marin-dna/zoonomia-v1-v4_ccre_enhancer_centered-order

bolinas-dna/zoonomia-v1-v4_ccre_enhancer_centered-order

An enhancer-CENTERED training set for issue #351, built by the `snakemake/zoonomia_projection_dataset` pipeline (`workflow/rules/centered.smk`) at commit `8127acfea5aa`.

Provenance

Each training window is defined directly from an ENCODE cCRE V4 enhancer (dELS + pELS): one 255 bp window centered on the cCRE midpoint (make_enhancer_anchors, keep-all — clustered enhancers each keep their own window), rather than a uniform genome tile labelled by cCRE overlap. Windows are then conservation-filtered (phyloP-447way proportion_conserved ≥ 0.20) and exon-subtracted (no CDS / UTR / ncRNA-exon / TSS overlap) with the same machinery as the tiled arm, and natively projected via halLiftover to the 19-species one-per-order Zoonomia cohort — its own scoped projection, not a subset of the 108-family set.

Contains 116,162 distinct human enhancer anchors → 4,237,620 training samples (anchors × up to 19 mammals × reverse-complement augmentation) — the exact row count across all JSONL.zst shards.

This is the centered arm of exp351; it is compared head-to-head against the tiled control `bolinas-dna/zoonomia-v1-v4_ccre_noexon_enhancer-order` (exp326's curated v4_ccre_noexon_enhancer at the same one-per-order cohort) — the two differ only in window anchoring (uniform grid tile vs enhancer-centered).

Schema

A single train split of JSONL.zst shards at data/train/shard_NNNN.jsonl.zst; key field sequence is exactly 255 bp, strand-aware.

Source code

• —Pipeline: snakemake/zoonomia_projection_dataset (latest)
• —Pinned to this dataset's build: commit `8127acfea5aa`
• —Centering + filters: `workflow/rules/centered.smk` → region_labels.make_enhancer_anchors / select_anchors_noexon
• —Tiled control: `bolinas-dna/zoonomia-v1-v4_ccre_noexon_enhancer-order`