theGreatHerrLebert/timstof-dda-pasef-cc0

Claudius timsTOF DDA-PASEF PSM corpus

A large, CC0, multi-workflow corpus of peptide-spectrum matches from timsTOF DDA-PASEF runs, built to train and benchmark peptide-property predictors (fragment intensity, ion mobility / CCS, retention time, charge) and to configure timsTOF simulators.

Design principle — a reference corpus, not a pre-baked training set. We include broadly, filter minimally at build, and label richly so you can reproduce any filtering decision yourself. Every confidence/quality signal is a column; nothing you might want to filter on was thrown away.

What's novel

To our knowledge the first public timsTOF DDA-PASEF PSM resource that ships:

1. Dual-engine (Sage + FragPipe) agreement labels per PSM and per fragment;
2. per-precursor RT and ion-mobility peak shapes (FWHM, σ, skew, fit R², trace SNR) — not just apex point estimates;
3. per-precursor, per-PASEF-event collision energies (raw volts);
4. matcher-consistent dual-engine b/y fragment intensities.

Cohort

58 datasets, 100% CC0 (PRIDE-verified, per-dataset evidence + retrieval dates retained), timsTOF family instruments (Pro / Pro2 / HT / SCP / fleX / Ultra), DDA / PASEF acquisition. Tryptic, HLA / immunopeptidomics, and several PTM-enriched workflows.

Tier	rows	train / validation / test
Tier 1 — precursors	33,717,207	27,128,597 / 3,205,642 / 3,382,968
Tier 3 — b/y fragments	504,097,214	404,005,449 / 48,374,948 / 51,716,817

(Per-dataset counts + provenance: manifest.json.)

Tiers

• tier1_psms — one row per accepted precursor identification (raw_file, precursor_id): sequence + modifications, charge, m/z, RT apex (raw + Sage cross-run-aligned rt_aligned), ion mobility, per-PASEF-event collision energy, RT/IM peak-shape labels (ms1_{rt,im}_{fwhm,sigma,skew,r2,snr}), isotope envelope, per-engine scores, engine-agreement, and derived width_reliable / strict presets.
• tier3_fragments — one row per matched b/y fragment (imspy-rematched for both engines under identical settings): calculated + experimental m/z, intensity, ppm error, per-engine match flags, Sage-native provenance.

The full field-by-field schema (units, null semantics) is in SCHEMA.md.

Filtering & certainty policy (read this before filtering)

• Inclusion floor: rank-1 target PSMs accepted by at least one engine at that engine's reported q ≤ 0.01. ⚠️ This is a union of two engine-specific 1%-FDR sets — not a guaranteed corpus-level 1% FDR. Treat the per-engine q/pep as the confidence signals.
• Filter up, freely: every score is exposed (sage_qvalue, sage_pep, fragpipe_qvalue, fragpipe_pep, fragpipe_probability, hyperscores, n_engines, sage_cosine, isotope_cosim). Applying a stricter reported-q threshold selects higher-confidence PSMs — it does not recompute FDR; post-hoc combinations don't retain a nominal FDR.
• Engine agreement is a label, not a gate — single-engine PSMs are kept and flagged (n_engines == 1). Want the high-confidence subset? Use the documented strict boolean column.
• Peak-shape reliability: width labels carry ms1_*_r2 + ms1_*_snr + the *_width_reliable flags. ⚠️ The SNR-reliability band is provisional; gate width labels on width_reliable for peak-shape work.

Splits

Peptide-hash on sequence_normalized → ~80 / 10 / 10 train / val / test (seed 0) — no peptide leakage across splits. Dataset/group-level (sibling PXD) splitting is deferred to a later version. The split column is in every row.

Validation — a model trained on this corpus

Training a unified peptide-property model directly on these parquets (from a cross-instrument pretrained checkpoint, all 58 datasets) reaches, on the held-out test split:

Head	Metric
MS2 b/y intensity	spectral angle 0.744
Ion mobility (CCS)	normalized MAE 0.0184
Retention time	Pearson 0.876 (native rt_aligned)
Charge	accuracy 0.809

The model fine-tunes to timsTOF-ready in a few epochs from pretraining — the corpus carries trainable signal across all four heads, at parity with an equivalent model trained from the raw Sage outputs. (Using raw rt_seconds instead of rt_aligned costs ~3 pp RT Pearson — prefer rt_aligned.)

Usage

from datasets import load_dataset

# precursor properties (RT / IM / charge / CE / peak shapes)
t1 = load_dataset("claudius-proteomics/timstof-psms-cc0", "tier1_psms", split="train")

# high-confidence subset (one boolean)
strict = t1.filter(lambda r: r["strict"])

# b/y fragment intensities (spectrum prediction)
t3 = load_dataset("claudius-proteomics/timstof-fragments-cc0", "tier3_fragments", split="train")

Or read the parquet directly with pyarrow / polars (the files are plain zstd parquet; use predicate pushdown on accession / split for subsets).

Limitations

• The inclusion floor is a per-engine q-union, not a corpus FDR (see above).
• b/y fragments only (no neutral-loss / immonium / internal ions).
• Fragment matching uses a fixed mass→UNIMOD converter; fragments for unmapped modifications are skipped — PTM-heavy datasets have reduced fragment coverage (tracked, not silently dropped).
• The SNR-reliability band is provisional; rt_aligned is Sage's per-dataset cross-run alignment (not a universal iRT).
• Decoys are not shipped in this release (a decoy companion for FDR recomputation / rescorer training is planned).

Versioning

Semantic, immutable revision tags on this repo (v0.1, v0.2, …); a published revision is never edited in place. Each revision ships a manifest.json (dataset list, build-pipeline commit, filter parameters, schema version) so the release is reproducible from the source artifacts.

License & citation

All datasets are CC0 (PRIDE-verified; per-dataset license + retrieval date in manifest.json). Please cite the originating PRIDE accessions (DOIs listed in the dataset metadata) and the Claudius processing pipeline (citation TBD).