marin-dna/zoonomia-v1-val_promoter

bolinas-dna/zoonomia-v1-val_promoter

Conservation pre-filtered, case-encoded human-genome validation set from the `snakemake/zoonomia_projection_dataset` pipeline (commit `main`).

This is one of seven per-recipe validation parquets (val_cds, val_utr5, val_utr3, val_ncrna, val_promoter, val_enhancer, val_tss_pc) built from the same human-anchored phyloP_447m scoring used to create the cross-mammal training sets `bolinas-dna/zoonomia-v1-v1` and `bolinas-dna/zoonomia-v1-v2`.

Recipe (val_promoter)

Promoter regions defined by ENCODE cCRE V4 Promoter-Like Signature (PLS) class. Each cCRE is centered and resized to exactly 255 bp. No subtraction — PLS sits at the TSS by definition, so its overlap with 5' UTR is the intended biology, and any overlap with CDS in genes with very short 5' UTRs is accepted as part of the natural PLS distribution rather than filtered out. For comparison, val_tss_pc is a gene-centric annotation-driven alternative that uses canonical protein_coding TSSes instead of ENCODE cCREs.

Schema

Sequence case encoding

For each base in seq:

• —Uppercase (A / C / G / T / N): the phyloP447m score at this position is `>= 2.2162` (the calibrated conservation threshold for the 447-way Cactus alignment, set to match phyloP241m's passing-base proportion at 2.27).
• —Lowercase (a / c / g / t / n): either the phyloP score is below the threshold, or there is no alignment at this position (NaN in the bigWig — NaN >= t is False in NumPy, so unaligned bases are encoded as lowercase).

Uppercase positions are confidently conserved across mammals; lowercase positions are either non-conserved or unaligned, and the encoding does not distinguish those two cases.

Construction

1. 1.Build region BED from Ensembl release 115 (canonical transcripts, where applicable) or ENCODE cCRE V4 — see Recipe above.
2. 2.Intersect with defined regions (genome minus N).
3. 3.Tile into 255 bp non-overlapping windows (bedtools makewindows -w 255 -s 255).
4. 4.Score each window against phyloP_447m (UCSC 447-way Cactus, Zoonomia + primates).
5. 5.Conservation pre-filter: keep only windows with proportion_conserved >= 0.2 (i.e. at least 51 of the 255 bases pass the phyloP_447m threshold of 2.2162).
6. 6.Deterministic subsample to ≤ 16,384 rows (seed 42).
7. 7.Extract sequences from hg38 (Ensembl r115) via twoBitToFa -bedPos.
8. 8.Case-encode each base by its per-position phyloP_447m value.

No reverse-complement augmentation — each genomic region appears exactly once. (RC augmentation is a training concern; for evaluation we want deterministic per-locus rows.)

Caveats

• —Recipes are independent probes, not a partition. A small number of bases may appear in multiple recipes (e.g. a val_promoter PLS overlapping a val_utr5 5' UTR; an add_flank(20) on CDS extending into intronic splice signal that abuts a 3' UTR). Don't concatenate the seven parquets without dedup.
• —Lowercase is ambiguous. Lowercase letters could mean either "phyloP score is below threshold" or "no alignment at this position". The encoding does not distinguish.
• —Only autosomes + chrX + chrY are included; the mitochondrial contig is excluded.
• —The val_promoter and val_enhancer cCRE recipes are not restricted to canonical transcripts — those classes are transcript-independent and come directly from ENCODE SCREEN cCRE V4.

Source code

• —Pipeline: snakemake/zoonomia_projection_dataset (latest)
• —Pinned to this dataset's build: commit `main`
• —Library: bolinas.zoonomia_projection_dataset.validation
• —Conservation track: phyloP_447m (`hg38.phyloP447way.bw`)