Effects of direct modeling on false-positives and accuracy. (a) Explicit modeling for differences in cell type reduces false-positive rate. Boxplots of test statistics for the difference in means based on linear regression estimation from models M1 and M2. Eighty percent of regions from M1 show a statistically significant difference in overall mean (at level 0.05); 16% and 12% of regions from M2 show a statistically significant difference in neurons or glia, respectively (at level 0.05). (b) Explicit modeling of neuronal methylation differences improves estimation accuracy. Comparison of gold-standard mean difference in methylation in neuron-specific DMRs to the estimated mean difference from models M1 (left) and M2 (right), along with the linear regression fit to the data (95% CI for the slope of the regression line of M1 = (0.29, 0.44), for M2 = (0.68, 0.95).
Montaño et al. Genome Biology 2013 14:R94 doi:10.1186/gb-2013-14-8-r94