bolinas-dna/genomes-v5-validation-intervals-v30_255_255

Conservation-encoded human validation probe for the v30 (projected conserved enhancers) region recipe, matched to the bolinas-dna/genomes-v5-genome_set-*-intervals-v30_255_128 training sets. Produced by the snakemake/training_dataset pipeline (commit 8db58254831f).

Size

16,384 sequences in validation.parquet (a subsample of up to 16,384 human windows). Exact row count — HuggingFace's auto-generated count can be wrong.

Region recipe: v30

Cross-species enhancers: human ENCODE cCRE ELS with ≥20 conserved bp (phyloP-241way ≥ 2.27), projected onto each target genome by mmseqs2 best-hit alignment (-s 7.5 --max-accept 1), resized to 255 bp and intersected with scannable (defined minus low-quality-excluded exons). On human this is the native set; on other genomes it is the alignment projection.

Built on the human genome (GRCh38 / GCF_000001405.40).

Schema

A single validation.parquet. Each row:

Field	Type	Description
id	str	{chrom}:{start}-{end} — 0-based half-open, bare RefSeq chrom names.
seq	str	255 bp DNA with conservation-encoded case (see below).

Sequence case = phyloP conservation

Each base's case is overwritten to encode conservation (the source soft-masking is discarded):

• Uppercase — phyloP-241way (Cactus 241-way alignment) score ≥ 2.27 at this position (conserved).
• Lowercase — score < 2.27, or there is no alignment / the score is missing (NaN, which compares false). Non-conserved and unaligned bases are not distinguished.

⚠️ This is the opposite convention to the training sets, where lowercase means repeat-masked, not non-conserved.

Construction

1. Build the v30 interval set on the human genome and tile into 255 bp / 255 bp (non-overlapping) windows.
2. Subsample up to 16384 windows (seed 42).
3. Re-encode each base's case by phyloP-241way ≥ 2.27.

Provenance

• Pipeline: snakemake/training_dataset @ commit 8db58254831f
• Matched training family: bolinas-dna/genomes-v5-genome_set-*-intervals-v30_255_128