Machine learning-driven discovery of multicomponent pharmaceutical solid forms via DualNet: confidence-aware prediction and ranking of salts and cocrystals

Salts and cocrystals are vital multicomponent entities for tuning pharmaceuticals’ solid-state properties, yet their experimental screening is labor-intensive and often inefficient. We introduce a DualNet Ensemble algorithm, a multi-class classification model that integrates molecular graph embeddings with curated physicochemical descriptors to predict the formation of salts, cocrystals, or physical mixtures while estimating predictive uncertainty. The proposed DualNet was trained on 70 % of a curated dataset containing 22,298 experimentally validated entries. Evaluated on a held-out 15 % test set, it achieved a macro-averaged Recall of 0.952 and F1-score of 0.940, demonstrating strong generalization. Additionally, it showed excellent calibration efficiency (ECE = 0.0161) and significantly outperformed the traditional empirical ΔpKa rule in distinguishing salts from cocrystals. The model demonstrated strong generalizability across six high-frequency compounds—salicylic acid, nicotinamide, succinic acid, benzoic acid, 2-butenoic acid, and oxalic acid—achieving a mean F1 score of 0.96. In a prospective ciprofloxacin case study, it successfully ranked three confirmed salts as the top three candidates and the only formed cocrystal as the fifth. The results clearly demonstrate the utility of the proposed methodology as a robust, interpretable, and experimentally reliable tool for accelerating multicomponent solid-form screening.