DSAI seminar with Janosch Menke

Abstract

The field of drug discovery utilizes de novo design to create new molecules that possess specific desired properties, without relying solely on known compounds. This approach allows researchers to quickly generate ideas for new compounds, speeding up the drug discovery process and facilitating the development of innovative therapeutics.

A popular approach is using a user-defined scoring function within a reinforcement learning context, which evaluates the goodness of the generated molecules. However, these scoring functions are usually underspecified, as chemists struggle to translate their knowledge and expertise into a parametrized function. Consequently, the de novo agents produce suboptimal molecules that require extensive post-processing.

In this project, our objective is to enhance the quality of generated molecules by addressing the issue of underspecified scoring functions through human-in-the-loop reinforcement learning. We actively gather context-specific feedback from chemists regarding the generated molecules to elucidate their implicit knowledge. This information is then utilized to develop a user-model of the chemists, which can be integrated into existing de novo agents. The incorporation of this user-model aims to generate molecules that better align with the ideas and insights of a chemist.

About the speaker

Janosch is a Post Doc at Data Science and AI Division working on improving generative models for drug design. He obtained his Ph.D. at the University of Münster working on improving virtual screening for molecules by enhancing molecular representations.

This is a seminar from the DSAI seminars series usually held every Monday at 14:00 by the Data Science and AI division. The seminars are usually hybrid.