# TE-Seq Data v1

Transposable Element locus sequences from the human genome (GRCh38).

## Dataset Summary

| Property | Value |
|----------|-------|
| Total records | 4,693,511 |
| File size | ~894 MB |
| Sequence column | `sequence` (consume verbatim; no padding trimming) |
| Labels | `family` (1,180 unique), `class` (13 unique) |

## Schema

| Column | Type | Description |
|--------|------|-------------|
| `sequence` | string | DNA sequence (consume as-is) |
| `family` | string | TE family label (primary training target) |
| `class` | string | TE class (e.g., SINE, LINE, LTR, DNA) |
| `chrom`, `start`, `end` | string/int | Genomic coordinates |
| `strand` | string | + or - |
| `subfamily` | string | Subfamily label |
| Other columns | | Locus metadata (ID, loci, group, N1, N2, TE_chrom, TE_start, TE_end) |

## File Status: Corrected Splits vs Legacy Root

Use the corrected split shards for all current DeepGenopix training, simulator, and benchmark jobs:

- `train/te_seqdata.parquet`
- `val/te_seqdata.parquet`
- `test/te_seqdata.parquet`
- `split_summary.json`
- `te_seqdata.recovered.parquet` for corrected raw sibling access

Do not use the root `te_seqdata.parquet` for new jobs. It is retained only as a legacy snapshot for backward compatibility and was intentionally left untouched when the corrected split was uploaded. The corrected split excludes the remaining 405 unresolved unlabeled rows.

Current canonical split-input revision for reproducible jobs: `f107d46c5a61087eb3fb19b4e3e75fdef0b74fdd`.

## Usage

Use this dataset as the raw input for `run_parquet_etl()` in DeepGenopix:

```python
from deepgenopix.etl import run_parquet_etl
run_parquet_etl(
    "te_seqdata.parquet",
    output_dir="data/processed/te_visuals/<run_id>",
    sequence_col="sequence",
    label_col="family",  # or "class"
)
```

## Class Distribution

- SINE: 1,770,903
- LINE: 1,516,226
- LTR: 720,177
- DNA: 483,994
- Satellite: 7,018
- Other: ~2,000 each or fewer

## Splits

The ETL builds stratified train/val/test splits automatically:
- Tier 1 (< 30 samples): all → train
- Tier 2 (30–200): floor-20 val, 10% test, rest → train
- Tier 3 (> 200): 10% val, 10% test, rest → train