---
pretty_name: "GAMBA ATG Contexts"
license: other
tags:
- biology
- genomics
- dna
- genome-language-model
- benchmark
- translation-initiation
- hg38
- parquet
configs:
- config_name: causal
  data_files:
  - split: test
    path: atg-gamba-causal.parquet
- config_name: bidi
  data_files:
  - split: test
    path: atg-gamba-bidi.parquet
---

# GAMBA ATG Contexts

This dataset packages GAMBA's translation-initiation representation
benchmark. It samples 2,000 complete MANE Select transcript examples from
`chr1`–`chr22`, exactly balanced across strand and stratified by chromosome;
each example contributes five candidate sites:

1. the true translation start;
2. a nearby noncoding ATG, 2–5 kb away;
3. a far noncoding ATG, at least 100 kb away;
4. a same-transcript in-frame methionine;
5. a same-transcript out-of-frame ATG motif.

The five rows share one `pair_id`, yielding 10,000 rows per configuration.

## Loading

```python
from datasets import load_dataset

bidi = load_dataset("Taykhoom/atg-gamba", "bidi", split="test")
causal = load_dataset("Taykhoom/atg-gamba", "causal", split="test")
```

## Choosing a context

Use the context geometry appropriate for the evaluated model:

| Config | Context placement | Intended models |
|---|---|---|
| `causal` | The three-base ATG is end-anchored after strand orientation. | Evo2 and other left-to-right/autoregressive models. |
| `bidi` | The three-base ATG is centered in the context. | GAMBA encoders, GPN-Star, PhyloGPN, distilled students, and other bidirectional models. |

Every released sequence is exactly 2,048 bp. Using the wrong configuration changes
which flanking bases the model can see and does not reproduce the paper
protocol.

## Evaluation protocol

The source data has five labels:

```text
start
noncoding_near
noncoding_far
inframe_methionine
outframe_atg
```

For the paper leaderboard, merge:

```text
noncoding_near -> noncoding
noncoding_far  -> noncoding
```

Then evaluate frozen representations with:

- cosine distance;
- leave-one-out 1-nearest-neighbor classification;
- balanced accuracy.

Exclude all other rows sharing `context_group_id` with the query from its
leave-one-out neighbor candidates. This groups the few exact sequence
duplicates without changing the five-row transcript `pair_id`.

GAMBA's extraction code also reports a separate five-way Euclidean
diagnostic. That is not the four-class paper leaderboard.

## Dataset size

| Label | Rows |
|---|---:|
| `start` | 2,000 |
| `noncoding_near` | 2,000 |
| `noncoding_far` | 2,000 |
| `inframe_methionine` | 2,000 |
| `outframe_atg` | 2,000 |
| **Total** | **10,000** |

Every row has `split == "test"` because ATG is a representation benchmark,
not a supervised train/test dataset.

## Columns

| Column | Description |
|---|---|
| `split` | Always `test`. |
| `sequence` | Strand-oriented causal or bidirectional context, exactly 2,048 bp. |
| `label`, `label_id` | Source five-way class and numeric ID. |
| `pair_id` | Shared by the five sites from one transcript example. |
| `context_group_id` | Evaluator leakage group joining transcript examples that share any exact released context. |
| `category`, `scope` | Constant `ATG` category and `roi` pooling scope. |
| `context_policy` | `causal` or `symmetric`. |
| `chrom`, `start`, `end`, `source_strand` | Valid contiguous zero-based, half-open hg38 motif interval and biological strand. |
| `sequence_orientation` | Explicit `+`/`-` orientation used for sequence geometry and reverse complementation. |
| `context_start`, `context_end` | Forward-genome coordinates represented by `sequence`. |
| `roi_start`, `roi_end` | Three-base site offsets in the strand-oriented sequence. |
| `pool_start_in_window`, `pool_end_in_window` | Exact representation-pooling span; identical to the three-base ROI. |
| `name` | Site identifier. |
| `transcript_id`, `gene_id` | Source MANE Select transcript and gene. |
| `motif_genomic_positions` | The three genomic base positions in transcript order. |
| `transcript_offset` | CDS offset for start/in-frame/out-of-frame motifs; null for noncoding controls. |
| `delta_bp` | Transcript offset for coding motifs; oriented first-base displacement for noncoding controls; zero for `start`. |
| `phylop_*` | Six float32 Zoonomia phyloP summaries over the three-base ATG. |
| `phylop_context_*` | Six float32 summaries over GAMBA's symmetric 2,048 bp phyloP context. |

## Strand convention

Minus-strand CDS blocks are reverse-complemented individually while retaining
true transcript order. Minus-strand motif positions therefore run from higher
to lower genomic coordinates, while `start`/`end` stores their enclosing
contiguous interval. Noncoding controls are strand-aware (`ATG` on `+`,
genomic `CAT` on `-`). Near/far controls exclude the union of every GENCODE
v47 CDS, while coding anchors remain MANE Select. Motifs spanning a CDS splice junction are excluded
deterministically rather than represented as a false contiguous interval.
Independent verification confirms every emitted oriented ROI equals `ATG`.

## PhyloP

PhyloP comes from the Zoonomia 241-mammalian track used by GAMBA.
Uncovered bases become zero, scores are rounded to two decimals before
summarization, and outputs are float32.

## Processing and verification

The source TSV reconciliation, transcript sampling, context construction,
phyloP calculation, output hashes, and exact reference comparison are
available at the processing reference below. The corrected
chromosome-by-strand cohort retains 1,208 transcript groups from the preceding
release and adds 792. It contains 1,000 examples per strand. Three exact
contexts shared across six rows are assigned to common leakage groups, and
2,548 detected junction-spanning source motifs are excluded:

- [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)

## 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.