<br/>In recent years, graph neural networks have proven to be highly successful for molecular property prediction, finding utility in multiple drug discovery contexts. However, it remains difficult for the majority of cheminformatics and machine learning scientists to leverage SOTA graph networks given that most advances stem from highly specialized research in graph theory. We aim to address this challenge with our GOLI library. We implement the most recent and high-performing graph models in a library that is both easy to use and flexible, while also incorporating rich atomic and bond properties rather than simple one-hot encodings. GOLI was built with a strong focus on molecular tasks, ensuring that the library works out-of-the-box with any SMILES dataset and comes readily compatible with common cheminformatics datasets such as Tox21, HIV inhibition, PCBA, and a subset of ZINC. Custom datasets can be added with only minor changes to a configuration file. GOLI further enables pre-training models on large datasets to facilitate transfer learning between tasks. By making GOLI open source, we hope to empower scientists working at the intersection of machine learning and drug discovery by making it easier to write, train, and tune highly performative, graph-based models.

GOLI: An open graph library for molecular property prediction

ACS Fall 2021

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