Datasets and Accessions

Datasets and ENCODE accessions

All inputs are public: chromatin & Hi-C from ENCODE, expression labels from Xpresso/FANTOM (Zenodo), genome hg38.

ENCODE accessions (per cell)

DNase (accessibility) + H3K27ac (activity; 2 filtered bio-reps where available) + H3K27ac narrowPeak (the encoder's summit source) + cell-specific Hi-C:

Cell (Roadmap)	DNase	H3K27ac rep(s)	narrowPeak	Hi-C (.hic)
K562 (E123)	ENCFF257HEE	ENCFF232RQF	ENCFF544LXB	ENCFF621AIY
GM12878 (E116)	ENCFF729UYK, ENCFF020WZB	ENCFF269GKF, ENCFF201OHW	ENCFF023LTU	ENCFF318GOM
H1 (E003)	ENCFF761ZRE	ENCFF860ABR, ENCFF693IFG	ENCFF689CJG	— (none in ENCODE — 5C/ChIA-PET only)
HepG2 (E118)	ENCFF691HJY	ENCFF862NDZ, ENCFF926NHE, ENCFF745JCH	ENCFF392KDI*	ENCFF805ALH
HUVEC (E122)	ENCFF091KTX	ENCFF374DGO, ENCFF609TUB	ENCFF077LGZ	ENCFF091YKP
NHEK (E127)	ENCFF117RNM	ENCFF770JWP, ENCFF051NTC	ENCFF666UYC	ENCFF776JNR

• K562 ships single-rep — its H3K27ac reps differ hugely in depth, so pooling dilutes.
• GM12878 uses both filtered reps per assay.
• HepG2 has 3 filtered H3K27ac reps (mean-pooled); its narrowPeak (*) is inferred, not cited.
• Wired in config/samples.tsv, config/encoder_narrowpeaks.json, and the per-cell BAM manifests under config/.

Other inputs

• Genome: hg38.fa → data/reference/hg38/hg38.fa.

• ABC reference: bash scripts/download_abc_reference.sh data/reference/hg38.

• Expression labels + 12-fold CV split: Zenodo 13232430 → data/.

• ABC average Hi-C (hg38; used automatically when a cell-specific .hic is absent or unavailable): ENCODE ENCFF134PUN (annotation ENCSR382HAW, 5 kb). Split it before running the pipeline:

  python scripts/split_avg_hic.py --in /path/to/ENCFF134PUN.bed.gz \
    --out data/reference/abc_avg_hic/by_chrom

  This is the default reference.average_hic_dir in config/config.yaml; a valid cell-specific .hic takes precedence.

Download destination and pinned inputs

The general downloader defaults to downloads/, but config/samples.tsv expects BAM and Hi-C files under data/{cell}/.... Always specify the destination:

python scripts/download_encode_data.py --cell-types K562,GM12878 --output-dir data

The live ENCODE query is discovery-oriented and may select different files as ENCODE changes. This reproduction pins exact accessions in config/samples.tsv; verify the expected K562 inputs before Stage 1:

test -f data/K562/DNase/ENCFF257HEE.bam
test -f data/K562/H3K27ac/ENCFF232RQF.bam
test -f data/K562/HiC/ENCFF621AIY.hic

Do not silently substitute a newly selected accession for a documented reproduction input.