Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis
Bogdan Zagribelnyy, Ivan Ilin, Nikita Bondarev, Maksim Kuznetsov, Mathieu Reymond, Vladimir Aladinskiy, Alex Aliper, Alex Zhavoronkov
cs.LG, cs.AI, cs.CE, cs.CL
2026-08-19
A 2.6B model trained Top-K on 45.6M reactions, then ChemCensor GRPO, hits Av. PT-Top-10 1.37 on URSA-expert-2026, ahead of LocalRetro at 1.22.
Single-step retrosynthesis (SSRS) is the inner loop of computer-aided synthesis planning: given a target molecule, propose reactions and precursors. The task is one-to-many. A product often admits several plausible disconnections. Standard Top-K accuracy still scores a single annotated answer, so diversity is treated as noise. Insilico's earlier URSA study, using ChemCensor as a plausibility proxy, found general LLMs promising but still behind conventional SSRS tools such as LocalRetro and MHNreact.
This paper puts the multi-answer protocol into both evaluation and training. It fine-tunes a 2.6B chemistry language model on tens of millions of verified reactions, then reinforces ChemCensor and novelty, and reports that the resulting model beats those conventional tools on an out-of-distribution expert set.
Two prompting modes. Top-1 wraps the product SMILES in 15 natural-language templates and asks for one reactant set each. Top-K appends "Give me 15 different answers" to those templates, samples 3 templates per target, scores each set, and averages. Conventional models emit 15 reactions per product via Syntheseus, so the comparison is aligned.
The training set, CREED-CCV-2+USPTO-XL, has 45,649,785 unique reactions over 3,680,906 unique products. CREED-CCV-2 enumerates reactions on ChEMBL compounds with a virtual synthesis engine, merges the previous CREED, and filters with ChemCensor v1.1.1 (36M reactions). USPTO-XL re-enumerates USPTO products to patch the public patent set's one-reaction-per-product bias (10.6M). The base is Liquid AI's LFM2 2.6B, SFT for 50,000 steps already in Top-K mode, with predicted ChemCensor order.
Reinforcement uses GRPO, group size 8, temperature 1, KL 0.1. The reward is a weighted sum: thinking format 0.1, valid SMILES 0.5, ChemCensor score 1.0 (rescaled from 0-5 to 0-1), Top-K uniqueness 0.2, count matching 0.1. A novelty term of weight 1.0 pays 1 when a reactant is absent from the exhaustive training list and ChemCensor is positive. The final checkpoint is C3LM-LFM2-RFT-CC-NR.
The headline benchmark is URSA-expert-2026: 100 expert-confirmed, publicly disjoint novel molecules. USPTO-50K-test-mini is a 497-target slice of USPTO-50K-test. ChemCensor's default precedent store is public USPTO-full.
On URSA-expert-2026 in Top-K mode, Av. PT-Top-K CC averages ChemCensor over the top-K unique predictions per target, then across targets.
| Method | Max | Top-3 | Top-5 | Top-10 |
| Gemini 3.1 Pro | 1.91 | 1.69 | 1.46 | 1.08 |
| GPT 5.5 | 1.94 | 1.68 | 1.46 | 1.05 |
| LocalRetro | 2.11 | 1.85 | 1.59 | 1.22 |
| MHNreact | 2.05 | 1.84 | 1.62 | 1.28 |
| C3LM SFT, Top-1 | 1.62 | 1.06 | 0.72 | 0.38 |
| C3LM SFT, large-set Top-K | 2.04 | 1.81 | 1.59 | 1.27 |
| C3LM RFT-CC-NR | 2.16 | 1.94 | 1.73 | 1.37 |
Switching the same data from Top-1 to Top-K training adds +0.30/+0.62/+0.70/+0.60 on Max/@3/@5/@10. Scaling the set more than 6x adds +0.12/+0.13/+0.17/+0.29. ChemCensor RL adds another 0.04; novelty another 0.08. RFT-CC-NR then leads the listed conventional models on this OOD set.
On USPTO-50K-test-mini, conventional Max scores sit near 4.8-4.9. The authors treat that as likely leakage: 481 of 497 products have at most 3 reactions in public USPTO-full, so Av. PT-Top-10 is the less contaminated column. RFT-CC-NR scores 1.85 there, just behind MHNreact at 1.90.
Uniqueness versus MHNreact is the more useful cut. Only the scaled C3LM (+0.4 unique plausible reactions per target) and RFT-CC-NR (+0.3) beat MHNreact on exclusive hits; Gemini 3.1 Pro is -2.7 and GPT 5.5 is -3.0. Pooling all 30 models pushes Av. PT-Top-10 to 1.81 against a best-single of 1.37, which is the argument for ensembles.
Two practical changes for synthesis-planning teams. When scoring LLMs on SSRS, stop ranking Top-1 single answers; asking for K distinct reactions reshuffles the leaderboard. Train in Top-K as well, and diversity metrics more than double. A 2.6B specialist with a plausibility reward can beat both frontier general models and template or graph SSRS tools on an expert OOD set. LLM and conventional proposals overlap only partly, so mixing them covers more chemistry.
The cost is explicit: both optimization and evaluation are tied to ChemCensor. Circularity is a design choice the authors own, not an accident.
Optimization and evaluation share ChemCensor, so some of the gain is fitting that proxy. The score ignores solvents, conditions, and purification. Precedents are mostly patents. Diversity is exact SMILES match, with no reaction-class or mechanism comparison. Training data comes from a virtual engine over about 3k expert templates, so the model inherits those patterns and will miss genuinely novel transforms.
URSA-expert-2026 has 100 molecules; OOD claims will be noisy. The novelty reward pays for "unseen in training and ChemCensor-positive", which can favor cuts the proxy likes and a chemist would not. The ethics section flags dual-use CASP risk and decontaminates against known hazardous structures, but the safeguard is policy language, not a technical gate.