Predicting both accurate and reliable solubility value has long been a crucial but challenging task. In this work, a novel method has been developed to accurately predict the solubility of two molecules (solute and solvent) through machine learning. Current study suggests optimal physicochemical properties and demonstrates that predicting accurate solubility with a small number of datasets is possible<br/>For construction of descriptors, we used the molecular fingerprint and physicochemical properties. In this approach, molecules are encoded into molecular fingerprints;166-dimensional MACCS keys. Physicochemical properties that are easily obtainable without additional complex computations are selected; molecular weight, volume, and melting point. Since data loss occurred in the process of properties acquisition, three variations of the original dataset are considered depending on availability of each property. Using different variations of datasets helped us determine whether it is worth to use additional descriptors at the expense of data loss.<br/>Two prediction tasks were conducted: 1) the solubility value (regression) and 2) the class (classification). In the latter task, solubility values are classified as three groups; low, medium, and high. LightGBM is implemented for both cases. Crossvalidation results by selecting datasets in a random fashion exhibit no difference between models; confirming the rigidity and functionality of trained model. The r-squared score reached 0.85 in regression and 0.81 (F1 score) in the classification. More importantly, the r-squared value reached 0.73 when trained only with 20% of the data. Such result indicates that present model can be employed to predict even wider unexplored chemical space. Current study clearly demonstrates that achieving high performance is possible by adding simple but significant descriptors to molecular fingerprints, even in small datasets.

Solubility prediction using physicochemical features via machine learning

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.