MIT's MOSAIC Beats the Best Chemistry AI by 23 Points
A single flow matching transformer learns molecules and three spectra jointly, beating specialist spectra-to-structure baselines by 23.2 points on QM9S.
Takeaways
- MOSAIC is a NeurIPS 2026 paper from Lila Sciences on any-to-any molecule and spectra generation.
- Single flow matching transformer jointly handles 3D structure plus UV, IR, and Raman spectra.
- Achieves 88.8% graph recovery on QM9S, beating the top spectra-to-structure baseline by 23.2 points.
- Also beats a dedicated structure-to-spectra specialist at generating spectra from structures.
- Random masking of target and conditioning modalities turns one experiment into many training tasks.
- Generated IR spectra match experimental EUROCHAMP references for small molecules, enabling retrieval-based identification.
Chemistry experiments usually produce a bundle of measurements. A sample run through spectrometers may yield ultraviolet, infrared, and Raman spectra alongside an existing structural hypothesis. Most molecular generative models consume one modality and predict another, leaving the remaining relationships unused. A NeurIPS 2026 paper from Lila Sciences and MIT proposes a single model that learns mappings among molecular structures and all three spectral modalities.
MOSAIC combines self-supervised modality encoders with a Multimodal Diffusion Transformer, or MMDiT, trained through flow matching. The same checkpoint can infer a structure from available spectra, predict spectra from a structure, complete a missing modality, or jointly generate a molecule and its spectra from noise.
One record, many objectives
Flow matching trains a model on intermediate points between random noise and real data, asking it to predict the velocity that carries each point toward the data distribution. At every optimization step, MOSAIC randomly assigns its four modalities, 3D structure, UV, IR, and Raman, to conditioning and target sets. Target streams are noised and reconstructed while the remaining streams provide context, turning each experimental record into several possible learning tasks.
Separate self-supervised encoders convert each modality into tokens before the MMDiT generator processes them jointly. Structure tokens contain element identities and 3D coordinates, while spectral tokens represent intensity values along each spectral axis. The transformer preserves modality-specific streams and allows attention across them, giving structural and spectral evidence a shared representation.