Machine learning in drug discovery: what can go wrong? - Vikram Sundar

Part of the CHUtalks: Ideas. Research. Discoveries series from the Churchill College Postgraduate and Fellowship community.
Drug discovery today is an expensive and time-consuming process. In this talk, Vikram focuses on the first step: the identification of small molecule ligands that tightly bind to target proteins of interest. Experimental methods like high-throughput screening are prone to failure since they can only feasibly screen a small portion of chemical space.
Due to the emergence of large databases of screening data, machine learning approaches to this problem have become popular. However, many machine learning approaches that supposedly worked very well, in theory, have failed in practice.
Vikram outlines some of the more common machine learning approaches to this problem and explain some of the pitfalls that caused problems for these approaches.