Uncertainty-Aware Deep Reinforcement Learning Approach for Computational Molecular Design

Designing molecules with tailored properties is crucial for the development of functionally superior and environmentally sustainable chemical products and processes. However, existing computational screening methods often overlook the crucial impact of property uncertainties in computational molecular design (CMD), potentially leading to solutions that violate design constraints and underperform in practice. Herein, we introduce a novel deep reinforcement learning (DRL) framework, DRL-CMD, that uniquely integrates graph-based fragment representation and property uncertainty quantification to systematically and confidently navigate the vast chemical space. To assess its efficacy, DRL-CMD is evaluated against literature-reported candidates in four molecular design cases, aiming to generate superior molecules with reduced uncertainties in molecular properties. DRL-CMD substantially improves molecular design outcomes, reducing constraint violations by an average of 39% and uncertainty margins by 27% compared with literature-reported molecules. Such improvements are particularly notable in complex design cases with limited data sets and extreme property ranges. These findings underscore DRL-CMD’s potential to accelerate the development of multiscale chemical products and processes by explicitly quantifying and managing uncertainties.