Taykhoom/eclip-seq-glm-tang

GLM-Tang Task 6: K562 eCLIP Binding

This dataset contains ten binary RNA-binding-protein tasks from Task 6 of Tang et al. Given an exact 200 nt sequence from a transcribed K562 region, predict whether it is centered on an eCLIP peak or a mock-input negative.

The DNA alphabet is used for storage (T rather than U), but the biological context is transcribed RNA.

Loading

from datasets import load_dataset

dataset = load_dataset("Taykhoom/eclip-seq-glm-tang")
ptbp1_train = dataset["train"].filter(lambda row: row["rbp"] == "PTBP1")
ptbp1_test = dataset["test"].filter(lambda row: row["rbp"] == "PTBP1")

Dataset size

Split	Rows
train	87,109
validation	12,443
test	24,895
Total	124,447

RBPs:

HNRNPK, PTBP1, PUM2, QKI, RBFOX2, SF3B4, SRSF1, TARDBP, TIA1, U2AF1

Published task construction

Tang et al. describe:

1. filtering two K562 eCLIP peak replicates;
2. requiring overlap across replicates;
3. applying a replicate-signal consistency filter;
4. extracting 200 nt positive windows around retained peak centers;
5. calling negatives from mock-input BAMs with Piranha;
6. removing negatives overlapping any unfiltered replicate peak;
7. downsampling negatives to approximately the positive count;
8. randomly splitting rows 70/10/20.

The exact historical ENCODE file manifest, mock-input BAMs, negative downsampling seed, and split seed were not released. The Tang HDF5 files are therefore the archive-exact boundary for membership and splitting.

Nine-channel source input

Each HDF5 stores (rows, 9, 200) features.

• Channels 1–4 are verified strict one-hot A/C/G/T and define sequence.
• Channels 5–9 are continuous but undocumented.

The five unknown channels remain in the byte-identical local HDF5 copies. They are deliberately not assigned speculative names or copied into the model-ready Parquet.

Sequence and target contract

sequence is the exact 200 nt string decoded from channels 1–4.

label is:

1 = eCLIP peak
0 = mock-input negative peak

The released models use binary cross-entropy and report accuracy, AUROC, and AUPR independently for each RBP.

Columns

Column	Description
split	Released train, val, or test partition.
sequence	Exact 200 nt archived input decoded from A/C/G/T channels.
label	Binary eCLIP-peak label.
example_id	Stable unique RBP/split/source-row identifier.
source_index	Zero-based source HDF5 row within the RBP and split.
rbp	RNA-binding protein defining the task.
cell_line	Always K562.
chrom, start, end, strand	Selected zero-based, half-open exact hg38 sequence match.
coordinate_match_count	Number of distinct exact hg38 loci reported by the pinned minimap2 invocation.
primary_match_count	Number of reported exact hits on chr1–22, X, Y, or M.
nonprimary_match_count	Number of reported exact hits on alternate/random/unplaced contigs.
coordinates_unique	Whether minimap2 reported exactly one exact locus.
selected_coordinates_unique	Whether the preferred primary/non-primary candidate tier contains one locus.
contig_class	primary or nonprimary for the selected locus.
coordinate_resolution	Exact-match selection method and selected-tier ambiguity.
sequence_matches_hg38	Always true for the selected exact locus.
genome_build	Always GRCh38/hg38.
locus_group_id	Strand-independent RBP/locus identifier used for leakage checks.
split_leakage	Whether the same RBP locus appears across archived partitions.

Coordinate ambiguity

All 120,599 unique sequences have at least one exact hg38 match reported by the pinned minimap2 command.

• 3,620 rows have multiple minimap2-reported exact loci.
• 123 rows remain ambiguous within the preferred candidate tier.
• The selected locus is deterministic, but coordinates_unique should be checked before treating it as the historical experimental locus.

These counts describe the exact alignments emitted with the pinned minimap2 options (-x sr -N 100 --secondary=yes -p 0 -c). They are not claimed to be a mathematically exhaustive enumeration of every exact occurrence in hg38.

No sequence is modified during coordinate annotation.

Split leakage

The archived split contains 1,477 same-RBP, strand-independent loci spanning multiple partitions, affecting 2,954 rows. split_leakage exposes these rows without replacing the released benchmark split.

Processing and verification

The HDF5 byte checks, channel validation, minimap2-reported exact-hit accounting, coordinate re-extraction, row-order verification, and release contract are documented at stable repository links:

• Processing repository: https://github.com/TaykhoomDalal/GLM-Tang-Processing/tree/main
• Task documentation: https://github.com/TaykhoomDalal/GLM-Tang-Processing/blob/main/eclip-seq/README.md
• Reproducibility manifest: https://github.com/TaykhoomDalal/GLM-Tang-Processing/blob/main/manifest.json

Sources and citation

Tang, Z. et al. Genome Biology 26, 203 (2025). https://doi.org/10.1186/s13059-025-03674-8

Van Nostrand, E. L. et al. Robust transcriptome-wide discovery of RNA-binding protein binding sites with enhanced CLIP (eCLIP). Nature Methods 13, 508–514 (2016). https://doi.org/10.1038/nmeth.3810

License

The Tang archive is CC BY 4.0. ENCODE data retain ENCODE's data-use and attribution policy. No separate license is assigned to the undocumented auxiliary HDF5 channels.