Skip to content

decemedip

Hierarchical Bayesian modeling for cell type deconvolution of immunoprecipitation-based DNA methylome — a Bioconductor package that infers relative cell type and tissue abundances from MeDIP-Seq data, and can leverage reference datasets from alternative technologies (microarray, WGBS) to improve deconvolution.

Talk at EuroBioC2025

  • Speaker: Ning Shen
  • Day / session: Day 2 (Thu Sep 18, 2025) — morning short-talk session, after the Sharpe keynote
  • Talk title: "decemedip for DNA methylation deconvolution"
  • Slides (Zenodo DOI): to verify after publish
  • Video: not recorded — short talks at EuroBioC2025 were not part of the plenary YouTube playlist
  • Abstract / schedule: eurobioc2025 schedule

What it does

decemedip fits a hierarchical Bayesian model to MeDIP-Seq signal to estimate the relative contribution of each reference cell type or tissue in a mixture sample. It can borrow strength from reference panels generated on other methylation platforms (microarray, WGBS), bridging across technologies rather than requiring matched-platform references.

How it works

Core idea. A hierarchical Bayesian model couples a logit-normal prior on the cell-type proportion simplex with a generalized additive model (GAM) that maps fractional methylation level and CpG density to expected MeDIP-Seq read counts. The simplex prior keeps proportions on the unit simplex while letting cell types covary; the GAM absorbs MeDIP-specific enrichment bias.

Inputs / outputs. Input: MeDIP-Seq read counts at a fixed CpG-site panel, a reference methylation matrix (fractional methylation for each reference cell type), and per-site CpG-density values. Output: posterior cell-type / tissue proportions with credible intervals, plus MCMC convergence diagnostics (Rhat, effective sample size).

Key innovation. Two complementary reference panels are used jointly: ~2,500 cell-type-specific CpG sites (cell-type discrimination) and ~1,000 "anchor" methylation-universal sites (calibration of the read-count-to-methylation curve). The anchor sites let the model learn MeDIP enrichment bias separately from biology, which is what makes references generated on microarray or WGBS usable to deconvolute MeDIP-Seq.

Parameters worth knowing. Reference panel selection (hg19.ref.cts.se cell-type sites + hg19.ref.anc.se anchor sites). MCMC iterations (vignette uses 500 for the toy example; ~2,000 recommended). Prior on the simplex (logit-normal). Choice of cell types in the reference (controls which mixtures the model can resolve).

Canonical example. The vignette runs decemedip on a prostate-tumor PDX MeDIP-Seq sample (LuCaP_147CR), producing posterior tissue proportions across 25 reference tissues. Convergence diagnostics confirm the run is stable at the demo iteration count.

Where it fits in the corpus

  • AACR 2026: no current dossier; relevant axis is epigenomics — methylation-based deconvolution is a recurring theme in liquid-biopsy / cfDNA work
  • GBCC 2025: queued
  • Nextflow Summit 2026: not mentioned

Notes

First Bioconductor release. The platform-bridging reference design — using microarray or WGBS panels to deconvolute MeDIP-Seq — is the methodological signature and unusual enough to be worth tracking.