Cocrystal Formation Prediction: Hybrid GIN-Mordred Model Outperforms DFT-Based Methods

Abstract

Cocrystals offer significant potential across various industries, especially pharmaceuticals, by addressing the poor solubility of new drug candidates. However, traditional experimental screening for cocrystal formation is expensive and time-consuming, highlighting the need for predictive models. In this study, we compared four cocrystal prediction approaches: two deep learning (DL) models based on DFT-driven data (PointNet for electrostatic potential (ESP) maps and a novel LSTM for sequential hydrogen bond parameters), a novel hybrid model combining graph isomorphism networks (GIN) with Mordred descriptors, and the empirical Hydrogen Bond Energy (HBE) method. To perform this comparison, we compiled and carried out DFT calculations for 14,790 molecules (7395 pairs of successful and unsuccessful cocrystals). Notably, the GIN-Mordred model outperformed all other methods, achieving the highest balanced accuracy (BACC: 0.916), F1-score (0.956), recall (0.932), and AUC (0.97), with superior segregation performance in distinguishing between cocrystallization outcomes. Importantly, the GIN-Mordred model does not require costly DFT calculations, demonstrating that a combination of graph-based and descriptor-based molecular representation provides an efficient and accurate alternative for cocrystal prediction. This model significantly streamlines the process of tuning the physicochemical properties of crystalline materials for various applications.