Lila Sciences' first NeurIPS paper: MOSAIC any-to-any flow matching for molecules and spectra

CatAstro_Piyush · x · 2026-11-21

Lila Sciences will present its first NeurIPS paper (in Sydney) introducing MOSAIC, a multimodal flow matching transformer over molecular 3D structure and UV/IR/Raman spectra. The key insight: experiments are expensive and data-limited, but each yields multiple coupled measurements; random masking lets any modality serve as conditioning input or prediction target, extracting multiple supervised denoising tasks from a single experiment. On QM9S, MOSAIC achieves 88.8% molecular graph recovery, beating the strongest spectra-to-structure baseline by 23.2% in structure generation while also outperforming a structure-to-spectra specialist model; ablations show the any-to-any objective improves structure generation across model sizes.

Original post →

More from Research

Research channel →