Learning from tandem mass spectra at scale with a self-supervised foundation model for proteomics
- 1Technical University of Denmark
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 1.63 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.
Related
- Interpreting Short Tandem Repeat Variations with Plant Genomic Foundation Models in Olive
- ProtCell: a foundation model for single-cell proteomics
- A self-supervised DNA foundation model with collapse-resistant multimodal fusion
- MtReNIRS: A Self-Supervised Foundation Model for Transferable Neurovascular Representations
- Learning from Error Predictions Rather Than Prediction Errors: A Perspective on Self-Supervised Learning in the Brain
BibTeX
@misc{nieuwoudt2026learning,
title = {{Learning from tandem mass spectra at scale with a self-supervised foundation model for proteomics}},
author = {M. Nieuwoudt and M. Reverenna and D. Patel and R. Catzel and I. H. J. Houngue and J. Daniel and K. Eloff and A. Santos and N. Lopez Carranza and T. P. Jenkins and J. Van Goey and K. Kalogeropoulos},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.03.747733},
url = {https://www.biorxiv.org/content/10.64898/2026.09.03.747733v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Learning from tandem mass spectra at scale with a self-supervised foundation model for proteomics
AU - M. Nieuwoudt
AU - M. Reverenna
AU - D. Patel
AU - R. Catzel
AU - I. H. J. Houngue
AU - J. Daniel
AU - K. Eloff
AU - A. Santos
AU - N. Lopez Carranza
AU - T. P. Jenkins
AU - J. Van Goey
AU - K. Kalogeropoulos
PY - 2026
DA - 2026/09/03
PB - bioRxiv
DO - 10.64898/2026.09.03.747733
UR - https://www.biorxiv.org/content/10.64898/2026.09.03.747733v1
AB - 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 1.63 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.
ER -