Semiparametric Tests for Identifying Differentially Methylated Loci With Case–Control Designs Using Illumina Arrays

Yong Chen, Yang Ning, Chuan Hong, Shuang WangGenetic 2013 · DOI

TL;DR

This paper introduces new statistical tests designed to identify differences in DNA methylation between cancer and normal cells. Unlike existing methods that primarily focus on average differences, these new tests can also detect variations in methylation patterns, which are crucial for understanding cancer development. The proposed methods demonstrate improved power in detecting these differences, especially when variances are not equal between groups, and successfully identified previously missed methylation sites in real-world cancer data.

What Problem Does This Paper Address?

DNA methylation is a key epigenetic mechanism that plays a significant role in cancer development. Identifying specific locations (loci) in the genome where methylation levels differ between cancer and normal cells is a fundamental step in understanding how these changes contribute to the disease. Researchers often use high-throughput technologies like Illumina methylation arrays to generate this data. However, existing statistical methods for identifying these differentially methylated loci primarily focus on differences in the mean methylation levels between cancer and normal groups. This overlooks a critical aspect: methylation levels can also differ in their variability (variance) between these groups. This paper addresses the limitation of current methods by proposing a new statistical framework that can detect differences in both the mean and variance of DNA methylation, thereby providing a more comprehensive analysis and potentially uncovering new insights into cancer biology.

Methods

The core of this paper's contribution lies in a novel statistical approach based on a generalized exponential tilt model. This model is designed to capture differences in both the mean and variance of methylation levels between two groups (cancer and normal).

Here's a breakdown of the key methodological aspects:

  • Generalized Exponential Tilt Model: This model assumes that the ratio of the probability density functions of methylation values in the cancer group to the normal group can be expressed as an exponential function of prespecified functions of the methylation value (e.g., log(x) and log(1-x) for beta-distributed data, or x and x^2 for normal-distributed data). This allows for flexible modeling of methylation distributions without assuming a specific parametric form for the underlying "normal" distribution. The parameters of this model directly relate to differences in means and variances.
  • Semiparametric Tests: To ensure robustness and avoid strict parametric assumptions, the authors derive semiparametric tests. This means they don't assume a specific distribution for the methylation data in the normal group.
  • Pseudolikelihood Approach: A key innovation is the use of a pseudolikelihood-based test. This method cleverly eliminates the need to know the exact distribution of the normal group by conditioning on order statistics. This approach is computationally more feasible than traditional empirical likelihood methods, especially for the large number of loci analyzed in methylation studies.
  • Proposed Tests: The paper introduces three specific tests derived from this pseudolikelihood framework: the Wald test ($T_W$), the score test ($T_S$), and the pseudolikelihood ratio test ($T_L$). The $T_L$ test is highlighted for its simple asymptotic distribution under the null hypothesis, making it easier to compute p-values.

Datasets

The authors evaluated their proposed methods using simulated data and a real-world dataset:

  • Simulated Data: They generated methylation data under various scenarios to mimic different patterns of differences between cancer and normal groups. These scenarios included:
  • Differences in means only.
  • Differences in variances only.
  • Differences in both means and variances.
  • Data generated from beta distributions, normal distributions, mixtures of normal distributions, and even t-distributions (to test robustness to model mis-specification).
  • Real Data: The proposed methods were applied to data from the United Kingdom Ovarian Cancer Population Study (UKOPS). This dataset involved 96 cancer samples and 136 normal samples, with methylation B-values measured across 22,951 loci.

Key Insights and Ideas

  • Variance Matters: The most significant insight is that differences in methylation variance between cancer and normal cells are not just noise but can be biologically informative and are missed by standard mean-focused tests.
  • Beyond the Mean: Traditional t-tests and regression-based methods are powerful for detecting mean differences but are blind to variance differences. The proposed exponential tilt model and associated tests offer a more complete picture.
  • Computational Efficiency: The pseudolikelihood approach is a clever way to handle the unknown reference distribution, making the analysis practical for large-scale genomic studies where computational burden is a major concern.
  • Robustness: The proposed tests show good control of Type I errors even when the underlying data distribution deviates from the assumed model (e.g., using t-distributions), suggesting a degree of robustness.
  • Balanced Design is Optimal: For maximizing statistical power, a balanced design (equal number of samples in case and control groups) is recommended.

Key Results

  • Improved Power: The proposed semiparametric tests ($T_L$, $T_W$, $T_S$) demonstrated significantly higher power than the t-test and logistic regression-based tests when the cancer and normal groups differed in variances only, or in both means and variances.
  • Competitive Performance: When groups differed only in means, the t-test was marginally more powerful, but the proposed tests were still competitive.
  • Type I Error Control: The proposed tests generally maintained well-controlled Type I errors across various simulation scenarios, including those with smaller sample sizes and potential model mis-specification. The $T_L$ test was particularly recommended for small sample sizes due to better Type I error control.
  • Real Data Discovery: In the UKOPS ovarian cancer data, the proposed $T_L$ test identified additional methylation loci that were missed by the standard t-test. Analysis of these uniquely identified loci revealed they were characterized by significant variance differences, validating the utility of the new approach.

Limitations and Future Work

  • Function Specification: The proposed tests require the specification of functions $h_1(\cdot)$ and $h_2(\cdot)$ within the exponential tilt model. The paper acknowledges that the effect of mis-specifying these functions on the tests' performance is an area for further investigation.
  • Higher-Order Differences: The current model captures differences in means and variances. The authors suggest that extensions to capture higher-order differences (like kurtosis) are possible but might lead to increased complexity and potential power loss.
  • Further Robustness Studies: While some robustness was shown, continued investigation into the effect of mis-specification on Type I errors and power is planned.

Overall, this paper presents a valuable advancement in the statistical toolkit for analyzing DNA methylation data, offering a more sensitive and comprehensive approach to identifying biologically relevant differences between disease and control groups.

Summary updated

Abstract

DNA methylation plays an important role in the development of many types of cancer. Identifying differentially methylated loci between cancer and normal patients is one of the central tasks to understand the contributions of the methylation process on cancer development. Through investigation of the methylation measurements generated by the Illumina methylation arrays, we notice that the methylation measurements of the cancer and normal groups could differ not only in means but also in variances. Therefore, we propose a generalized exponential tilt model to capture the differences in both means and variances between the cancer and normal groups. We derive the semiparametric tests to obtain model robustness. Through simulation studies, we demonstrate the feasibility of the proposed tests and a much improved power of the proposed tests than that of the t-test and the regression-based tests when the cancer and normal groups are different in variances only or in both means and variances. Hence the proposed tests can serve as useful complements to the standard tests that only test differences in means. We also illustrate the proposed methods by applying to a real methylation data from a recent study on ovarian cancer where the proposed methods identified additional methylation loci that were missed by the existing method.
BibTeX
@inproceedings{chen2013semiparametric,
  author = {Yong Chen and Yang Ning and Chuan Hong and Shuang Wang},
  title = {{Semiparametric Tests for Identifying Differentially Methylated Loci With Case–Control Designs Using Illumina Arrays}},
  booktitle = {Genetic},
  year = {2013},
  doi = {10.1002/gepi.21774},
}
Figure