A paper about our molecular foundation model work MIST is now on arXiv! MIST is a family of models that spans small molecules, organometallics, mixtures/formulations, and even isotopes, relying on our Smirk tokenizer to incorporate stereochemical, isotopic, and electronic information. Using MIST, we screen over 8 million candidate electrolyte solvents in eight hours, identifying 63 Pareto-optimal candidates, and achieve state-of-the-art accuracy for mixture properties across temperature and composition, surpassing our previous GNN-based DiffMix models.
MIST also performs strongly on olfactory perception prediction on a dataset curated by Alex Wiltschko and the Google DeepMind team, and its learned token embeddings align with chemical “meaning,” revealing clean clusters of d- and p-block elements, features related to Lipinski’s Rule of Five, aromaticity, and PAH condensation.
Read more about MIST in our pre-print here or and try out Smirk.
