🧬 biology

Unifying the statistical foundations of variant and methylation calling for enabling sequencing platform and calling scenario independence within Varlociraptor

This paper presents a generalized Bayesian statistical framework within Varlociraptor that unifies the detection of genetic variants and methylation rates across diverse sequencing platforms, enabling more accurate, uncertainty-aware, and joint analysis of genomic and epigenomic data.

Original authors: Adrian Dominic Prinz, Johannes Köster

Published 2026-09-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Adrian Dominic Prinz, Johannes Köster

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Inside the nucleus of nearly every cell, a complex system of chemical tags works like a dimmer switch for our genes, turning them up, down, or off entirely. One of the most important of these tags is a tiny methyl group that attaches to specific parts of the DNA molecule, a process known as methylation. This chemical modification does not change the genetic code itself, but it dictates how that code is read, influencing everything from how an embryo develops to how a cell responds to its environment. When this system malfunctions, it can lead to serious diseases, including cancer. To understand these biological processes, scientists must be able to read these chemical tags with extreme precision. However, reading them is difficult because the tools used to sequence DNA are imperfect, introducing noise and uncertainty that can obscure the true biological signal.

For years, researchers have relied on different tools to read these tags depending on the type of sequencing machine they use. Some machines read short snippets of DNA, while others read much longer strands. The software designed to interpret the data from these machines has historically been separate, with each program built to handle only one specific technology. This fragmentation meant that scientists could not easily compare results across different platforms or combine data from multiple samples to get a clearer picture. Furthermore, existing methods often treated the data as a simple count of methylated versus unmethylated reads, ignoring the many sources of error that occur during the sequencing process, such as imperfect alignment of the DNA strands to the reference genome.

A team of researchers at the University of Duisburg-Essen has developed a new approach that unifies these separate worlds. They extended a statistical framework originally designed for finding genetic mutations to also handle methylation, creating a single, flexible system that works across all major sequencing technologies. This new method, an update to a tool called Varlociraptor, treats the reading of methylation not as a simple tally, but as a probability problem. Instead of just counting how many times a tag appears, the software calculates the likelihood that a tag is truly there, taking into account the quality of the DNA read, how well it fits into the genome, and whether it might be a mistake caused by the machine. By using a mathematical model that accounts for these uncertainties, the system can distinguish between a genuine biological signal and random noise with greater accuracy than previous methods.

The researchers tested this new system on real-world data from a well-studied human reference sample, using data generated by three different sequencing platforms: Illumina, which uses short reads; PacBio, which uses long reads; and Oxford Nanopore, which reads DNA as it passes through a tiny pore. They also created simulated data where the true methylation levels were known, allowing them to see exactly how close the software's predictions came to reality. In these tests, the new method consistently outperformed existing tools. When comparing results from different runs of the same sample, the new software showed much higher agreement, meaning it was less likely to produce conflicting results due to random errors. It also managed to filter out false positives more effectively, ensuring that the methylation sites it reported were highly reliable.

One of the most powerful features of this new system is its ability to look at multiple samples at once. In traditional analysis, scientists often analyze each sample individually and then try to compare the results later, a process that can accumulate errors. The new software allows researchers to define a "calling scenario," which tells the program how the samples are related. For example, a researcher can instruct the software to assume that two different samples from the same person should have the same methylation pattern at a specific location. The software then uses this shared information to strengthen the evidence, effectively borrowing confidence from one sample to clarify the results in another. This approach allows for the detection of subtle biological changes that might be missed when looking at a single sample in isolation, and it provides a way to control the rate of false discoveries without relying on arbitrary technical thresholds.

The study also revealed that the software could identify specific types of errors that occur during sequencing, such as when a DNA fragment is mapped to the wrong location or when the machine has a bias toward reading certain strands of DNA more often than others. By modeling these biases explicitly, the system can adjust its calculations to compensate for them, leading to a more accurate final result. In simulations where the true methylation levels were known, the new method produced predictions that were closer to the truth than those from other leading tools. This suggests that the Bayesian approach, which continuously updates the probability of a finding as new evidence is considered, is a robust way to handle the messy reality of biological data.

While the researchers noted that the software requires more computational time than some simpler tools, the trade-off is a significant gain in accuracy and the ability to handle complex experimental designs. The system is open-source and can be installed by other scientists, allowing the broader community to apply this unified statistical foundation to their own work. By bringing variant calling and methylation calling under one roof, the researchers have provided a tool that can adapt to the increasing complexity of modern genomics. This flexibility is crucial as scientists begin to study methylation across different tissues, over time, and in families, where understanding the precise uncertainty of each measurement is just as important as the measurement itself. The work demonstrates that by acknowledging and modeling the imperfections of our tools, we can extract a clearer, more reliable picture of the biological processes that govern life.

Technical Summary: Unifying the Statistical Foundations of Variant and Methylation Calling in Varlociraptor

Problem Statement
Methylation is a critical epigenetic biomarker influencing gene expression, development, and disease states such as cancer. However, accurately calling methylation rates from sequencing data is challenging due to multiple sources of uncertainty and bias inherent in sequencing and alignment processes. Existing methylation callers are typically platform-specific: tools like Bismark, BSMAP, and MethylDackel are designed for short-read bisulfite or enzymatic conversion data, while tools like Modkit, Nanopolish, and pb-CpG-tools are tailored for long-read technologies (Oxford Nanopore and PacBio) that detect methylation directly via electrical or fluorescent signals. Crucially, no existing approach comprehensively captures the uncertainty of methylation measurements while remaining independent of the sequencing platform. Furthermore, traditional methods often rely on point estimates and technical thresholds, failing to provide a unified probabilistic framework for joint analysis across samples or technologies.

Methodology
The authors present a generalization of the statistical model underlying Varlociraptor, a previously published framework for genomic variant calling, to support methylation rate calling. The core of the method is a Bayesian latent variable model that quantifies uncertainties arising during sequencing and read alignment.

Unified Statistical Model: The model treats the "alteration fraction" ( $\theta$ ) as a latent variable, which represents either an allele frequency (for variants) or a methylation rate (for methylation). It operates on a vector of sequencing fragments ( $Z$ ) overlapping a specific locus.
Latent Variables: The model accounts for two primary uncertainties for each fragment:
1. Alignment Uncertainty ( $\omega$ ): Modeled as a Bernoulli variable derived from mapping quality (MAPQ), indicating whether a fragment originates from the locus of interest.
2. Typing Uncertainty ( $\xi$ ): Modeled as a Bernoulli variable representing whether the fragment stems from an allele (or methylation state) of interest, adjusted for sampling bias.
Bias Modeling: The framework incorporates a bias vector ( $\beta$ ) to model various sequencing artifacts (e.g., strand bias, read position bias, softclip bias, alternative locus bias). The likelihood is computed by considering all possible realizations of these biases.
Platform Adaptation:
- Long Reads (Nanopore/PacBio): Methylation probabilities ( $p_i$ ) are extracted directly from the MM and ML tags in BAM/CRAM files.
- Short Reads (Bisulfite/Enzymatic): Methylation status is inferred from base calls (C vs. T) after conversion, calculating probabilities ( $p_i$ and $a_i$ ) based on base quality scores and confusion factors for sequencing errors.
Calling Scenarios: A key feature is the "calling scenario" grammar, which allows users to define complex relationships between samples (e.g., inheritance patterns, contamination, or joint analysis of orthogonal technologies). This enables the calculation of posterior probabilities for user-defined boolean events (e.g., "Sample A has a higher methylation rate than Sample B") rather than just point estimates.
Filtering: The method controls the local false discovery rate (FDR) by filtering loci based on the posterior probabilities of "present," "absent," or "artifact" events, rather than arbitrary coverage or quality thresholds.

Key Contributions

Platform Independence: The model is the first to unify variant and methylation calling across short-read (Illumina) and long-read (PacBio, Oxford Nanopore) technologies within a single statistical framework.
Uncertainty Quantification: By employing a Bayesian approach, the method provides posterior distributions for methylation rates and posterior probabilities for specific events, explicitly capturing uncertainty rather than providing single point estimates.
Joint Multi-Sample Analysis: The scenario-based grammar allows for the joint assessment of multiple samples (e.g., replicates, time-series, or pedigrees) without the uncontrolled accumulation of per-sample calling errors.
Bias Integration: The model explicitly estimates and accounts for various sequencing biases, improving the robustness of calls in the presence of artifacts.

Results
The authors evaluated Varlociraptor on both real-world data (HG002 Genome in a Bottle samples across Illumina, PacBio, and Nanopore) and simulated data with known ground truth.

Accuracy and Consistency: On Illumina data, Varlociraptor demonstrated lower absolute and relative discordance between technical replicates compared to state-of-the-art callers (Bismark, MethylDackel, BisSNP, BSMAPz). It also achieved comparable or better performance on PacBio and Nanopore data against tools like Modkit and pb-CpG-tools.
Robustness: The local FDR control effectively filtered uncertain calls, resulting in higher consistency between replicates. The method successfully distinguished true signals from noise, particularly at low methylation rates where uncertainty is higher.
Multi-Sample Utility: When combining orthogonal technologies (e.g., Nanopore + PacBio) in a joint scenario, Varlociraptor increased the number of reliably called loci (improved recall) while maintaining high concordance, demonstrating the benefit of leveraging information across technologies.
Performance: While Varlociraptor has a longer runtime than some single-step tools (due to its complex statistical model and two-step process of preprocessing and joint calling), it maintains comparatively low peak memory usage. Its preprocessing step is reusable across multiple scenarios, enhancing efficiency for downstream analyses.

Significance and Claims
The paper claims that Varlociraptor represents a significant step toward a unified, universally applicable model for detecting both genetic and epigenetic variations. By moving beyond platform-specific heuristics and point estimates, the Bayesian framework allows researchers to:

Assess methylation biomarkers in a fully uncertainty-aware and interpretable manner.
Jointly analyze complex experimental designs (e.g., across time, space, or pedigrees) while rigorously controlling the local false discovery rate.
Avoid the accumulation of errors inherent in sequential, independent sample calling.

The authors emphasize that this approach is particularly valuable for precision medicine, where accurate assessment of methylation biomarkers requires robust handling of uncertainty and the ability to integrate diverse data types without relying on arbitrary technical thresholds. The work lays the foundation for future extensions, such as incorporating species-specific priors and jointly analyzing conversion-based data for variant calling.

More like this