Beyond Search: Direct Model Guidance for Steerable Synthesis Planning

Najwa Laabid, Vikas K Garg

Keywords: retrosynthesis, synthesis planning, steerable chemical synthesis, classifier-guidance, sequence-to-sequence models

TL;DR: We apply classifier guidance to a seq2seq retrosynthesis model to steer synthesis planning toward desired properties, enabling exploration of new chemical spaces without retraining.

Abstract:

Synthetic chemists often need to incorporate domain-specific constraints and preferences into synthesis plans,
such as preferred reaction types or available starting materials, motivating the development of steerable synthesis planning methods.
While previous approaches navigate the search space with a frozen single-step model for plausible multi-step routes,
we demonstrate that directly guiding the single-step retrosynthesis model enables exploration of previously inaccessible chemical
spaces during generation. Specifically, we invoke guidance to modify the logits of an autoregressive seq2seq retrosynthesis model,
enabling conditioning on various properties without retraining.
Empirically, we demonstrate that while commonly used single-step models struggle to find routes with chemically feasible single-step
reactions throughout the entire synthesis plan, our method generates synthesis routes of equal or better quality than
template-based approaches while satisfying the specified constraints.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 2759