Transformer that reads peptide sequences straight from tandem mass spectra, paired with InstaNovo+, a diffusion model that refines its predictions.
No providers recorded yet. Browse all providers
Bottom-up proteomics identifies a peptide by matching its fragmentation spectrum against theoretical spectra generated from a protein database, which means it can only find what the database already contains. A nanobody from a phage-display library, a snake venom toxin from an unsequenced species, a bacterium contaminating a wound, or a peptide cut at an unexpected protease site all fall outside the search space, and every post-translational modification added to the search multiplies its cost. De novo sequencing reads the peptide directly from the spectrum instead, but before deep learning it was too error-prone to use at scale.
InstaNovo is an encoder-decoder transformer that translates the fragment-ion peaks of a tandem mass (MS/MS) spectrum into an amino-acid sequence, one residue at a time. It is paired with InstaNovo+, a multinomial diffusion model that takes a complete candidate sequence, either InstaNovo's prediction or random noise, and revises all positions together over a series of denoising steps. The authors frame this as mimicking how a human expert sequences a spectrum: anchor on the unambiguous ions first, then revisit the rest with the whole peptide in view. The two models find partly different peptides, so they are most useful run together.
InstaNovo was developed by InstaDeep with the Technical University of Denmark and Delft University of Technology. It was posted as a bioRxiv preprint in August 2023 and published in Nature Machine Intelligence in 2025. It is a different artifact from the later InstaNovo-FM, a self-supervised spectrum encoder from the same group.
InstaNovo has a 9-layer transformer encoder over the embedded peaks and a 9-layer causal decoder, each layer with 16 attention heads, hidden dimension 768 and feed-forward dimension 1,024, for about 95 million parameters. Precursor mass and charge are encoded and concatenated to the encoder output, which the decoder cross-attends to. InstaNovo+ pairs a similar encoder with a diffusion decoder that refines predictions over 20 steps, sampling five candidates per spectrum and keeping the most probable one that fits the precursor mass. The paper's models were trained on ProteomeTools synthetic human peptides: a high-confidence set of 2.6 million spectra and an all-confidence set of 28 million, both covering 742,000 peptides, split 80:10:10 by unique peptide sequence. The released v1.1 checkpoints were retrained on ProteomeTools, MassIVE-KB and additional datasets processed by the CompOmics group at Ghent University.
On held-out yeast in the nine-species benchmark, InstaNovo identified 39,830 peptide-spectrum matches at 5% FDR and InstaNovo+ 52,633, against 39,659 for Casanovo; together they reached 56,230, 41.78% more than Casanovo. Across the eight application datasets, mean peptide accuracy was 48.3% for InstaNovo and 51.5% for InstaNovo+. On a HeLa single-shot run the paper model recovered 49.6% of database-search peptides; the developers report 63% for v1.1.
InstaNovo addresses proteomics problems where the right database does not exist. Applied to 13 nanobodies, it raised the average number of unique peptides per nanobody from 5 to 40, and it extended coverage of the antibody herceptin to 92.87% of the heavy chain and 100% of the light chain. In an immunopeptidomics dataset it found 3,495 HLA peptides that target-decoy search missed, with the 9-mer motif expected of MHC ligands. In human wound exudates it detected peptides from bacteria, and the presence of E. coli and P. aeruginosa was confirmed by 16S rRNA PCR. It also surfaced candidate new species in a marine metaproteome, unreported snake venom sequences, and GluC cleavage products validated by targeted proteomics.
InstaNovo showed that de novo sequencing can add to database search rather than merely approximate it, producing new identifications in immunopeptidomics, antibody sequencing, metaproteomics and degradomics. It has grown into a product line: the InstaNovo-P phosphoproteomics model and the self-supervised InstaNovo-FM build on it, and it sits alongside AdaNovo and XuanjiNovo as a reference point for newer sequencers. Accuracy still varies widely between sample types, and the paper models were trained on synthetic tryptic human peptides with only three modification types. The code is Apache-2.0, but the checkpoints are CC BY-NC-SA 4.0, which excludes commercial use.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.