InstaDeepAI/InstaNovo-FM-embeddings

InstaNovo-FM spectrum embeddings

Frozen embeddings of held-out tandem mass spectra, produced by InstaNovo-FM — a self-supervised foundation model trained to reconstruct masked regions of MS/MS spectra without peptide-sequence labels.

Each row is one spectrum: a 768-dimensional mean-pooled embedding, the search-engine and acquisition metadata that identifies it, and the UMAP coordinates used in the paper.

Publication: Learning from tandem mass spectra at scale with a self-supervised foundation model for proteomics, bioRxiv, 3 September 2026 (v2). doi:10.64898/2026.09.03.747733

The two configs

from datasets import load_dataset

fig3 = load_dataset("InstaDeepAI/InstaNovo-FM-embeddings", "100k", split="train")
big  = load_dataset("InstaDeepAI/InstaNovo-FM-embeddings", "1M", split="train")

config	spectra	what it is
100k	100,000	the point set published as Figure 3. Carries figure3_umap_x / figure3_umap_y, the exact coordinates in the paper
1M	1,000,000	a larger draw from the same held-out split. Carries two 2-D and two 3-D UMAP layouts, not yet mentioned in v2 of the preprint

Both are the LCFM test split — held out from training — embedded with the released checkpoint instanovo-fm-v0.1.0 using mean_pool.

The two configs are not disjoint, and not redundant. The 100k spectra are also in the 1M set. They come from two separate extraction runs, and their embeddings were checked against each other: over 2,000 spectra sampled from the overlap, every vector is bit-identical (maximum absolute difference 0.0, minimum cosine similarity 1.0). Use 100k to reproduce the figure, 1M for anything that wants more data.

Columns

• embedding — 768 × float32, mean-pooled over the spectrum's peak representations.

• 52 metadata fields — usi, sequence, precursor_mz, precursor_charge, retention_time, hyperscore, expectation, search_project, experiment_name, protein, header (the instrument filter line), and so on. Several arrive from the eval harness as strings even when numeric; cast them rather than assuming a dtype.

• UMAP coordinates — figure3_umap_x/y in 100k; native_x/y, native3_x/y/z, transform_x/y, transform3_x/y/z in 1M.

• Five derived columns — sequence_length, n_peaks, peak_center_of_mass, peak_spread, top_duplicate_peptides. These are computed after extraction, so they are not in the model's own output; they were recomputed here and checked against the published Figure 3 table over all 100,000 rows:

  column	agreement with the published values
  sequence_length	100,000 / 100,000 exact
  n_peaks	100,000 / 100,000 exact
  peak_center_of_mass	bit-identical, maximum difference 0.0
  peak_spread	bit-identical, maximum difference 0.0
  top_duplicate_peptides	99,728 / 100,000 — see below

  n_peaks counts the model's 200 peak slots, not the full spectrum: the stored peak list runs to 800 peaks (median 235), so counting that would give a plausible number meaning something else.

  top_duplicate_peptides is the rank among the ten most repeated peptides, or 10 for everything else. It is population-dependent by definition, so a spectrum ranks differently in the two configs. The 272 rows that differ from the published column are all ties at equal counts — ranks 0-6 match exactly, ranks 7 and 8 are both count 70 and swap, and rank 9 is a tie at count 66 so one peptide falls either side of the cut. Ties are broken alphabetically here, which is deterministic; the original order was not recorded.

The 1M layouts differ in a way worth knowing before using them:

columns	layout
transform_*	the published Figure 3 layout extended to a million points. Figure 3's own spectra keep their exact published coordinates; the rest were placed into that same space with umap-learn's .transform()
native_*	an independent cuml.manifold.UMAP fit over all 1,000,000 embeddings. A different embedding, not a denser Figure 3 — Procrustes r = 0.858 against the transform layout globally, but about 11% 20-NN neighbourhood overlap

A 3-component UMAP is a separate optimisation, not an extra axis on a 2-D fit, so the *3_x/y/z triples are not the 2-D columns plus a third value. Use each triple together.

Seven columns the Figure 3 table has and this does not, yet

annotation_ratio, backbone_coverage, median_ppm_error, n_fragment_groups_metric, signal_intensity_ratio, match_metrics and spectrum_quality all require theoretical-spectrum generation, peak matching and quality scoring. They are absent from this revision rather than approximated: the annotation path needs the harness's own modified-sequence handling, which 31% of these rows require, and a reimplementation would risk plausible-but-wrong values for a third of the data. They are planned for a later revision, generated by the harness with the parameters this extraction recorded (20 ppm tolerance, a/b/y ions, H2O/NH3/SO3/H3PO4 losses, isotopes to 4).

What is not here

• The spectra themselves. Peak m/z and intensity arrays are excluded; the corpus is published properly at InstaDeepAI/InstaNovo, joinable on usi.
• Per-peak tensors (peak_mask, spectra_mask, targets, precursors) — model plumbing, reconstructible from the corpus.
• Unlabelled spectra. These are PSM-annotated rows only.

usi is not a unique key

1,000,000 rows carry 999,720 distinct usi values. The repeats are timsTOF spectra: their identifier is really the (frame, scan) pair, and the USI keeps only scan, so two spectra from different frames collapse onto one identifier. In the 100k config, 100,000 rows carry 99,997 distinct values.

Join on usi with first-occurrence semantics, or deduplicate first. Nothing here is ordered by usi, and sample_idx is a position within one extraction run — it is not comparable between the two configs.

Reproducing

The embeddings come from the released checkpoint, so they are reproducible from the model and the corpus:

from instanovo_fm.model.encoder import FoundationModel
model, config = FoundationModel.from_pretrained("instanovo-fm-v0.1.0")

The UMAP layouts, the figure and an interactive explorer over the same spectra are described in the repository, which also holds the script that produced these shards (scripts/release/prepare_embeddings.py).

Licence

CC BY-NC-SA 4.0. These are derived from public PRIDE submissions through a non-commercially-licensed model, so the model's terms carry over. The underlying spectra remain subject to EMBL-EBI terms of use.

Citation

@article{instanovofm,
  title   = {Learning from tandem mass spectra at scale with a self-supervised foundation
             model for proteomics},
  author  = {Nieuwoudt, Mechiel and Reverenna, Marco and Patel, Divanisha
             and Catzel, Rachel and Houngue, Isaac H.J. and Daniel, Jemma
             and Eloff, Kevin and Santos, Alberto and Lopez Carranza, Nicolas
             and Jenkins, Timothy P. and Van Goey, Jeroen
             and Kalogeropoulos, Konstantinos},
  year    = {2026},
  journal = {bioRxiv},
  doi     = {10.64898/2026.09.03.747733},
  url     = {https://doi.org/10.64898/2026.09.03.747733},
  note    = {Preprint}
}