AqSol-Blend: A CPU-Trained Descriptor Ensemble for Aqueous Solubility Prediction

AqSol-Blend investigates molecular representation changes within a reproducible CPU-trained ensemble for aqueous solubility prediction. The study combines controlled development comparisons, an overlap-screened external evaluation and fresh-environment training reproduction. The method integrates LightGBM, Mordredbased CatBoost and robust-scaled support-vector regression with fixed weights of 0.675, 0.225 and 0.100. On the Therapeutics Data Commons AqSolDB benchmark, the ensemble achieves a mean absolute error of 0.736 ± 0.014 across five runs. Substituting Mordred descriptors for the CatBoost branch’s 2,937 fingerprint and RDKit features reduces development MAE by 0.0078 and improves nine of ten splits. In contrast, seed averaging, alternative CatBoost losses and reduced LightGBM capacity do not improve the development baseline. On an overlap-screened external cohort of 360 compounds from BigSolDB, the frozen five-run average obtains MAE 1.008, compared with 1.017 for its predecessor and 1.065 for a fixed random-forest baseline. The external improvement over the predecessor remains uncertain. Development reused validation partitions and included four sequential benchmark test evaluations; external results are interpreted separately because measurement conditions and concentration conventions differ.