README.md

metadata

configs:
  - config_name: reference
    data_files:
      - split: train
        path: reference/train.parquet
      - split: valid
        path: reference/valid.parquet
      - split: test
        path: reference/test.parquet
  - config_name: personalized
    data_files:
      - split: train
        path: personalized/train.parquet
      - split: valid
        path: personalized/valid.parquet
      - split: test
        path: personalized/test.parquet
license: cc-by-4.0
task_categories:
  - tabular-regression
tags:
  - biology
  - genomics
  - dna-methylation
  - epigenetics
  - sequence-to-function
  - collaborative-cross
  - mouse
size_categories:
  - 10M<n<100M

S2F Mouse Methylation Dataset

Sequence-to-function dataset for predicting CpG DNA methylation from genomic sequence in nine genetically diverse Collaborative Cross mouse founder strains.

Dataset Description

This dataset contains approximately 20 million sequence-methylation examples derived from reduced representation bisulfite sequencing (RRBS) data from nine Collaborative Cross founder strains. Each example pairs a 2,000-nucleotide DNA sequence window centered at a CpG dinucleotide with the corresponding methylation proportion (averaged across three biological replicates per strain).

Two encoding configurations are provided:

• reference: All strains use the mm10 reference genome sequence, producing identical input sequences regardless of strain identity.
• personalized: Strain-specific genetic variants are incorporated using g2gtools and IUPAC ambiguity codes. For example, a heterozygous A/T SNP is represented by IUPAC symbol W and encoded as [0.5, 0, 0, 0.5].

Usage

from datasets import load_dataset

# Load reference encoding
ds = load_dataset("cbreenmachine/s2f-mouse-methylation", "reference", split="train")

# Load personalized encoding
ds = load_dataset("cbreenmachine/s2f-mouse-methylation", "personalized", split="train")

Column Descriptions

Column	Type	Description
feature	string	2,000-nucleotide DNA sequence centered at a CpG site. Reference config uses standard nucleotides (A, C, G, T, N). Personalized config additionally uses IUPAC ambiguity codes (M, R, W, S, K, Y, B, D, H, V) to represent strain-specific variants.
methylation	float	Mean methylation proportion across three biological replicates (range 0--1).
subject	string	Mouse strain identifier: 129S1_SvImJ, A_J, C57BL_6, CAST_EiJ, DBA_2J, NOD_ShiLtJ, NZO_HlLtJ, PWK_PhJ, or WSB_EiJ.
chrom	string	Chromosome number (1--19).
name	string	Locus identifier in mm10 reference coordinates, enabling cross-strain comparisons at the same genomic position.

Data Splits

Split	Chromosomes	Purpose
train	1--12	Model training
valid	14	Validation / early stopping
test	15--19	Held-out evaluation

All nine strains are represented in every split.

Mouse Strains

Nine Collaborative Cross founder strains spanning three Mus musculus subspecies:

Strain	Subspecies	Notes
C57BL/6J	M. m. domesticus	Reference strain (mm10 genome)
129S1/SvImJ	M. m. domesticus
A/J	M. m. domesticus
DBA/2J	M. m. domesticus
NOD/ShiLtJ	M. m. domesticus
NZO/HlLtJ	M. m. domesticus
CAST/EiJ	M. m. castaneus	Wild-derived
PWK/PhJ	M. m. musculus	Wild-derived
WSB/EiJ	M. m. domesticus	Wild-derived

Source Data

• Methylation data: Tyler et al. (2023). Reduced representation bisulfite sequencing of hepatocyte DNA from 12-week-old female mice. Three biological replicates per strain, averaged.
• Genetic variants: Svenson et al. (2012). SNPs and indels for Collaborative Cross founder strains.
• Reference genome: GRCm38/mm10.

Associated Paper

Breen C, Keles S. Variant-Aware Sequence Encoding Does Not Improve Deep Learning Prediction of DNA Methylation in Genetically Diverse Mice. (2026).

Citation

@article{breen2026variant,
  title={Variant-Aware Sequence Encoding Does Not Improve Deep Learning Prediction of DNA Methylation in Genetically Diverse Mice},
  author={Breen, Coleman and Kele{\c{s}}, S{\"u}nd{\"u}z},
  year={2026}
}

License

CC-BY-4.0

Repository Structure

s2f-mouse-methylation/
├── README.md
├── reference/
│   ├── train.parquet
│   ├── valid.parquet
│   └── test.parquet
└── personalized/
    ├── train.parquet
    ├── valid.parquet
    └── test.parquet

Uploading the Data

The parquet files are located on the training server at:

/storage10/cebreen/S2M-data-v3/data_derived/mouse/s2m_input/reference_merged/{train,valid,test}.parquet
/storage10/cebreen/S2M-data-v3/data_derived/mouse/s2m_input/personal_merged/{train,valid,test}.parquet

To populate this repo:

# Create directory structure
mkdir -p reference personalized

# Copy from server (adjust hostname as needed)
scp server:/storage10/cebreen/S2M-data-v3/data_derived/mouse/s2m_input/reference_merged/{train,valid,test}.parquet reference/
scp server:/storage10/cebreen/S2M-data-v3/data_derived/mouse/s2m_input/personal_merged/{train,valid,test}.parquet personalized/

# Initialize as HuggingFace dataset repo
huggingface-cli repo create s2f-mouse-methylation --type dataset
git init
git remote add origin https://huggingface.co/datasets/cbreenmachine/s2f-mouse-methylation
git lfs install
git lfs track "*.parquet"
git add .gitattributes README.md reference/ personalized/
git commit -m "Initial upload: reference and personalized encodings"
git push -u origin main

Note: Files >10MB are automatically handled by Git LFS on HuggingFace.