Applying machine learning for solubility prediction: comparing different representations of molecular data

Solubility is one of the crucial properties of drugs and is important to determine early in the drug development cycle. Artificial intelligence (AI) based algorithms offer a faster solution compared to much more computationally expensive methods that utilize energy calculations, quantum dynamics and calculation of molecular dynamics. In this work, several AI-based algorithms utilizing different molecular data representation approaches, namely convolutional neural networks, graph neural networks and decision tree based gradient boosting, are applied to an open dataset from a recently published Kaggle challenge on solubility with more than 70 000 compounds. Performance of the models is evaluated on a testing set provided by the Kaggle challenge, as well as on a locally created independent dataset. Results demonstrate superior performance by the gradient boosting model trained on tabular feature representation of the molecules. Developed model is also shown to be competitive with other solutions posted on the leaderboards of the Kaggle challenge.