SchNetPack Hyperparameter Optimization for a More Reliable Top Docking Scores Prediction.

Options to improve the extrapolation power of the neural network designed using the SchNetPack package with respect to top docking scores prediction are presented. It is shown that hyperparameter tuning of the atomistic model representation (in the schnetpack.representation) improves the prediction of the top scoring compounds, which have characteristically a low incidence in randomized data sets for training of machine learning models. The prediction robustness is evaluated according to the mean square error (MSE) and the entropy of the average loss landscape decrease. Admittedly, the improvement of the top scoring compounds' prediction accuracy comes with the penalty of worsening the overall prediction power. It is revealed that the most impactful hyperparameter is the cutoff (5 Å is reported as the optimal choice). Other parameters (e.g., number of radial basis functions, number of interaction layers of the neural network, feature vector size or its batch size) are found to not affect the prediction robustness of the top scoring compounds in any comparable way relative to the cutoff. The MSE of the best docking score prediction (below -13 kcal/mol) improves from ca. 3.5 to 0.9 kcal/mol, while the prediction of less potent compounds (-13 to -11 kcal/mol) shows a lesser improvement, i.e., a decrease of MSE from 1.6 to 1.3 kcal/mol. Additionally, oversampling and undersampling of the training set with respect to the top scoring compounds' abundance is presented. The results indicate that the cutoff choice performs better than over- or undersampling of the training set, with undersampling performing better than oversampling.

Compromise in Docking Power of Liganded Crystal Structures of Mpro SARS-CoV-2 Surpasses 90% Success Rate.

Herein, we present the capacity of three different molecular docking programs (AutoDock, AutoDock Vina, and PLANTS) to identify and reproduce the binding modes of ligands present in 247 covalent and 169 noncovalent complex crystal structures of the severe acute respiratory syndrome coronavirus 2 main protease (Mpro). The compromise in docking power is evaluated with respect to their ability to generate poses similar to the crystal structure binding mode (heavy atoms' root-mean-square deviation < 2 Å) and their ability to recognize the native binding mode with an included compensation for the scoring function error. Noncovalently bound inhibitors are best modeled by AutoDock Vina (90.6% success rate in the active site), while the most relevant results for covalently bound inhibitors are produced by PLANTS (93.0%). AutoDock shows acceptable performance for both types of ligands, 81.1 and 76.4% for noncovalent and covalent complexes, respectively. All three programs manifest worse performance when reproducing surface-bound ligands. Comparison with other works illustrates the importance of crystal structure processing (12% of noncovalent and 26% of covalent ligands had to be manually corrected), proper sampling protocol settings, and inclusion of root-mean-square deviation (RMSD)/scoring function error compensations in crystal structure pose identification. Results are analyzed with respect to a clustering scheme of the noncovalently bound ligands and the chemical reaction type of the covalent ligand bound to the Cys145 residue. A comparison of screening power based on the docking scores of noncovalent ligands from the crystal structures with a "Directory of Useful Decoys, Enhanced" set of known decoys (6562 compounds) and ZINC15 in vivo subset (60,394 compounds) is provided. Ligand and protein input files are provided for future benchmarking purposes.

Molecular docking and machine learning affinity prediction of compounds identified upon softwood bark extraction to the main protease of the SARS-CoV-2 virus.

Molecular docking of 234 unique compounds identified in the softwood bark (W set) is presented with a focus on their inhibition potential to the main protease of the SARS-CoV-2 virus 3CLpro (6WQF). The docking results are compared with the docking results of 866 COVID19-related compounds (S set). Furthermore, machine learning (ML) prediction of docking scores of the W set is presented using the S set trained TensorFlow, XGBoost, and SchNetPack ML approaches. Docking scores are evaluated with the Autodock 4.2.6 software. Four compounds in the W set achieve a docking score below -13 kcal/mol, with (+)-lariciresinol 9'-p-coumarate (CID 11497085) achieving the best docking score (-15 kcal/mol) within the W and S sets. In addition, 50% of W set docking scores are found below -8 kcal/mol and 25% below -10 kcal/mol. Therefore, the compounds identified in the softwood bark, show potential for antiviral activity upon extraction or further derivatization. The W set molecular docking studies are validated by means of molecular dynamics (five best compounds). The solubility (Log S, ESOL) and druglikeness of the best docking compounds in S and W sets are compared to evaluate the pharmacological potential of compounds identified in softwood bark.

Machine learning prediction of 3CLpro SARS-CoV-2 docking scores.

Molecular docking results of two training sets containing 866 and 8,696 compounds were used to train three different machine learning (ML) approaches. Neural network approaches according to Keras and TensorFlow libraries and the gradient boosted decision trees approach of XGBoost were used with DScribe's Smooth Overlap of Atomic Positions molecular descriptors. In addition, neural networks using the SchNetPack library and descriptors were used. The ML performance was tested on three different sets, including compounds for future organic synthesis. The final evaluation of the ML predicted docking scores was based on the ZINC in vivo set, from which 1,200 compounds were randomly selected with respect to their size. The results obtained showed a consistent ML prediction capability of docking scores, and even though compounds with more than 60 atoms were found slightly overestimated they remain valid for a subsequent evaluation of their drug repurposing suitability.

3CLpro and PLpro affinity, a docking study to fight COVID19 based on 900 compounds from PubChem and literature. Are there new drugs to be found?

The spread of a novel coronavirus SARS-CoV-2 and a resulting COVID19 disease in late 2019 has transformed into a worldwide pandemic and has effectively brought the world to a halt. Proteases 3CLpro and PLpro, responsible for proteolysis of new virions, represent vital inhibition targets for the COVID19 treatment. Herein, we report an in silico docking study of more than 860 COVID19-related compounds from the PubChem database. Molecular dynamic simulations were carried out to validate the conformation stability of compound-ligand complexes with best docking scores. The MM-PBSA approach was employed to calculate binding free energies. The comparison with ca. 50 previously reported potential SARS-CoV-2's proteases inhibitors show a number of new compounds with excellent binding affinities. Anti-inflammatory drugs Montelukast, Ebastine and Solumedrol, the anti-migraine drug Vazegepant or the anti-MRSA pro-drug TAK-599, among many others, all show remarkable affinities to 3CLpro and with known side effects present candidates for immediate clinical trials. This study reports thorough docking scores summary of COVID19-related compounds found in the PubChem database and illustrates the asset of computational screening methods in search for possible drug-like candidates. Several yet-untested compounds show affinities on par with reported inhibitors and warrant further attention. Furthermore, the submitted work provides readers with ADME data, ZINC and PubChem IDs, as well as docking scores of all studied compounds for further comparisons.