Loading…
Self-supervised encoder for tandem mass spectra that learns proteomics representations from 1.63 billion MS/MS scans with no peptide labels.
No providers recorded yet. Browse all providers
A shotgun proteomics run produces millions of tandem mass spectra, and a database search assigns a peptide to a minority of them; the rest are discarded before analysis begins. Every model trained on this data inherits that filter, because its labels come from the same search engines whose blind spots created it.
InstaNovo-FM removes the label from the training loop. It is an encoder-only transformer trained to reconstruct masked regions of an MS/MS spectrum from the surrounding peaks, so the supervision signal is the physics of fragmentation rather than a peptide sequence. The resulting embedding space organizes spectra by fragmentation method, instrument family, chemical labelling and modification state without ever being shown those categories, and the same frozen encoder feeds de novo sequencing, database-free identification, PTM detection and whole-run classification. It was released in September 2026 by InstaDeep with the Technical University of Denmark and Delft University of Technology.
It shares a product line with InstaDeep's earlier InstaNovo diffusion-based sequencer but is a distinct artifact with a distinct objective: InstaNovo predicts peptides, InstaNovo-FM predicts spectra. Its comparison class is the supervised large-scale proteomics models — XuanjiNovo, pUniFind, DIA-BERT — which need annotated spectra to pretrain.
InstaNovo-FM is an encoder-only transformer with model dimension 768, 12 layers, 12 attention heads and feedforward dimension 3072, totalling approximately 89.5M parameters. Each peak is embedded by a multi-scale sinusoidal encoder over m/z with learnable log-spaced frequencies, concatenated with its intensity and projected through an MLP; no learned positional encoding is applied, so the transformer treats a spectrum as an unordered set, and masked positions instead carry a Gaussian-blurred m/z value (sigma = 10 Da). The pretraining corpus was assembled from 92 PRIDE projects selected by an LLM-assisted curation pipeline over 26,603 submissions, uniformly reprocessed with FragPipe v22.0 and MSFragger v4.1, spanning 72 organisms. It is tiered: 1,625,276,573 MS/MS scans in all, of which 184.6 million are high-confidence peptide-spectrum matches at 1% FDR — and it is that annotated tier, not the full corpus, on which the deployed model was trained, for roughly 230,000 steps on four NVIDIA H100 GPUs.
Linear probes on frozen embeddings classify fragmentation method at macro-F1 0.855 and instrument family at 0.804, and regress precursor m/z at R-squared 0.929. Phosphorylation is detected at AUROC 0.988 and glycosylation at 0.9999, with a five-way N-glycan family probe reaching macro-AUROC 0.915. Run-level embeddings classify digestion enzyme across 61 runs with 100% accuracy. Against encoders trained with sequence labels, InstaNovo-FM leads on acquisition- and fragmentation-related probes while trailing on peptide-property and retrieval measures; fine-tuned for de novo sequencing it beats Casanovo on five of six held-out biological datasets.
The frozen encoder is a general-purpose spectrum featurizer. Nearest-neighbour retrieval over embedded reference spectra identifies peptides without a search database and rescues assignments a search left unidentified, validated by raw-spectrum cosine similarity of 0.960–0.996. Linear probes on the same embeddings expose PTM presence and coarse glycan composition, useful where modification-aware search is intractable. Run-level pooling classifies technical and biological conditions with no peptide or protein identifications at all — a quality-control and study-design tool rather than an identification one. Labs working on venoms, nanobodies, metaproteomes and other poorly-databased samples benefit most.
The demonstration that matters is not a leaderboard position but parity: label-free representations match label-trained ones on most probes, at a scale no prior annotation-free proteomics encoder has reached. The authors are candid that the deployed checkpoint still trains on the annotated tier, whose membership was itself defined by database search, so it partially inherits the corpus-selection bias the method is designed to escape; the full unlabelled tier is left to future work. The model is a preprint awaiting peer review. Code is Apache-2.0 and the eight checkpoints are CC BY-NC-SA 4.0, which excludes commercial use.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.