Hydration free energies from kernel-based machine learning: Compound-database bias

We consider the prediction of a basic thermodynamic property - hydration free energies - across a large subset of the chemical space of small organic molecules. Our in silico study is based on computer simulations at the atomistic level with implicit solvent. We report on a kernel-based machine learning approach that is inspired by recent work in learning electronic properties but differs in key aspects: The representation is averaged over several conformers to account for the statistical ensemble. We also include an atomic-decomposition ansatz, which offers significant added transferability compared to molecular learning. Finally, we explore the existence of severe biases from databases of experimental compounds. By performing a combination of dimensionality reduction and cross-learning models, we show that the rate of learning depends significantly on the breadth and variety of the training dataset. Our study highlights the dangers of fitting machine-learning models to databases of a narrow chemical range.