Creation and interpretation of machine learning models for aqueous solubility prediction

Abstract

Aim: Solubility prediction is an essential factor in rational drug design and many models have been developed with machine learning (ML) methods to enhance the predictive ability. However, most of the ML models are hard to interpret which limits the insights they can give in the lead optimization process. Here, an approach to construct and interpret solubility models with a combination of physicochemical properties and ML algorithms is presented. Methods: The models were trained, optimized, and tested in a dataset containing 12,983 compounds from two public datasets and further evaluated in two external test sets. More importantly, the SHapley Additive exPlanations (SHAP) and heat map coloring approaches were used to explain the predictive models and assess their suitability to guide compound optimization. Results: Among the different ML methods, random forest (RF) models obtain the best performance in the different test sets. From the interpretability perspective, fragment-based coloring offers a more robust interpretation than atom-based coloring and that normalizing the values further improves it. Conclusions: Overall, for certain applications simple ML algorithms such as RF work well and can outperform more complex methods and that combining them with fragment-coloring can offer guidance for chemists to modify the structure with a desired property. This interpretation strategy is publicly available at https://github.com/Pharmacelera/predictive-model-coloring and could be further applied in other property predictions to improve the interpretability of ML models.