AqSolDB Aqueous Solubility QSAR Model

What this research found

Aqueous solubility caps how much of an oral drug dose can dissolve, yet it is hard to predict from chemical structure alone. K-Dense built a quantitative structure-property relationship model on AqSolDB, a curated set of 9,982 compounds with measured solubility, and deliberately tested it on chemical scaffolds it had never seen. A Random Forest reached an R-squared of 0.765 on that harder out-of-distribution split versus 0.820 on a conventional random split, quantifying how much apparent accuracy is inflated when close analogs leak into training.

• On 1,844 held-out compounds with scaffolds absent from training, the Random Forest reached RMSE 1.192, MAE 0.841, and R-squared 0.765, beating a Ridge regression baseline on every metric (RMSE 1.420, MAE 1.064, R-squared 0.666).
• The scaffold split is uniformly harder than a random split. Random Forest RMSE rises from 0.990 to 1.192 and R-squared falls from 0.820 to 0.765, so reporting only the random-split number overstates prospective accuracy on novel chemistry.
• The best model places 70.0% of scaffold-held-out compounds within one log unit of the measured value, against 58.2% for Ridge.
• Two independently trained models agree on what drives poor solubility: lipophilicity and molecular size. Calculated octanol-water partition coefficient is the top Random Forest feature (importance 0.084), while molecular surface area carries the largest negative Ridge coefficient (-0.98), followed by lipophilicity (-0.78) and molecular weight (-0.73).
• A specific fragment effect was validated without the model: compounds containing an extended all-carbon chain motif (present in 645 of the 9,220 compounds) are on average 1.52 log units, roughly 30-fold, less soluble than those without it.
• Polar and hydrogen-bonding descriptors correctly carry positive weight — hydrogen-bond acceptors +0.31, donors +0.17, polar surface area +0.03 — indicating the model learned physically meaningful chemistry rather than spurious correlations.

How it was done

AqSolDB version 1.0 was retrieved from the Harvard Dataverse and verified by checksum, then curated: RDKit structure standardization, de-duplication by standard InChIKey with conflicting values averaged, and removal of 124 discordant compounds whose duplicate measurements disagreed by more than 1.5 log units, leaving 9,220 compounds. Each molecule became a 2,066-dimensional vector combining a 2,048-bit Morgan fingerprint of radius 2 with 18 RDKit 2D physicochemical descriptors, standardized using training-set statistics only. Ridge regression and a Random Forest were tuned by five-fold cross-validation and evaluated under two 80/20 partitions: a Bemis-Murcko scaffold split with zero scaffold overlap, and a random split as an in-distribution reference. Model interpretation combined signed Ridge coefficients, Random Forest impurity importances, and model-free comparisons of mean solubility for compounds with and without each candidate fragment.

Data sources

• AqSolDB v1.0, Harvard Dataverse doi:10.7910/DVN/OVHAW8, released 2019-07-18 — 9,982 compounds with experimental LogS
• Sorkun, Khetan & Er, Scientific Data 6:143 (2019) — the AqSolDB descriptor paper
• Delaney, ESOL, Journal of Chemical Information and Computer Sciences 44:1000 (2004)
• Yalkowsky & Valvani General Solubility Equation, Journal of Pharmaceutical Sciences 69:912 (1980)
• Wu et al., MoleculeNet, Chemical Science 9:513 (2018) — source of the scaffold-split convention