marin-dna/functional-enhancer
marin-dna/functional-enhancer Human-anchored 255 bp vertebrate sequences from the Zoonomia 447-mammal Cactus alignment and official version-matched UCSC hg38-to-target liftOver chains. This draft covers the enhancer region cohort with all species scope and preserves source FASTA/2bit letter case. Non-human rows project only the central human nucleotide and extract the 255 bp target window centered on its unique mapped locus. For the 28 non-mammalian targets, the stable… See the full description on the dataset page: https://huggingface.co/datasets/marin-dna/functional-enhancer.
marin-dna/functional-enhancer
Human-anchored 255 bp vertebrate sequences from the Zoonomia 447-mammal Cactus alignment and official version-matched UCSC hg38-to-target liftOver chains. This draft covers the enhancer region cohort with all species scope and preserves source FASTA/2bit letter case.
Non-human rows project only the central human nucleotide and extract the 255 bp target window centered on its unique mapped locus.
For the 28 non-mammalian targets, the stable ucsc_multiz100way alignment_source value identifies the pinned UCSC assembly cohort; the actual projection operation uses the matching checksum-pinned pairwise liftOver chain.
Anchor eligibility uses the pipeline's pinned phyloP conservation filter. Sequence case is independent of that filter: lowercase bases preserve source repeat masking, uppercase bases preserve source non-repeat-masked sequence, and conservation scores never rewrite emitted characters or case.
Produced by the commit-pinned vertebrate projection pipeline.
Human anchor provenance
Human anchors were constructed on GRCh38 from the complete Ensembl release 115 GTF with all qualifying transcripts; RefSeq and canonical-transcript-only filtering were not used. The enhancer input is ENCODE SCREEN Registry V4 dELS/pELS. The five disjoint classes use the base-priority order cds > utr3 > tss_region > ncrna > enhancer followed by the source-arm window-majority gate. The shared projection catalog requires at least 10% human bases above the pinned phyloP cutoff; this initial training dataset contains only the nested at-least-20% catalog. Projected 10–20% rows are excluded and retained as a separate producer artifact.
Coordinates and sequence semantics
Human source coordinates use GRCh38/hg38 primary chromosomes. All human and target coordinates are 0-based and half-open. Every emitted sequence is exactly 255 bases, preserves its source FASTA/2bit letter case, and is oriented to the human anchor.
The published format is zstd-compressed JSON Lines under data/train/ and data/validation/. The release is validated before upload against a producer-keyed manifest of file sizes and SHA-256 checksums, and the immutable Hub revision exposes each large-file SHA-256 identifier.
license: other preserves the source-specific terms of the public genome assemblies, annotations, and alignments rather than relicensing those inputs.
Splits
train: 25,364,652 rows after removing the validation sample and applying the configured reverse-complement augmentation.validation: 16,384 original-orientation rows sampled uniformly without replacement with seed 517 before augmentation (4,194,304 tokens including BOS).
The split is row-level and does not stratify by chromosome, species, or human anchor. Different species projections from one human anchor may occur on opposite sides of the split. The reverse complement of a selected validation row is excluded from training.
The selected target manifest contains 135 family-deduplicated projection targets; human reference rows are added separately once per anchor.
Schema
query_name:Stringsource_chrom:Stringsource_start:Int64source_end:Int64region_label:Stringspecies:Stringalignment_name:Stringassembly:Stringtaxonomy_id:Int64family:Stringclade:Stringphylogenetic_rank:Int64alignment_source:Stringt_chrom:Stringt_start:Int64t_end:Int64t_strand:Stringt_src_size:Int64pre_resize_t_start:Int64pre_resize_t_end:Int64fragment_count:Int64aligned_bases:Int64sequence:Stringaugmentation:String
Intended use and limitations
This dataset is intended for genomic language-model research and is not a clinical resource. Assembly quality, alignment gaps, repeat masking, family-deduplicated species selection, human conservation selection, and the center-nucleotide acceptance contract affect the observed sequence distribution. Projecting the center nucleotide does not establish that both 127 bp target flanks are homologous to the full human anchor.
