FRIGID: Scaling Diffusion-Based Molecular Generation from Mass Spectra at Training and Inference Time
This work addresses the problem of de novo structural elucidation from mass spectra, a key bottleneck in analytical chemistry, with a scalable approach that significantly outperforms existing methods.
FRIGID introduces a diffusion language model for generating molecular structures from mass spectra, using intermediate fingerprint representations and chemical formulae. It achieves 18% Top-1 accuracy on MassSpecGym and triples the Top-1 accuracy of leading methods on NPLIB1, with inference-time scaling showing log-linear performance improvements.
In this work, we present FRIGID, a framework with a novel diffusion language model that generates molecular structures conditioned on mass spectra via intermediate fingerprint representations and determined chemical formulae, training at the scale of hundreds of millions of unlabeled structures. We then demonstrate how forward fragmentation models enable inference-time scaling by identifying spectrum-inconsistent fragments and refining them through targeted remasking and denoising. While FRIGID already achieves strong performance with its diffusion base, inference-time scaling significantly improves its accuracy, surpassing 18% Top-1 accuracy on the challenging MassSpecGym benchmark and tripling the Top-1 accuracy of the leading methods on NPLIB1. Further empirical analyses show that FRIGID exhibits log-linear performance scaling with increasing inference-time compute, opening a promising new direction for continued improvements in de novo structural elucidation. FRIGID code is publicly available at https://github.com/coleygroup/FRIGID