TAG-X: Joint Multispectral Representations for Molecular Structure Elucidation

Effective molecular structure elucidation reconciles complementary evidence from spectra describing different aspects of chemical structure (proton/carbon electronic environments, functional groups, correlations, fragmentation...). We introduce TAG-X, a formula-conditioned spectra-to-structure framework that jointly encodes $^1$H NMR, $^{13}$C NMR, HSQC, IR, and MS/MS into a chemically supervised token bank that keeps modality-specific spectral information and mixes complementary evidence across modalities. A token-preserving interface conditions a pretrained autoregressive (AR) molecular decoder on this representation and an inverse solver reranks the resulting beam by predicting the spectra of each candidate structure and comparing them with the input spectra. Experiments on a large simulated multimodal corpus show that our joint encoding captures complementary information unavailable to independent modality representations, supports accurate molecular reconstruction, and enables spectral verification to recover correct structures that are generated but initially misranked to achieve a 86.13\% top-1 reconstruction accuracy on a 30000+ molecule test set of a simulated dataset.