I'm analyzing the HM450K array methylation data and I'm wondering why does some probes have this bimodal distribution? I've already removed probes that have SNPs or have SNPs within 10 bp. This is a histogram of beta values for methylation data at one CpG site. The axis is beta values ranging from 0 to 1.
Also are there any suggestions for modeling methylation with this distribution as the outcome? Thank you!
Methylation is generally a somewhat heritable property of a eukaryote genome - each site starts off either methylated or unmethylated in a gamete, and its status tends to be propagated. When you measure methylation in a population of cells, you are measuring some kind of average of all the cells, but in cases where inheritance is true, you are seeing the same level of methylation in all the cells in the pool.
Inevitably, in a diploid organism, some sites will be heterozygous for methylation status, so strictly speaking, methylation data is likely to be trimodel, but for in many (most?) organisms, heterozygosity for methylation is probably at a low level, so it looks bimodal at a gross scale. Perhaps this is what you are observing here.
There are of course many sites where methylation changes during development, in which case intermediate values would be observed. Also, since methylation is generally observed in a somatic tissue, there tend to be a proportion of cells where a methylated site has changed its methylation status due to unreliable copying, which is probably more faithful in the germline.
Modeling part: Suppose that linear model is used. Then assumption of the linear model is the distribution of error terms, estimated by residuals, instead of the distribution of response variable itself. So bimodal distribution displayed here cannot dis/prove anything. So maybe you can fit the linear model first, then check the distribution of residual. if still the residuals have bimodal distribution, try to fix the model based on mixture distribution.
