Evaluating and adapting modelling strategies for data-driven prediction of solvent effects on reaction barriers

Predicting solvent effects on reaction activation barriers is central to understanding chemical reactivity and reaction kinetics, and guiding solvent selection. The solvent-induced change in activation free energy (DDG_solv‡) provides a quantitative descriptor of this effect, but remains costly to evaluate across vast reaction-solvent spaces, using quantum mechanical methods. Recent data-driven models have enabled prediction of solvent effects. However, most typically rely on two-dimensional representation of reactions and do not explicitly encode sufficient reaction context, such as transition-state information, or three-dimensional structural changes along the reaction, resulting in limited generalizability and predictive accuracy. In this study, systematic evaluation is presented of modelling strategies for predicting DDG_solv‡, with a focus on the role of reaction-state representation, input-geometry fidelity, and input modality. Using a large reaction-solvent dataset, models based on two-dimensional condensed reaction graphs are compared with models incorporating three-dimensional geometries of reactants, transition states, and products. The sensitivity of geometry-based models to structural accuracy is assessed by replacing quantum-chemically optimized transition states with structures predicted by a generative model. In addition, a dual-modality architecture combining two-dimensional graph-based and three-dimensional geometry-based representations is examined. The results show that explicit inclusion of both reactant and transition-state geometries leads to improved prediction accuracy relative to representations based on reaction endpoints or transition states alone. However, model performance depends strongly on the fidelity of the input geometries, with substantial degradation observed when low-quality structures are used. The dual-modality approach partially mitigates this sensitivity by adaptively reweighting two-dimensional and three-dimensional information, leading to performance recovery under low-fidelity conditions.