ABSTRACT:BIO

Machine learning has become an indispensable aid to researchers developing the next generation of novel therapeutics. In this talk, I will discuss how some of the most important problems  in virtual screening for new potential drug molecules can be cast as black-box optimization problems, where the goal is to find molecules maximizing some desired property — for example, the binding affinity to a known drug target. By leveraging recent work on representation learning for molecules and high dimensional black-box optimization, we are able to achieve up to a 20x performance improvement over state of the art on several of the most widely used benchmarks for molecule design. I will then show how this powerful new approach reveals flaws in tools commonly used for computational molecule design. Even the most widely used docking simulators can be fooled by a sufficiently powerful optimizer producing molecules that could not plausibly exist in nature — a challenge reminiscent of adversarial image generation in computer vision. These findings can be mitigated to a degree through the use of constrained optimization, but also motivate adapting lessons from robust machine learning to the docking simulators themselves.