---
pretty_name: Cembra OA pseudo-SNP training data
language:
- en
license: other
task_categories:
- text-classification
tags:
- genomics
- bioinformatics
- osteoarthritis
- transfer-learning
- dna-sequence
- research
size_categories:
- 10K<n<100K
configs:
- config_name: supervised
  default: true
  data_files:
  - split: train
    path: data/supervised_model_input.jsonl.gz
- config_name: pretraining
  data_files:
  - split: train
    path: data/pretraining-*.jsonl.gz
---

# Cembra OA pseudo-SNP training data

Private, model-ready data for the [Cembra pseudo-SNP SVD64 model](https://huggingface.co/zzhaobz/cembra). This package contains the exact reference/alternate sequence pairs consumed by the selected development model; it contains no individual-level genotype or phenotype records.

## Configurations

### `supervised`

- 168 OA variant examples: 56 positives and 112 matched controls.
- 56 complete matched triplets in 49 guarded independence components.
- Exact five-fold assignment: 39, 30, 45, 27, and 27 rows.
- Each example contains paired 3,000 bp hg38 REF/ALT sequences differing only at zero-based index 1,500.
- `pip` and related fields document cohort selection; they were not model features.
- `data/supervised_full_metadata.tsv.gz` retains all source/audit columns while removing model prediction columns.

### `pretraining`

- 20,000 label-free pseudo-SNP sequence pairs sampled from hg38 autosomes.
- Seed `20260808`; OA guarded intervals plus 3 kb were excluded.
- No labels, PIP, phenotype, source identity, GWAS p-values, or GWAS effect sizes were read during pretraining.
- Coordinates are retained in `data/pseudo_snp_coordinates.tsv`.

## Minimal loading example

```python
from datasets import load_dataset

supervised = load_dataset("zzhaobz/cembra-training-data", "supervised", split="train")
pretraining = load_dataset("zzhaobz/cembra-training-data", "pretraining", split="train")
```

For outer fold `k`, train on rows where `outer_fold != k` and validate on rows where `outer_fold == k`. Keep each `matched_set_id` intact.

## Provenance

Supervised annotations derive from Supplementary Table 9 of Hatzikotoulas et al., *Translational genomics of osteoarthritis in 1,962,069 individuals*, DOI [10.1038/s41586-025-08771-z](https://doi.org/10.1038/s41586-025-08771-z), distributed under CC BY 4.0. Sequence windows use the [UCSC hg38 FASTA](https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.fa.gz). Exact hashes, exclusions, and file inventories are recorded in `DATASET_MANIFEST.json` and `provenance/`.

The upstream 1.1 GB GWAS archive and full 3.2 GB FASTA are not duplicated: the exact model-ready 3 kb sequences are included here. No UK Biobank, arcOGEN, or other participant-level genotypes or phenotypes are present.

## Intended use and limits

Research reproduction and method development only. Labels are fine-mapping proxy labels, not experimentally established causal truth. The dataset is development evidence, has no external confirmation claim, and is not suitable for clinical use.

See `DATA_USE.md`, `LICENSE`, and `verify_dataset.py` before reuse.

## Verification

`verify_dataset.py` checks the package manifest, sequence invariants, labels, folds, and matched triplets. `verify_reproduction.py` independently reconstructs the governed cohort/fold map and recomputes the full 20,000 × 9,216 pseudo-SNP feature matrix. The recorded result is in `VERIFICATION.md`.