Virtual screening method based on scoring function optimization and used for efficient screening of PAK4 inhibitor

Abstract

The invention provides a virtual screening method based on scoring function optimization. The virtual screening method is used for efficiently screening a PAK4 inhibitor. According to the method, molecular docking, scoring function optimization, machine learning and molecular dynamics simulation are combined, and the screening efficiency and accuracy are improved through a multi-level screening strategy. The method comprises the following steps: firstly, screening candidate compounds in a database by utilizing a molecular similarity algorithm, and reserving molecules similar to target binding pocket characteristics; and then, screening and optimizing the compound by adopting rapid molecular docking and key residues. Thirdly, using scoring functions such as RF-VS and GNINA as well as a deep learning method to optimize screening, and evaluating the performance of the model through an ROC curve and an enrichment factor; and finally, verifying a screening result by utilizing high-precision docking and 150ns molecular dynamics simulation, and confirming the stability and targeted binding capacity of the candidate molecules. The method is successfully applied to PAK4 inhibitor screening, the screening efficiency is remarkably improved, and an efficient and reliable solution is provided for research and development of PAK4 targeted drugs.

Description

Technical Field

[0001] The present invention relates to the technical field of drug virtual screening, specifically combining scoring function optimization, molecular docking and molecular dynamics methods, through the optimization of scoring function evaluation test set and the application of multi-level screening scheme, and taking PAK4 inhibitors as a research case, to explore new highly selective compounds targeting specific targets. Background Art

[0002] Virtual screening is a high-throughput screening method based on molecular docking, which is widely used in the field of drug development. However, the existing molecular docking technology evaluation standards are mainly based on conformational sampling and scoring function evaluation. In order to meet the needs of the increasingly large compound libraries, the sampling and scoring function evaluation have been simplified in algorithms, which can easily cause false positives and reduce the reliability of screening results. By optimizing the scoring function evaluation test set and applying multi-level screening schemes, the efficiency and accuracy of virtual screening can be improved, false positives can be reduced, and new highly selective compounds targeting specific targets can be explored using the PAK4 target as an example.

[0003] PAK4 (p21-activated protein kinase 4) is a representative subtype of type II in the PAK family and is closely related to the occurrence, invasion and migration of various cancers. However, most PAK4 inhibitors are still in the preclinical research stage due to poor bioavailability and insufficient selectivity, and the existing chemical structure is relatively simple. The present invention optimizes the virtual screening strategy, takes the PAK4 target as an example, improves the screening ability of PAK4 inhibitors, and thus quickly screens new, highly selective PAK4 inhibitors from large databases. The present invention aims to optimize the virtual screening technology process, speed up the screening of large databases, reduce the false positive rate, and at the same time improve the screening success rate and accuracy of active molecules. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of insufficient efficiency and accuracy in the prior art when screening drug molecules based on virtual screening methods. To this end, a virtual screening method based on a multi-level screening strategy is provided, which can improve the screening efficiency when computing power is limited, and effectively improve the discovery rate of high-quality candidate compounds. At the same time, the present invention combines deep learning with pharmacophore shape similarity screening to effectively improve the discrimination of the test set, thereby enhancing the reliability of the test results and optimizing the scoring function. The present invention also aims to overcome the problems of poor selectivity of PAK4 inhibitors in the prior art, high similarity of chemical structures, high false positives and low accuracy during virtual screening, and provide an efficient solution for targeted drug research and development. To achieve the above purpose, the present invention is implemented by the following technical solutions.

[0005] The virtual screening method based on the multi-level screening strategy of scoring function and structure includes the following steps:

[0006] S1) Data preparation and template screening

[0007] Through the known active template compounds, combined with the characteristics of the target binding pocket, the molecular similarity algorithm is used to screen candidate compounds from the target database, retaining molecules with similar 3D shapes but potential structural novelty.

[0008] S2) Rapid molecular docking and residue screening

[0009] A rapid molecular docking algorithm is used to screen the preliminary molecules, and the structural images are screened in combination with the action mode of key residues in the target protein to retain molecules that can form favorable interactions with important active amino acid residues and quickly filter out compounds that do not match the target.

[0010] S3) Test set construction and scoring function training

[0011] By screening molecules with similar shapes but significantly different pharmacophores, we construct a bait set and optimize the discrimination of the test set. We combine multiple algorithms, including machine learning and deep learning models, to train and optimize the scoring function. We use ROC curves and enrichment factors to comprehensively evaluate the performance of the scoring function, significantly improving the accuracy and reliability of the screening function.

[0012] S4) Fine Molecular Docking and Rescoring

[0013] A high-precision molecular docking algorithm is used to accurately dock candidate compounds, and the docking results are re-scored using an optimized scoring function based on a neural network to further improve screening accuracy.

[0014] S5) Kinetic verification

[0015] The candidate compounds finally screened were subjected to a 150ns molecular dynamics simulation, combined with RMSD and binding free energy analysis, to verify the stability of the molecules in the docking system and their ability to bind to the target.

[0016] As a further improvement of the present invention, in the template screening of step S1), a suitable template is selected according to the size characteristics of the target-specific hydrophobic pocket and the binding template, and based on the principle of molecular similarity, compounds with similar shapes are more likely to have similar activities.

[0017] As a further improvement of the present invention, the rapid molecular docking in step S2) is specifically: rapid docking and obtaining the conformation of the complex, and screening molecules that generate favorable interactions such as hydrogen bonds, van der Waals forces, π-π interactions or hydrophobic interactions with the key residues counted in the literature.

[0018] As a further improvement of the present invention, the scoring functions tested in step S3) include RF-VS, GNINA, DS (PLP1, PLP2, LigScore1, LigScore2, Jain, PMF, PMF04), (docking score,XPscore,Glide score).

[0019] The beneficial effects of the present invention are as follows:

[0020] The present invention proposes a virtual screening method based on a scoring function and a multi-level screening strategy. By combining rapid molecular docking, key residue screening, test set optimization and machine learning model optimization scoring function strategy, the efficiency and accuracy of molecular screening are significantly improved. In particular, by optimizing the test set, the ability to distinguish active and inactive molecules is improved, the false positive rate is significantly reduced, and the reliability of the screening results is enhanced. The screening optimization method of the present invention has been successfully applied to large-scale specific screening of PAK4 targets, providing an efficient and reliable solution for the development of targeted drugs for PAK4, and has a wide range of practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flowchart of the multi-layer virtual screening method based on scoring function optimization and structure provided by the present invention;

[0022] Figure 2 Chemical structures of active molecules used to screen template molecules and establish test sets;

[0023] Figure 3 Calculate the results of ROC and EF for each model and scoring function;

[0024] Figure 4 The top ten compound structures obtained after re-scoring and screening;

[0025] Figure 5 To screen the obtained compounds, RMSD results of dynamic simulation were obtained;

[0026] Figure 6 GNINA scoring results and MMGBSA results for the screened compounds. DETAILED DESCRIPTION

[0027] The present invention is further described in detail below with reference to the accompanying drawings and specific examples.

[0028] Refer to the attached Figure 1 As shown, the multi-conformation dynamics virtual screening method of the present invention comprises the following steps:

[0029] 1. Rapidly filter the data molecule library based on active molecule structure information

[0030] 1.1 Data preparation and template shape screening

[0031] This method selects the compound compd55, which is known to have PAK4 targeting activity, as a template. This molecule enters the specific hydrophobic pocket of PAK4 through the rigid structure of its alkyl side chain and forms stable interactions with key residues GLU366, GLU396, and LEU398. The shape similarity algorithm in the Schrödinger Shape Screening function is used to screen molecules with highly similar structures to the template compound. The specific parameters and databases for screening are as follows:

[0032] Target database: To ensure biological activity, natural products and their derivatives databases are preferred, such as TopScience's natural compound library and Analyticon Discovery's natural product library. A total of 2,128,437 molecules are included, and each group is divided into about 100,000 molecules to facilitate subsequent screening operations.

[0033] Screening parameters: The shape screening tool sets the shape similarity threshold to 0.6. Based on the scoring results, the top 10% of compounds in each molecular set are selected to ensure that the molecules have structural similarity after screening.

[0034] 1.2 Rapid molecular docking and screening of important residues

[0035] The fast molecular docking algorithm (LibDock) of Discovery Studio software was used to perform fast docking filtration on the compounds that had undergone shape screening. By calculating the docking score, compounds with a score greater than 110 were screened out. At the same time, in combination with the key amino acid residue Asp458 of the target, molecules that can form key favorable interactions with this residue (such as hydrogen bonds, hydrophobic interactions, van der Waals forces, Pi-Sigma, etc.) were screened out, thereby quickly excluding molecules that do not match the target or have low potential activity, reducing the number of databases, and speeding up subsequent screening.

[0036] Filters:

[0037] Docking score: The threshold was set to LibDock Score>110, which retained approximately 60% of the compounds.

[0038] Interaction: Focus on screening molecules that form favorable interactions with Asp458 (such as hydrogen bonds, hydrophobic interactions, etc.) to improve the screening efficiency of potential active compounds.

[0039] 2. Build a test set, train and test the scoring function

[0040] 2.1 Test set construction

[0041] In order to improve the accuracy of screening, the present invention constructed a test set consisting of active molecules and inactive molecules. The active molecules were derived from literature, PDB database, etc., and 18 compounds with PAK4 inhibitory activity were collected (see Appendix Figure 2 ). Decoy molecules were generated by DUD-E website tools, and the screening process was similar to that of the database. To avoid false positives, the pharmacophore characteristics of each active molecule were analyzed to ensure that the decoy molecules had similar shapes but significantly different pharmacodynamic characteristics, thereby improving the discrimination of the test set. The final test set consisted of 18 active molecules and 311 inactive molecules.

[0042] Active compound collection: Through literature research and database query, the molecular structures of 18 PAK4 activity inhibitors were obtained and sorted. The DUD-E web tool was used to import the SMILES structure of each active molecule and generate the corresponding bait molecule set.

[0043] Screening parameters of decoy molecules: Schrödinger Shape Screening algorithm was used for initial screening. Each decoy set used each active molecule as a template, and the decoy molecules with shape similarity greater than 0.6 were retained. The Auto Pharmacophore Generation function of Discovery Studio (DS) was further used to calculate the pharmacophore model of the active template, and the Ligand Pharmacophore Mapping function was used to perform pharmacophore matching analysis on the decoy molecules. Based on the FitValue and the number of matching features, the decoy molecules with lower pharmacophore similarity were selected as part of the final test set.

[0044] 2.2 Scoring function training

[0045] In order to screen highly active and selective PAK4 inhibitors, a variety of scoring functions were used for training and evaluation, including: RF_VS, GNINA (CNN_affinity, CNN_score, affinity, CNN_VS), DS built-in functions (-PLP1, -PLP2, Lig1score, Lig2score, Jain, PMF, PMF04), The Glide docking scoring function (docking score, XP score, Glide score) is used, where CNN_VS is the product of CNN_affinity and CNN_score, and multiple indicators of model performance are combined to optimize the screening method. First, active molecules and inactive molecules after screening are marked as 1 and 0, respectively, for supervised learning of the model. After training, the ROC curve is used to reflect the discrimination ability of the model, and the enrichment factor (EF) is used to evaluate the ability of the model to enrich active molecules in the early stage of screening, focusing on the analysis of the EF2% and EF5% indicators. The results are shown in the attached figure. Figure 4 The following is the specific construction and evaluation process:

[0046] RF_VS model training: RF_VS is a virtual screening tool based on the random forest (RF) algorithm. The training process uses a large-scale data set from the DUD-E database, including more than 15,000 active molecules and 900,000 decoy molecules. Rf-vs uses supervised learning and the random forest algorithm to build a classifier that learns the mapping relationship between molecular features and active labels. During the training process, the model gradually reduces misclassification and improves prediction accuracy by optimizing the construction and adjustment of decision trees.

[0047] GNINA model training: Based on a deep learning framework, GNINA uses a 3D convolutional neural network (3D-CNN) to process the three-dimensional spatial structure information of molecules and predict the binding ability of molecules to targets. The model repeatedly optimizes network parameters and combines the spatial characteristics of input compounds to ensure that the model output results have high classification accuracy and prediction ability.

[0048] 3. Fine Molecular Docking and Rescoring

[0049] 3.1 High-precision molecular docking

[0050] The compounds initially screened were subjected to high-precision molecular docking using The Glide XP mode in the software was used for fine docking. After preliminary screening, the remaining 134,918 compounds were used for further docking. The crystal structure of PAK4 (PDB code: 7CP4) was used The Protein Preparation Wizard module of the PCR product was used to process the protein structure, remove water molecules and metal ions, add hydrogen atoms, and repair missing amino acid residues. At the same time, amino acids were protonated at pH 7, and the protein crystal structure was energy minimized using the OPLS3 force field. According to the position of the original ligand, the docking region was set to The center coordinates are (-13.8, 17.53, -7.2).

[0051] 3.2CNN-VS Re-scoring

[0052] According to the attached Figure 4 Scoring function evaluation results show that although the CNN-VS function performs generally well in terms of area under the ROC curve, it performs best in terms of EF2% and EF5% indicators, and can efficiently identify top molecules. Based on the comprehensive score of conformational accuracy and docking score, CNN-VS was selected as the final scoring function for virtual screening to achieve accurate screening of highly selective inhibitors. The conformations of the docking results were rescored using GNINA, and the compounds were re-ranked based on the product of CNN_score and CNN_affinity (i.e., CNN_VS). After sorting, the top ten molecules were selected for 150ns molecular dynamics simulation.

[0053] 4. Kinetic Verification and Result Analysis

[0054] Gromacs2021.6 and MMGBSA were used to perform molecular dynamics simulation and binding free energy calculation on the 10 complex systems with higher scores. The stability of the candidate compounds and their ability to bind to the target were verified. By analyzing the RMSD and MMGBSA binding free energy results, the candidate compound STOCK7S-56165 with higher activity was finally screened out. The results are detailed in the attached Figure 6 The following are the specific parameter settings and calculation methods:

[0055] Kinetic parameter setting: AMBER14ffSB force field was used for protein, and the topology file was generated based on GAFF after ligand optimization by Gaussian16. The complex was placed in TIP3P water box (side length ) and added Cl- to neutralize the system charge. The simulation used the SHAKE method to constrain hydrogen bonds and the PME method to calculate long-range electrostatic interactions. The time step was set to 2fs, the temperature was constant at 298.5K, and the pressure was 1bar. The system was minimized and the energy was reduced to 100kJ / mol in 8000 steps. Then a 150ns simulation was run. The RMSD analysis results are shown in the attached Figure 5 , a stable time period is selected for further calculations.

[0056] MMGBSA calculation settings: extract 75-125ns stable trajectory based on RMSD results, extract one frame every 50ps, and extract 1000 frames for each system for calculation. Adjust the script print_res parameters to calculate the distance ligand The amino acid energy in the range of ΔG total represents binding free energy; ΔG GAS Including electrostatic energy (ΔG ele) and van der Waals energy (ΔG vdw );ΔG solv represents the solvent contribution, including the polar solubility free energy (ΔG GB ) and nonpolar dissolution free energy (ΔG surf ).

Claims

1. A virtual screening method based on scoring function optimization for screening PAK4 inhibitors, characterized in that: The following steps are involved: S1) Data preparation and template screening: Based on known PAK4 active template compounds and the characteristics of the target binding pocket, the molecular similarity algorithm is used to screen candidate compounds from the target database, and molecules with structural similarity and potential structural novelty are retained; S2) Rapid molecular docking and residue screening: The preliminary selected molecules are screened using a rapid molecular docking algorithm, and the molecular conformations are screened in combination with the action mode of the key residues of the target, and the molecules that can form favorable interactions with the active important amino acid residues are retained; S3) Test set construction and scoring function training: Molecules with similar shapes but different pharmacophores are used to generate bait sets, and scoring functions are trained using machine learning and deep learning models. Model performance is evaluated using ROC curves and enrichment factors. S4) Fine molecular docking and rescoring: Use a high-precision molecular docking algorithm to accurately dock the candidate compounds, and re-score the docking results in combination with a neural network-based scoring function; S5) Dynamic verification: The candidate molecules finally screened were subjected to 150 ns molecular dynamics simulation, combined with RMSD and binding free energy analysis to verify the stability of the molecules and their ability to bind to the target.

2. The method according to claim 1, wherein: In the step S1, a molecular similarity threshold is set to screen molecules having a similarity score greater than 0.6 with the template active molecule.

3. The method according to claim 1, wherein: The molecular docking in step S2 is performed using LibDock of Discovery Studio to screen molecules that can form hydrogen bonds, van der Waals forces, π-π interactions or hydrophobic interactions with key residues of the target, and to filter out molecules with poor docking scores.

4. The method according to claim 1, wherein: The scoring functions used in step S3 include RF-VS, GNINA, DS and Built-in functions.

5. The method according to claim 1, wherein: The molecular docking used in step S4 adopts In Glide-XP mode, the rescoring function is calculated using the neural network-based GNINA model.

6. The method according to claim 1, wherein: The 150 ns molecular dynamics simulation in step S5 was performed using Gromacs 2021.6 software, and the binding free energy was calculated in combination with the MMGBSA method.

7. The method according to claim 1, wherein: The test set includes 18 active molecules and 642 inactive molecules. The inactive molecules are generated by DUD-E. After screening for shape and pharmacophore similarity, a total of 329 molecules are screened. The active molecules are marked as 1 and the inactive molecules are marked as 0 for supervised learning.

8. The method according to claim 1, wherein: The enrichment factor (EF) was used to evaluate the model's ability to screen active molecules, and the ROC curve was used to evaluate the model's classification performance. Finally, the scoring function with the best EF performance was selected for virtual screening.

9. The method according to claim 1, wherein: The PAK4 inhibitor screening is used for drug development targeting PAK4 protein.

Images