README.md

---
license: cc-by-4.0
pretty_name: "SYTO v1.0 — data-driven soft labeling for whole-body cell-type deconvolution"
viewer: false
tags:
  - biology
  - genomics
  - epigenomics
  - dna-methylation
  - cell-type-deconvolution
  - read-classification
---

# SYTO v1.0 publication data

Data underlying *"Data-driven soft labeling scales DNA read classification to
whole-body cell-type deconvolution"*. SYTO is a framework for read-level DNA
methylation classification and whole-body cell-type deconvolution. This deposit
contains the trained models, generated pseudobulk mixtures, training data,
source reads and final result tables behind every figure and table in the paper.

> **This repository is a mirror.**
> The canonical, citable version of record is TBD.
> This mirror exists to make the data easier
> to fetch programmatically; contents are identical to the planed citable deposit.

- **Size:** 190.4 GB across 89 files (84 `.zip` archives + 4 acompanying metadata files + 1 csv file with Syto variants and baselines ranking)
- **Unpacked:** 82,010 files
- **Largest file:** 5.66 GB

## Contents

The deposit is organised in five tiers of decreasing necessity, so you can take
only the depth you need.

| Tier | Contents | Size |
|---|---|---|
| `tier0-results/` | Final result tables underlying important published figures and tables | 0.5 GB |
| `tier1-models/` | Trained classifiers, deconvolvers and calibrators | 24.0 GB |
| `tier2-pseudobulks/` | Generated pseudobulk mixtures (columnar Parquet) | 141.5 GB |
| `tier3-training-data/` | Marker atlases, target proportions, training datasets | 12.6 GB |
| `tier4-source-data/` | Staged and recovered reads, hg19/hg38 reference genomes | 11.8 GB |

**`tier0-results/` alone (0.5 GB) is enough to inspect the final results of the paper.**
Tiers 1–4 exist so the analysis can be independently re-run and verified.

Metadata files are readable without downloading anything large:

| File | Purpose |
|---|---|
| `MANIFEST.csv` | One row per archive: description, size, file count, SHA-256 |
| `MANIFEST_CONTENTS.csv` | One row per file *inside* the archives — inspect an archive's contents without downloading it |
| `runs.csv` | One row per experiment run, linking tier1 ↔ tier2 ↔ tier3 paths |
| `FILE_NAMING_CONVENTION.txt` | Full naming scheme and every abbreviation used |

## Downloading

Archives are stored as-is; there is no dataset viewer. Fetch selectively.

Inspect what exists first — both manifests are small:

```python
import pandas as pd
pd.read_csv("hf://datasets/CompEpigen/syto.1.0/MANIFEST.csv")
pd.read_csv("hf://datasets/CompEpigen/syto.1.0/MANIFEST_CONTENTS.csv")
```

Just the results (0.5 GB):

```bash
hf download CompEpigen/syto.1.0 --repo-type=dataset \
  --include "tier0-results/*" "*.csv" "*.txt" --local-dir ./syto-data
```

One specific run:

```bash
hf download CompEpigen/syto.1.0 --repo-type=dataset \
  --include "tier1-models/ood-rrbs/SYTO_tier1_models_oodrrbs_dismir_softlabelpooled_v1.zip" \
  --local-dir ./syto-data
```

Everything (190 GB):

```bash
hf download CompEpigen/syto.1.0 --repo-type=dataset --local-dir ./syto-data
```

## Unpacking

**Always unpack from the deposit root.** Archive members are stored with paths
relative to the root, so unpacking an archive from inside its own folder nests
the tree a second time (`tier1-models/ood-rrbs/tier1-models/...`) 

```bash
cd ./syto
find . -name 'SYTO_*.zip' -exec unzip -o -q {} ';'
```

The reference-genome archives contain relative symlinks
(`genome.fa.gz -> hg38.fa.gz`) matching the `wgbs_tools` layout. Unpack them
with a tool that preserves symlinks — `unzip` on Linux and macOS does.

## Verifying

`MANIFEST.csv` carries a SHA-256 for every archive:

```bash
python - <<'PY'
import csv, hashlib, pathlib
for r in csv.DictReader(open("MANIFEST.csv")):
    p = pathlib.Path(r["item"])
    if not p.exists():
        continue
    h = hashlib.sha256()
    with p.open("rb") as f:
        for chunk in iter(lambda: f.read(1 << 20), b""):
            h.update(chunk)
    print("OK " if h.hexdigest() == r["sha256"] else "BAD", p)
PY
```

## Naming convention

See `FILE_NAMING_CONVENTION.txt` for the complete scheme, including feature
sets, atlas names and reference-genome codes.

## What is not included

Redundant, regenerable and training-only artefacts were removed before deposit: 
duplicated inputs, per-calibrator predictions
regenerable from the calibrator plus features, optimizer states, intermediates
superseded by the deposited aggregate tables, and diagnostic plots not part of
the published results.

Two inputs are referenced by the published configs but not redistributed here,
marked by placeholders:

- `${LOYFER_RECOVERED_READS}` — Loyfer recovered read tables; outputs of
  [CompEpigen/wgbs_atlas_simulation](https://github.com/CompEpigen/wgbs_atlas_simulation)
- `${SYTO_MLFLOW}` — an intermediate store for the `*_predicted.pkl` files
  (data splits enriched with trained-classifier outputs, used as pseudobulk
  inputs). The pseudobulk pipeline runs in `predictions_only` mode to produce
  them; two-stage configs are included where relevant.

Experiments were originally recorded in a local MLflow registry. MLflow-specific
files were stripped for release, but each `tier1-models` element retains its
recorded metrics, parameters and tags.

**A note on configs:** published YAML configs use paths relative to the deposit
root and were automatically redacted from their original HPC paths. Experiments
were not re-run after repackaging, so treat the configs as authoritative for
*parameters* while expecting to adjust *paths* to match your unpacking layout.
For MethylBERT you would need to re-point corresponding configs to the 
pretrained foundational model (which is not included in this dataset). 

## Citation

```bibtex
@article{rizdvanetskyi2026data,
  title={Data-Driven Soft Labeling Scales DNA Read Classification to Whole-Body Cell-Type Deconvolution},
  author={Rizdvanetskyi, Dmytro and Roos, Nathan and Lutsik, Pavlo},
  journal={arXiv preprint arXiv:2607.04987},
  year={2026}
}
```

Thee dataset DOI is TBD. 

## License

Released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).