RetroMol: Knowledge Distillation of Energy-Based Retrosynthesis Models
4 affiliations
- 1Dept. of Chemistry & Chemical Biology
- 2McMaster University
- 3Dept. of Applied Science & Engineering
- 4University of Toronto
Abstract
Retrosynthesis prediction is a central task in computer-aided synthesis planning, and synthesizability remains a persistent bottleneck in AI-driven molecular and materials discovery. We introduce RetroMol, a family of compact text-based models that transfers an energy-based critic’s ranking of candidate reactant sets into the generator itself. RetroBase, an 11M-parameter transformer encoderdecoder trained on two-reactant USPTO reactions, achieves Top-1 and Top-5 accuracies of 46.8% and 67.8% on a held-out test split. RetroERM, an energybased model over molecular fingerprint counts, scores RetroBase’s candidates, and its ranking preferences are distilled back into the generator. The distilled model, RetroDistil, reaches 52.6% Top-1 and 73.5% Top-5 under identical data and evaluation, improving on RetroBase by 5.7–6.8 points at Top-1/3/5 with no reranking stage at inference and no increase in model size. Distilling a critic is therefore a practical way to improve a compact retrosynthesis generator without larger or more complicated architectures.
Related
- Beyond Chemistry Knowledge: Embedding Pedagogical Content Knowledge through an AI-Augmented Retrosynthesis Program for Chemistry Education
- Analysis of fine-tuned foundation interatomic potentials and their knowledge distillation into a cluster expansion for Ni–Pt ordering thermodynamics
- Chemist-aligned retrosynthesis by ensembling diverse inductive bias models
- Improving computer-aided retrosynthesis of chemically intuitive disconnections
- Steerable Single-step Retrosynthesis with Electron-pushing Discrete Flow Matching and Feynman-Kac Steering
BibTeX
@misc{sondhi2026retromol,
title = {{RetroMol: Knowledge Distillation of Energy-Based Retrosynthesis Models}},
author = {Rudra Sondhi and Edwin Chacko and Rodrigo A. Vargas-Hernández},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010571/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010571/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - RetroMol: Knowledge Distillation of Energy-Based Retrosynthesis Models
AU - Rudra Sondhi
AU - Edwin Chacko
AU - Rodrigo A. Vargas-Hernández
PY - 2026
DA - 2026/10/16
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010571/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010571/v1
AB - Retrosynthesis prediction is a central task in computer-aided synthesis planning, and synthesizability remains a persistent bottleneck in AI-driven molecular and materials discovery. We introduce RetroMol, a family of compact text-based models that transfers an energy-based critic’s ranking of candidate reactant sets into the generator itself. RetroBase, an 11M-parameter transformer encoderdecoder trained on two-reactant USPTO reactions, achieves Top-1 and Top-5 accuracies of 46.8% and 67.8% on a held-out test split. RetroERM, an energybased model over molecular fingerprint counts, scores RetroBase’s candidates, and its ranking preferences are distilled back into the generator. The distilled model, RetroDistil, reaches 52.6% Top-1 and 73.5% Top-5 under identical data and evaluation, improving on RetroBase by 5.7–6.8 points at Top-1/3/5 with no reranking stage at inference and no increase in model size. Distilling a critic is therefore a practical way to improve a compact retrosynthesis generator without larger or more complicated architectures.
ER -