InstaNovo-FM: a foundation model for tandem mass spectra

This is the pretraining objective itself, run live. Hide peaks in a spectrum and the model predicts which cell of a 0.2 Da grid each one belongs to, using only the peaks around it. A de novo model has no way to answer this question, because it emits sequences, not spectra.

Two things to be clear about. A hidden peak is not erased: the encoder replaces it with its true m/z blurred by Gaussian noise of σ = 10 Da plus a learned "masked" bias, and keeps its true intensity. And because peaks arrive sorted by m/z, the model also knows a hidden peak lies between its nearest still-visible neighbours. Its job is to sharpen all of that into one 0.2 Da bin, so the baseline it is measured against uses the same two hints.

Step 2: explain one reconstruction

Every peak the model could still see is deleted in turn and the reconstruction re-run, so what comes out is the evidence the model needed, not merely what it attended to. The mass gap to each peak that mattered is then matched against residue masses, isotope spacings and neutral losses, which is arithmetic the model was never taught.

The manuscript's Fig. 4 is what loads by default, so reproducing it takes three button presses and no setup. The spectrum is TGVELGKPTHFTVNAK, HCD on a Q Exactive at precursor charge 3+, and step 1 is already set to hide peaks 91, 92: its y5 ion at 532.308 m/z together with the isotope co-masked alongside it. Press Step 1, then Explain it here, then Step 3 below. The peak that matters is y6 at 679.379, one phenylalanine step of 147.07 Da away, which is the ladder neighbour the figure reports.

To score the model over many spectra instead, load one of the other example files and switch to a masking mode that hides more than one chosen peak.

Reading the lower panel: it has one entry for every peak that was deleted. A grey dot on the zero line means deleting that peak changed nothing, so it carried no evidence. A green bar means the correct bin became less likely, and the bar's height is how much probability it lost. A single tall green bar in a field of grey dots means the model was leaning on one specific peak, and the label says which chemical relationship connects the two. Probability on the correct bin is used rather than added error because a rich spectrum carries redundant evidence: deleting one peak of a hundred rarely moves the answer out of the right bin, but it still shows in how strongly the model backs it.

Step 3: the three views of Fig. 4A

Panel A of the manuscript's Fig. 4 draws one spectrum three times, and the three together say more than any one of them. Ion type comes from matching the known peptide, so it is ground truth and owes nothing to the model. Attribution is integrated gradients against a zero-intensity baseline, the same quantity the paper plots, and says what the reconstruction drew on. Confidence is what the model thinks it knows about every peak, hidden or not.

Reading down a column is how you separate "the model used the fragment ladder" from "the model happened to be right". On the bundled Fig. 4 spectrum the most attributed peak is the y ion at 679.379, which is the ladder neighbour the figure reports.

The figure's remaining panels follow underneath. B lays the same attribution along the peptide backbone, b ions above and y ions below, so you can see where in the sequence the evidence sits. C sums it by how each peak relates to the hidden one. D plots it against intensity, which is the figure's check that the model is not merely following the tallest peaks. E is a Lorenz curve showing how few peaks carry most of the attribution.

Search by representation, not by sequence. Embed a set of already-identified spectra as anchors, embed your unassigned spectra as queries, and transfer the identity of the nearest anchor. The peptide never has to be in a local FASTA file; it only has to have been observed and identified somewhere.

This is how unassigned spectra get rescued. On a held-out project the model never saw during pretraining, the paper reports rescued matches at embedding similarity ≥ 0.998 confirmed by raw-spectrum cosine of 0.96 to 0.996 and 59 to 78% peak alignment.

Check one match against the raw spectra

One encoder, any spectrum, no labels. Each spectrum becomes a unit-length 768-dimensional vector you can cluster, index, probe or feed to a downstream model.

In the paper these vectors separate by fragmentation method at the coarsest scale, then by instrument, then by labelling chemistry and peptide identity: an entire hierarchy that emerges without any of those labels at training time. A linear probe on frozen embeddings reaches macro-F1 0.855 for fragmentation method and detects phosphorylation at AUROC 0.988.

Upload several files and colour by source file to see the run-level version: whole runs acquired under different conditions separate with no peptide identifications at all. Note that source file is as close as this interface gets to acquisition mode. An MGF carries no field for fragmentation method or acquisition, so those cannot be offered as colourings; each bundled example file happens to be a single run, which makes source file a stand-in for it here but not in general.