The logarithm of the octanol-water partition coefficient (logP) is an essential physiochemical property and helps determine the drug-likeness of a chemical compound. Here, we propose a method to predict the logP values, referred to as “ClassicalGSG.” This method employs neural networks (NNs) to train logP predictor models on molecular features that are generated using a recently developed technique called Geometric Scattering for Graphs (GSG). The GSG method transforms atomic attributes into symmetry-invariant molecular features using a graph representation of molecules where each atom has a set of attributes. Here, the atomic attributes are molecular mechanics force field parameters such as partial charges, Lennard-Jones well depth, Lennard-Jones radius, and atomic type. We examine the performance of the logP predictor models trained using atomic attributes generated either by small molecule force fields such as General AMBER Force Field (GAFF), CHARMM Generalized Force Field (CGenFF) or the Merck Molecular Force Field 94 (MMFF94). These are applied to a range of independent test sets, including predicting logP values for the SAMPL7 physical property challenge. For the latter example we find that ClassicalGSG outperformed all other entries, obtaining 1st, 3rd, 7th and 9th place using predictors obtained with different subsets of a master dataset. When used with open source force fields such as Merck Molecular Force Field 94 (MMFF94), ClassicalGSG provides an end-to-end predictor that uses SMILES strings as input. Pre-trained predictors, datasets and code to train new predictors are available at https://github.com/ADicksonLab/ClassicalGSG .

ClassicalGSG: Prediction of logP using molecular mechanics force field parameters and geometric scattering for graphs

ACS Spring Meeting 2021

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