Interpretable Deep Supramolecular Language Modeling for Cocrystal Design: PCA-Augmented DeepCocrystal Screening of Picolinamide, Nicotinamide, and Isonicotinamide Systems

Cocrystallization is an effective strategy for modifying the physicochemical properties of active pharmaceutical ingredients (APIs), yet rational coformer selection remains challenging. In this work, we propose an interpretable extension of DeepCocrystal by integrating Principal Component Analysis (PCA) with supramolecular language-based prediction. Picolinamide, nicotinamide, and isonicotinamide were selected as models for APIs due to their diverse hydrogen-bonding capabilities but structural similarity. A curated coformer library was screened computationally, and the molecular descriptor space was analyzed to identify dominant structural drivers of cocrystal compatibility. PCA revealed that cocrystal propensity in the studied systems is primarily governed by hydrogen-bond complementarity, polarity matching, and aromatic surface interactions. The integrated framework is intended to enhance mechanistic interpretability without altering the predictive output of DeepCocrystal. In this setting, the PCA layer functions as a post hoc, unsupervised readout of the descriptor space, so the original ranking and scoring of the predictions remain unchanged. This approach provides a scalable and chemically transparent methodology for rational coformer prioritization and supports data-driven crystal engineering strategies. In addition, a simpler post-filtering strategy based on pKa values was also applied to the DeepCocrystal predictions to attempt a classification that can exclude salts. Following the pKa-based post-filtering rules, from the 137 API–coformer pairs classified by DeepCocrystal as positive or as uncertain, that is, all pairs not confidently excluded, 46 pairs retained their classification as high-probability cocrystal-compatible, 39 pairs were assigned the classification of uncertain but salt formation excluded, 11 pairs retained the classification of negative with large uncertainty, 34 pairs were reclassified as belonging to the salt/cocrystal “gray zone”, and 7 pairs were classified as salt-risk candidates.