We describe a novel approach to employing AI in drug discovery: using AI to help guide scientists' work.  Much the way that consumer systems suggest movies or music, our systems suggest related work, relevant publications, and interesting compounds within the context of a scientist's regular work activities.  Rather than a wholesale change in tactics, we seek to build upon existing data and systems to suggest fruitful new directions to our scientists.<br/><br/>In this presentation, we'll describe the development of five different recommender systems.  Chem Recommender monitors lab notebook entries to suggest similar syntheses.  NIBR Hops uses deposition of macromolecular crystal structures to recommend other 3D-similar molecules.  Assay Sommelier suggests the most relevant assays for a structure based upon a combined activity profile of nearest neighbors.  Transformer uses matched molecular pair analysis to prioritize transformations with higher confidence of improving some properties while limiting impact on other properties.  And Every Well Counts suggests other compounds to fill unused wells in assay plates.  Since recommenders in drug discovery are new, our approach has been to focus on high-value use cases--those that provide immediate benefit to our scientists and are well-supported by our existing systems.<br/><br/>Though common in consumer systems, there are challenges to deploying recommenders in a scientific enterprise.  How should recommendations be delivered?  How do we evaluate the efficacy of our recommendations?  What are the best algorithms for generating recommendations?  How do we combine recommendations with systems that never anticipated them?  These are challenging questions but ones we feel are worth investigating.  As the industry races to employ artificial intelligence to the challenges of drug discovery, we believe that using AI techniques to augment human work practices in the lab represents an untapped opportunity.

AI-based recommendations for drug discovery

ACS Fall 2020 Virtual Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.