Genome-wide hg38 reference-base importance (dynseq)

Sources: ENCFF564PMV (HepG2 ATAC ChromBPNet profile contributions),
ENCFF183YWR (liver right lobe ATAC ChromBPNet profile contributions).
These are the two sources supplied in the CVH spec, not the earlier
ENCFF829DSC DNase counts demonstration.

Each .bgz is a BGZF-compressed BED6 with matching .bgz.tbi:
chrom, start, end, reference A/C/G/T, absolute contribution, contribution sign.
Coordinates are zero-based, half-open. Column 6 is score polarity, not DNA strand.
Every scored, finite, nonzero reference A/C/G/T base is retained, on every
source chromosome/contig with scored data. Missing model coverage is not zero.
The original signed value is column 5 times the sign in column 6.
.bgz avoids the .gz suffix sometimes associated with HTTP content decoding.

CVH_Liver_dynseq.gosling.json replaces the two supplied prediction tracks,
removes the separate reference sequence track, and keeps the other tracks.
The initial chr4 window is only a starting view, not a data restriction.
Both dynseq tracks display positives only, with 300px height, letters at
<=400bp and bars at wider views. Zoom to 50-100bp for wider letters.
Each track autoscales independently: compare the numeric y axes, not letter
heights across the two models. Raw scores are not transformed or clipped.

patch_cvh_spec.py INPUT.json OUTPUT.json applies only these three edits to
another copy of your original spec, preserving all other fields.

Generation: python build_bed.py ACCESSION.bigWig ACCESSION.hg38.dynseq.bgz
The converter uses pybigtools, pysam, and bgzip (HTSlib). The exact local reference and
per-chromosome counts are recorded in the matching summary JSON.

BED parser compatibility: Gosling converts strand to numeric 1/-1.
The positive filter must use oneOf: [1], not ["+"]. TSV sources retain strings.
CVH_Liver_dynseq.closeup.gosling.json opens at chr4:73446530-73446580
so letters are immediately visible. The standard spec keeps the 100kb overview.