Improving computer-aided retrosynthesis of chemically intuitive disconnections
- 1School of Chemistry
- 2University Park
- 3School of Computer Science, University of Nottingham, Jubilee Campus, Nottingham, United Kingdom, NG8 1BB
Abstract
Reliability in computer-assisted retrosynthesis relies on each architectural component. In template-based retrosynthesis, defining components include the precursor library, the trained model, and the templates. In this work, we evaluated open-source retrosynthesis tools (with default parameters), against a benchmark of 19 common disconnection strategies. An accompanying open-source dataset, template library, and trained model were developed to target simple chemistries. We investigated how precursor libraries, training data and network underfitting affect the reliability of retrosynthetic routes. A method for enumerating reaction SMIRKS into reaction SMILES is also described, providing a route to generate additional training data. The combination of an improved template library, retrained model, and expanded stock file increased the proportion of correctly solved simple case examples from 37% to 84%, outperforming all the other open-source tools. Analysis of the results showed that template presence alone is insufficient: successful retrosynthesis also depends on whether relevant templates are ranked within the search space and whether the search favours the expected route over shorter alternatives. These results highlight the importance of treating template coverage, model ranking, stock definition, and search behaviour as interacting components of retrosynthetic performance. The stock file, template library and model have been implemented into our open-source electronic lab notebook (ELN), AI4Green.
Related
- Synthesis of Arylpiperazine-Based CNS Ligands: Batch Disconnections and Their Continuous-Flow Implementation
- Advancement of triphos-Ru catalyzed C–O bond disconnections and its challenges
- RetroMol: Knowledge Distillation of Energy-Based Retrosynthesis Models
- Beyond Chemistry Knowledge: Embedding Pedagogical Content Knowledge through an AI-Augmented Retrosynthesis Program for Chemistry Education
- Steerable Single-step Retrosynthesis with Electron-pushing Discrete Flow Matching and Feynman-Kac Steering
BibTeX
@misc{blackshaw2026improving,
title = {{Improving computer-aided retrosynthesis of chemically intuitive disconnections}},
author = {Ton M. Blackshaw and Joseph Heeley and Kristian T. Spoerer and Jonathan D. Hirst},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010695/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010695/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Improving computer-aided retrosynthesis of chemically intuitive disconnections
AU - Ton M. Blackshaw
AU - Joseph Heeley
AU - Kristian T. Spoerer
AU - Jonathan D. Hirst
PY - 2026
DA - 2026/10/20
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010695/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010695/v1
AB - Reliability in computer-assisted retrosynthesis relies on each architectural component. In template-based retrosynthesis, defining components include the precursor library, the trained model, and the templates. In this work, we evaluated open-source retrosynthesis tools (with default parameters), against a benchmark of 19 common disconnection strategies. An accompanying open-source dataset, template library, and trained model were developed to target simple chemistries. We investigated how precursor libraries, training data and network underfitting affect the reliability of retrosynthetic routes. A method for enumerating reaction SMIRKS into reaction SMILES is also described, providing a route to generate additional training data. The combination of an improved template library, retrained model, and expanded stock file increased the proportion of correctly solved simple case examples from 37% to 84%, outperforming all the other open-source tools. Analysis of the results showed that template presence alone is insufficient: successful retrosynthesis also depends on whether relevant templates are ranked within the search space and whether the search favours the expected route over shorter alternatives. These results highlight the importance of treating template coverage, model ranking, stock definition, and search behaviour as interacting components of retrosynthetic performance. The stock file, template library and model have been implemented into our open-source electronic lab notebook (ELN), AI4Green.
ER -