Synthesis-constrained molecular design with direct optimization of reaction conditions

Victor Sabanza Gil, Jeff Guo, Jeremy Scott Luterbacher, Philippe Schwaller

Keywords: Synthesis-constrained generative molecular design, reaction condition prediction, language-models, reinforcement learning

TL;DR: Directly optimizing reaction conditions from synthetic routes of generated molecules in molecular design tasks.

Abstract: Generative molecular design promises to accelerate drug and materials discovery through the proposal of optimized candidates for given molecular properties. However, current models still struggle to incorporate key synthetic constraints like reaction conditions in their designs, which hampers experimental validation. In this work, we show it is possible to directly optimize reaction conditions as part of the molecular optimization process. We utilize a reaction condition prediction model to annotate the retrosynthetic pathways of generated molecules. The generative model is then rewarded based on the alignment of these predicted conditions with user-defined constraints. We demonstrate the ability to avoid restricted solvents such as dimethylformamide (DMF) and dichloromethane (DCM) or enforce the presence of desired solvents like water compared to experiments without reaction condition constraints. Our results indicate that the model learns a distribution shift toward reaction classes compatible with these constraints without compromising property optimization. Finally, we show the utility of the framework in an in silico experiment where we design potential fungicides following green chemistry principles. Overall, our contribution helps to bring molecular design closer to laboratory reality and accelerate the validation of proposed molecules.

Submission Track: Full Paper

Submission Category: AI-Guided Design + Automated Synthesis

Submission Number: 43