Question on ''penalized for reporting multiple clusters for a given gene' and pvalues for the ROC curves

[clin...@umn.edu]

Hello Dr Davidson,

I like Corset and its algorithm paper (https://genomebiology.biomedcentral.com/articles/10.1186/s13059-014-0410-6).

Thank for the explain on the 'Unique True positive' https://groups.google.com/forum/#!topic/corset-project/JdliHRN0AbQ.  I have two more questions:

1. Your paper page 6 stated 'Firstly, we examined the cumulative number of unique true positive clusters as a function of the total number of clusters (Figure 4)....This metric penalized for reporting multiple clusters for a given gene (that is, poor recall)'  

My understanding is the X axis of Figure 4 only counts the top ranked clusters. If multiple clusters (no matter 2 cluster or 10) were reported for a given gene, the X axis only counts the top ranked one. Could you kindly explain how the 'metric penalized for reporting multiple clusters for a given gene'?  

2. I did not completely understand how to plot a ROC. When you plot the ROC (Fig. 4 and Fig.5), did you use a range of pvalues, which from the cuffdiff2 and edgeR, to get list of number of unique true positives and list of number of top ranked clusters? 

If pvalue used. Could you kindly explain what range of p used?

Thank you very much.

Best wishes,

Chen

[Nadia Davidson]

Hi Chen,

1.  The X axis of Figure 4 includes all clusters, not just the "unique" ones. It's the Y axis that has the extra clusters that represent a gene removed. So if a DE gene had say 6 clusters, the first ranked one would increase the number of unique true positives by one, where as the other 5 would not.

2.The ROC curves were made using the p-value ranked genes from cuffdiff2. We used all p-values. The same plots were also made with EdgeR just to check that the results didn't depend on how DE testing was done. These aren't shown in the main paper, but are in the supplementary material.

Cheers,

Nadia.