Interpolation versus extrapolation in solubility prediction: Benchmarking machine learning against linear models for furosemide in mono-solvents

Following an earlier work reporting the experimental solubility and a number of computational results, to provide a complementary computational investigation, in this work a number of solute and solvent parameters were employed to provide correlative/predictive models for the solubility of furosemide in mono-solvents at various temperatures. The models include the Abraham, Hansen, KAT-LSER, Catalan and Laurence solubility approaches along with a multi-parameter model. The accuracy of the investigated models was assessed by average absolute percentage deviation (AA%D) where the best correlative analysis with the AA%D of 9.6% and predictive analysis employing one solubility datum in each solvent at room temperature with the 29.6% were obtained. In parallel, multiple linear regression (MLR), CatBoost, and TabPFN were evaluated using three validation strategies representing interpolation and extrapolation tasks: random five-fold cross-validation (5CV), leave-one-solvent-out cross-validation (LOSOCV), and temperature extrapolation validation (TEV). Under interpolation conditions, TabPFN and CatBoost outperformed MLR, yielding overall AA%D of 2.2%, 4.4%, and 10.7%, respectively. However, under more rigorous extrapolation scenarios – LOSOCV and TEV – MLR demonstrated superior or comparable predictive accuracy (AA%D = 23.2% and 22.5%, respectively) relative to CatBoost (27.7% and 44.7%) and TabPFN (26.2% and 44.9%), with notably narrower error distributions and fewer outliers.