Trading-off Multiple Properties for Molecular Optimization
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
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: Molecular Optimization, Multiple Properties
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Molecular optimization, a critical research area in drug discovery, aims to enhance the properties or performance of molecules through systematic modifications of their chemical structures. Recently, existing Multi-Objective Molecular Optimization (MOMO) methods are extended from Single-Objective Molecular Optimization (SOMO) approaches by employing techniques such as Linear Scalarization, Evolutionary Algorithms, and Multi-Objective Bayesian Optimization. In Multi-Objective Optimization, the ideal goal is to find Pareto optimal solutions over different preferences, which indicate the importance of different objectives. However, these straightforward extensions often struggle with trading off multiple properties due to the conflicting or correlated nature of certain properties. More specifically, current MOMO methods derived from SOMO are still challenged in finding preference-conditioned Pareto solutions and exhibit low efficiency in Pareto search. To address the aforementioned problems, we propose the \textbf{P}reference-\textbf{C}onditioned \textbf{I}nversion (PCI) framework, efficiently ``inverting'' a pre-trained surrogate oracle under the guidance of a non-dominated gradient, to generate candidate Pareto optimal molecules over preference-conditioned distributions. Additionally, we provide theoretical guarantees for PCI's capability in converging to preference-conditioned solutions. This unique characteristic enables PCI to search the full Pareto front approximately, thereby assisting in the discovery of diverse molecules with varying ratios of properties. Comprehensive experimental evaluations show that our model significantly outperforms state-of-the-art baselines in multi-objective molecular optimization settings.
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: 3745
Loading