A SIMILARITY-AGNOSTIC REINFORCEMENT LEARNING APPROACH FOR LEAD OPTIMIZATION

Abhor Gupta, Sean Current, Balaraman Ravindran, Rohit Batra, srinivasan parthasarathy, Karthik Raman

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Drug Design, Molecular Optimisation, Reinforcement Learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: A new approach for lead optimization (molecular optimization) using goal-conditioned reinforcement learning.

Abstract: Lead optimization in drug discovery is a pivotal phase in identifying promising drug candidates for further development. Traditionally, lead optimization in the machine learning community has been treated as a constraint optimization problem where methods like generative models and reinforcement learning(RL) have been widely employed. However, these methods often rely on molecular similarity metrics to define constraints, which poses significant challenges due to the inherently ambiguous nature of molecular similarity. In this work, we present a similarity-agnostic approach to lead optimization, which we term "Lead Optimization using Goal-conditioned Reinforcement Learning" or LOGRL. Contrary to conventional methods, LOGRL is uniquely trained on a distinct task: source-to-target path prediction. This allows LOGRL to produce molecules with significantly higher Tanimoto similarity to target molecules, even without direct exposure to this metric during training. Furthermore, we incorporate a beam search strategy during the molecule generation process. This strategy empowers us to generate a substantial number of candidate molecules, facilitating further curation to meet desired properties. Notably, our unique approach permits us to leverage the Euclidean distance between learned action representations as a surrogate for molecular similarity during beam search.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8186