π-SPECFormer: A Mass Spectrometry-based Foundation Model for Decoding the Dark Proteome

preprint · bioRxiv · 2026

preprint · bioRxiv · 2026. Tianze Ling et al. In mass-spectrometry-based proteomics, a substantial proportion of acquired spectra remain unidentified…
Date 2026-11-09
Type preprint
Venue bioRxiv
Publisher Cold Spring Harbor Laboratory
Contribution algorithm
DOI 10.64898/2026.11.06.764812

Abstract

In mass-spectrometry-based proteomics, a substantial proportion of acquired spectra remain unidentified, concealing a vast reservoir of biological variations commonly termed the “dark proteome”, potentially obscuring sequence mutations, non-canonical peptides and post-translational modifications. Existing deep-learning methods largely rely on supervised training with peptide-spectrum matches (PSMs), limiting their ability to learn from the much larger pool of unannotated spectra. Here we introduce π-SPECFormer, a mass-spectrometry foundation model pre-trained on 291 million spectra through a two-stage framework that combines self-supervised and PSMs-guided representation learning. π-SPECFormer transfers effectively to de novo peptide sequencing, spectrum clustering and modification detection, consistently outperforming task-specific baselines. When applied to proteomic profiles from 103 individuals with lung adenocarcinoma, the model identified nine high-confidence mutated peptides, including six that were absent from the reference databases. π-SPECFormer thus enables unbiased decoding of molecular complexity beyond the boundaries of the canonical proteome.

Authors

  1. Tianze Ling · Beijing Institute of Lifeomics, National Center for Protein Sciences (Beijing), State Key Laboratory of Medical Proteomics, Tsinghua University
  2. Tao Liu · National Center for Protein Sciences (Beijing)
  3. Fan Xu · Peng Cheng Laboratory
  4. Yiming Zhang (NCPSB) · National Center for Protein Sciences (Beijing)
  5. Jun Xia · The Hong Kong University of Science and Technology, The Hong Kong University of Science and Technology (Guangzhou), Westlake University
  6. Xiansong Huang · Peng Cheng Laboratory
  7. Shaorong Chen · Westlake University, Zhejiang University
  8. Linhai Xie · Beijing Institute of Lifeomics, International Academy of Phronesis Medicine (Guangdong), National Center for Protein Sciences (Beijing), State Key Laboratory of Medical Proteomics
  9. Rong Xu · Guangzhou National Laboratory
  10. Tingpeng Yang · Peng Cheng Laboratory, Tsinghua Shenzhen International Graduate School, Tsinghua University
  11. Kaixuan Li · National Center for Protein Sciences (Beijing)
  12. Guibin Wang · Beijing Institute of Lifeomics, National Center for Protein Sciences (Beijing)
  13. Zhendong Liang · Peng Cheng Laboratory, Tsinghua Shenzhen International Graduate School, Tsinghua University
  14. Jun Zhu (Tsinghua) · National Center for Protein Sciences (Beijing), Tsinghua University
  15. Stan Z. Li · Westlake University
  16. Fuchu He · Beijing Institute of Lifeomics, International Academy of Phronesis Medicine (Guangdong), National Center for Protein Sciences (Beijing), State Key Laboratory of Medical Proteomics
  17. Yu Wang · Peng Cheng Laboratory
  18. Cheng Chang · Beijing Institute of Lifeomics, International Academy of Phronesis Medicine (Guangdong), National Center for Protein Sciences (Beijing), State Key Laboratory of Medical Proteomics

Methods and tools

  • π-SPECFormer: Mass-spectrometry foundation model: a 21M-parameter Transformer spectrum encoder pre-trained contrastively (MoCo-style, InfoNCE over masked-peak views) on 178M unidentified spectra, then refined on 114M PSMs grouped by precursor. For de novo sequencing the frozen encoder feeds a task-specific Transformer encoder and autoregressive peptide decoder; the same embeddings serve spectrum clustering and, fine-tuned, PTM detection. Applied to mine single-amino-acid variants from dark spectra in a 103-patient lung adenocarcinoma cohort.

Trained on

Evaluated on

Seen in the charts

Back to the full map

Back to top