Cell-type Deconvolution Using CellMethMarker Markers

Estimate cell-type proportions in bulk methylation samples

This case study demonstrates how CellMethMarker custom markers can be converted into a platform-matched RRBS reference atlas for estimating cell/tissue contributions in bulk methylation or cell-free DNA methylation samples.

Cell-type deconvolution Bulk methylation RRBS reference atlas UXM-compatible benchmark
CellMethMarker / Case Studies / cell-type-deconvolution

Summary

Take-home message. CellMethMarker custom markers can be used to construct a platform-matched RRBS reference atlas for UXM-based methylation deconvolution. In this benchmark, the CellMethMarker-derived RRBS atlas achieved lower RMSE in 7 of 9 tested cell types and reduced the average RMSE from 0.365 to 0.144 compared with the UXM default WGBS atlas.
Tested cell types 9 Breast, Kidney, Liver, Muscle, Nerve, Ovary, Pancreas, Stom and Thyroid
Cell types with lower RMSE 7 / 9 Using the CellMethMarker-RRBS atlas
Average RMSE 0.365 → 0.144 Lower values indicate smaller prediction error
Average R2 -0.591 → 0.701 UXM default atlas to CellMethMarker-RRBS atlas

Background

Reference-based methylation deconvolution estimates cell or tissue composition by comparing mixed methylomes with a reference atlas of cell/tissue-specific methylation signatures. Its accuracy depends on marker coverage, reference quality and compatibility between the input assay and the reference atlas.

UXM provides a fragment-level deconvolution framework with a default WGBS-derived atlas. Because RRBS captures a more restricted CpG-rich feature space, we tested whether an RRBS-matched reference atlas built from CellMethMarker custom markers could improve UXM performance for RRBS benchmark samples.

Scope. This case study demonstrates a marker-to-reference workflow. The atlas used here is presented as a benchmark example rather than a universal pre-built reference for all datasets.

Benchmark design and methods

We split CellMethMarker-associated RRBS samples into independent testing and reference-building sets. A subset of samples was held out for evaluation, while non-overlapping quality-controlled samples were used to construct a UXM-compatible RRBS reference atlas. Atlas construction followed the reference-generation logic of the UXM deconvolution workflow: https://github.com/nloyfer/UXM_deconv.

  • Input markers: CellMethMarker RRBS custom marker regions.
  • Reference atlas: high-quality RRBS reference samples converted into a UXM-compatible atlas.
  • Test set: held-out RRBS samples from the same benchmark cell/tissue categories.
  • Comparison: UXM was run with either the default WGBS atlas or the CellMethMarker-derived RRBS atlas.
  • Metrics: RMSE, R2 and Pearson correlation between expected and predicted cell/tissue fractions.

Results

The CellMethMarker-derived RRBS atlas reduced prediction error for most benchmark cell types. It achieved lower RMSE in seven of nine tested cell types and improved both average RMSE and average R2 compared with the UXM default atlas.

UXM default WGBS atlas

Average RMSE 0.365
Average R2 -0.591
Average Pearson correlation 0.962
Cell types with lower RMSE 2 / 9

UXM + CellMethMarker-RRBS atlas

Average RMSE 0.144
Average R2 0.701
Average Pearson correlation 0.975
Cell types with lower RMSE 7 / 9
Key observation: In this benchmark, the platform-matched CellMethMarker-RRBS atlas reduced average RMSE by approximately 60.4% relative to the UXM default atlas.

Visual evidence

RMSE directly measures prediction error; lower values indicate better agreement between expected and predicted cell/tissue fractions.

RMSE comparison between UXM default atlas and CellMethMarker-RRBS atlas
RMSE comparison across benchmark cell types. The CellMethMarker-derived RRBS atlas achieved lower RMSE than the UXM default atlas in Breast, Kidney, Nerve, Ovary, Pancreas, Stom and Thyroid. The default atlas performed better for Liver and Muscle in this benchmark.

Cell-type-level performance

Cell type RMSE: UXM default atlas RMSE: CellMethMarker-RRBS atlas Lower RMSE
Breast0.4400.028CellMethMarker-RRBS
Kidney0.3110.123CellMethMarker-RRBS
Liver0.3200.367UXM default
Muscle0.0990.155UXM default
Nerve0.5400.115CellMethMarker-RRBS
Ovary0.5500.069CellMethMarker-RRBS
Pancreas0.2090.041CellMethMarker-RRBS
Stom0.5800.313CellMethMarker-RRBS
Thyroid0.2360.103CellMethMarker-RRBS
Average0.3650.144CellMethMarker-RRBS

Suggested workflow for users

  • Select relevant cell/tissue types and marker regions from CellMethMarker.
  • Prepare reference methylation samples matched to the input assay, preferably RRBS for RRBS input data.
  • Construct a UXM-compatible reference atlas from CellMethMarker markers and user-selected reference samples.
  • Run UXM deconvolution using the customized platform-matched atlas.
  • Validate predictions using known mixtures, control samples or expected tissue composition when available.

Notes and limitations

This result supports the use of CellMethMarker markers for constructing deconvolution references, but it does not imply that one fixed atlas will outperform all existing references in every dataset. Performance may vary with sequencing platform, marker coverage, cell/tissue composition, reference sample quality, disease state and batch effects.

Users are encouraged to build study-specific reference atlases using CellMethMarker markers and their own reference samples when applying methylation deconvolution to new datasets.

Related method: UXM deconvolution reference-building workflow, https://github.com/nloyfer/UXM_deconv.