Invited talk from Peter Frazier.
Title: Bayesian Optimization in Biochemistry
Abstract: Excitement is growing in the chemical sciences about Bayesian optimization and other methods for AI-based adaptive experimentation for accelerating the design of drugs, materials, and chemical formulations. Query-efficient methods like BayesOpt seem well-suited to these problems because measuring chemical and biological properties is often time-consuming and/or expensive, because they can incorporate scientific expertise into Bayesian prior distributions, and because the automation they provide allows getting the full benefits of robotics and high-throughput experimentation.
At the same time, barriers remain: designing a drug, material or consumer product is a multi-stage process involving many goals and evaluation methods, not a single optimization problem; poorly performing recommended molecules can destroy trust; training data often lacks negative examples; humans are hard to beat; experimentalists can have unrealistic expectations; and disciplinary differences in approach, expertise, and language can hinder collaboration.
The talk will focus on the speaker's experience from 3 biochemical design problems in which experimentalists performed wet lab experiments based on algorithmic recommendations from a BayesOpt or other AI algorithm. Drawing on these experiences, we suggest approaches for overcoming these challenges, including grey-box BayesOpt methods, methods for BayesOpt with preference learning, and human-focused approaches supporting effective collaboration with chemists.