chrisagrams/ms2-peptide-replicate-retrieval

MS2 Peptide-Replicate-Retrieval Benchmark

A benchmark for evaluating spectrum-embedding models for tandem mass spectrometry (MS2). It is a set of real experimental MS2 spectra, each labelled with the peptide it was identified as (a peptide-spectrum match, PSM), pooled so that every peptide is represented by many replicate acquisitions. The task: does an embedding map replicate spectra of the same peptide close together?

• 15,649 spectra
• 1,000 unique peptide/charge labels (≈ 15 replicate spectra per label)
• precursor charges 1–5

Why this exists

Spectrum-embedding models compress an MS2 spectrum into a single dense vector so that spectra can be clustered, deduplicated, or searched by nearest-neighbour instead of by exhaustive spectral matching. The quality that matters for those uses is replicate retrieval: two acquisitions of the same peptide should land near each other in embedding space. This set isolates that property — many labelled replicates per peptide — so embeddings can be compared with the same three metrics regardless of how they were produced.

Metrics

Given one L2-normalised vector per spectrum, with peptide/charge as the label:

metric	definition
Hit@1	leave-one-out: fraction of spectra whose nearest (cosine) neighbour shares the same label
MAP	mean average precision of that same leave-one-out ranking
PairF1	pairwise F1 of a spherical k-means clustering (k = #labels) against the ground-truth labels

A reference implementation of all three (matching denominators, singleton handling, and spherical-k-means details) lives in msdelta/replicate_retrieval.py in the model repo.

Schema

Each parquet row is one spectrum:

column	type	notes
dataset_id	string	source acquisition/dataset id
spectrum_number	int32	scan index within the source
peptide	string	identified peptide sequence (label, with charge)
charge	int16	precursor charge (1–5)
score	float	identification score
appearance_count	int32	# replicate acquisitions of this peptide
ret_time	float	retention time
precursor	float	observed precursor m/z
mz	list<float>	peak m/z values
intensity	list<float>	peak intensities (same length as mz)

The label used by the benchmark is f"{peptide}/{charge}".

Usage

from datasets import load_dataset
ds = load_dataset("<your-username>/ms2-peptide-replicate-retrieval", split="test")
print(ds[0]["peptide"], ds[0]["charge"], len(ds[0]["mz"]))