Skip to main content
Back to timeline
arXivSource publication:

One model learns molecular structures and three spectra at once: MOSAIC reaches 88.8% graph recovery on QM9S

Synopsis

MOSAIC places a molecule's 3D structure and its simulated UV, IR and Raman spectra inside a single flow-matching generative model, and during training randomly decides which modalities act as conditions and which as prediction targets, so one model generates any subset of modalities from any disjoint subset of the others. On QM9S, given all three spectra it recovers 88.8% of molecular graphs, 23.2% above the strongest spectra-to-structure baseline, while also beating a dedicated structure-to-spectra predictor in the reverse direction. The result rests on QM9-sized molecules and simulated DFT spectra; experimental noise, larger molecules and drug-like sizes remain untested.

Source-provided article image: Every Measurement, Every Direction, All at Once: Multimodal Flow Matching for Molecules and Spectra
arXiv

Interpretation

A molecule's structure, UV, IR and Raman spectra are coupled measurements of one physical entity, and MOSAIC lets any of them serve as condition or prediction target, so a single model covers spectra-to-structure, structure-to-spectra and generation between arbitrary disjoint subsets. Earlier conditional generative models typically learn one direction, such as generating structures from spectra, and discard the supervision carried by the remaining measurements. On QM9S, with all three spectra as input, MOSAIC reaches 88.8% molecular graph recovery, 23.2% above the strongest spectra-to-structure baseline, and exceeds the spectral cosine similarity of the dedicated structure-to-spectra predictor DetaNet on the same dataset.

The method encodes structures as per-atom latent representations and each spectrum as a sequence of patches, then a non-equivariant multimodal diffusion transformer denoises any target subset under a rectified-flow objective. During training, per-modality Bernoulli sampling chooses the target set, and unselected modalities enter the encoder as clean context, so one set of weights handles structure generation, spectral generation and joint generation. Ablations compare the multimodal objective with a structure-only objective at S, M and L decoder sizes, where the multimodal version scores roughly 5% higher graph recovery at each size; the advantage persists as the training set grows from 1k to 110k molecules.

Perspective

All structure-generation results are obtained on QM9-sized molecules with a limited heavy-atom count and on simulated DFT spectra, evaluated on the fixed QM9S train, validation and test split. Structure-to-spectra generation is used to retrieve the matching spectrum from a fixed library of experimental gas-phase IR spectra, and that experiment evaluates retrieval only, not structure generation from experimental spectra.

Structure generation has not been tested under experimental noise, larger conformational diversity or drug-like molecular sizes; the scaling conclusions come from QM9S, whose training and test sets cover similar chemical space. The variable-count variant removes the atom-count input but keeps a fixed maximum slot budget and is evaluated on the same test split. UV alone as conditioning recovers structures markedly worse than IR and Raman, which the authors attribute to electronic transitions constraining connectivity more weakly, and that remains a hypothesis.