Loading…
Peptide tandem mass spectrum prediction across the full fragment ion series, with neutral losses and modification-specific peaks beyond b/y ions.
No providers recorded yet. Browse all providers
A peptide colliding with gas in a mass spectrometer does not break in only one place. Alongside the backbone b and y ions that database search engines score, it sheds water, ammonia and phosphate, throws off immonium ions from individual side chains, and produces losses specific to whatever chemical modification it happens to carry. A predictor restricted to b and y ions therefore accounts for only 35–45% of the total fragment intensity in a real spectrum. The remaining peaks are measured, reproducible, and simply unexplained — signal that a matching algorithm cannot use because nothing tells it what those peaks should be.
UniSpec is an attention-driven neural network that predicts the whole picture: an annotated spectrum spanning a fixed dictionary of 7,919 fragment peaks, generated from a peptide sequence, its precursor charge and the collision energy. Of those 7,919 entries, 5,712 are neutral-loss peaks and 2,310 correspond to losses that only occur on a modified residue. It was built at the Mass Spectrometry Data Center of the National Institute of Standards and Technology in Gaithersburg by Joel Lapin — also of the Department of Physics at Georgetown University — with Xinjian Yan and Qian Dong, and published in Analytical Chemistry in February 2024.
The model is trained once, on NIST's own reference libraries, then applied to peptides it has never seen. That matters because the headline application is not rescoring someone else's search output: UniSpec predicts a complete in-silico spectral library for a proteome, which is then searched directly as a stand-alone identification resource.
UniSpec stacks nine transformer blocks built on talking-heads attention, a variant of multi-head attention that applies learned linear maps across the head dimension before and after the softmax. The running width is 256 channels with 64 attention heads of 16-dimensional queries, keys and values, batch normalization after each attention and feed-forward layer, and a learned positional embedding over peptides of up to 40 residues at charges 1–8. A penultimate projection to 512 channels feeds a sigmoid output layer that emits the 7,919 intensities, averaged across sequence positions. The released checkpoint holds roughly 14.9 million parameters. Training used 1.8 million unique high-quality MS2 spectra from 0.8 million unique peptide ions, drawn from NIST reference library collections acquired on Orbitrap Elite, Q Exactive, LTQ Velos and Fusion Lumos instruments and including synthetic-peptide and phosphopeptide sets. Folding UniSpec's predictions into a shotgun proteomics workflow raised tryptic peptide identifications by 48% at 1% false discovery rate and by 60% at the stricter 0.1% FDR. Head-to-head comparison with Prosit found each model strong on the ion series it was trained to emit, with UniSpec distinguished by the complexity and annotation diversity of the spectra it generates.
The immediate users are shotgun proteomics groups running data-dependent acquisition. A predicted library can replace or supplement an experimental one, extending coverage to peptides no experimental library contains, and the richer set of predicted peaks supplies more discriminating features for peptide-spectrum match rescoring — particularly for phosphopeptides, where site localization hinges on exactly the neutral losses that b/y-only predictors omit. The task is the inverse of Casanovo's: a sequence becomes the spectrum it should produce, rather than a spectrum becoming a sequence.
UniSpec pushed spectral prediction past the b/y series the field had settled on, and did so with reference data of established provenance: NIST's libraries are the same curated resource many proteomics workflows already benchmark against. Code and trained weights sit in a public NIST repository, and a version returning the top 200 predicted peaks is served through the Koina prediction platform, so the model can be used without local training. The scope is bounded in ways worth stating plainly: it covers collision-induced and higher-energy collisional dissociation only, the 73–77% coverage figure is measured against the NIST reference libraries specifically, and the supported chemistry is the seven modifications listed above rather than the open-ended modified proteome.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.