Chemical Language Models for Prediction of Radiolytic Transformation Products from Molecular Structure
Assessing the fate of organic compounds under ionizing radiation requires knowing which degradation products form, yet identifying them experimentally is slow and costly. We ask how well machine-learning models can generate those products from a parent compound’s structure alone. Because reported product distributions depend on irradiation conditions that studies record inconsistently, we target a condition-agnostic screening task: recovering the set of products reported for a parent under any condition. We compile a curated benchmark of 446 experimentally reported reactant–product pairs across 128 parent compounds from the primary radiolysis and photolysis literature, and cast prediction as one-to-many SMILES generation under reagent-level repeated cross-validation. GT4SD achieved the highest mean Hit@10 among the models evaluated under shared cross-validation, recovering at least one reported product for 59.4% of parent compounds, compared with 41.1% for nearest-neighbor retrieval. These results demonstrate the potential of chemistry-pretrained language models to generate radiolytic transformation-product candidates for experimental follow-up. Open-weight models further enable local deployment and control over molecular data, supporting their practical use in radiolysis research.
Authors
- Mantas Vaškevičius (ORCID: https://orcid.org/0000-0002-7130-9465)
- Dominykas Milašius (ORCID: https://orcid.org/0009-0003-9982-309X)
Institutions
- Vytautas Magnus University (LT)
Publication Details
- Journal
- Journal of Chemical Information and Modeling
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1021/acs.jcim.6c03614
- Primary Topic
- Computational Drug Discovery Methods
- Type
- article
- Field-Weighted Citation Impact
- 0.00