADMET property evaluation is performed simultaneously. The ADMET properties include absorption, distribution, metabolism, excretion, and toxicity. After standardizing the molecular docking score, graph neural network prediction score, and ADMET score, the weights are adjusted according to the target type and weighted fusion is performed to obtain a comprehensive score. The molecules with the highest comprehensive score and the first preset proportion are selected to enter the next layer.
[0011] S400, Third Layer Validation: Perform molecular dynamics simulations on the candidate molecules after fine screening to calculate the binding free energy of the protein-ligand complex; perform a final ranking based on the kinetic stability and binding free energy, and output a list of candidate molecules and an analysis report;
[0012] S500, Feedback Optimization: Collect experimental verification data, update the parameters of the graph neural network model through an incremental learning mechanism, optimize the scoring fusion weights, and add the new data to the knowledge base.
[0013] Furthermore, the formula for calculating the complexity score is as follows:
[0014] ;
[0015] in, The numerator complexity fraction. The total number of atoms in the molecule. is the number of rotatable bonds in the molecule. The number of rings in the molecule. , and These are the weighting coefficients corresponding to the number of atoms, the number of rotatable bonds, and the number of rings, respectively, all of which are positive real numbers;
[0016] Based on complexity score The molecules are divided into at least two categories, and a different batch size is set for each category, where the batch size for the category with lower complexity is larger than that for the category with higher complexity.
[0017] Furthermore, the graph neural network model sequentially includes a graph convolutional layer, an attention mechanism layer, a graph pooling layer, and a fully connected layer. The graph convolutional layer is used to extract local features of atoms and chemical bonds, the attention mechanism layer is used to learn key interactions between proteins and ligands, the graph pooling layer is used to aggregate molecular-level features, and the fully connected layer is used to output the predicted binding affinity value. The GPU batch inference adopts mixed-precision computing and accelerates matrix operations through a parallel computing library to achieve pipelined parallelism of computation and data transmission.
[0018] Furthermore, the GPU-parallel molecular docking employs an improved genetic algorithm, specifically including: generating multiple initial conformation populations in parallel on the GPU; evaluating the docking scores of all conformations in parallel on the GPU, wherein the docking scores comprehensively consider van der Waals interaction energy, electrostatic interaction energy, hydrogen bond interaction energy, desolvation energy, and entropy change contribution; performing population evolution through tournament selection, single-point crossover, and Gaussian mutation; and setting an adaptive termination condition, stopping early when the optimal solution remains unchanged for N consecutive generations, where N is a preset positive integer, dynamically determined based on the population size and convergence speed.
[0019] Furthermore, the comprehensive score is calculated using the following formula:
[0020] ;
[0021] in, For comprehensive scoring; The molecular docking score is the result of standardization. The graph neural network predicts scores after standardization. The standardized ADMET property score; , , These are the corresponding weighting coefficients; the weighting coefficients are adaptively adjusted according to the target type.
[0022] Furthermore, the molecular dynamics simulations sequentially include:
[0023] System preparation: Place the protein-ligand complex in the solvent box and add counterions to neutralize the system;
[0024] Energy minimization can be achieved using the steepest descent method or the conjugate gradient method.
[0025] The equilibrium simulation first performs NVT ensemble equilibrium for a first preset duration, and then performs NPT ensemble equilibrium for a second preset duration. The first and second preset durations are obtained based on the system size and force field type.
[0026] Production simulation is performed for a third preset duration, during which the system is maintained at thermodynamic equilibrium.
[0027] Trajectory analysis was performed to calculate the root mean square deviation (RMSD), root mean square fluctuation (RMSF), hydrogen bond occupancy, and contact area. The binding free energy was calculated using either the molecular mechanics-Poisson-Boltzmann surface area (MM-PBSA) method or the molecular mechanics-generalized Born surface area (MM-GBSA) method.
[0028] Furthermore, the first preset duration and the second preset duration are 100ps to 200ps respectively, and the third preset duration is 10ns to 100ns.
[0029] Furthermore, the incremental learning mechanism includes:
[0030] Collect experimentally determined compound activity data and convert the activity values into standardized pIC50; prioritize adjusting the top-level fully connected parameters of the graph neural network model, and selectively adjust the bottom-level graph convolution parameters according to the scale of the new data; use a preset ratio of the original training learning rate as the fine-tuning learning rate, the preset ratio being determined based on the ratio of the scale of the new data to the scale of the original training data; perform a preset number of training iterations, the preset number being determined based on convergence and validation set performance; verify the accuracy of the updated model on an independent test set, and if the accuracy improves, replace the current model with the updated model.
[0031] Furthermore, the method also includes intelligent task scheduling, specifically:
[0032] The computation time of each docking task is predicted based on the molecular complexity score C and historical statistical data; the shortest job first or shortest remaining time first algorithm is used to schedule and sort the tasks; a CPU-GPU data pipeline is implemented, so that while the GPU is executing the current batch of computation, the CPU completes the preprocessing of the next batch of data; GPU utilization is dynamically monitored, and the batch processing size is increased when the GPU utilization is lower than a preset threshold; dynamic load balancing is supported in multi-GPU environments, and tasks are allocated according to the real-time load of each GPU.
[0033] A virtual screening acceleration system for targeted drug molecules includes:
[0034] The data input module receives three-dimensional structural data of target proteins and compound library data; performs structural preprocessing on target proteins and identifies active sites; performs chemical structure normalization on each molecule in the compound library; calculates the complexity score of each molecule, which is obtained by weighting the number of atoms, rotatable bonds, and rings of the molecule; and allocates molecules to different processing queues according to the complexity score, and sets a batch size for each queue that matches the complexity level.
[0035] The first coarse screening module converts the target protein active sites and candidate molecules into graph representations, where nodes represent atoms and edges represent chemical bonds. The graph neural network model is used to perform GPU batch inference on the molecules in each processing queue to predict the protein-ligand binding affinity. The molecules with the highest predicted affinity are selected to enter the next layer.
[0036] The second-layer fine screening module is used to perform GPU parallel molecular docking calculations on candidate molecules after coarse screening, and at the same time to evaluate ADMET properties. After standardizing the molecular docking score, graph neural network prediction score and ADMET score, the weights are adjusted according to the target type and weighted and fused to obtain a comprehensive score. The molecules with the highest comprehensive scores are selected to enter the next layer.
[0037] The third-layer verification module is used to perform molecular dynamics simulations on the candidate molecules after fine screening, calculate the binding free energy of the protein-ligand complex, and perform a final ranking based on the kinetic stability and binding free energy, outputting a list of candidate molecules and an analysis report.
[0038] The feedback learning module is used to collect experimental verification data, update the parameters of the graph neural network model through an incremental learning mechanism, optimize the scoring fusion weights, and add new data to the knowledge base.
[0039] The technical solution provided by this invention has the following advantages compared with the known prior art:
[0040] This invention constructs a layered virtual screening process consisting of graph neural network coarse screening, parallel molecular docking fine screening, and molecular dynamics simulation verification. It also introduces a differentiated batch processing mechanism based on molecular complexity scores and an intelligent task scheduling strategy. This allows for dynamic adjustment of GPU computing resource allocation according to molecular structural characteristics. While ensuring fine processing of complex molecules, it fully leverages the large-scale parallel throughput capability of graphics processors for simple molecules, significantly shortening the screening cycle of ultra-large-scale compound libraries and achieving a several-fold increase in virtual screening efficiency.
[0041] This invention utilizes a graph neural network model incorporating an attention mechanism to capture key protein-ligand interactions in the coarse screening stage. In the fine screening stage, an improved genetic algorithm is used to search for the optimal docking conformation in parallel, and adaptive termination conditions are combined to reduce redundant calculations. At the same time, molecular docking scores, graph neural network prediction scores, and ADMET property scores are weighted and fused in an adaptive manner to overcome the problem of large screening bias caused by a single scoring standard, effectively improving the overall hit rate and drug potential of candidate molecules. Finally, rigorous verification is carried out through molecular dynamics simulations and combined free energy calculations to ensure the reliability of the screening results and the reproducibility of the experiments. Attached Figure Description
[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0043] Figure 1 This is a schematic diagram of the method flow of the present invention;
[0044] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation
[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0046] The present invention will be further described below with reference to embodiments.
[0047] Example:
[0048] Reference Figure 1 The method for accelerating virtual screening of targeted drug molecules provided by this invention includes the following specific steps:
[0049] S100. Data Preprocessing: Receive the 3D structural data of the target protein and the compound library data to be screened. Perform structural preprocessing on the target protein, including removing water molecules, adding hydrogen atoms, and repairing missing residues, and use existing tools to identify active site pockets. Perform chemical structure normalization on each molecule in the compound library, such as standardizing protonation states, removing salt ions, and generating 3D conformations.
[0050] Specifically, a molecule complexity score can be introduced to differentiate the allocation of subsequent computing resources. The complexity score for each molecule is calculated using the following formula:
[0051] ;
[0052] in, The numerator complexity fraction. The total number of atoms in the molecule. is the number of rotatable bonds in the molecule. The number of rings in the molecule. , and These are the weighting coefficients corresponding to the number of atoms, the number of rotatable bonds, and the number of rings, respectively. All are positive real numbers and can be pre-calibrated according to the computing power environment; for example, taking... , , .
[0053] Based on a complexity score C, molecules are divided into at least two processing queues. For example, molecules with C values below a preset threshold are classified as low-complexity molecules, and those above the threshold are classified as high-complexity molecules. Different batch sizes are set for different queues: low-complexity queues are matched with larger batch sizes (e.g., 512) to fully utilize GPU throughput; high-complexity queues are matched with smaller batch sizes (e.g., 64) to prevent GPU memory overflow and maintain computational efficiency. This hierarchical strategy is a crucial foundation for accelerating coarse screening.
[0054] S200, First coarse screen
[0055] The target protein active sites and candidate molecules are converted into graph representations, where nodes represent atoms and edges represent chemical bonds. A graph neural network model is used for GPU batch inference on molecules in each processing queue to predict protein-ligand binding affinity. This graph neural network model consists of: a graph convolutional layer for extracting local chemical features of atoms and chemical bonds; an attention mechanism layer for learning the key interaction weights between protein and ligands; a graph pooling layer for aggregating and generating molecular-level overall features; and a fully connected layer for finally outputting the predicted binding affinity value.
[0056] During batch inference, mixed-precision computing (FP16 / FP32) is employed, and a parallel computing library is used to accelerate matrix operations, achieving pipelined parallelism of computation and data transmission, further improving throughput. After inference is completed, molecules are sorted according to predicted affinity, and a pre-defined percentage (e.g., the top 10%) are selected to proceed to the next layer. This layer quickly eliminates the vast majority of low-activity molecules with extremely low time cost.
[0057] S300, the second layer of fine screening, performs GPU-parallel molecular docking calculations on candidate molecules after coarse screening, while simultaneously evaluating their ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties. Parallel molecular docking employs an improved genetic algorithm: multiple initial conformation populations are generated in parallel on the GPU; docking scores for all conformations are evaluated in parallel, taking into account van der Waals interaction energy, electrostatic interaction energy, hydrogen bonding energy, desolvation energy, and entropy change contribution; then, population evolution is carried out through tournament selection, single-point crossover, and Gaussian mutation. Specifically, an adaptive termination condition is set: evolution ends early when the optimal solution remains unchanged for N consecutive generations, with N dynamically determined based on population size and convergence speed, thus significantly reducing computational costs while maintaining accuracy. ADMET property evaluation can be performed in parallel using existing rule-based models or machine learning models. After obtaining the molecular docking score D, graph neural network prediction score P, and ADMET score A, the three scores are first standardized (e.g., Z-score standardization to eliminate dimensional differences) to obtain... , , Then, a weighted fusion is performed according to the following comprehensive scoring formula:
[0058] ;
[0059] in, For comprehensive scoring; The molecular docking score is the result of standardization. The graph neural network predicts scores after standardization. The standardized ADMET property score; , , These are the corresponding weighting coefficients; the weighting coefficients are adaptively adjusted according to the target type, for example, for kinase targets, the binding ability can be emphasized and the weighting coefficients increased. and It can enhance the ability to target targets that need to penetrate the blood-brain barrier. Molecules are sorted according to their comprehensive score S, and a predetermined percentage, such as the top 20%, is selected before entering the final validation layer.
[0060] S400, Third-Level Validation
[0061] Molecular dynamics simulations were performed on the candidate molecules after fine screening to obtain the actual binding stability and free energy. The following steps were performed sequentially:
[0062] System preparation involves placing the protein-ligand complex in a solvent box and adding counterions for neutralization. Energy minimization is performed using either the steepest descent method or the conjugate gradient method. Equilibrium simulations are conducted, first with a canonical ensemble (NVT, i.e., constant particle number, volume, and temperature) for 100–200 ps, followed by isothermal-isobaric ensemble (NPT, i.e., constant particle number, pressure, and temperature) for another 100–200 ps. Production simulations involve trajectory sampling over 10–100 ns. The trajectory is then analyzed, and the root mean square deviation (RMSD), root mean square fluctuation (RMSF), hydrogen bond occupancy, and contact area are calculated. Binding free energy is calculated using either molecular mechanics-Poisson-Boltzmann surface area (MM-PBSA) or molecular mechanics-generalized Born surface area (MM-GBSA) methods. Finally, the kinetic stability and binding free energy are combined to rank the molecules, outputting a list of potentially active candidate molecules and a detailed analysis report. This layer serves as final validation to ensure the reliability of the results.
[0063] S500 feedback optimization involves collecting actual compound activity data from subsequent experiments and converting the activity values into the negative logarithm of the half-maximal inhibitory concentration (pIC50). The graph neural network model is updated using an incremental learning mechanism: top-level parameters, such as fully connected layers, are adjusted first; bottom-level graph convolutional parameters are selectively adjusted based on the scale of the new data. The fine-tuning learning rate is set to a preset proportion of the original training learning rate, determined by the ratio of the new data to the original training data size, for example, 1 / 10. A preset number of training iterations are performed until convergence on the validation set. The model accuracy is validated using an independent test set; if it shows improvement over the old model, the updated model is used to replace the old one, while simultaneously optimizing the scoring fusion weights and incorporating the new data into the knowledge base. In this way, the system's predictive ability continuously improves as the selection process progresses.
[0064] Throughout the screening process, especially in the computationally intensive docking and simulation stages, this invention integrates a task scheduling mechanism. Based on the molecular complexity score C and historical statistical data, the computation time for each docking task is predicted. Shortest job first or shortest remaining time first algorithms are used to schedule and sort tasks to reduce average waiting time. A CPU-GPU data pipeline is constructed: while the GPU executes the current batch computation, the CPU asynchronously completes the preprocessing and transmission of the next batch of data. GPU utilization is dynamically monitored; when utilization falls below a preset threshold, the batch processing size is automatically increased to improve resource utilization. In a multi-GPU environment, tasks are dynamically allocated based on the real-time load of each GPU to achieve load balancing. This mechanism is implemented throughout the batch parallelization process of S200 and S300, maximizing hardware throughput efficiency.
[0065] Reference Figure 2 The targeted drug molecule virtual screening acceleration system includes a data input module, a first-layer coarse screening module, a second-layer fine screening module, a third-layer validation module, and a feedback learning module.
[0066] The data input module (S100) handles the input, preprocessing, and complexity-based sorting of protein and compound libraries. The first-layer coarse screening module (S200) utilizes a graph neural network for rapid pre-screening. The second-layer fine screening module (S300) performs parallel docking, ADMET evaluation, and multi-index weighted fusion. The third-layer validation module (S400) generates molecular dynamics simulations and a sorting report. The feedback learning module (S500) incrementally updates the model and weights using experimental data. These modules work collaboratively to form a closed-loop optimized virtual screening pipeline.
[0067] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for accelerating virtual screening of targeted drug molecules, characterized in that, The specific steps include: S100. Data Preprocessing: Receive the three-dimensional structure data of the target protein and the compound library data; perform structural preprocessing on the target protein and identify active sites; perform chemical structure normalization on each molecule in the compound library; calculate the complexity score of each molecule, which is obtained by weighting the number of atoms, the number of rotatable bonds, and the number of rings; allocate molecules to different processing queues according to the complexity score, and set a batch size for each queue that matches the complexity level; S200, First coarse screening: The active sites of the target protein and candidate molecules are converted into graph representations, where nodes represent atoms and edges represent chemical bonds; GPU batch inference is performed on molecules in each processing queue using a graph neural network model to predict protein-ligand binding affinity; the molecules with the highest predicted affinity are selected to enter the next layer. S300, Second-layer fine screening: GPU parallel molecular docking calculations are performed on candidate molecules after coarse screening, and ADMET property evaluation is performed simultaneously. The ADMET properties include absorption, distribution, metabolism, excretion, and toxicity. After standardizing the molecular docking score, graph neural network prediction score, and ADMET score, the weights are adjusted according to the target type and weighted fusion is performed to obtain a comprehensive score. The molecules with the highest comprehensive score and the first preset proportion are selected to enter the next layer. S400, Third Layer Validation: Perform molecular dynamics simulations on the candidate molecules after fine screening to calculate the binding free energy of the protein-ligand complex; perform a final ranking based on the kinetic stability and binding free energy, and output a list of candidate molecules and an analysis report; S500, Feedback Optimization: Collect experimental verification data, update the parameters of the graph neural network model through an incremental learning mechanism, optimize the scoring fusion weights, and add the new data to the knowledge base.
2. The method for accelerating virtual screening of targeted drug molecules according to claim 1, characterized in that, The formula for calculating the complexity score is as follows: ； in, The numerator complexity fraction. The total number of atoms in the molecule. is the number of rotatable bonds in the molecule. The number of rings in the molecule. , and These are the weighting coefficients corresponding to the number of atoms, the number of rotatable bonds, and the number of rings, respectively, all of which are positive real numbers; Based on complexity score The molecules are divided into at least two categories, and a different batch size is set for each category, where the batch size for the category with lower complexity is larger than that for the category with higher complexity.
3. The method for accelerating virtual screening of targeted drug molecules according to claim 1, characterized in that, The graph neural network model sequentially includes a graph convolutional layer, an attention mechanism layer, a graph pooling layer, and a fully connected layer. The graph convolutional layer is used to extract local features of atoms and chemical bonds, the attention mechanism layer is used to learn key interactions between proteins and ligands, the graph pooling layer is used to aggregate molecular-level features, and the fully connected layer is used to output the predicted binding affinity value. The GPU batch inference uses mixed-precision computing and accelerates matrix operations through a parallel computing library to achieve pipelined parallelism of computation and data transmission.
4. The method for accelerating virtual screening of targeted drug molecules according to claim 1, characterized in that, The GPU-parallel molecular docking employs an improved genetic algorithm, specifically including: generating multiple initial conformation populations in parallel on the GPU; evaluating the docking scores of all conformations in parallel on the GPU, whereby the docking scores comprehensively consider van der Waals interaction energy, electrostatic interaction energy, hydrogen bond interaction energy, desolvation energy, and entropy change contribution; performing population evolution through tournament selection, single-point crossover, and Gaussian mutation; and setting an adaptive termination condition, stopping early when the optimal solution remains unchanged for N consecutive generations, where N is a preset positive integer dynamically determined based on the population size and convergence speed.
5. The method for accelerating virtual screening of targeted drug molecules according to claim 1, characterized in that, The comprehensive score is calculated using the following formula: ； in, For comprehensive scoring; The molecular docking score is the result of standardization. The graph neural network predicts scores after standardization. The standardized ADMET property score; , , These are the corresponding weighting coefficients; the weighting coefficients are adaptively adjusted according to the target type.
6. The method for accelerating virtual screening of targeted drug molecules according to claim 1, characterized in that, The molecular dynamics simulations include, in sequence: System preparation: Place the protein-ligand complex in the solvent box and add counterions to neutralize the system; Energy minimization can be achieved using the steepest descent method or the conjugate gradient method. The equilibrium simulation first performs NVT ensemble equilibrium for a first preset duration, and then performs NPT ensemble equilibrium for a second preset duration. The first and second preset durations are obtained based on the system size and force field type. Production simulation is performed for a third preset duration, during which the system is maintained at thermodynamic equilibrium. Trajectory analysis was performed to calculate the root mean square deviation (RMSD), root mean square fluctuation (RMSF), hydrogen bond occupancy, and contact area. The binding free energy was calculated using either the molecular mechanics-Poisson-Boltzmann surface area (MM-PBSA) method or the molecular mechanics-generalized Born surface area (MM-GBSA) method.
7. The method for accelerating virtual screening of targeted drug molecules according to claim 6, characterized in that, The first preset duration and the second preset duration are 100ps to 200ps, respectively, and the third preset duration is 10ns to 100ns.
8. The method for accelerating virtual screening of targeted drug molecules according to claim 1, characterized in that, The incremental learning mechanism includes: Collect experimentally determined compound activity data and convert the activity values into standardized pIC50; prioritize adjusting the top-level fully connected parameters of the graph neural network model, and selectively adjust the bottom-level graph convolution parameters according to the scale of the new data; use a preset ratio of the original training learning rate as the fine-tuning learning rate, the preset ratio being determined based on the ratio of the scale of the new data to the scale of the original training data; perform a preset number of training iterations, the preset number being determined based on convergence and validation set performance; verify the accuracy of the updated model on an independent test set, and if the accuracy improves, replace the current model with the updated model.
9. The method for accelerating virtual screening of targeted drug molecules according to claim 1, characterized in that, The method also includes intelligent task scheduling, specifically: The computation time of each docking task is predicted based on the molecular complexity score C and historical statistical data; the shortest job first or shortest remaining time first algorithm is used to schedule and sort the tasks; a CPU-GPU data pipeline is implemented, so that while the GPU is executing the current batch of computation, the CPU completes the preprocessing of the next batch of data; GPU utilization is dynamically monitored, and the batch processing size is increased when the GPU utilization is lower than a preset threshold; dynamic load balancing is supported in multi-GPU environments, and tasks are allocated according to the real-time load of each GPU.
10. A virtual screening acceleration system for targeted drug molecules, characterized in that, include: The data input module is used to receive target protein three-dimensional structure data and compound library data; Structural preprocessing of target proteins and identification of active sites; Chemical structure standardization was performed on each molecule in the compound library; Calculate the complexity score for each molecule, which is a weighted average of the number of atoms, rotatable bonds, and rings in the molecule; assign molecules to different processing queues based on their complexity scores, and set a batch size for each queue that matches the complexity level; The first coarse screening module converts the target protein active sites and candidate molecules into graph representations, where nodes represent atoms and edges represent chemical bonds. The graph neural network model is used to perform GPU batch inference on the molecules in each processing queue to predict the protein-ligand binding affinity. The molecules with the highest predicted affinity are selected to enter the next layer. The second-layer fine screening module is used to perform GPU parallel molecular docking calculations on candidate molecules after coarse screening, and at the same time to evaluate ADMET properties. After standardizing the molecular docking score, graph neural network prediction score and ADMET score, the weights are adjusted according to the target type and weighted and fused to obtain a comprehensive score. The molecules with the highest comprehensive scores are selected to enter the next layer. The third-layer verification module is used to perform molecular dynamics simulations on the candidate molecules after fine screening, calculate the binding free energy of the protein-ligand complex, and perform a final ranking based on the kinetic stability and binding free energy, outputting a list of candidate molecules and an analysis report. The feedback learning module is used to collect experimental verification data, update the parameters of the graph neural network model through an incremental learning mechanism, optimize the scoring fusion weights, and add new data to the knowledge base.