Taykhoom/functional-upstream-gamba

GAMBA Functional Regions: Feature vs Upstream

This paired binary representation benchmark asks whether a model can distinguish an annotated functional region from a strand-aware, equal-length control located 2 kb upstream.

Each biological feature contributes:

• one feature row;
• one matched upstream row;
• a shared pair_id.

Loading

from datasets import load_dataset

bidi = load_dataset(
    "Taykhoom/functional-upstream-gamba",
    "bidi",
    split="all",
)

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

The Hugging Face split all means “load the physical Parquet.” The Parquet's own split column contains GAMBA's chromosome partition.

Dataset size

Scope	Rows	Held-out test rows
Complete release	184,932	36,794
Corrected paper categories (category != "noncoding_regions")	165,936	33,000

Labels are balanced:

Label	Rows
feature	92,466
upstream	92,466

The held-out chromosomes are chr2, chr3, chr16, and chr22; all other included chromosomes are marked train.

Functional categories

Category	Feature rows
repeats	9,983
UCNE	4,169
vista_enhancer	691
promoters	9,961
UTR5	9,809
UTR3	9,713
coding_regions	9,904
exons	9,926
introns	9,026
upstream_TSS	9,786
noncoding_regions extension	9,498

noncoding_regions is an explicit extension needed by later GAMBA evaluation code. It is the complement of all GENCODE v47 exons and CDS annotations; intronic sequence is allowed only where it does not overlap an exon or CDS. Exclude it to reproduce the corrected ten-category paper task.

Choosing a context

Config	Geometry
causal	ROI end-anchored after strand orientation for left-to-right models.
bidi	ROI centered for bidirectional/masked models.

All 184,932 bidi sequences are 2,048 bp. In causal, 164,972 rows are 2,048 bp and 19,960 long-feature rows use GAMBA's 1,000 bp truncation.

Evaluation

Use the held-out paper subset:

split == "test" and category != "noncoding_regions"

Pool representations over:

[pool_start_in_window, pool_end_in_window)

Then use cosine leave-one-out 1-nearest-neighbor balanced accuracy. pair_id supports paired analyses but the paper representation protocol uses the pooled embeddings and labels.

Columns

Column group	Description
split, sequence, label, pair_id	Chromosome partition, model input, binary class, and matched pair.
context_group_id	Exact model-input leakage group within this physical config.
category, repeat_class, scope	Biological source category, nullable RepeatMasker class, and full pooling scope.
context_policy	causal or symmetric.
chrom, start, end, source_strand	Zero-based, half-open hg38 coordinates and biological strand (. means unknown).
sequence_orientation	Explicit +/- orientation used for context geometry and reverse complementation.
context_start, context_end	Forward-genome interval represented by sequence.
roi_start, roi_end	Feature/control offsets inside the strand-oriented sequence.
pool_start_in_window, pool_end_in_window	Representation pooling offsets.
name	Source feature/control identifier.
phylop_*	Six float32 phyloP summaries over the task pooling span.
phylop_context_*	Six float32 summaries over the symmetric 2,048 bp context.

RepeatMasker correction

The source RepeatMasker table stores genomic strand in UCSC field 9 and repeat class in field 11. An earlier processing release placed repeat class in the strand column. This release uses field 9 as source_strand, preserves field 11 as repeat_class, and regenerates every repeat-derived sequence and control.

Unknown biological strand remains . in source_strand; its explicit sequence_orientation is +.

VISTA rows require hg38 and normalized positive expression. The 1,367 raw positive hg38 records collapse to 1,240 unique element-coordinate records; 1,522 non-positive hg38 and 1,750 non-hg38 rows are excluded without liftover. Ambiguous inputs and conflicting-label exact-input groups are removed. Remaining same-label exact duplicates share context_group_id, and no exact input crosses the chromosome train/test split. ATG test inputs are reserved globally, so the conflicting chr7 coding pair is absent from functional data.

Processing and verification

The pinned GAMBA commit, source/reference hashes, corrected RepeatMasker conversion, deterministic controls, FASTA/ROI reconstruction, phyloP checks, and output hashes are available at:

• Processing repository
• Processing and validation documentation

Citation

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.