Learning from tandem mass spectra at scale with a self-supervised foundation model for proteomics
preprint · bioRxiv · 2026
| Date | 2026-09-03 |
| Type | preprint |
| Venue | bioRxiv |
| Publisher | Cold Spring Harbor Laboratory |
| Contribution | algorithm |
| DOI | 10.64898/2026.09.03.747733 |
| Citations (OpenAlex) | 0 |
Abstract
Mass spectrometry-based proteomics increasingly relies on machine learning, yet existing models are trained for defined supervised tasks such as peptide identification, de novo sequencing or fragment intensity prediction, limiting transfer across datasets, instruments and acquisition methods. Here we present InstaNovo-FM, a self-supervised foundation model for bottom-up proteomics trained to reconstruct masked regions of tandem mass spectra. We assemble a diverse training corpus spanning 1.47 billion MS/MS spectra and 184.6 million high-confidence annotations. We train an encoder-only transformer on the annotated tier using a physics-aware masked reconstruction objective. We demonstrate that the InstaNovo-FM embeddings encode fundamental experimental and biological properties, including fragmentation method, sequence properties and post-translational modifications, without requiring peptide labels. Furthermore, this foundation model directly enables diverse downstream applications, including de novo peptide sequencing, database-free identification and analytical run classification. InstaNovo-FM establishes a unified representation space for peptide fragmentation spectra, enabling robust transferability across the proteomics ecosystem.
Methods and tools
- InstaNovo-FM: Self-supervised foundation model for bottom-up proteomics: an encoder-only transformer trained to reconstruct masked regions of tandem mass spectra under a physics-aware objective, over a corpus of 1.47 billion MS/MS spectra with 184.6 million high-confidence annotations. The embeddings capture fragmentation method, sequence properties and post-translational modifications without peptide labels, and support de novo sequencing, database-free identification and analytical run classification as downstream tasks.