FullMS2Former

A peak sitting 61 Da below y7 in an HCD spectrum is not noise. It reproduces across runs, it tracks the peptide that produced it, and its height depends on the sequence around the cleavage. What it lacks is a name: no fragment-ion nomenclature accounts for it, so a spectrum predictor built on a curated list of ion types has no output slot for it, and a search engine scoring against that prediction discards the intensity.

FullMS2Former widens the vocabulary rather than the network. It predicts an intensity for every entry in a dictionary that takes each canonical b and y fragment — at every cleavage position and fragment charge — and pairs it with a set of integer mass offsets: losses of 17, 18, 28, and on down past 92 Da, together with gains such as +17 and +28. Immonium ions and precursor-derived species round the dictionary out. The unnamed peak below y7 now has a slot, and the model learns it directly from annotated collision-induced dissociation spectra instead of waiting for someone to work out which chemistry produced it. This is the difference from a fixed fragment dictionary: the offsets are enumerated mechanically from the canonical ions, so coverage does not depend on anyone having catalogued the loss first.

The model was built by Justin R. Zhang at the University of California, Santa Barbara with Zhongqi Zhang at Amgen, and published in Analytical Chemistry in August 2026. Its closest relative is UniSpec, the NIST fragment-dictionary predictor it benchmarks against; both go well beyond the b/y-restricted intensity predictors such as Prosit-PTM that most proteomics pipelines still rely on.

#Key Features

• Offset-annotated ion dictionary: The released dictionary enumerates 22,499 annotated ions, covering peptides up to 40 residues and fragment charges to 5+, with each canonical fragment carrying its own family of integer mass offsets.
• Talking-head attention: The encoder uses attention with learned linear projections across heads both before and after the softmax, letting heads exchange information rather than attending independently.
• Masked cosine objective: Ions marked as unmeasured in a reference spectrum are excluded from the per-spectrum cosine loss, so a missing annotation is not trained as a zero intensity.
• Explicit instrument conditioning: Precursor charge and normalized collision energy are projected into a global embedding added at every sequence position, so one model spans collision-energy settings instead of requiring one model per setting.
• Ensemble prediction: Inference averages several independently trained checkpoints and thresholds intensities below 1e-4 to zero, trading compute for stability on low-abundance ions.

#Technical Details

The network is compact: one transformer encoder layer with 256-dimensional embeddings and 64 talking-heads, a two-layer feedforward block, and a 2048-unit penultimate projection, followed by a sigmoid and mean pooling across sequence positions. The repository's specification puts the total at roughly 56M parameters, though that table assumes a 26,916-entry dictionary and the dictionary actually shipped declares 22,499 ions, which works out nearer 47M. Either way the distribution is lopsided — nearly all of the parameters sit in the single output projection onto the ion dictionary, so it is the dictionary's width, not the encoder's depth, that sets the model's scale. Input is a one-hot peptide sequence over a 22-character alphabet, capped at 40 residues, plus charge and normalized collision energy. Training uses a subset of the UniSpec dataset released by Lapin, Yan and Dong — a synthetic-peptide Orbitrap Lumos split and a Q Exactive human/CHO split, drawn in turn from NIST higher-energy collisional dissociation reference libraries — with held-out validation and test libraries kept separate from the training files. A C++ preprocessor converts .msp libraries into the annotated feature vectors the model consumes, but the authors' own Zenodo deposit ships those annotated files directly. Against UniSpec, FullMS2Former reaches comparable spectral similarity on validation data and higher similarity on most test sets, and it recovers rare sequence-dependent ions that fall outside conventional annotation.

#Applications

Predicted spectra feed the standard bottom-up proteomics workflow: generating in-silico spectral libraries for database search, rescoring peptide-spectrum matches to lift identification rates at a fixed false-discovery rate, and adjudicating ambiguous matches where the informative peaks are the uncommon ones. The wider ion coverage matters most where spectra are interrogated peak by peak — biopharmaceutical characterization, sequence-variant and degradation analysis, and any setting where an unexplained peak has to be either assigned or ruled out.

#Impact

FullMS2Former reframes near-complete spectrum prediction as a vocabulary problem, and the demonstration that a model can learn ions nobody has named is the durable contribution: it turns the unexplained fraction of a spectrum into trainable signal rather than a known limitation. The practical caveats are real. No trained checkpoints are distributed, so reproducing the results means retraining — though the authors' CC-BY Zenodo deposit ships the annotated training files, so that retraining skips both the NIST download and the Windows/MSVC C++ preprocessing build. The code carries no license file, and the ion dictionary is frozen at training time, so extending it to new ion types requires re-deriving the dictionary and a new output layer.