Boosting Predictability: Towards Rapid Estimation of Organic Molecule Solubility

The water solubility of organic molecules is critical for optimizing the performance and stability of aqueous flow batteries, as well as for various other applications. Although relatively straightforward to measure in some cases, the theoretical prediction of the solubility remains a considerable challenge. To this end, machine learning algorithms have become increasingly important tools in the past decade. High-quality data and effective descriptors are essential for constructing reliable data-driven estimation models. We systematically investigate the effectiveness of enhanced structure-based descriptors and an outlier detection procedure for improving aqueous solubility predictability. We train and evaluate random forest regression models using various descriptors to predict experimental solubility. Outliers are identified through an iterative maximum-error deletion procedure. We discover that descriptors derived from hydration free energy and weighted fingerprints, along with other established features, are effective. Notably, solvation energy, octanol-water partition coefficient, atomic charge polarizability interactions, and the presence of a full-carbon aromatic ring are critical for solubility prediction. Furthermore, the effectiveness of the outlier detection protocol is validated by improving the performance of the model and detailed analysis of the dataset. This study significantly improves the predictive capacity of supervised machine learning for molecular properties, enabling advancements in various technological applications.