Taykhoom/functional-random-noannot-gamba

GAMBA Functional Regions: Feature vs Annotation-Free Random

This paired binary representation benchmark asks whether a model can distinguish an annotated functional region from a chromosome- and length-matched random sequence selected outside all retained functional annotation categories.

This is a stricter background than functional-random-gamba, whose controls only avoid anchors from the corresponding category.

Loading

from datasets import load_dataset

causal = load_dataset(
    "Taykhoom/functional-random-noannot-gamba",
    "causal",
    split="all",
)

paper_test = causal.filter(
    lambda row: row["split"] == "test"
    and row["category"] != "noncoding_regions"
)

Dataset size

Scope	Rows	Held-out test rows
Complete release	184,978	36,808
Corrected paper categories	165,972	33,012

Labels are balanced: 92,489 feature and 92,489 random-noannot rows. Every pair shares pair_id.

The row-level split uses held-out chromosomes chr2, chr3, chr16, and chr22. Hugging Face split all loads the entire file.

Categories and controls

Source features cover:

repeats, UCNE, vista_enhancer, promoters, UTR5, UTR3,
coding_regions, exons, introns, upstream_TSS, noncoding_regions

For each control, the sampler avoids the union of retained annotation intervals and previously selected annotation-free controls. The control has the same chromosome and length as its feature. Independent verification repeats this check without using the sampler and finds zero overlaps across all 92,489 controls. It also confirms every causal and bidirectional control context contains only uppercase A/C/G/T.

Exclude noncoding_regions for the corrected ten-category paper view.

Context and evaluation

Use:

• causal for left-to-right models;
• bidi for bidirectional/masked models.

All 184,978 bidi contexts are 2,048 bp. In causal, 165,012 rows are 2,048 bp and 19,966 long-feature rows are 1,000 bp. For paper evaluation:

1. filter held-out test rows;
2. exclude noncoding_regions;
3. pool from pool_start_in_window to pool_end_in_window;
4. use cosine leave-one-out 1-nearest-neighbor balanced accuracy.

Columns

Column group	Description
split, sequence, label, pair_id	Partition, exact input, binary class, and matched pair.
context_group_id	Exact model-input leakage group within this physical config.
category, repeat_class, scope, context_policy	Source class, nullable repeat class, pooling scope, and context geometry.
chrom, start, end, source_strand	Zero-based, half-open hg38 coordinates and biological strand (. means unknown).
sequence_orientation	Explicit +/- orientation used for sequence geometry and reverse complementation.
context_start, context_end	Forward-genome context interval.
roi_start, roi_end	Feature/control offsets inside oriented sequence.
pool_start_in_window, pool_end_in_window	Model pooling span.
name	Source/control identifier.
phylop_*, phylop_context_*	ROI and symmetric-context phyloP features.

Corrected repeat processing

The release separates RepeatMasker strand from repeat class and regenerates repeat-derived features/controls with correct orientation. repeat_class remains available as metadata. Unknown source_strand stays ., while sequence_orientation is always explicit + or -.

Release hg38-positive-exact-input-v5.2 retains only VISTA rows explicitly annotated as hg38 with normalized positive expression (1,367 raw records, 1,240 unique elements). It excludes 1,522 non-positive hg38 and 1,750 non-hg38 rows without liftover and removes every ambiguous feature/control group and every conflicting-label exact-input group. Same-label duplicates remaining within a config share context_group_id; independent checks find zero exact inputs crossing train/test and zero conflicting-label groups. The held-out ATG benchmark has priority over an exact duplicate functional training input, so the affected chr7 coding pair is removed consistently.

Processing and citation

Processing and verification:

• Processing repository
• Processing and validation documentation

Consens, M. E. et al. Predicting evolutionary rate as a pretraining task improves genome language model representations. bioRxiv (2026). https://doi.org/10.64898/2026.02.02.703275

License

The processing code derived from GAMBA is MIT licensed under the processing repository's LICENSE. This generated dataset is marked other: incorporated reference sequence, annotations, and phyloP-derived values retain their upstream terms, so no blanket MIT license is asserted for the Parquets.