Taykhoom/chip-seq-glm-tang

GLM-Tang Task 2: GM12878 ChIP-seq Binding

This dataset contains ten binary transcription-factor binding tasks from Task 2 of Tang et al. Given an exact 200 nt sequence from GM12878, predict whether it is centered on a factor-specific ChIP-seq peak or a matched DNase-accessible negative.

Each transcription factor is an independent task identified by factor.

Loading

from datasets import load_dataset

dataset = load_dataset("Taykhoom/chip-seq-glm-tang", "corrected")
ctcf_train = dataset["train"].filter(lambda row: row["factor"] == "CTCF")
ctcf_test = dataset["test"].filter(lambda row: row["factor"] == "CTCF")

Archive dataset size

Split	Rows
train	262,228
validation	37,459
test	74,929
Total	374,616

The archive config contains exactly 187,308 positive and 187,308 negative rows with the distributed partitions unchanged.

Use corrected for scientific evaluation. It excludes 752 negative rows whose 200-nt interval overlaps a factor-specific positive source peak by at least one base, then assigns interval-connected components wholly to one split per factor. It contains 373,864 rows: 261,705 train, 37,385 validation, and 74,774 test. Contaminated negatives are excluded, never relabeled. The one-base threshold is deliberate: any shared genomic base duplicates input content, and transitive overlaps remain in the same component.

Factors:

ATF2, Arid3, BACH1, CTCF, ELK1, GABPA, MAX, REST, SRF, ZNF24

Arid3 is the archived filename label for the ARID3A experiment.

Task construction

• Positive rows are 200 nt windows centered on the corresponding factor-specific GM12878 ChIP-seq peak.
• Negative rows are 200 nt windows centered on GM12878 DNase hotspots.
• The source study downsampled negatives to balance the classes.
• The archive supplies 70/10/20 train/validation/test partitions.

Every row is tied to the exact published source BED:

• ten factor peak files from Majdandzic Supplementary Table S1;
• GM12878 DNase file ENCFF235KUD for negatives.

The selected source peak accession, source BED line, original peak interval, and number of matching source candidates are retained.

Sequence and target contract

sequence is always the exact archived 200 nt benchmark input decoded from the HDF5 one-hot array. It is never replaced by an assembly-normalized sequence.

label is:

1 = factor-specific ChIP-seq peak
0 = GM12878 DNase hotspot negative

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

Columns

Column group	Description
split, sequence, label, example_id	Selected config partition, exact input, target, and unique archive row ID.
archive_split	Original partition; present in corrected only.
source_index, factor, cell_line	Source HDF5 row, task identity, and constant GM12878.
sequence_group_id, archive_sequence_split_leakage	Reverse-complement-canonical same-factor sequence group and whether it spans archive splits.
interval_component_id, archive_interval_component_split_leakage	Same-factor connected component of source intervals linked by at least 1 bp overlap, and whether it spans archive splits.
negative_overlaps_positive_peak	Whether a negative 200-nt interval overlaps any pinned positive source peak for that factor.
source_chrom, source_start, source_end, source_strand	Zero-based, half-open hg19 source window selected from the published BED.
source_coordinate_count, source_coordinates_unique	Number of distinct matching coordinate tuples and whether exactly one exists.
source_bed_record_count	Number of matching BED records, counted separately from coordinate uniqueness.
source_peak_accession, source_peak_line	ENCODE file accession and one-based decompressed BED line.
source_peak_start, source_peak_end	Original source peak/hotspot interval before 200 nt centering.
source_genome_build	Always GRCh37/hg19.
source_id	Stable source-window identifier.
chrom, start, end, strand	Nullable strict center-lifted hg38 200 nt window.
liftover_start, liftover_end	Nullable raw lifted two-base source-center anchor.
liftover_match_count, coordinates_unique, selected_coordinates_unique	Strict liftover candidate count and uniqueness flags.
coordinate_resolution, coordinates_available	Mapping status and whether an hg38 annotation was retained.
hg38_sequence	Nullable 200 nt sequence extracted from the strict hg38 center lift.
hg38_sequence_matches_source, hg38_sequence_edit_distance	Comparison between sequence and the explicit hg38 derivative.
genome_build	GRCh38/hg38 when coordinates are available; null otherwise.

All source-BED candidates, including ambiguous alternatives, are retained in the processing release's coordinate-recovery sidecar. That intermediate distinguishes the original 200-nt hg19 source window (source_window_start/end) from the two-base source center anchor (source_center_start/end) and its lifted hg38 anchor (liftover_start/end). The dataset columns above remain the final 200-nt source and hg38 windows.

hg38 annotation

The hg38 fields are optional annotations, not benchmark replacements.

• 374,586 rows have a strict unique source-center liftover.
• 30 rows remain unresolved and have null hg38 fields.
• 1,374 mapped rows differ between hg19 benchmark input and hg38 derivative.
• Those rows require 3,743 total edit operations; the maximum is 50.

Archive and corrected splits

The archive split has 1,651 interval-connected components spanning partitions, affecting 3,384 rows. Exact-sequence grouping alone detects only 272 affected rows. The corrected config uses the stronger interval components and has zero cross-split source-interval overlap.

The associated one-hot evaluator must transpose GradientCorrection arrays from (rows, 4, 200) to (rows, 200, 4) exactly once before Conv1D. Corrected evaluator outputs identify this protocol explicitly; legacy metrics without that marker used ambiguous orientation.

Processing and verification

The HDF5 checksum verification, exact source-BED matching, strict center liftover, candidate sidecar, row-order checks, and the root manifest 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/chip-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

Majdandzic, A., Rajesh, C. & Koo, P. K. Correcting gradient-based interpretations of deep neural networks for genomics. Genome Biology 24, 109 (2023). https://doi.org/10.1186/s13059-023-02956-3

License

The Tang and GradientCorrection Zenodo releases are CC BY 4.0. ENCODE data retain the ENCODE data-use and attribution requirements.