TL;DR
This paper tackles the challenge of converting continuous DNA methylation and gene expression data into binary "on/off" signals for easier analysis. They developed a novel bivariate extension of the StepMiner algorithm to find optimal thresholds for this dichotomization, revealing that these thresholds are highly specific to individual CpG sites and that methylation-gene expression relationships are often tissue-specific.
What Problem Does This Paper Address?
DNA methylation is a crucial epigenetic mechanism that regulates gene expression, often by silencing genes. While the biological effect is often described as a binary switch (methylation on/off leading to gene expression off/on), the actual measurements from high-throughput technologies like microarrays and sequencing yield continuous values (beta values for methylation, RPKM for expression). To bridge this gap and analyze these relationships using a binary framework, appropriate thresholds are needed to convert these continuous values into discrete states. This paper addresses the problem of how to identify these optimal thresholds for both methylation and gene expression data, and how to use them to investigate the relationship between them, particularly focusing on the hypothesis that methylation silences gene expression.
Methods
The core of this paper's methodology lies in extending the StepMiner algorithm. Originally designed for univariate analysis of gene expression data, StepMiner fits a step function to sorted data to identify a threshold that maximizes a signal-to-noise ratio (SNR).
The authors introduce StepMiner2D, a bivariate extension that considers both methylation and gene expression simultaneously. Instead of a simple sorted list, they map observations to a 3D space where coordinates represent ranks of methylation and expression, and the z-axis represents a combination of these ranks. A bivariate step function is then fitted to this surface to identify thresholds for both variables. This approach allows for a joint analysis of the two data types.
To assess the biological significance of the identified thresholds and the methylation-gene expression relationship, they employ a hypergeometric test. This test specifically evaluates whether the observed count of samples in the "methylated and low expression" quadrant is significantly lower than expected by chance, thus testing the hypothesis that methylation leads to gene silencing.
Datasets
The study utilizes data from The Cancer Genome Atlas (TCGA) project. Specifically, they downloaded:
- DNA methylation data: Generated by the Illumina Infinium Human DNA Methylation27 array platform, covering 27,578 CpG sites in proximal promoter regions of 14,475 genes.
- RNA-seq gene expression data: Processed data (RPKM values) from Illumina GA II and HiSeq platforms.
The analysis was performed on 992 overlapping samples across five cancer types:
- Breast cancer (BRCA)
- Colon and rectal cancer (COAD/READ)
- Kidney renal clear cell carcinoma (KIRC)
- Squamous cell lung cancer (LUSC)
- Uterine corpus endometrioid carcinoma (UCEC)
The authors also provide their Matlab scripts for automatically downloading and processing TCGA data, available at http://odin.mdacc.tmc.edu/~pqiu/software/DownloadTCGA/.
Key Insights and Ideas
- Thresholds are CpG-site specific: The study found that the optimal threshold for dichotomizing methylation data is not universal but varies significantly across different CpG sites. This highlights the need for site-specific threshold determination rather than a one-size-fits-all approach.
- Methylation-expression relationships are tissue-specific: A major finding is that associations between methylation and gene expression are often not consistent across all cancer types. What appears significant when analyzing all samples together can become insignificant when looking at individual cancer types, suggesting that biological context (tissue type) plays a critical role.
- Pan-cancer analysis offers power but requires careful interpretation: Pooling samples from multiple cancer types allows for the identification of signals that might be too weak in individual tissues. However, this can also lead to identifying associations that are driven by tissue differences rather than direct mechanistic interactions.
- StepMiner2D provides a robust framework for joint analysis: The bivariate extension of StepMiner offers a principled way to find optimal thresholds for two related variables simultaneously, enabling a more nuanced investigation of their interplay.
Key Results
- After preprocessing and filtering, 11,189 CpG sites annotated to 7,344 unique genes were analyzed.
- A pan-cancer analysis identified 2,976 significant negative associations between methylation and gene expression (p < 0.01 with Bonferroni correction).
- However, when analyzed individually, only a fraction of these associations remained significant, with 2,072 becoming insignificant across all five cancer types.
- The study identified genes like SOX8, where methylation-expression association was significant in some cancer types (BRCA, KIRC) but not others, underscoring tissue specificity.
- Conversely, genes like BST2, GSTT1, and GSTM1 showed significant associations across multiple or all individual cancer types, suggesting more robust mechanistic links.
Limitations and Future Work
- Limited CpG coverage: The methylation data used was from the Illumina Methylation27k platform, which covers a subset of CpG sites. The authors plan to re-analyze the data using the Illumina Methylation 450k platform, which measures approximately 20 times more CpG sites, to identify more associations.
- Potential for tissue-driven associations: The authors acknowledge that associations significant only in pan-cancer analysis might be induced by tissue differences rather than direct regulatory relationships. Future work should focus on distinguishing these.
- Extension to other modalities: The proposed analysis strategy can be applied to other molecular data types, such as microRNA expression, DNA copy number variation, and protein expression, to explore multi-modal associations.
Summary updated