ARUNA: Slice-based self-supervised imputation enables upscaling of sequencing-based DNA methylation assays

Overview

Whole-genome bisulfite sequencing (WGBS) provides near-comprehensive, base-resolution maps of DNA methylation, but its cost limits large-scale studies. Reduced representation bisulfite sequencing (RRBS) and related protocols offer cost-effective alternatives, but measure only a sparse subset of CpGs, creating substantial coverage mismatches across assays.

ARUNA is a self-supervised denoising convolutional autoencoder designed to upscale sparse, sequencing-based methylomes to whole-genome resolution. ARUNA operates on methylation slices: spatially stacked genomic windows that preserve local CpG correlation and cross-sample structure, allowing it to generalize across assays, donors, tissues, and datasets.

This repository provides:

1. Core ARUNA model and data-processing code.
2. Precomputed example datasets (GTEx chr21 subset).
3. Example notebooks for data preparation, inference, and training.
4. A pretrained model checkpoint for demonstration.

Quick Navigation

• Installation
• Repository Structure
• Data and Preprocessing
• Running the Examples
  • Data creation
  • Inference with a Pretrained Model
  • Training a Model
• Notes on Reproducibility
• Citation

---

Installation

# Clone the repository:
git clone https://github.com/ylaboratory/ARUNA.git
cd ARUNA

# Create and activate a Python environment (Python ≥ 3.9 recommended):
conda create -n aruna python=3.11
conda activate aruna

# Install dependencies:
pip install -r requirements.txt

This repository is not intended to be installed as a published Python package as of now. All scripts and notebooks assume execution from the repository root.

Repository Structure

ARUNA/
├── aruna/                  # Core model and data-processing code
├── checkpoints/            # Pretrained model checkpoints
├── configs/                # Example configuration YAML files
├── data/
│   ├── gtex_subset/        # Example GTEx WGBS input data (chr21 only)
│   ├── metadata/           # Reference genome and RRBS metadata    
├── notebooks/              # Example notebooks 
├── scripts/                # Inference helpers
├── results/                # Output storage
└── requirements.txt

Large intermediate files and derived datasets are intentionally excluded from version control.

Data and Preprocessing

Bioinformatics pipeline

A reference WGBS preprocessing pipeline used to generate the input .cov files from FASTQ inputs is provided in: assets/bioinfopipe_pairedWGBS.sh.

Patch Creation

ARUNA operates on a patch-centric representation of the methylome and also creates chrom-centric data in the preprocessing pipeline.

Input data

Input data are BED-like CpG-level methylation files (e.g. Bismark *.cov) produced from the initial bioinformatics pipeline, organized as:

data/gtex_subset/
└── <sample_id>/
    └── *.cpgMerged.CpG_report.merged_CpG_evidence.cov

An example GTEx subset (16 samples and chromosome 21 only) is included for demonstration.

Chrom-centric

• Shape: (num_cpgs, num_samples)
• Canonical CpG set derived from the reference genome (hg38)
• Stores fractional methylation and read depth

Patch-centric

• Fixed-size windows defined by number of CpGs per patch
• Enables convolutional modeling of local methylation structure

Reference metadata

Located under: data/metadata/

Includes:

• Canonical CpG coordinates (hg38, 0-indexed)
• Chromosome lengths (hg38)
• Optional RRBS CpG observation probabilities for realistic rrbs-like missingness simulation

Running the Examples

Data creation

The subsequent step converts sample-wise methylation files into chrom-centric and patch-centric formats.

Relevant code:

• aruna/process_dataset.py
• aruna/patch_metadata.py

Example notebook: notebooks/data_prep.ipynb

Running this notebook produces:

data/gtex/
├── chrom_centric/
│   ├── true/
│   ├── mcar_90/
│   └── rrbs_sim/
└── patch_centric/
    └── numCpg128/
        ├── true/
        ├── mcar_90/
        └── rrbs_sim/

Inference with a Pretrained Model

• A pretrained ARUNA model is provided in: checkpoints/trained_model.pth.
• You can run inference using: notebooks/example_infer.ipynb.
• Predicted methylation values and evaluation artifacts will be written to: results/.

Training a Model

An example training workflow is provided in: notebooks/example_train.ipynb.

This notebook demonstrates:

• Loading and batching training data.
• Initializing and training ARUNA from scratch on simulated sparse methylomes.
• Saving model checkpoints for downstream inference to: checkpoints/.

Note: Training is computationally intensive and was performed on GPU hardware for the experiments reported in the paper.

Reproducibility Notes

• Example data are restricted to chromosome 21 for tractability.
• Noise simulation for inference is performed once and saved to disk.
• Exact Python dependencies are specified in requirements.txt.

Citation

Coming Soon