arxiv:2608.18940

Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis

Published on Aug 19

· Submitted by

Maksim Kuznetsov on Aug 20

Insilico Medicine

Upvote

Authors:

Maksim Kuznetsov ,

Abstract

Top-K prompting and plausibility-aware training improve diverse reaction prediction in single-step retrosynthesis, yielding state-of-the-art results on a large verified reaction dataset and motivating ensemble systems.

Generated by thinkingmachines/Inkling-Small

Single-step retrosynthesis is a central component of computer-aided synthesis planning, yet its intrinsically one-to-many nature is poorly captured by single-answer evaluation and benchmarking protocols. To address this, we introduce Top-K prompting as a robust training and inference paradigm to better capture diverse, plausible reaction predictions. We compile CREED-CCV-2+USPTO-XL, an ultra-large-scale dataset of ~45.6 million verified reactions to train the C3LM (Chemistry Constraint-Consistent Language Model). By integrating fine-tuning with ChemCensor-based and novelty-oriented rewards, our model achieves state-of-the-art performance on the OOD URSA-expert-2026 benchmark. Further analysis of reaction uniqueness shows that LLMs and conventional models explore complementary reaction spaces, motivating ensemble-based retrosynthesis systems. Overall, our results establish Top-K, plausibility-aware training as a practical new direction for robust future LLM-based synthesis planning.

View arXiv page View PDF Add to collection

Community

maksimkuznetsov

Paper author Paper submitter 1 day ago

This work extends the previously introduced ChemCensor framework for single-step retrosynthesis (https://arxiv.org/abs/2602.03554). It addresses the one-to-many nature of retrosynthetic prediction, where a target molecule may have several chemically valid synthetic routes, making single-answer and exact-match evaluation insufficient. The study introduces Top‑K prompting and training to generate multiple diverse reaction proposals rather than a single predicted disconnection.

A new C3LM version is trained on CREED-CCV-2+USPTO-XL, an expanded dataset of approximately 45.6 million verified reactions derived from expert-coded templates. Combined with the Top‑K strategy, this larger training set enables C3LM to compete with - and in some settings surpass - strong conventional, non-LLM retrosynthesis models on the URSA-expert-2026 benchmark.

Moreover, the study shows that the chemical spaces of reactions generated by the new C3LM and the best conventional models are complementary rather than intersecting, making the contribution of the new model more substantial if the model is harnessed into a synthetic planning engine.