Taykhoom/functional-upstream-gamba

GAMBA Functional Regions: Feature vs Upstream

This dataset packages GAMBA's binary functional-region benchmark comparing annotated genomic features with strand-aware, chromosome- and length-matched upstream controls. Controls are separated from their feature by 2 kb and do not overlap retained anchors from the same category.

Configs

Config	Train rows	Test rows	Total
causal, bidi	150,728	37,834	188,562

Each config includes the noncoding_regions extension. Filter category != "noncoding_regions" to recover the exact 169,674-row, ten-category GAMBA paper dataset without storing those rows twice.

Choosing a context

The GAMBA paper uses context geometry matched to the model. Choose *-causal for Evo2 or another left-to-right model: the ROI is at the end of the strand-oriented window, so all preceding bases are usable context. Choose *-bidi for GAMBA encoders, the distilled student, GPN-Star, PhyloGPN, or another masked/bidirectional model: the ROI is centered so the model sees both flanks. Using a causal file for a bidirectional model removes its right-side context and is not the paper protocol.

Loading

from datasets import load_dataset

dataset = load_dataset(
    "Taykhoom/functional-upstream-gamba",
    "bidi",
    split="all",
)
paper = dataset.filter(lambda row: row["category"] != "noncoding_regions")
test = paper.filter(lambda row: row["split"] == "test")

Here, split="all" tells Hugging Face to load the single physical parquet file. The parquet's own split column records the biological chromosome split (train or test). Keeping one file avoids duplicating the dataset; filter the column as shown above.

Columns

Column	Description
split, label	Probe/fine-tuning train or held-out test; labels are feature or upstream.
sequence	Strand-oriented causal or bidirectional context of at most 2,048 bp.
context_policy	causal for *-causal; symmetric for *-bidi.
pair_id	Shared identifier for a feature and its matched upstream control.
category, scope	Functional class and pooling scope.
chrom, start, end, strand	Zero-based, half-open hg38 feature/control coordinates.
context_start, context_end	Forward-genome coordinates of sequence.
roi_start, roi_end	Feature/control offsets within sequence.
pool_start, pool_end	Offsets to pool for representation evaluation.
name	Source annotation name.
phylop_mean, phylop_std	Mean and population standard deviation over GAMBA's baseline ROI.
phylop_frac_pos, phylop_frac_neg	Fractions of ROI bases with positive or negative phyloP scores.
phylop_mean_pos, phylop_mean_neg	Means over positive or negative ROI scores, or zero when absent.
phylop_context_mean, phylop_context_std	Mean and population standard deviation over GAMBA's symmetric 2,048 bp phyloP context.
phylop_context_frac_pos, phylop_context_frac_neg	Fractions of context bases with positive or negative scores.
phylop_context_mean_pos, phylop_context_mean_neg	Means over positive or negative context scores, or zero when absent.

Data source and processing

The processing code reproduces the ten-category BED output from Microsoft GAMBA commit e83984e byte-for-byte, then freezes all GAMBA autosomes plus chrX in parquet. The GAMBA models were pretrained on the chromosomes marked train. The paper's zero-shot result used the frozen models on chr2, chr3, chr16, and chr22, which are marked test; no task-specific model was fitted on these benchmark rows. Use only test to reproduce that paper subset. The train rows are provided for probing/fine-tuning or broader chromosome evaluation.

PhyloP values come from the Zoonomia 241-mammalian track used by GAMBA. Uncovered positions are zero and scores are rounded to two decimal places before the float32 summaries are computed.

Citation

Please cite the GAMBA paper:

Consens, M. E. et al. Predicting evolutionary rate as a pretraining task improves genome language model representations. bioRxiv (2026). doi:10.64898/2026.02.02.703275. A conference version is available on OpenReview.

@article{consens2026predicting,
  title = {Predicting evolutionary rate as a pretraining task improves genome language model representations},
  author = {Consens, Micaela Elisa and Yang, Kevin K. and Hall, Jimmy and Conard, Ashley Mae and Wang, Bo and Crawford, Lorin and Moses, Alan and Lu, Alex X.},
  journal = {bioRxiv},
  year = {2026},
  doi = {10.64898/2026.02.02.703275},
  url = {https://doi.org/10.64898/2026.02.02.703275}
}

License

GAMBA and its released benchmark code/data are distributed under the MIT license. Upstream annotation resources retain their original terms.