Designing interpretable molecular property predictors

Abstract

Complex neural models often suffer from a lack of interpretability, i.e., they lack methodology for justifying their predictions. For example, while there have been many performance improvements in molecular property prediction, these advances have come in the form of black box models. As deep learning and chemistry are becoming increasingly intertwined, it is imperative that we continue to investigate interpretability of associated models. We propose a method to augment property predictors with extractive rationalization, where the model selects a subset of the input, or rationale, that it believes to be most relevant for the property of interest. These rationales serve as the model's explanations for its decisions. We show that our methodology can generate reasonable rationales while also maintaining predictive performance, and propose some future directions.