File size: 3,654 Bytes
1328edc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b9dcd69
1328edc
b9dcd69
1328edc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
---
license: mit
language:
- en
tags:
- dna
- variant-effect-prediction
- biology
- genomics
- non-coding
- regulatory-variants
configs:
- config_name: mendelian_traits
  data_files:
  - split: test
    path: mendelian_traits_vep.parquet
- config_name: complex_traits
  data_files:
  - split: test
    path: complex_traits_vep.parquet
---

# TraitGym + 8,192 bp pre-extracted windows

This dataset is a **repackaging** of [`songlab/TraitGym`](https://huggingface.co/datasets/songlab/TraitGym) (Benegas, Eraslan & Song, bioRxiv 2025.02.11.637758), with one extra step: for every variant we pre-extract the 8,192 bp window centered on the variant from the hg38 reference, plus the same window with the alt allele substituted.

The variants, labels and matched controls are **identical** to the original `songlab/TraitGym` `_matched_9` configs.

## Configs

- **`mendelian_traits`** (n = 3,380): 338 putative-causal non-coding variants from 113 monogenic Mendelian diseases (curated from OMIM), matched 9:1 against gnomAD common-variant controls on chromosome × consequence × TSS-distance.
- **`complex_traits`** (n = 11,400): 1,140 putative-causal non-coding variants from 83 polygenic UK BioBank traits (PIP > 0.9 from statistical fine-mapping), matched 9:1 against PIP < 0.01 controls on chromosome × consequence × TSS-distance × MAF × LD score.

## Schema

| column | description |
|--------|-------------|
| `chrom` | chromosome (`1``22`, `X`, `Y`) |
| `pos` | 1-indexed hg38 position |
| `ref`, `alt` | single bases (SNVs only) |
| `label` | int 0/1 (1 = causal/positive, 0 = matched control) |
| `class` | `"LOF"` (label=1) or `"FUNC/INT"` (label=0) — added so the standard `brca_eval.py` AUROC code path works without changes |
| `consequence` | molecular consequence from the source dataset |
| `tss_dist` | distance to nearest TSS |
| `match_group` | matched-controls group ID (preserve from source) |
| `score` | continuous PIP for `complex_traits`; `None` for `mendelian_traits` |
| `ref_seq`, `var_seq` | **8,192 bp window centered on the variant** (variant at index 4096), forward strand from chr-fasta-hg38 (UCSC). `var_seq` is `ref_seq` with the alt base substituted at index 4096. Reverse-complement is computed at eval time when `--rev_comp_avg` is requested (matches TraitGym's `run_vep_evo2.py` strand-symmetric scoring). |

## Eval methodology

Same recipe as our other VEP evals (BRCA1, BRCA2, ClinVar): `delta = LL(var_seq) − LL(ref_seq)` from a centered 8,192 bp window. Score per-variant; AUROC / AUPRC / `AUPRC_by_chrom_weighted_average` (the TraitGym leaderboard convention) against the binary label. Optional `--rev_comp_avg` averages the LLR computed on the forward window and on its reverse-complement, exactly as TraitGym does.

Eval scripts:
- [`reproduction-evo2-evals/brca/brca_eval.py`](https://github.com/huggingface/carbon/tree/evo2-evals/evaluation/reproduction-evo2-evals/brca/brca_eval.py) — gene-agnostic centered+full-LL eval
- Sharded variant: [`reproduction-evo2-evals/traitgym/`](https://github.com/huggingface/carbon/tree/evo2-evals/evaluation/reproduction-evo2-evals/traitgym/) — array-job sharding for Evo2 7B+/40B

## Citation

If you use TraitGym, cite the original paper:

```
@article{benegas2025traitgym,
  title  = {Benchmarking DNA Sequence Models for Causal Regulatory Variant Prediction in Human Genetics},
  author = {Benegas, Gonzalo and Eraslan, Gokcen and Song, Yun S.},
  journal = {bioRxiv},
  year   = {2025},
  doi    = {10.1101/2025.02.11.637758}
}
```

Leaderboard: [`songlab/TraitGym-leaderboard`](https://huggingface.co/spaces/songlab/TraitGym-leaderboard) on HF Spaces.