Hi, I am trying to plot a few ChIP-Seq signals on a region set using deepTools plotProfile. I have been using this tool but it is only recently I observed this unusual behavior. When I plotted identical data 4 times the outputs are very different. When I checked the corresponding output bed files, indeed the regions are clustered differently. This is true with the plotHeatmap as well. And, the clustered output bed file from plotProfile is not matching with that from plotHeatmap (in terms of cluster members). This is happening with the kmeans clustering option. Following are the codes I used and results.
computeMatrix reference-point \
-R regionset.bed \
-S H3K27ac.bw H3K4me1.bw H3K27me3.bw H3K4me3.bw \
--referencePoint center \
-b 3000 -a 3000 \
--verbose \
--missingDataAsZero \
-p max/2 \
--sortRegions descend \
--sortUsingSamples 1 \
--skipZeros \
-o "matrix.gz"
plotProfile -m matrix.gz --perGroup --kmeans 8 --numPlotsPerRow 4 -out "profile.png" --outFileSortedRegions "profile_K8.bed"
plotHeatmap -m matrix.gz \
--boxAroundHeatmaps no \
--whatToShow 'plot, heatmap and colorbar' \
--legendLocation none \
--colorList white,red \
--kmeans 8 \
--outFileNameMatrix "hpMatrix.tab" \
--outFileSortedRegions "heatmap_k8.bed" \
--outFileName "heatmap.png"
profile plots of 4 run with identical data:
Is it expected that the clustering will be different every time? or Am I missing something here?
Please help.
Thank you
•
written
by
Ekalavya •
The clustering can be different each time with kmeans if the initial cluster centroids are chosen randomly. I don't see any information in the deepTools documentation about what initializing method is used, but based on your results it seems like they may be assigned randomly. You can try a different number of clusters to see if you get a more stable result, or try the hierarchical clustering option which should produce the same results every time.
Thank you very much colin.kern! It is the random starting points of the cluster. It will be good if can have control over this. I am using hierarchical clustering. I am also trying to find out the optimal number of clusters.
Log in to answer this question.
Sorry, somehow the image is not uploaded. it can be found here. https://i.ibb.co/c1ggNJc/profile-Plots.jpg
You should have used this link and not the link to the website containing the image (https://ibb.co/xDCCMhS, which you had used originally). I've made the necessary changes (and also formatted your post better) now.