README.md

---
pretty_name: "GAMBA Functional Regions: Feature vs Category-Matched Random"
license: other
tags:
- biology
- genomics
- dna
- regulatory-genomics
- genome-language-model
- benchmark
- hg38
- parquet
configs:
- config_name: causal
  data_files:
  - split: all
    path: functional-random-gamba-causal.parquet
- config_name: bidi
  data_files:
  - split: all
    path: functional-random-gamba-bidi.parquet
---

# GAMBA Functional Regions: Feature vs Category-Matched Random

This paired binary representation benchmark asks whether a model can
distinguish an annotated functional region from a chromosome- and
length-matched random control.

For this dataset, a random control avoids retained anchors from the **same
functional category**. It may overlap annotations from other categories.
Use the annotation-free random dataset if controls must avoid every retained
annotation category.

Each feature/control pair shares `pair_id`.

## Loading

```python
from datasets import load_dataset

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

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

## Dataset size and splits

| 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` rows.

Rows on `chr2`, `chr3`, `chr16`, and `chr22` are marked `test`; remaining
included chromosomes are `train`. The Hugging Face split `all` loads the
whole physical file.

## Categories

The complete release contains 92,489 source features across:

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

`category != "noncoding_regions"` yields the corrected ten-category paper
view. `repeat_class` is populated on the 19,974 repeat-derived rows
(feature and matched control).

## Context configurations

| Config | Geometry |
|---|---|
| `causal` | ROI end-anchored after strand orientation for autoregressive models. |
| `bidi` | ROI centered for bidirectional/masked models. |

All 184,978 `bidi` sequences are 2,048 bp. In `causal`, 165,012 rows are
2,048 bp and 19,966 long-feature rows are truncated to 1,000 bp. Every released random
control context contains only uppercase `A/C/G/T`.

## Evaluation

For the paper protocol:

1. filter `split == "test"`;
2. exclude `category == "noncoding_regions"`;
3. pool each representation over
   `[pool_start_in_window, pool_end_in_window)`;
4. evaluate cosine leave-one-out 1-nearest-neighbor balanced accuracy.

## Columns

| Column group | Description |
|---|---|
| `split`, `sequence`, `label`, `pair_id` | Chromosome partition, exact input, `feature`/`random` class, and matched pair. |
| `context_group_id` | Exact model-input leakage group within this physical config. |
| `category`, `repeat_class`, `scope` | Source category, nullable repeat class, and `full` 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 sequence geometry and reverse complementation. |
| `context_start`, `context_end` | Forward-genome sequence context coordinates. |
| `roi_start`, `roi_end` | Feature/control offsets inside oriented `sequence`. |
| `pool_start_in_window`, `pool_end_in_window` | Evaluation pooling offsets. |
| `name` | Source/control identifier. |
| `phylop_*`, `phylop_context_*` | Six ROI and six symmetric-context float32 phyloP summaries. |

## Corrected repeat processing

RepeatMasker genomic strand and repeat class are separate fields. The release
uses UCSC `rmsk` field 9 as `source_strand`, field 11 as `repeat_class`, and
rebuilds repeat-derived controls and sequence orientation. Labels remain
unchanged.

VISTA requires hg38 and normalized positive expression: 1,367 raw records
collapse to 1,240 unique elements; 1,522 non-positive hg38 and 1,750 non-hg38
rows are excluded without liftover. All ambiguous feature/control groups and
all conflicting-label exact-input groups are removed. 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.
Every preserved random candidate is also rechecked against current
same-category anchors; invalid candidates are deterministically resampled.

## Processing and citation

Processing and verification:

- [Processing repository](https://github.com/TaykhoomDalal/Gamba-Processing/tree/main)
- [Processing and validation documentation](https://github.com/TaykhoomDalal/Gamba-Processing/blob/main/README.md#validation)

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.