<div>Here, we present a collection of publicly available<br>intrinsic aqueous solubility data of 829 drug-like<br>compounds. Four different machine learning algorithms<br>(random forest, light GBM, partial least squares and<br>LASSO) coupled with multi-stage permutation<br>importance for feature selection and Bayesian hyperparameter optimization were employed for the<br>prediction of solubility based on chemical structural<br>information. Our results have shown that LASSO<br>yielded the best predictive ability on an external test set<br>with and RMSE(test) of 0.70 log points and 105 features<br>in the model. Taking into account the number of<br>descriptors as well, an RF model achieved the best<br>balance between complexity and predictive ability with<br>an RMSE(test) of 0.72 with only 17 features. We<br>propose a ranking score for choosing the best model, as<br>test set performance is only one of the factors in creating<br>an applicable model. The ranking score is a weighted<br>combination of generalization, number of features<br>involved and test set performance
<br></div><div><br></div><div><br></div><div><br></div><div>The data related to this paper can be downloaded from 10.5281/zenodo.3968754</div>
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.