Molecular Graph Representation Learning and Generation for Drug Discovery

Abstract

Machine learning methods have been widely pervasive in the domain of drug discovery, enabling more powerful and efficient models. Before deep models, modeling molecules was largely driven by expert knowledge; and to represent the complexities of the molecular landscape, these hand-engineered rules prove insufficient. Deep learning models are powerful because they learn the important statistical features of the problem–but only with the correct inductive biases. We tackle this important problem in the context of two molecular problems: representation and generation. Canonical success of deep learning is deeply rooted in its ability to map the input domain into a meaningful representation space. This is especially poignant for molecular problems, where the “right” relations between molecules is nuanced and complex.

The first part of this thesis will focus on molecular representation, in particular, property and reaction prediction. Here, we explore a transformer-style architecture for molecular representation, providing new tools to apply these models to graph-structured objects. Moving away from the traditional graph neural network paradigm, we demonstrate the efficacy of prototype networks for molecular representation, which allows us to reason over learned property prototypes of molecules. Lastly, we look at the molecular representations in the context of improving reaction predictions.

The second part of this thesis will focus on molecular generation, which is crucial in drug discovery as a means to propose promising drug candidates. Here we develop a new method for multi-property molecule generation, by first learning a distributional vocabulary over molecular fragments. Then, using this vocabulary, we survey efficient exploration methods over the chemical space.