Dataset Viewer
Auto-converted to Parquet Duplicate
n_proteins
int64
mask_mode
string
Q5
float64
per_class
dict
SOV
dict
topology_accuracy
float64
sp_tm_confusion
dict
bootstrap
dict
9,999
split
0.917483
{ "i": { "precision": 0.8911648384108761, "recall": 0.980002025212898, "f1": 0.9334745750572272, "support": 1994852 }, "o": { "precision": 0.9581757123986336, "recall": 0.8664895988959304, "f1": 0.9100291254170542, "support": 1536135 }, "H": { "precision": 0.919460176760887...
{ "i": 0.9745243080666607, "o": 0.8160562328827037, "H": 0.881367466778817, "B": 0.7297525624175409, "S": 0.9783007094501648, "all": 0.9005172645268766 }
0.740974
{ "goldSP_predTM": 279, "goldTM_predSP": 973 }
{ "n_boot": 1000, "alpha": 0.05, "Q5": { "mean": 0.9176560440658355, "lo": 0.9121205207443579, "hi": 0.9230444816255653 }, "topology_accuracy": { "mean": 0.7411314131413141, "lo": 0.7323732373237324, "hi": 0.7494774477447745 }, "recall": { "i": { "mean": 0.980108119250946...

ralle TM topology dataset (v6.1)

Harmonized per-residue transmembrane topology dataset used to train and evaluate ralle, a single-sequence per-residue TM topology predictor (frozen ESMC embeddings + CNN + light-attention + grammar-CRF head).

This revision supersedes all earlier revisions (see git history / "Files and versions"). Current build: v6.1 — rebuilt 2026-08-25; see audit_rebuild_v6.json for the integrity audit.

Contents

file contents
records_clean_v6.jsonl one JSON per protein: id, sequence, labels (per-residue), tier (evidence 1–3), source, meta
splits_v6.json `id -> "train"
clusters.tsv.gz MMseqs2 cluster assignment used for the leak repair
audit_rebuild_v6.json automated integrity audit (schema, duplicates, leakage fixpoint)
build_summary_rebuild.json build funnel + repair counts
anchorify_report.json benchmark-anchor recovery report

Label scheme

Per-residue, same length as sequence: H TM alpha-helix, B TM beta-strand, S signal peptide, i inside/cytoplasmic, o outside; . = unannotated (ignore index). DeepTMHMM conventions: globular proteins without a signal peptide are all-i; residues after a signal-peptide run are o.

Numbers (v6.1)

records (final) 193,886
train / val / test 91,806 / 7,662 / 94,418
all-vs-all pairs >=30% id / >=50% cov 6,184,249
records moved by leak repair 15,132
residual train<->eval pairs at fixpoint 0

Per-source records: uniprot 92,449, uniprot_globular 51,101, pdb_globular 19,866, htp 13,304, pdbtm 8,065, topdb 5,529, deeptmhmm 3,571, uniprot_beta 1

Provenance & decontamination

Sources (evidence-tiered): PDBTM (structure-verified) > TOPDB v3.3 (experimental topology) > UniProtKB/Swiss-Prot reviewed annotations; human TM proteome consensus (CCTOP); PDB SEQRES globular chains and reviewed signal-peptide proteins as negative/signal sets; DeepTMHMM legacy set reserved for evaluation. Complete all-vs-all MMseqs2 decontamination with union-find component repair — zero train<->eval similar pairs at fixpoint. Full details: docs/SETUP.md, docs/REBUILD_DESIGN.md, and docs/references.md in the GitHub repo.

License & citation

CC-BY-4.0. If you use this dataset, cite the ralle repo and the underlying source databases (PDBTM, TOPDB, UniProtKB/Swiss-Prot, CCTOP, DeepTMHMM) — see docs/references.md for the ranked reference list.

Downloads last month
42