Can Self-supervised Models Make a Difference in Real-world Drug Discovery?

Bharath Ramsundar , Deep Forest Sciences ABSTRACT: When starting a drug discovery project, there often isn’t much data available. How can we bootstrap machine learning approaches to drug discovery in the presence of limited data? Self-supervised approaches provide a systematic methodology to solve the low data problem in real world drug discovery. I will explore a variety of different self-supervised strategies to learn in the absence of experimental data, including some very recent work leveraging large language models such as GPT3 for chemistry. I will share some results from past and on-going work at Deep Forest Sciences working to extend self-supervision to real-world drug discovery.