InstaNovo-FM v0.1.0

First public release: an encoder-only transformer trained to reconstruct masked regions of tandem mass spectra without any peptide-sequence labels, whose frozen embeddings transfer across datasets, instruments and acquisition methods.

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

Checkpoints

Attached below and loadable by id — they download on first use and cache under ~/.cache/instanovo-fm/.

from instanovo_fm.model.encoder import FoundationModel

model, config = FoundationModel.from_pretrained("instanovo-fm-v0.1.0")
FoundationModel.describe_pretrained()      # what each one is

Foundation models. 12 layers, d=768, 12 heads, feedforward 3072, ~89.5M parameters, 230K steps on LCFM, except MCFM as noted.

id	masking	PA bias
instanovo-fm-v0.1.0	Thompson-span	no	the published model — every TS-noPA number in the paper
instanovo-fm-lcfm-ts-pa-v0.1.0	Thompson-span	yes	factorial cell
instanovo-fm-lcfm-sa-nopa-v0.1.0	signal-aware	no	factorial cell
instanovo-fm-lcfm-sa-pa-v0.1.0	signal-aware	yes	factorial cell
instanovo-fm-mcfm-90k-v0.1.0	Thompson-span	no	corpus-scale baseline on MCFM: 9 layers, feedforward 1024, ~40M, 90K steps

The four LCFM models are the 2×2 factorial over masking strategy and pairwise attention bias, the published model is its TS-noPA cell.

De novo sequencers. The foundation encoder plus an InstaNovo decoder, 2.5M steps at batch size 128, warming up over 100K steps to a learning rate of 5e-5.

id	encoder
instanovo-fm-denovo-v0.1.0	fine-tuned	the published sequencer — benchmarked against IN v1.2, Casanovo and XuanjiNovo; unfreezes the encoder at step 100K
instanovo-fm-denovo-frozen-v0.1.0	frozen	retains on average 85% of the fine-tuned peptide recall
instanovo-fm-denovo-scratch-v0.1.0	from scratch	the no-pretraining control

The checkpoints are CC BY-NC-SA 4.0.

Also released

• Corpus: InstaDeepAI/InstaNovo on HuggingFace: 92 public PRIDE submissions reprocessed with FragPipe, in nested confidence tiers. The released models are trained on LCFM, 184.6M PSMs at 1% FDR.
• Corpus-production code: Figshare, 10.6084/m9.figshare.33368752.
• Embedding explorer: instadeepai.github.io/InstaNovo-FM, a UMAP browser over ~100,000 held-out LCFM spectra.