Virtual Screening of Pharmaceutical Cocrystals Using Machine Learning Algorithms
Roshni Jayachandiran, Suneesh Jacob Akkarapakam, T. Karthick · Crystal Growth & Design · 2025
In the crystal engineering field, the design and synthesis of pharmaceutical cocrystals are crucial for enhancing the physicochemical properties of active pharmaceutical ingredients (APIs) and for promoting multidrug treatments. However, selecting suitable coformers is a crucial and challenging task in developing cocrystals. The present work reports the suitability of theoretical machine learning (ML) models for successfully predicting coformers that are likely to form cocrystals. In this study, a data set containing 2476 data points, each with 1922 features, was considered. Four machine learning (ML) models, such as Artificial Neural Network (ANN), Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting (GB), were considered for predicting cocrystal formation between various APIs and coformers. The molecular descriptors from the Mordred descriptor calculator and drug profile parameters from the SwissADME web server tool were used to train these ML models. The GB model outperformed compared to other models with an accuracy of ∼98%, which stands out higher than the accuracy reported in the literature so far for predicting cocrystal formation. Other models, such as RF and SVM, also showed improved accuracies compared with the existing literature. The output performance of the ML models implemented in this work was evaluated using various metrics, and sensitivity analysis was performed to tune the hyperparameters. The k-fold cross-validation demonstrated that the trained models were not prone to high overfitting. It is worth mentioning that our proposed ML models were the first in the literature to incorporate drug profiles into cocrystal prediction. For reproducing the results obtained, the source codes, data set, and extracted features are made available at https://github.com/kathickphy/ml_cocrystal_screening.git .