Integrating Low-Cost xTB Electronic Priors for Predicting Electronic and Excited-State Properties of Organic Conjugated Molecules.
Accurate prediction of electronic and excited-state properties is important for the computational screening of organic conjugated molecules. However, this task remains challenging because these properties depend on both local chemical environments and global structural and electronic information. In this work, we develop a dual-tower molecular representation framework that combines a graph tower for local atom-bond connectivity with a global tower for global structural information and extended tight-binding (xTB)-derived electronic priors. Evaluated on the OCELOT chromophore dataset, the fusion of graph representations with global structural features consistently outperforms the graph-only baseline. Incorporating xTB descriptors into the global tower further reduces the average MAE. The best fusion model reaches an average MAE of 0.148 ± 0.001 eV compared with 0.156 ± 0.002 eV for the graph-only model. Ablation studies across different fusion strategies show that the main benefit arises from combining local atom-bond connectivity information with global molecule-level information rather than from any single fusion operator. These results demonstrate that global representations combining RDKit descriptors, Morgan fingerprints, and xTB-derived electronic descriptors provide useful support for graph neural networks (GNNs) to predict electronic and excited-state properties of organic conjugated molecules.