Datasets:
Download README.md from SinitcynLab/electron_fragmentation: direct link, hf CLI and curl.
- Browser
- Download file 5.93 kB
-
https://huggingface.co/datasets/SinitcynLab/electron_fragmentation/resolve/refs%2Fpr%2F1/README.md
- Command line
-
hf download hf://datasets/SinitcynLab/electron_fragmentation@refs/pr/1/README.md
-
curl -L -o README.md https://huggingface.co/datasets/SinitcynLab/electron_fragmentation/resolve/refs%2Fpr%2F1/README.md
pretty_name: Electron fragmentation MS/MS scan tables (ETD, EThcD, ETciD)
size_categories:
- 10M<n<100M
tags:
- mass-spectrometry
- proteomics
- de-novo-sequencing
- etd
- ethcd
- etcid
- tabular
configs:
- config_name: scan_table_v1
default: true
data_files:
- split: etd
path: scan_table_v1/fragmentation=ETD/*.parquet
- split: ethcd
path: scan_table_v1/fragmentation=EThcD/*.parquet
- split: etcid
path: scan_table_v1/fragmentation=ETciD/*.parquet
- config_name: spectra_only_wave6
data_files:
- split: etd
path: spectra_only_wave6/fragmentation=ETD/*.parquet
- split: ethcd
path: spectra_only_wave6/fragmentation=EThcD/*.parquet
- split: etcid
path: spectra_only_wave6/fragmentation=ETciD/*.parquet
Electron fragmentation MS/MS scan tables
This repository contains experimental electron-fragmentation MS2 spectra prepared for a later Casanovo de novo sequencing study. It is an evidence-preserving scan table, not a filtered training set or a claim that every PSM is correct. No model was fine-tuned as part of this release.
| Configuration | Spectra | Native runs | Source studies | ETD | EThcD | ETciD | Annotation status |
|---|---|---|---|---|---|---|---|
scan_table_v1 |
52,761,622 | 2,487 | 18 | 24,511,137 | 18,139,687 | 10,110,798 | Authors, Sage and MSFragger fields when available; missing values retained |
spectra_only_wave6 |
1,195,849 | 288 | 18 | 301,275 | 314,228 | 580,346 | All three annotation fields are null pending source-aware validation and local searches |
Total: 53,957,471 spectra from 36 source accessions in 22 Zstandard-compressed Parquet shards (125.8 GiB). Each configuration has three mode partitions named etd, ethcd, and etcid in the Hugging Face loader. These are fragmentation modes, not train, validation, or test splits. The two configurations have the same schema but represent different annotation stages.
What each row contains
One row corresponds to one native electron-mode MS2 scan, including scans without any sequence assignment. mz and intensity are full experimental centroid peak arrays; no Casanovo top-peak truncation has been applied. The table also stores the source accession and run, native scan ID and USI, retention time, precursor m/z and charge, species/sample origin/enzyme context, instrument and analyzer, native filter and activation metadata, and ETD/supplemental activation settings.
The nested authors, sage, and msfragger structs hold one selected rank-one PSM per scan when evidence exists. They contain sequence, modification, target/decoy, score, PSM/peptide q-values where available, method, evidence source, FDR scope, and exact/I-L-collapsed FASTA membership flags. High-q PSMs are retained, and a missing q-value is unknown rather than zero. q-values can have different scopes and are not a collection-wide FDR guarantee. See metadata/scan-table-v1-policy.json for the precise interpretation. search_fasta_sha256 records the assigned search reference where one was used.
ETciD means ETD with supplemental CID activation. etd_setting is the instrument's ETD setting and is not a collision energy. hcd_energy, cid_energy, collision_energy, and supplemental_collision_energy are populated only where recorded by the native file. Some runs include specialized methods such as activated-ion ETD; consult the source catalogue and native activation fields before grouping them with ordinary ETD.
Use and provenance
from datasets import load_dataset
rows = load_dataset(
"SinitcynLab/electron_fragmentation", "scan_table_v1", split="etd", streaming=True
)
first_scan = next(iter(rows))
For a local Parquet subset, scan only the desired mode and use dataset_id/species filters with PyArrow or DuckDB. The mzml_path, search_fasta, and PSM source_file fields record paths in the preparation environment and are provenance strings, not downloadable paths. Source study links, species and protease context, row counts, and native run counts are in metadata/source_catalogue.csv. Original vendor RAW, mzML and search FASTA files are not included here; obtain them from their source repositories.
The repository includes metadata/release.json with the shard SHA-256 checksums and row counts, source assembly manifests, a schema description, and an integrity report. The 52.8-million-scan table passed its complete-scope integrity check. The newer 1.2-million-scan expansion passed its manifest run/count checks, but its author/search annotations are still awaiting reconciliation. More recent search reruns and recalibration sidecars in the preparation workspace have not been merged into this immutable v1 table.
Limits and intended filtering
Sequence assignments from library search are conditional on the search FASTA, chemistry and FDR method. An I/L choice in a protein database does not prove that the spectrum contains a diagnostic w ion; the table makes no I/L accuracy claim. Many spectra have incomplete backbone coverage, and no quality metric or train/evaluation selection is applied here. Before training or evaluating, filter target status, label agreement, q-value/PEP, FASTA membership, supported modifications, precursor match, backbone coverage, source-study leakage, and peptide-group overlap. Treat spectra_only_wave6 as unlabeled until its source-specific annotations are validated.
This is a compilation of spectra from third-party public deposits. No blanket license for the underlying source spectra is asserted here. Follow the terms and attribution requirements of each original accession, listed in the source catalogue, when redistributing or reusing their data. Please cite the original studies as well as this compilation.