Chemical Signatures of Descriptor Performance in Band Gap Prediction for Organic Molecules

Machine-learning (ML) models for molecular property prediction depend critically on the representation used to encode molecular structure, namely the descriptor. For properties such as the band gap (HOMO-LUMO gap), descriptor choice is difficult to anticipate because the target is electronic in origin but remains strongly shaped by molecular composition and structure. Here, we compare four molecular representations for band gap prediction on a large dataset of organic molecules extracted from the CSD: the geometry-based descriptors SOAP and SLATM, the quantum-informed descriptor MODA, and embeddings extracted from the pretrained MACE-OFF model. In addition to global error metrics, we featurize the molecules to associate descriptor errors to interpretable molecular motifs, enabling chemically resolved analyses. MACE gives the best overall performance, followed by SLATM, MODA, and SOAP, but no descriptor is uniformly optimal: even lower-ranked descriptors are often the best for substantial subsets of molecules. The chemically resolved analysis reveals that descriptor failures are not random, but are associated with specific molecular motifs responsible for creating -or amplifying- a mismatch between the information encoded by a descriptor, and the electronic factors controlling the property (i.e. the band gap). A frontier-molecular-orbital variant of MODA, MODA_FMO, supports this interpretation by showing that part of the noise introduced by ”spectator” motifs can be anticipated and reduced. Overall, these results show that descriptor performance in band gap prediction is best understood as a chemistry-dependent tradeoff, and that error analysis can provide chemical insight into the performance of molecular representations.