Machine learning of free energies in chemical compound space using ensemble representations: Reaching experimental uncertainty for solvation

Free energies govern the behavior of soft and liquid matter, improving their predictions could have a large impact on development drugs, electrolytes or homogeneous catalysts. Unfortunately, it is challenging to devise an accurate description effects governing solvation such as hydrogen-bonding, van der Waals interactions, conformational sampling. We present energy Machine Learning (FML) model applicable throughout chemical compound space based representation that employs Boltzmann averages account for approximated sampling configurational space. Using FreeSolv database, FML's out-of-sample prediction errors experimental hydration free decay systematically with training set size, uncertainty (0.6 kcal/mol) reached after 490 molecules (80\% FreeSolv). Corresponding FML are also par state-of-the art physics approaches. To generate input new query compound, requires approximate short molecular dynamics runs. showcase its usefulness through analysis 116k organic (all force-field compatible in QM9 database) identifying most least solvated systems, rediscovering quasi-linear structure property relationships terms simple descriptors hydrogen-bond donors, number NH OH groups, oxygen atoms hydrocarbons, heavy atoms. accuracy maximal when temperature used simulation averaged samples same compounds. The time converges rapidly respect error.