Machine Learning in Prediction of Intrinsic Aqueous Solubility of Drug-like Compounds: Generalization, Complexity or Predictive Ability?

Here, we present a collection of publicly available
intrinsic aqueous solubility data of 829 drug-like
compounds. Four different machine learning algorithms
(random forest, light GBM, partial least squares and
LASSO) coupled with multi-stage permutation
importance for feature selection and Bayesian hyperparameter optimization were employed for the
prediction of solubility based on chemical structural
information. Our results have shown that LASSO
yielded the best predictive ability on an external test set
with and RMSE(test) of 0.70 log points and 105 features
in the model. Taking into account the number of
descriptors as well, an RF model achieved the best
balance between complexity and predictive ability with
an RMSE(test) of 0.72 with only 17 features. We
propose a ranking score for choosing the best model, as
test set performance is only one of the factors in creating
an applicable model. The ranking score is a weighted
combination of generalization, number of features
involved and test set performance

The data related to this paper can be downloaded from 10.5281/zenodo.3968754