Cross-Radical Knowledge Sharing via Graph Neural Networks for Unified Prediction of Aqueous Organic Contaminant Second-Order Radical Reaction Rate Constants

Abstract

Abstract Selecting among hydroxyl (HO•), sulfate (SO4•–), and carbonate (CO3•–) radical advanced oxidation processes requires reliable intrinsic second-order rate constants, yet the corresponding curated data sets contain 1250, 493, and 248 records, respectively. Reframing this imbalance as a cross-radical few-shot learning problem, we introduce maco, a radical-conditioned graph neural network that uses shared molecular representation learning, oxidant/pH cross-attention, residual adapters, and a two-stage curriculum to transfer structural knowledge from the data-rich HO• task to the lower-data SO4•– and CO3•– tasks. On similarity-stratified held-out tests, maco achieved R2 = 0.855 (95% CI 0.711–0.906) for SO4•– and 0.815 (0.269–0.944) for CO3•–; on the more stringent scaffold-disjoint Murcko tests, the corresponding values were 0.652 (0.447–0.779) and 0.715 (0.500–0.825). Temperature-conditioned sensitivity analyses showed no statistically discernible generalization gain under the highly asymmetric temperature coverage. Functional-group attention analysis identified radical- and pH-dependent reactive-site patterns, while applicability-domain assessment and ensemble uncertainty quantified prediction reliability for target-radical scaffold-novel contaminants. maco provides a reproducible screening input for prioritising radical–contaminant measurements and process-specific evaluation. A browser-accessible WebUI further provides single-compound predictions, uncertainty estimates, and applicability-domain flags without local installation.