When relationships matter more than scale: fragment distances improve chromophore prediction Two molecules can contain similar chemical groups and still behave very differently because those groups sit in different places within the molecular structure. That is the idea behind TDiMS. Instead of asking a large pretrained model to discover every long-range relationship from scratch, Hamada and coauthors explicitly represent pairs of molecular fragments and how far apart they are along the molecular graph. For chromophores, that turns out to be especially useful. Their optical properties depend not only on which functional groups are present, but also on how those groups are arranged relative to one another. For predicting the Stokes shift, the difference between the light a molecule absorbs and later emits, this simple distance-based representation improves over the second-best descriptor by 7.7% in one solvent and 14.1% in another, outperforming representations from pretrained models such as MoLFormer and MolCLR. The nice part is that the important features remain chemically meaningful. The model highlights specific pairs of molecular groups whose separation is linked to how electrons move through the molecule. For AI for Science, the useful point is broader: a good representation should encode the relationships that the science tells you matter. Sometimes that can be more valuable than simply scaling the model. Hamada et al., Nature Computational Science (2026), CC BY 4.0. doi.org/10.1038/s43588-026-0…