alchemy: systematic enumeration and thermodynamic screening of small-molecule radical reaction networks
Abstract
Automated construction of chemical reaction networks is limited by the combinatorial growth of possible molecules, intermediates, and formal transformations. We present alchemy, an open-source Python workflow for systematic enumeration and screening of small-molecule radical reaction networks. Within an explicitly bounded graph space, alchemy generates connected, neutral C/N/H molecular graphs with single, double, or triple bonds and specified radical counts; RDKit sanitisation and canonicalisation then remove graph-invalid and duplicate records. Reaction candidates are constructed without learned reaction templates by combining bond-dissociation records into rearrangement and transfer stoichiometries. Parallel generation, disk-backed storage, machine-learning prediction of molecular Gibbs free energies, and Kuzu graph representation make the resulting formal network searchable. In a trimethylamine case study limited to at most five heavy atoms and one unpaired electron per reaction side, the workflow generated 41,735 unique molecular graphs and 60,035 first-generation radical-transfer records. A PaiNN model trained and tested on OpenBabel-generated QM9 geometries supplies rapid ranking scores, but QM9 contains no open-shell species and therefore cannot validate predictions for the radical chemistry enumerated here. alchemy therefore provides a reproducible route to efficient reaction-space enumeration while retaining practical filters for chemical plausibility and downstream mechanistic analysis.
Related
- Optimal Control of Chemical Reaction Networks with Thermodynamic Cost
- Thermodynamic Space of Chemical Reaction Networks
- High-throughput thermodynamic screening of oxide-scale adhesion across the CoCrFeMnNiAl high-entropy alloys
- Generative Design and Graph Neural Network Screening of Novel Bimetallic Alloy Catalysts for the Hydrogen Evolution Reaction
- Intractable enumeration problems are like Russian nesting dolls: structural properties of monomer-dimer coverings on two-dimensional quadratic lattices
BibTeX
@misc{vu2026alchemy,
title = {{alchemy: systematic enumeration and thermodynamic screening of small-molecule radical reaction networks}},
author = {Le Minh Vu and Priya Sharmesh and Mega Kar and Truyen Tran and Sherif Abdulkader Tawfik},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008558/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008558/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - alchemy: systematic enumeration and thermodynamic screening of small-molecule radical reaction networks
AU - Le Minh Vu
AU - Priya Sharmesh
AU - Mega Kar
AU - Truyen Tran
AU - Sherif Abdulkader Tawfik
PY - 2026
DA - 2026/09/09
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008558/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008558/v1
AB - Automated construction of chemical reaction networks is limited by the combinatorial growth of possible molecules, intermediates, and formal transformations. We present alchemy, an open-source Python workflow for systematic enumeration and screening of small-molecule radical reaction networks. Within an explicitly bounded graph space, alchemy generates connected, neutral C/N/H molecular graphs with single, double, or triple bonds and specified radical counts; RDKit sanitisation and canonicalisation then remove graph-invalid and duplicate records. Reaction candidates are constructed without learned reaction templates by combining bond-dissociation records into rearrangement and transfer stoichiometries. Parallel generation, disk-backed storage, machine-learning prediction of molecular Gibbs free energies, and Kuzu graph representation make the resulting formal network searchable. In a trimethylamine case study limited to at most five heavy atoms and one unpaired electron per reaction side, the workflow generated 41,735 unique molecular graphs and 60,035 first-generation radical-transfer records. A PaiNN model trained and tested on OpenBabel-generated QM9 geometries supplies rapid ranking scores, but QM9 contains no open-shell species and therefore cannot validate predictions for the radical chemistry enumerated here. alchemy therefore provides a reproducible route to efficient reaction-space enumeration while retaining practical filters for chemical plausibility and downstream mechanistic analysis.
ER -