Utilization of advanced machine learning models for analysis of pharmaceutical cocrystals by prediction of solubility parameters using Dragonfly algorithm optimization.

Pharmaceutical cocrystals have emerged as an effective strategy for enhancing the solubility and bioavailability of poorly water-soluble drugs. The successful design of cocrystals largely depends on the selection of suitable coformers, which requires an understanding of the molecular interactions governing cocrystal formation during the process. The objective of this study was to develop robust machine learning models for predicting Hansen solubility parameters of pharmaceutical cocrystal using molecular descriptors generated from COSMO-RS and group contribution methods. A dataset consisting of 181 samples and 86 input features was utilized, incorporating molecular descriptors related to hydrogen-bonding capability, van der Waals interactions, and structural functional groups. Data preprocessing included outlier detection using the Isolation Forest algorithm and feature selection through Sequential Floating Forward Selection (SFFS). Three tree-based ensemble learning models, namely Random Forest (RF), Extra Trees (ET), and Gradient Boosting Regression Trees (GBRT), were developed and optimized using the Dragonfly algorithm for hyperparameter tuning. The obtained results demonstrated that the Extra Trees model consistently outperformed the RF and GBRT models in predicting all three Hansen solubility parameters. For the three target outputs, the ET model achieved test-set R2 values of 0.9175, 0.8661, and 0.9815, respectively, while also exhibiting the lowest RMSE and MAE values. Feature importance analysis revealed that both group contribution descriptors and COSMO-RS-derived molecular properties play significant roles in determining the Hansen solubility parameters of coformers. Overall, the proposed machine learning framework provides an efficient and accurate approach for predicting Hansen solubility parameters and screening pharmaceutical cocrystals. The findings highlight the superior predictive capability of the Extra Trees model and demonstrate the potential of combining molecular thermodynamic descriptors with advanced machine learning techniques for pharmaceutical cocrystal design.