I have a question about the correlation heat map plot which generation by diffBind.

When we employing diffBind on those commands, it will generate heat map, such as dba(), dba.count().dba.analyze(). And I found that within the same sample set, the correlation heat map is so different. Is there a way that I could know what's the dataset be used for generated the correlation map? I guess the plot from dba() is based on the peak file. Could you comment on that? Thank you.

Y

The correlation heatmaps are generated using the correlations between the binding patterns for each sample.

After the peaks are initially loaded, using dba(), the peaks are merged and the scores (normalized to 0..1) are retained. If a peak was not called for a sample, it gets a score of -1. So each sample has a vector of scores for each peak, which may contain -1 scores for peaks not associated with that sample.

After a call to dba.count(), the reads are counted in every sample for every peak, and a score computed. The default score is a normalized read count (after subtracting control counts). So each sample has a vector of scores for each consensus peak. Generally there are at least some counts for each peak, so the scores are more uniform. I find that if there is a big difference in clustering between the peak-based and read-based correlation heatmaps, this may be due to the noisy nature of how peaks are called or left out for individual samples. The count-based scores generally exhibit higher correlations.

After a call to dba.analyze(), a correlation heatmap is plotted using only differentially bound peaks. As these peaks differentiate between two conditions, one would expect to see the samples clustering into those two conditions (although this will not necessarily be the case).

Hope this helps!

Cheers-

Rory

Hi Rory,

Thank you so much for the nice explanation. It is really helpful.

Y