I'm analysing low cell number ChIP-seq data (3 ChIP replicates / 3 Input replicates). The replicates are highly variable due to the low amount of starting material and the number of PCR cycles used for amplification. I am counting reads into windows along the genome and quantile normalising the counts to try and overcome some of this technical variation. I decided to use limma-voom to test for differential windows between the ChIP factor and the Input chromatin (mainly because it allows me to use quantile normalisation, unlike edgeR or DESeq2, please correct me If I'm wrong).
I have uploaded an image of the mean-variance plot produced by limma-voom ( https://ibb.co/eDfSxw ). In my opinion there seems to be three distinct components to the data, which I have circled in a copy of the image ( https://ibb.co/eEkGqG ). To me the windows in red are the low count - high variance windows (These generally correspond to windows which have been highly amplified randomly in one of the replicates). The windows in green are the increasing count - decreasing variance windows (These seem to be windows containing genuine binding). And the windows in orange are the low to medium count - constant variance windows (This seems to be a mix of the other two windows).
My problem is that the windows in the lower half of the orange circle which have a low variance are (I think) being squeezed to the trend line and therefore the actual variance of those windows is inflated. From prior knowledge we know that some of these windows contain genuine binding. The topTable reports a positive logFC, but the FDR is non-significant (I'm using FDR < 0.1 as a threshold).
Overall I think the trend isn't a very good fit to my data, and would like to know if there is anything else I can do to solve this? I should mention I am already using the robust = TRUE and trend = TRUE arguments to eBayes.