TL;DR
This paper introduces ABBA, a novel Bayesian approach for analyzing whole-genome bisulfite sequencing (WGBS) data to identify differentially methylated regions (DMRs). ABBA excels at adapting to complex genomic correlations and data quality variations, outperforming existing methods in simulations and revealing disease-associated methylation changes in a rat model of glomerulonephritis, highlighting a new role for the Ifitm3 gene.
What Problem Does This Paper Address?
DNA methylation is a crucial epigenetic modification influencing gene regulation and disease susceptibility, but its precise role is still being unraveled. Whole-genome bisulfite sequencing (WGBS) provides high-resolution data, but accurately identifying meaningful methylation changes (differentially methylated regions or DMRs) across the entire genome is challenging. Existing methods often struggle with the complex, spatially correlated nature of methylation data, are sensitive to data quality variations (like sequencing depth or noise), and can be heavily reliant on user-defined parameters for smoothing. This makes it difficult to reliably pinpoint disease-associated methylation changes and generate robust biological hypotheses.
Methods
The core of this paper is the ABBA (approximate Bayesian bisulphite sequencing analysis) approach. ABBA is a fully Bayesian method that models WGBS data using a latent Gaussian model. Key innovations include:
- Probabilistic Smoothing: ABBA automatically adapts to the correlation structures of DNA methylation levels between CpG sites, considering the distance between them as a covariate. This is a significant departure from methods that rely on fixed window sizes or kernel bandwidths.
- Random Effects for Replicates: It explicitly models the variability between experimental replicates using a random effect, improving robustness.
- Integrated Nested Laplace Approximation (INLA): To overcome the computational burden of fully Bayesian inference, ABBA utilizes INLA, a fast and accurate approximation technique.
- Data-Driven Parameterization: Unlike many methods that require user-defined smoothing parameters, ABBA assigns prior distributions to these parameters, allowing the model to learn them directly from the data.
- False Discovery Rate (FDR) Calculation: A robust FDR procedure is implemented to identify significant DMRs.
Datasets
The authors used simulated WGBS datasets to benchmark ABBA against existing methods. These simulations were designed to mimic various real-world scenarios, varying parameters such as:
- Number of replicates
- Average read depth per CpG
- Level of noise
- Methylation probability difference between groups
- Random switching of methylation status of CpG sites
For empirical validation, they generated new WGBS data from primary macrophages of an experimental rat model of crescentic glomerulonephritis (CRGN), comparing Wistar Kyoto (WKY) rats (disease-prone) and Lewis (LEW) rats (control). They also integrated RNA-seq and ChIP-seq data from the same system.
Key Insights and Ideas
- Adaptive Smoothing is Crucial: The paper strongly emphasizes that the ability of a method to automatically adapt its smoothing parameters based on the data's genomic context and distance between CpGs is critical for robust DMR detection. Methods relying on fixed parameters are less generalizable.
- Bayesian Approach Handles Uncertainty Well: By incorporating prior distributions and using a fully Bayesian framework, ABBA effectively accounts for uncertainty in the data and model parameters, leading to more reliable estimates.
- Biological Hypotheses from Data: The study demonstrates how a robust DMR detection method like ABBA can be a powerful tool for generating novel biological hypotheses, as seen with the Ifitm3 gene in glomerulonephritis.
- Data Quality Matters, But ABBA is Robust: ABBA's performance is shown to be less sensitive to variations in data quality (e.g., sequencing coverage, noise) compared to other methods, making it suitable for diverse experimental settings.
Key Results
- Superior Performance in Simulations: ABBA was the best-performing method in 64% of 216 simulated scenarios, significantly outperforming other methods like DSS (12%) and BSmooth (10%). It showed particular strength in detecting biologically meaningful methylation changes (Ameth ≥ 30%) and when error rates were low.
- Identification of Disease-Associated DMRs: In the rat glomerulonephritis model, ABBA identified over 1,000 disease-associated DMRs.
- Discovery of Ifitm3 Role: Analysis of these DMRs pinpointed a 600bp region in the promoter of the Ifitm3 gene. This was further supported by RNA-seq and ChIP-seq data, suggesting that promoter hypermethylation in WKY rats leads to Ifitm3 repression, potentially contributing to glomerulonephritis pathogenesis.
- Computational Efficiency: Despite being a fully Bayesian approach, ABBA's use of INLA makes it computationally efficient, allowing for genome-wide analysis within a reasonable timeframe.
Limitations and Future Work
- Computational Cost for Very Large Genomes: While efficient, the paper notes that analyzing extremely large genomes might still require splitting into smaller chunks to manage computational resources, although results are consistent.
- Further Integration of 'Omics' Data: The authors suggest that integrating ABBA's differential methylation results with additional transcriptional and epigenetic data (beyond RNA-seq and ChIP-seq) could further refine hypotheses and elucidate specific regulatory mechanisms.
- Extending to Other Data Types: While ABBA is tailored for WGBS, its underlying Bayesian framework could potentially be adapted for other types of epigenetic data.
Summary updated