There has been a recent explosion in the use of deep machine learning (ML) in computational chemistry and drug discovery. ML models can predict many important drug properties but rely on the underlying training data and chemical featurization. Another popular method in drug discovery and soft condensed matter physics is atomistic molecular dynamics (MD) simulations. However, these simulations are extremely resource intensive, making it infeasible to use simulations at each step in a virtual drug screening campaign. We are working to address how ML can be effectively combined with MD and other physics-based modeling data to develop more predictive models. In the first application, we use MD data and MD features to train ML models that predict small molecule partitioning and hydrophobicity, which are critical properties in drug discovery. We calculated 15,000 small molecule free energies of transfer using atomistic MD simulations. This data was then used to train ML models to predict the free energies of transfer, comparing shallow learning from the 2D molecular fingerprint to spatial-graph neural networks (NN), and 3D convolutional NN. The accuracy of the best model was lower than that of the expected error from the MD calculation. The second application is predicting accurate protein–ligand binding affinities, which is critical in drug discovery for prioritizing molecules for subsequent experimental validation. Traditional biophysics methods are either too resource intensive or are not reliable proxies for binding affinity. We have undertaken a comprehensive study of deep ML models trained on the experimentally derived PDBbind dataset, comparing the performance of 3D-CNNs, SG-CNNs, and their use in combination (“fusion”) against docking scoring and MM/GBSA rescoring. Our results show that the fusion of the two models more accurately predicts protein-ligand binding affinity than docking scores or MM/GBSA rescoring with greater computational efficiency than the MM/GBSA method. In an effort to improve predictions, we have generated a large atomistic MD dataset of all the complexes in PDBbind 2019. Chemical and dynamic features across the entire dataset are computed for ML modeling and characterization of protein-ligand interactions. The synergistic use of ML and MD methods are ongoing and can improve the efficiency, accuracy, and scale of ML in drug discovery.

Informing deep learning methodologies for drug discovery with biophysical simulation

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.