A comparison of reference-based algorithms for correcting cell-type heterogeneity in Epigenome-Wide Association Studies

• Published in: BMC bioinformatics Vol. 18; no. 1; p. 105
• Main Authors: Teschendorff, Andrew E., Breeze, Charles E., Zheng, Shijie C., Beck, Stephan
• Format: Journal Article
• Language: English
• Published: London BioMed Central 13.02.2017; Springer Nature B.V
• Subjects: Algorithms; Bioinformatics; Biomedical and Life Sciences; Blood; Breast cancer; Computational Biology/Bioinformatics; Computer Appl. in Life Sciences; Construction; Datasets; Deoxyribonucleic acid; DNA; DNA methylation; Epigenomics - methods; Estimates; Gene expression; Genetic Markers; Genome-Wide Association Study; Heterogeneity; Humans; Life Sciences; Methodology; Methodology Article; Methods; Microarrays; Noise levels; Results and data; EWAS; DNA methylation; Cellular heterogeneity
• ISSN: 1471-2105; 1471-2105
• DOI: 10.1186/s12859-017-1511-5

Background Intra-sample cellular heterogeneity presents numerous challenges to the identification of biomarkers in large Epigenome-Wide Association Studies (EWAS). While a number of reference-based deconvolution algorithms have emerged, their potential remains underexplored and a comparative evaluation of these algorithms beyond tissues such as blood is still lacking. Results Here we present a novel framework for reference-based inference, which leverages cell-type specific DNAse Hypersensitive Site (DHS) information from the NIH Epigenomics Roadmap to construct an improved reference DNA methylation database. We show that this leads to a marginal but statistically significant improvement of cell-count estimates in whole blood as well as in mixtures involving epithelial cell-types. Using this framework we compare a widely used state-of-the-art reference-based algorithm (called constrained projection) to two non-constrained approaches including CIBERSORT and a method based on robust partial correlations. We conclude that the widely-used constrained projection technique may not always be optimal. Instead, we find that the method based on robust partial correlations is generally more robust across a range of different tissue types and for realistic noise levels. We call the combined algorithm which uses DHS data and robust partial correlations for inference, EpiDISH ( Epi genetic D issection of I ntra- S ample H eterogeneity). Finally, we demonstrate the added value of EpiDISH in an EWAS of smoking. Conclusions Estimating cell-type fractions and subsequent inference in EWAS may benefit from the use of non-constrained reference-based cell-type deconvolution methods.