Radical-Net: A chemistry-enhanced transformer for elementary radical reactions in pollutant chemistry

Radical reactions are central to environmental remediation and pollutant degradation, yet their prediction remains challenging due to complex mechanisms and transient intermediates. Traditional approaches suffer from limited generalizability (template-based) or lack of interpretability (pure machine learning). Here, we introduce Radical-Net, a T5 transformer-based network for predicting elementary radical reactions. Trained on > 5300 curated elementary reactions from RMechDB, achieving 74.53 % top-1 and 82.85 % top-5 accuracy on test datasets, outperforming vanilla T5 by + 2.5 % via a novel chemically informed scoring system that fuses AI outputs with atomic balance, charge conservation, and validity checks. We developed a chemistry-aware Byte Pair Encoding to retain key substructures (e.g., radicals and functional groups) and applied gradient-based attribution for mechanistic insights, pinpointing reactive sites and electron flows. UMAP dimensionality reduction and minimum spanning tree mapping of the reaction space revealed clustered patterns by reaction type, facilitating pathway exploration. Spatial analysis of the reaction embedding topology shows that prediction confidence correlates with local density and k-NN distances, enabling uncertainty quantification without ensemble methods. Demonstrated on hydroxyl radical additions to alkenes and perfluorinated pollutant transformations, Radical-Net identifies multiple pathways and interpretable mechanisms. This framework bridges AI with chemical principles, accelerating predictions of radical-mediated pollutant fate and novel remediation strategies in hazardous environments.