Ultra-Fast Analysis of Sparse DNA Methylome via Recurrent Pattern Encoding

Methods for analyzing DNA methylation data via Most Recurrent Methylation Patterns (MRMPs). Supports cell-type annotation, spatial deconvolution, unsupervised clustering, and cancer cell-of-origin inference. Includes C-backed summaries for YAME ".cg/.cm" files (overlap counts, log2 odds ratios, beta/depth aggregation), an XGBoost classifier, NNLS deconvolution, and plotting utilities. Scales to large spatial and single-cell methylomes and is robust to extreme sparsity.

MethScope

MethScope is an R package for ultra-fast analysis of sparse DNA methylome data using Most Recurrent Methylation Patterns (MRMPs).

It supports downstream analysis for:

• Cell type annotation
• Cell type deconvolution
• Unsupervised clustering
• Cancer cell-of-origin prediction
• Missing value imputation

Why MethScope?

Sparse single-cell and spatial methylome data are often too sparse to analyze directly. MethScope compresses methylation signals into MRMP-based embeddings so you can run robust and scalable downstream tasks with standard analysis workflows.

Method overview

MethScope converts high-dimensional methylation atlas signals into compact MRMP features and applies these features across multiple analysis tasks.

Core workflow:

• Binarize methylation atlas profiles and consolidate recurrent patterns
• Select top recurrent methylation patterns (MRMPs)
• Encode each sample, cell, or pixel into an MRMP-based representation
• Run downstream modeling for annotation, deconvolution, imputation, and representation learning

Use cases supported in the current pipeline:

• Cell-type annotation in sparse single-cell methylome profiles
• Mini-bulk deconvolution for mixed-cell samples
• Missing-value imputation for sparse CpG measurements
• Representation learning for clustering and embedding analysis

System Requirements

Software Dependencies

• R >= 4.0 (tested on R 4.5.2 and R 4.6.0)
• Operating systems tested: macOS, Linux (Ubuntu)
• System library: zlib
• R package dependencies (installed automatically): xgboost, dplyr, tidyr, stringr, caret, doParallel, parallel, ggplot2, uwot, magrittr, FNN, data.table, nnls

Hardware Requirements

No non-standard hardware is required. MethScope runs on a standard laptop or desktop CPU. No GPU is needed.

Installation

Install from CRAN:

install.packages("MethScope")

Or install the development version from GitHub:

# install.packages("devtools")
devtools::install_github("zhou-lab/MethScope")

Example Data

The GitHub repository includes a full example query file, inst/extdata/example.cg, for testing MethScope end to end. After cloning this repository, run the example from the repository root.

CRAN packages have size limits, so the CRAN release contains only tiny toy files. Use the GitHub example data below for functional testing and cell-type prediction.

GitHub reference .cm files are named by genome build and source dataset:

• inst/extdata/mm10_Liu2021.cm: mouse brain MRMP reference
• inst/extdata/hg38_Zhou2025.cm: human atlas MRMP reference
• inst/extdata/hg38_Loyfer2023.cm: human atlas MRMP reference from Loyfer et al.

Quick Start

library(MethScope)

# Run this from the root of a cloned zhou-lab/MethScope repository.
example_file      <- "inst/extdata/example.cg"
reference_pattern <- "inst/extdata/mm10_Liu2021.cm"

input_pattern <- GenerateInput(example_file, reference_pattern)

model <- Liu2021_MouseBrain_P1000()
prediction_result <- PredictCellType(model, input_pattern)

umap_plot <- PlotUMAP(input_pattern, prediction_result)

Expected output: a cell-by-MRMP matrix (input_pattern) and a data frame of predicted cell type labels with confidence scores (prediction_result). The full mm10_Liu2021.cm reference contains more than 1000 MRMPs; the built-in mouse brain model uses the first 1000 patterns. The UMAP plot will display cells colored by predicted cell type.

Tutorials and documentation

• Documentation website: zhou-lab.github.io/MethScope
• End-to-end tutorial: MethScope-Tutorial
• Preparing MethScope input files: MethScope-Input
• Building MRMP references: MethScope-MRMP
• C command-line implementation: methscope-cli

C command-line implementation

For users who want a standalone command-line workflow without an R runtime, we also provide methscope-cli, a pure-C implementation of the main MethScope workflows. methscope-cli uses YAME for .cg/.cm input/output and links libxgboost for model inference.

methscope-cli repository: zhou-lab/methscope-cli

methscope-cli is useful for large-scale batch processing, command-line pipelines, and environments where users want to run MRMP feature generation, cell-type prediction, deconvolution, or upscaling directly from the shell.

Agent skill

This repository includes a reusable MethScope agent skill under agent-skills/methscope/.

Codex

Install the skill under $CODEX_HOME/skills/methscope/ with this layout:

$CODEX_HOME/skills/methscope/
  SKILL.md        <- copy from agent-skills/methscope/codex/SKILL.md
  core/

Copy:

• agent-skills/methscope/codex/SKILL.md to $CODEX_HOME/skills/methscope/SKILL.md
• agent-skills/methscope/core/ to $CODEX_HOME/skills/methscope/core/

Then invoke the skill when working on MethScope package usage, vignettes, .cg and .cm inputs, MRMP embeddings, prediction, training, deconvolution, or visualization.

Claude

If you already keep a repository-level CLAUDE.md, copy the contents of agent-skills/methscope/claude/CLAUDE.md into it or reference that file from your existing Claude project instructions.

If you do not already have a project CLAUDE.md, use agent-skills/methscope/claude/CLAUDE.md as the starting project context for this repository.

Keep agent-skills/methscope/core/ in the repository, because the Claude instructions point to those shared files.

Shared files

• agent-skills/methscope/core/INSTRUCTIONS.md
• agent-skills/methscope/core/WORKFLOWS.md
• agent-skills/methscope/core/REFERENCES.md

Data resources

• Example .cg, .cm, and reference .rds files are bundled in inst/extdata
• .cg generation and preprocessing: YAME
• Pattern interpretation: knowYourCG

License

This project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0).

Copyright (c) 2025 Hongxiang Fu and Wanding Zhou ([email protected])

For commercial use or if the AGPL-3.0 restrictions are not suitable for your use case, please contact us for a commercial license: [email protected]

Citation

If you use MethScope, please cite (coming soon):

Fu H*, Xu H*, Lee CN, Cloud C, Deng Y, Zhou W. MethScope: Ultra-Fast Analysis of Sparse DNA Methylome via Recurrent Pattern Encoding.