A Bayesian framework for genome-wide inference of DNA methylation levels

Marcel Hirt, Axel Finke, Alexandros Beskos, Πέτρος Δελλαπόρτας, Stephan Beck, Ismail Moghul, Simone EckerarXiv 2022 · DOI · arXiv

TL;DR

Researchers have developed a new Bayesian "change-point" model that uses Sequential Monte Carlo (particle filtering) to map DNA methylation across the entire genome. This framework efficiently identifies differences between healthy and diseased tissues by accounting for spatial correlations between neighboring sites, offering better accuracy and error control than current industry-standard tools.

What Problem Does This Paper Address?

DNA methylation is a crucial "on/off" switch for our genes. When these switches malfunction, it can lead to cancer, aging, and developmental disorders. Scientists use Whole-Genome Bisulphite Sequencing (WGBS) to measure these levels at millions of individual sites (CpGs) across the genome.

However, analyzing this data is a statistical nightmare. First, there are over 10 million sites to track. Second, methylation levels at one site are usually highly correlated with their neighbors. Third, the data is often "noisy" or has missing gaps. Existing tools often struggle to balance computational speed with the ability to accurately detect "Differentially Methylated Regions" (DMRs)—specific areas where a "case" group (like cancer patients) differs significantly from a "control" group.

Methods

The authors introduce a Bayesian methylome change-point model. Instead of looking at each site in isolation, the model assumes the genome is a sequence of "regimes" (e.g., highly methylated, lowly methylated, or "chaotic").

Key technical innovations include:

  • Regime Modeling: The single-group model uses six predefined biological regimes based on mean methylation and standard deviation.
  • Case-Control "Split/Merge" Logic: For comparing two groups, the model uses a latent binary indicator ($z_t$). If $z_t=1$, the groups are "merged" (identical); if $z_t=0$, they have "split" and evolved separately. This allows the model to detect exactly where and how two biological conditions diverge.
  • Sequential Monte Carlo (SMC): To handle the massive scale of the genome, the authors use particle filtering and smoothing. This "survival-of-the-fittest" algorithm for statistical hypotheses allows the computation time to grow linearly with the number of sites, making genome-wide inference feasible on standard hardware.
  • Local FDR Control: They implement a principled way to control the False Discovery Rate (FDR), ensuring that the "discoveries" made by the model are statistically reliable.

Datasets

The framework was evaluated using 100 simulated methylation datasets.

  • Each dataset contained 400,000 CpG sites.
  • They simulated up to 10 samples per group with varying "read depths" (averaging 10 reads per site).
  • Why simulations? By using simulated data where the "ground truth" is known, the authors could intentionally introduce "model misspecification"—scenarios where the data doesn't perfectly match the model's assumptions—to test how robust the tool is in messy, real-world conditions.

Key Insights and Ideas

  • Spatial Awareness is Key: Most existing methods either ignore the distance between sites or smooth the data too aggressively. This model treats the genome as a continuous process where "change points" occur, preserving sharp boundaries between different biological regions.
  • Joint Modeling vs. Independent Fitting: Instead of analyzing the control group and the case group separately and then subtracting them, this model looks at them simultaneously. This joint approach is much more efficient at picking up subtle signals in small sample sizes.
  • Linear Scaling: By using discrete particle filters, the authors achieved $O(T)$ complexity (where $T$ is the number of sites). This is a significant "aha moment" for bioinformatics, as many complex Bayesian models are usually too slow for whole-genome applications.
  • Imputation by Default: Because the model understands the "regime" of a region, it can naturally guess (impute) methylation levels for sites where data is missing based on the surrounding context.

Key Results

The authors benchmarked their method against popular tools like DSS (Dispersion Shrinkage for Sequencing) and DMRSEQ.

  • Superior Error Control: In testing for different means, the Bayesian framework yielded significantly lower False Discovery Proportions (FDP) than DSS.
  • Higher Power: The model achieved lower False Non-discovery Rates (FNP), meaning it missed fewer real biological signals, especially when sample sizes were greater than one.
  • Flexibility: Unlike many tools that only look for changes in the average methylation, this model successfully detected changes in variance (how "chaotic" the methylation is), which is an emerging marker in cancer research.
  • Region-wise Accuracy: When identifying entire regions of difference (DMRs), the model maintained tight control over the FDR even when 99% of the sites in a region were differentially methylated—a scenario where other tools often produced high rates of false positives.

Limitations and Future Work

While the model is robust, the authors noted that the realized False Discovery Rate can slightly exceed the target level when the data generation process is significantly different from the model's assumptions.

Future directions include: * Adapting the model specifically for twin studies (where the control and case are genetically identical). * Incorporating more genomic context, such as the actual physical distance (in base pairs) between CpG sites, into the "sojourn time" (how long a regime lasts). * Applying the framework to real-world datasets, specifically comparing newborns to adults and studying type-I diabetes-discordant twins.

Summary updated

Abstract

DNA methylation is an important epigenetic mark that has been studied extensively for its regulatory role in biological processes and diseases. WGBS allows for genome-wide measurements of DNA methylation up to single-base resolutions, yet poses challenges in identifying significantly different methylation patterns across distinct biological conditions. We propose a novel methylome change-point model which describes the joint dynamics of methylation regimes of a case and a control group and benefits from taking into account the information of neighbouring methylation sites among all available samples. We also devise particle filtering and smoothing algorithms to perform efficient inference of the latent methylation patterns. We illustrate that our approach can detect and test for very flexible differential methylation signatures with high power while controlling Type-I error measures.
BibTeX
@misc{hirt2022bayesian,
  author = {Marcel Hirt and Axel Finke and Alexandros Beskos and Πέτρος Δελλαπόρτας and Stephan Beck and Ismail Moghul and Simone Ecker},
  title = {{A Bayesian framework for genome-wide inference of DNA methylation levels}},
  year = {2022},
  eprint = {2211.07311},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2211.07311},
}
Figure