3D Convolutional Neural Network for Predicting Free Energies of Partitioning

Abstract

Accurate predictions of small molecule hydrophobicity and oil/water partitioning is a critical part of drug discovery. The hydrophobicity of a drug will determine how effectively it can cross the hydrophobic cellular membrane to reach their intracellular targets. In addition to traversing the many heterogenous chemical environments within a cell, hydrophobicity is also an important driving force for drug-protein binding. Traditionally, atomistic molecular dynamics simulations are used to calculate free energies of small molecules crossing lipid membranes and solvation. Machine learning models and empirical methods have also been used, but simulations are expensive and machine learning models trained on experimental data have limited domains of applicability. Here, we combine data generated from atomistic molecular dynamics simulations with machine learning to predict free energy of partitioning. We’ve simulated over 15,000 small molecules to calculate the relative free energy in three environments: water, at an interface, and in bulk hydrocarbon. The simulation data is fed into a 3D convolutional neural network to predict free energies for transfer. To ensure a large domain of applicability, we compare random splits of the data with scaffold splitting. Our trained models are tested on previously unseen, and reasonably different, set of data generated from a subset of world approved drugs. We compare our results against shallow models trained using chemical fingerprints. This work is a step towards integrating molecular dynamics simulations and machine learning, with many new avenues of research to pursue. This work performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. Release number: LLNL-ABS-792042.