Message passing neural networks for molecular property prediction

Abstract

Developing new drugs relies heavily on understanding the various molecular properties of potential drug candidates. While experimental assays performed in the lab are the best source of information about molecular properties, these assays are slow and expensive. For this reason, there has been great interest in the potential of machine learning models to predict molecular properties without the need for experimental assays. However, recent literature has not yet clearly determined which machine learning models are optimal for molecular property prediction. In this thesis, I apply the Direct Message Passing Neural Network (D-MPNN) from [47, 48] to 19 publicly available property prediction datasets, and I demonstrate that it consistently outperforms prior machine learning models. Additionally, I introduce several optimizations to the D-MPNN which further enhance its performance and lead to new state-of-the-art results.