Research map / Large language models / AI and society

Trustworthy Retrosynthesis: Eliminating Hallucinations with a Diverse Ensemble of Reaction Scorers

NeurIPS 2026 · Large language models › AI and society · approach: Data commons and human baselines

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RetroTrim, a retrosynthesis system that pairs a BART-based single-step generator with a MetaScorer ensemble (Reaction Prior, graph-based plausibility, and reference-reaction retrieval) to prune hallucinated reactions, evaluated via expert chemist labeling on 32 drug-like targets.

Problem

Single-step retrosynthesis generators produce hallucinated, chemically unsound reactions that invalidate multi-step synthetic routes, and automated metrics fail to reliably assess route plausibility.

Approach

Combine three complementary reaction scorers (a BART-based Reaction Prior with regioselectivity and reaction-center terms, a GAT graph plausibility classifier trained with template-generated negatives, and a retrieval-based precedent score) into a MetaScorer that filters reactions during Retro* search, plus a fine-grained expert labeling protocol for evaluation.

Key results

• RetroTrim is reported as the only evaluated system with zero Nonsense (hallucinated) pathways across the 32 targets while finding the most flawless routes.
• On reaction-level prediction, the MetaScorer achieves higher ROC-AUC and PR-AUC than the individual scorers (RP, RGP, RRS), with RP the strongest single scorer.
• RetroTrim won first place in the Standard Industries $1 million Retrosynthesis Challenge.

Surprising findings

• Individual scorers similar to those in the literature (RGP feasibility classifier and forward scorers) were not sufficient to achieve correct pathways.
• RetroChimera, despite ensembling generators, still produces a substantial number of Nonsense pathways.

Limitations

• Core training data (Pistachio) is proprietary, limiting reproducibility.
• Evaluation focuses on top-1 paths and drug-like targets; diversity of paths not analyzed.
• Expert labeling is time-consuming and relies on a limited number of chemists and targets.

Experimental setup

Datasets: Pistachio (2024Q3) eMolecules

Benchmarks: 32 novel drug-like retrosynthesis targets labeled reaction dataset of over 4,500 SSR-generated reactions

Metrics: expert-labeled path confidence (Safe bet, Worthwhile, Rather ROC-AUC PR-AUC false positive counts path coverage

Baselines: AiZynthFinder IBM RXN RetroChimera LocalRetro RootAligned without scoring RootAligned with forward scorer

Builds on: Retro* BART Graph Attention Network Molecular Transformer RetroChimera

Closest papers by meaning

• RetroOOD: Understanding Out-of-Distribution Generalization in Retrosynthesis Prediction (AAAI 2024 · similarity 0.89)
• URSA: Chemistry-Aware Benchmark for Utilitarian Retrosynthesis Assessment (NeurIPS 2026 · similarity 0.89)
• GDiffRetro: Retrosynthesis Prediction with Dual Graph Enhanced Molecular Representation and Diffusion Generation (AAAI 2025 · similarity 0.88)
• LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.88)
• Retro-fallback: retrosynthetic planning in an uncertain world (ICLR 2024 · similarity 0.88)
• RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.88)