Skip to content

Dog peptides: MHC ligands and the Canine PeptideAtlas

Two dog peptide datasets record which peptides of known dog proteins mass spectrometry has observed. Neither searched for noncanonical ORFs, so they don't add ORFs to the dog proteome (why dog has no ncORF catalogue).

For a verified paired RNA/protein study and the latest bounded search for dog Ribo-seq, see the canine source audit.

Dataset What it holds How to get it
kaabinejadian-dog-dla-ligands DLA class I ligands from four dog tumours and three DLA-88 monoallelic cell lines orfdata datasets download (CC0, 25 MB)
peptideatlas-dog-peptides 149,546 tryptic peptides from 18 Labrador tissue, plasma and urine samples, with genome locations orfdata datasets build (built locally from 71 MB of sources)

DLA class I ligands

Kaabinejadian et al. (2026) eluted peptides from DLA class I molecules with the BB7.6 or H58A antibody. PRIDE PXD074485 deposits their PEAKS Studio 11.5 peptide tables (1% FDR), which this dataset imports in full: 76,550 rows, each with its PEAKS area, intensity and spectrum count.

Sample Tumour Breed DLA-88 typing Antibodies
Lola osteosarcoma Black Russian Terrier *501:01 BB7.6, H58A
163828A anal sac adenocarcinoma English Cocker Spaniel *004:02 BB7.6
Lily anal sac adenocarcinoma English Cocker Spaniel *501:01, *501:02, *004:02 H58A
Bogey oral melanoma American Cocker Spaniel *003:02, *006:01, *017:01 BB7.6
HCT116-DLA88… human HCT 116 cells, one DLA-88 allele each — *003:02, *012:01 or *501:01 BB7.6

Each tumour's spectra were searched four times: against UniProt's dog proteome, Basenji and Great Dane (Zoey) Ensembl 104 proteins, and a merged database of four dog assemblies. Each search is its own assay (Lola-BB76:UniProt9615, Lola-BB76:MergedCanFams, ...). A peptide found in two searches of one sample is one observation, not two. The HCT 116 lines were searched against human SwissProt, and their peptides are human proteins, so they sit in the dataset's separate human reference context.

Database search finds 11,363 distinct dog peptides in the UniProt searches and 12,948 across all four databases, leaving out iRT standards (below). For the cell lines, the counts of unique 8–14-mers from database search (591 for DLA-88*003:02, 1,465 for *501:01) match the article's. The article's tumour counts come from its curated supplementary lists, which aren't imported (they are CC BY-NC-ND), and match no single table.

Every row keeps its PEAKS identification type in provenance as found_by: DB Search, DeepNovo (de novo sequencing) or Homolog (SPIDER homology search). Only database-search peptides are guaranteed to occur in a listed protein, so filter on it for a conservative ligand list.

iRT standards

Every run was spiked with Biognosys iRT retention-time peptides, and their spectra appear in every search, often in disguise. 10,629 rows are marked irt_standard, and irt_rule names the first rule that marked them:

irt_rule Rows Marks a row that
irt_sequence 1,179 is a kit peptide or a fragment of one, with I and L equal
kit_only 1,097 PEAKS attributes only to the kit (homology reads of kit peptides)
irt_precursor 1,276 sits at the precursor of a row marked by the first two rules in the same immunoprecipitation: same charge, m/z within 20 ppm, retention time within a minute
de_novo_irt 7,077 is a de novo sequence that PEAKS matches to the kit, or that shares 6 or more consecutive residues with a kit peptide

The precursor rule applies to de novo and homology rows. It applies to database-search rows only in the merged database, which lacks the kit, and only when no search with the kit found the same peptide. Elsewhere PEAKS could choose the kit peptide and didn't.

  • De novo rows: half of them (8,448 of 16,830) are iRT reads, often a kit peptide with residues added or changed (KATLFLQFGAQGSPFLK). Don't take an unmarked de novo sequence for a new ligand without checking its spectrum.
  • Homology rows: 8 are marked by precursor, such as TGFLLDPVLNR, which has exactly the mass of GTFIIDPGGVIR and co-elutes with it.
  • Database-search rows: apart from 1,075 kit peptides, only one is marked: IFLAFQQKAYVAP, matched to dog proteins in the merged database at the precursor of LFLQFGAAGSPFLK.
  • Chance matches: few. Shuffled kit sequences share 6 residues with about 3 de novo rows, and shifting the kit's retention times by 10–20 minutes puts about 14 de novo rows at a kit precursor.

A mapping is unmapped when any of its rows is marked. Release 2026-10-16.1 (importer kaabinejadian-dla-v1) marked only exact kit sequences and kit-only attributions (2,242 rows), so it missed the disguised reads.

from orfdata import DatasetRegistry, ScopedTarget

registry = DatasetRegistry()
ligands = registry.open("kaabinejadian-dog-dla-ligands")  # After downloading it.
dog = next(r for r in ligands.manifest.references if r.taxon_id == 9615)

# Where a peptide was eluted, and how it was identified.
peptide = ScopedTarget(dog, "peptide", "AIKKGLEEL", "amino-acid-sequence")
for m in ligands.measurements(target=peptide, quantity="spectrum_count"):
    print(m.sample_id, m.assay_id, m.value, m.provenance["found_by"])

# Lola's database-search ligands (BB7.6, UniProt search).
lola = {m.target.identifier for m in ligands.measurements(
            reference=dog, assay_id="Lola-BB76:UniProt9615", quantity="spectrum_count")
        if m.provenance["found_by"] == "DB Search" and not m.provenance["irt_standard"]}

# Why a row is marked as an iRT standard.
for m in ligands.measurements(target=peptide, quantity="spectrum_count"):
    print(m.assay_id, m.provenance["irt_standard"], m.provenance["irt_rule"])

Targets are unmodified sequences; each row's modified form (for example M(+15.99)) is in its provenance as peptide. mappings() lists, for each search and identification type, the proteins PEAKS reported for a peptide: unique or ambiguous for database-search and homology matches, and unmapped with a reason for iRT standards and de novo sequences, whose listed proteins stay in the mapping's evidence. The search FASTAs weren't deposited, so no accession can be checked against a sequence.

Runs missing from the deposit

PEAKS reports, for each peptide, the raw file and scan of its best spectrum, and sums its spectra and area over all of the sample's runs. The authors' PEAKS projects for two immunoprecipitations included runs that the deposit lacks, so each row records raw_file_deposited. For 1,695 rows the best spectrum comes from a missing run:

  • Lily (H58A): a second injection of the remaining eluate, 22-1107 "1uL Residue" (fractions 12–30), 299 rows. The first injection, 22-1104, is deposited.
  • DLA-88*012:01: the 20240929 acquisition (fractions 14–25), 1,396 of its 1,635 rows. The line's archive holds only its 20241013 runs (fractions 26–28 in triplicate, and blanks), though the SDRF names the 20240929 run as the assay.

Those identifications come from the authors' tables, but their spectra can't be checked against the deposit. The raw files of 163828A's H58A run have no deposited results and aren't included.

Canine PeptideAtlas

The Canine PeptideAtlas (Kusebauch et al. 2026) reprocessed 138 runs from Labrador Retrievers (13 tissues, plasma and urine; PRIDE PXD052932) at a PSM FDR of 0.001. This dataset imports its 2024-01 build (573):

  • PSM counts for 149,546 peptides in each of the 18 samples where they were observed (528,272 values);
  • every protein each peptide maps to, with its position and, for Ensembl proteins, its ROS_Cfam_1.0 coordinates (exon-spanning peptides have one location per exon);
  • the build's 17,795 protein calls with presence levels (6,423 canonical). Of its 17,307 UniProt and Ensembl calls, 16,959 map to one Ensembl 110 protein with an identical sequence, 151 to several and 197 to none; contaminant and decoy calls aren't mapped.

PeptideAtlas states no licence for its downloads, so orfdata never redistributes this dataset. build downloads the build's 61 MB export and Ensembl 110's dog proteins, checks their SHA-256, and builds the dataset locally (about a minute; 540 MB on disk). After that, open works offline.

orfdata datasets build peptideatlas-dog-peptides

from pyensembl import EnsemblRelease
from orfdata import DatasetRegistry, ScopedTarget

atlas = DatasetRegistry().open("peptideatlas-dog-peptides")  # After building it.
source, ensembl = atlas.manifest.references

# PSMs of one peptide in each sample.
peptide = ScopedTarget(source, "peptide", "AAHEEICTTNEGVMYR", "amino-acid-sequence")
for m in atlas.measurements(target=peptide):
    print(m.sample_id, m.value)

# The PeptideAtlas call for a PyEnsembl dog protein (albumin).
transcript = EnsemblRelease(110, species="dog").transcript_by_id("ENSCAFT00845010439")
protein = ScopedTarget(ensembl, "protein", transcript.protein_id, "ensembl")
for mapped in atlas.mapped_measurements(protein, require_sequence=True):
    call = mapped.measurement
    print(call.target.identifier, call.value, call.provenance["presence_level"])

The ensembl context is exactly PyEnsembl's dog release 110 (ReferenceContext.from_genome(EnsemblRelease(110, species="dog"))), so PyEnsembl protein IDs work directly. The dog RNA dataset uses Ensembl 110 IDs too, in a context that also pins its GTF. Ensembl 105–111 have identical sequences for all 25,710 Ensembl proteins in PeptideAtlas's search set.

Each peptide's mapping (mappings(), keyed by PeptideAtlas accession) keeps its locations as [protein, start, end, chromosome, strand, chromosome_start, chromosome_end, exact]. exact is false for the 4,767 locations where the protein has isoleucine for leucine or the reverse, which PeptideAtlas maps as equal. 204 peptides map only to decoy sequences; they pass the FDR threshold but are probably false identifications, and are unmapped with that reason.

A protein call's PSM count includes peptides it shares with other proteins, so calls don't add up to the build's 13,781,115 PSMs. PeptideAtlas's web page reports 245,884 peptides and 13,716,253 PSMs for the build; its export, which this dataset reproduces, has 149,546 peptides (197,285 distinct modified forms) and 13,781,115 PSMs.