π-SPECFormer 学习 2.91 亿张质谱图，挖掘蛋白变异线索

Pi-SPECFormer: A Mass Spectrometry-based Foundation Model for Decoding the Dark Proteome 1. The study frames a core bottleneck in LC–MS/MS proteomics: typically <50% of acquired MS/MS spectra get confident peptide assignments, leaving a large “dark matter” space that may contain mutations, non-canonical peptides, and unexpected PTMs missed by database search constraints. 2. π-SPECFormer is introduced as a mass-spectrometry foundation model pre-trained on 291 million spectra using a two-stage strategy designed to learn from the vast pool of unannotated spectra rather than relying only on peptide–spectrum matches (PSMs). 3. Stage 1 is fully self-supervised contrastive learning: two masked views of the same spectrum are created by random peak masking; an online encoder and a momentum encoder (with a queue of negatives) are trained with an InfoNCE objective to make embeddings robust to missing peaks. 4. Stage 2 adds PSM-guided contrastive refinement: spectra assigned to the same precursor/peptide group are treated as positives and different assignments as negatives, but peptide identities are only used to construct pairs (not provided as model inputs), aiming to improve peptide-level separability while keeping representation learning spectrum-centric. 5. The same pre-trained encoder is transferred to three tasks that probe complementary capabilities: de novo peptide sequencing (sequence recovery without databases), spectrum clustering (peptide-level similarity structure), and PTM detection (modification-associated fragmentation signatures). 6. On a nine-species de novo benchmark, π-SPECFormer improves complete-sequence correctness (peptide precision) over common tools (PEAKS, DeepNovo, PointNovo, Casanovo); Stage 2 is reported as best across all nine datasets, supporting the claim that pretraining on unlabeled spectra plus PSM-guided refinement benefits downstream decoding. 7. For spectrum clustering, embeddings from π-SPECFormer outperform GLEAMS across HeLa, cerebellum, and nine-species datasets, yielding improved clustered-spectrum proportion and cluster completeness at matched error rates; Stage 2 further improves separation, consistent with the PSM-guided refinement objective. 8. For PTM detection, fine-tuning π-SPECFormer with a classification head improves phosphorylation-spectrum detection over AHLF and over training-from-scratch baselines, with particularly large gains reported on a LUAD phosphoproteomic dataset—evidence that large-scale spectrum pretraining captures transferable PTM-relevant features. 9. In a clinical application to 103 lung adenocarcinoma (LUAD) patients, a de novo-based workflow identifies 5,103 candidate single-amino-acid substitutions across 1,712 proteins; integrated prioritization (tumor/normal abundance patterns, recurrence, survival association, and predicted functional/structural impact) highlights candidates including PYGL p.A303S, where experiments in LUAD cell lines suggest the mutant counteracts the proliferation-promoting effect of wild-type PYGL. 10. The paper also builds a mutation–drug-target interface hypothesis resource by screening predicted interactions between prioritized mutant proteins and FDA-approved oncology drug targets using AlphaFold 3 complex modeling (ipTM filtering) plus FoldX/PRODIGY-style interface perturbation metrics; synthetic peptide spectra support nine high-confidence mutated peptides (six not found in referenced public databases), positioning the approach as a route from “dark spectra” to testable mutation and interface hypotheses. 📜Paper: https:// biorxiv.org/content/10.648 98/2026.11.06.764812 … #Proteomics #MassSpectrometry #FoundationModels #SelfSupervisedLearning #Transformers #DeNovoSequencing #PTM #CancerProteogenomics #LUAD #Bioinformatics