Abstract

Early approaches to molecular property prediction have largely relied on hand crafted features such as molecular fingerprints and physical-chemical descriptors. Although successful in many applications, these features can often fail when applied to difficult tasks and out-of-domain data. Inspired by work in other AI fields, the study of molecular representation learning has recently demonstrated the potential to overcome these challenges, resulting in more robust and generalisable modelling. In this presentation, we explore a novel method to obtain learned molecular representations. In particular, we leverage the large amount of publicly available quantum mechanical data to pre-train graph neural networks and use the resulting representations in downstream applications for ADMET predictive modelling. We explore a variety of approaches to exploit these representations, from combining them with classical features in classical models, to transfer-learning directly onto ADMET data. We also explore the pre-trained models in a multitask learning context. Overall, we see notable improvements in model performance on our internal and public data.