Multimodal Fusion-Based Prediction and Preparation of Cocrystal: A Case Study of Diosmetin

Abstract Cocrystals offer promising approaches for the targeted regulation of raw material properties. Nevertheless, the accurate screening of cocrystal co-formers remains a critical bottleneck in the rational preparation of cocrystals. To address this challenge, a cocrystal prediction model based on multimodal fusion was constructed by integrating molecular descriptors, fingerprints, SMILES strings and molecular graphs to screen potential co-formers. The model incorporates a dynamic attention fusion mechanism and a tailored data-balancing strategy to address unimodal limitations and class imbalance, thereby enhancing classification performance under the present evaluation framework. It demonstrated strong classification performance and outperformed the evaluated unimodal models. Diosmetin was selected as the model compound, and systematic virtual screening identified promising co-formers, leading to the successful synthesis of diosmetin-4,4′-bipyridine and diosmetin-isonicotinamide cocrystals. Structural characterization revealed a 2:1 supramolecular network of diosmetin-4,4′-bipyridine stabilized by O−H···N hydrogen bonds. The two cocrystals showed lower moisture uptake than diosmetin under the tested high-humidity conditions of 40 °C/75% RH and 40 °C/95% RH. This study presents a multimodal fusion strategy for cocrystal prediction and highlights its potential in the rational design of solid forms.