Supervised Source-Task Pretraining for Low-Data Electrochemical Molecular Property Prediction

Electrochemical molecular property data obtained from experiments or high-level quantum chemical calculations are often limited in size, which constrains the direct use of graph neural networks for predicting oxidation potentials and redox-reaction-related energies. Here, we investigate supervised source-task pretraining as a candidate strategy for low-data molecular prediction by using QM9 as the source domain, pretraining a GINE molecular graph encoder on HOMO, LUMO, HOMO-LUMO gap, and u0, and transferring the resulting encoders to OxPot oxidation-potential prediction and RedDB reaction-energy-related prediction. Across random and scaffold splits and multiple downstream labeleddata budgets, pretraining improved performance only under specific conditions: OxPot benefited more consistently from HOMO and LUMO source tasks, whereas RedDB showed a more dispersed transfer pattern in which gap and u0 could also be competitive. Pretraining-scale and pretraining-seed analyses further supported a sample-size-dependent stage hypothesis in which negative transfer is more likely at extreme-low-data budgets, a positive-transfer window can emerge at low-to-mid data budgets, and gains may diminish or become negative again at higher budgets depending on source-task alignment, downstream difficulty, split difficulty, and source-domain scale.