Global descriptors informed learning for organic melting-point prediction

Machine learning-assisted melting-point prediction remains challenging because topology-derived representations can not explicitly encode all whole-molecule physicochemical factors. Herein, we evaluate a global-informed framework by incorporating efficient molecular physical descriptors through a conditioning branch, achieving MAEs of 21.45 ± 0.08 °C and 23.06 ± 0.25 °C across five random-split and five scaffold-split folds. A matched comparison with the topology-only model shows a small but consistent contribution from the descriptor branch under scaffold-aware validation. This work highlights the value of global physicochemical information in complementing topology-derived representations for physical-property prediction and of evaluation under reduced structural overlap.