Machine learning approach to predict Hansen solubility parameters of cocrystal coformers via integrating group contribution and COSMO-RS

Hansen Solubility Parameters (HSPs) have been a hot topic on predicting the tendency of pharmaceutical cocrystals formation and cocrystal coformers (CCFs) screening. However, the limitation of such application is the lack of models to accurately predict the values of HSPs for drug CCFs with more structural complexity. Accordingly, three ML (machine learning) models, i.e. ANN (Artificial Neural Network), XGBoostRegressor (Extreme Gradient Boosting Regressor) and LGBMRegressor (Light Gradient Boosting Machine Regressor), were developed for predicting the HSPs on CCFs screening for drugs. The HSPs database for 181 CCFs (containing alcohols, alkenes, aromatics, haloalkanes, amines, ketones, ethers, amides, esters, pharmaceuticals, alkanes, acids, nitroalkanes) were established and classified into the training set (140 compounds) and the test set (41 compounds with various functional polarity and groups, covering solid reagents and solvents). The prediction molecular descriptors were combined from the GC (Group Contribution) methods, the COSMO-RS (the Conductor-like Screening Model for Real Solvents) sigma-moments and energy descriptors. The results showed that ANN and XGBoostRegressor beat out LGBMRegressor in predicting HSPs for CCFs. Finally, SHapley Additive exPlanations (SHAP) was employed to visualize and explain the most important characteristics and effects on predicting HSPs via XGBoostRegressor, indicating that CH3, M2 and MHbdon3 had a significant influence and high contribution to the prediction of δd, δp and δh, respectively. The coupled GC and COSMO-RS strategy had been proven as a promising tool to predict HSPs through XGBoostRegressor for screening, designing, and selecting CCFs for drugs.