Solubility prediction in water and organic solvents through a combination of chemometrics and computational chemistry

Accurate solubility prediction is crucial across a range of scientific disciplines including drug
discovery, protein engineering, drug and agrochemical process design, biochemistry, route
prediction, crystallisation, and extraction. We herein report a successful approach to predicting
solubility, not only in water but also in organic solvents (ethanol, benzene, and acetone), using a
combination of machine learning and computational chemistry. Our new approach, named Causal
Structure Property Relationship (CSPR), allowed examination of the physical chemistry behind
dissolution to choose a small number of chemically relevant descriptors to produce highly
interpretable models. These models gave significantly more accurate predictions than leading
open-source and commercial solubility prediction tools, achieving accuracy (60-80 %) close to
the expected level of noise in the training data (LogS±0.7). By reproducing the physicochemical
relationship between solubility and molecular properties in different solvents, rational
improvements to the models were explored. Subsequent improvements to the models included
modifying the solvation energy and combining machine learning methods to provide a consensus
prediction. A larger dataset in water provided the basis for the discussion of pKa and speciation
in water. We conclude that gathering accurate solubility data across a range of solvents is crucial
to expanding this work and promoting sustainable chemistry in the future. It is our hope that this
methodology will be applied to other problems in chemistry and that our open-access datasets
(the first of its kind for benzene and acetone) will stimulate further research in this field.