Transformer that reads phosphopeptides directly from tandem mass spectra and places each phosphate on its serine, threonine or tyrosine.
No providers recorded yet. Browse all providers
A phosphate group adds 79.97 Da to a serine, threonine or tyrosine, and a database search finds phosphopeptides by allowing that mass shift on every candidate residue. Each allowed modification multiplies the search space, the engine must still decide which nearby residue carries the phosphate, and peptides absent from the reference proteome cannot be found at all. General-purpose de novo sequencers read the peptide straight from the spectrum, but they were trained mostly on unmodified peptides and have no vocabulary for a phosphorylated residue.
InstaNovo-P is a phosphorylation-specific version of the InstaNovo encoder-decoder transformer. Three new tokens (phosphoserine, phosphothreonine and phosphotyrosine) were added to the decoder's vocabulary, so the model writes the modified residue as it translates the fragment-ion peaks, and the phosphosite is localized as part of the sequence rather than assigned afterwards. It was then fine-tuned on 2.76 million phosphopeptide spectrum matches with a schedule that unfreezes the encoder before the decoder, which let it learn phosphorylation-specific fragmentation without forgetting unmodified peptides.
The model was developed at the Technical University of Denmark with InstaDeep, CompOmics at the VIB-UGent Center for Medical Biotechnology and Ghent University, the Vrije Universiteit Brussel, the University of Manchester and Delft University of Technology. It was posted to bioRxiv in May 2025 and published in Nature Communications in 2026. It is a different model from the self-supervised spectrum encoder InstaNovo-FM, whose phosphorylation results come from a linear probe on frozen embeddings.
InstaNovo-P keeps the architecture of the base InstaNovo v0.1.4 checkpoint: nine encoder and nine decoder layers, 16 attention heads, model dimension 768, about 95 million parameters, with knapsack beam search to enforce the precursor mass. The head and sequence embedding were reinitialized to hold the extended vocabulary. Fine-tuning data came from 29 PRIDE phosphoproteomics projects reprocessed with IonBot in the Scop3P resource, filtered to localization probability above 0.80: 2,760,939 PSMs covering 74,686 unique peptides, mostly human, with phosphoserine at 84.5% of sites, phosphothreonine 12.0% and phosphotyrosine 3.4%. GraphPart split the peptides 70/10/20 by sequence homology. Training ran for 10 epochs with a slanted triangular learning-rate schedule restarted at each unfreezing phase.
On unmodified peptides, peptide recall rose from 64.0% to 68.1% on the ProteomeTools test set and from 70.0% to 71.4% on the unmodified portion of the synthetic 21PTM set. On 21PTM phosphopeptides, held out entirely, recall was 81.4%, against 68% reported for PrimeNovo after it was fine-tuned on that dataset. On an FGFR2 phosphoproteomics dataset searched with FragPipe, recall was 44.1% against 30.7% for the PrimeNovo-PTM checkpoint. Across the full FGFR2 evaluation set at 5% FDR, recall was 44.4%, 30.2% and 22.6% for phosphoserine, phosphothreonine and phosphotyrosine peptides.
InstaNovo-P is a complement to database search in phosphoproteomics studies of kinase signalling. In T47D breast cancer cells with and without FGFR2, treated with its ligand FGF7, running it on every scan raised the average number of phosphopeptides from 3,068 to 5,436 and mapped them to 21.6% more proteins. Targeted proteomics confirmed predicted sites that database search missed, including S926 on LMO7 and S75 on kinectin-1. Immunoblotting showed that RSK1 S732 phosphorylation depends on FGFR2, and cell counts under an RSK1 inhibitor linked RSK1 activity to FGF7-driven proliferation. On G3BP1 the model placed the phosphate on S232 where database search had chosen S230, and site-determining fragment ions confirmed the model's assignment. Of 42 predicted phosphopeptides tested without phospho-enrichment, 23 were detected by targeted assays and four more by DIA.
InstaNovo-P brought de novo sequencing into phosphoproteomics, and the paper presents its fine-tuning recipe as a template for other modifications with a fixed mass shift, such as ubiquitination and dimethylation. Its limits are clear. It works only on data-dependent acquisition, because it needs precursor mass and charge. Recall falls with peptide length and with the number of phosphosites. Tyrosine sites are recovered least often, reflecting their scarcity in the training data. On an Orbitrap Astral dataset it recovered only 9.4% of phosphopeptides. The training code is Apache-2.0, while the checkpoint is released under CC BY-NC-SA 4.0 and cannot be used commercially.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.