The increasingly pervasive role of AI/ML in chemistry has introduced new (and reintroduced old) opportunities to design new drug candidates, plan synthetic routes, and elucidate the structures of molecules from spectral data. While chemistry is unlikely to be immune to Sutton’s Bitter Lesson in the limit of infinite data, many of our methods must cope with the imperfections and finite sizes of experimental datasets. I will describe some of the ways that we have sought to integrate molecular intuition and an understanding of chemistry into the deep learning models we use for these various tasks.
