Something for nothing: improved solvation free energy prediction with Δ-learning

Molecular solubility is among the key properties that determine the clinical performance of a drug candidate because poor molecular solubility often indicates inadequate bioavailability. Using the CombiSolv-Exp database, we test several models (Gaussian process regression, decision trees, k-nearest neighbors) for hydration free energies by integrating Δ$$\Delta$$-learning and a universal quantum-chemistry continuum solvation model, SMD. The optimal model is Gaussian process regression with MAE of 0.63 kcal/mol. The reported models improve the accuracy of SMD, but have negligible additional computational cost.