A statistical method for excluding non-variable CpG sites in high-throughput DNA methylation profiling
1 Altria Client Services, Research Development & Engineering, 601 E. Jackson Street, Richmond, VA 23219, USA
2 Venebio Group, LLC, Virginia Bio-Technology Research Park, Richmond, Virginia, USA
3 Center for Biomarker Research and Personalized Medicine, School of Pharmacy, Virginia Commonwealth University, Richmond, VA 23298, USA
4 Memorial Sloan-Kettering Cancer Center, 415 East 68th Street, New York, NY 10021, USA
5 Bon Secours Virginia Health System, 14331 Roderick Ct., Midlothian, VA 23113, USA
6 American Type Culture Collection, 10801 University Blvd, Manassas, VA 20110, USA
BMC Bioinformatics 2010, 11:227 doi:10.1186/1471-2105-11-227Published: 5 May 2010
High-throughput DNA methylation arrays are likely to accelerate the pace of methylation biomarker discovery for a wide variety of diseases. A potential problem with a standard set of probes measuring the methylation status of CpG sites across the whole genome is that many sites may not show inter-individual methylation variation among the biosamples for the disease outcome being studied. Inclusion of these so-called "non-variable sites" will increase the risk of false discoveries and reduce statistical power to detect biologically relevant methylation markers.
We propose a method to estimate the proportion of non-variable CpG sites and eliminate those sites from further analyses. Our method is illustrated using data obtained by hybridizing DNA extracted from the peripheral blood mononuclear cells of 311 samples to an array assaying 1505 CpG sites. Results showed that a large proportion of the CpG sites did not show inter-individual variation in methylation.
Our method resulted in a substantial improvement in association signals between methylation sites and outcome variables while controlling the false discovery rate at the same level.