I have one-dimentional array (human genome). Also I have two annotations for it, we can think about them as different peaks (it's nucleosome and secondary structures). How can we find correlation and/or causation between this annotations? Currently we are forming hypothesis like "one element is always before other" and statistically testing them. Can we be smarter and use known tools or at least automate this?
6
-
1$\begingroup$ Do you mean with something like a Markov model? $\endgroup$– bennCommented Nov 17, 2017 at 20:26
-
$\begingroup$ Yeah, but how will you implement them? $\endgroup$– D MCommented Nov 17, 2017 at 21:32
-
$\begingroup$ I am also interested in applying such models to multiple 'omics data, I hope this question will get some useful answers. $\endgroup$– bennCommented Nov 18, 2017 at 8:33
-
1$\begingroup$ Could you specify which kind of hypothesis do you want to test? What I have seen is centering all the positions you are interested in and observe if there is a common pattern on the other variables observed (annotations, CpG, mutation rates...) which is not noise. See here an example $\endgroup$– llrsCommented Nov 21, 2017 at 11:44
-
1$\begingroup$ Perhaps phrase this as time series analysis? Does one event precede another in a statistical sense? I think if you look for tools in this domain, you'll find some good solutions. $\endgroup$– DanCommented Dec 19, 2017 at 10:34
|
Show 1 more comment
2 Answers
$\begingroup$
$\endgroup$
You can express the annotations of the genome with a dot plot, x and y represent the two annotations along the genome. The you can summary the result with Spearman correlation coefficient(s) based on whole set of annotations or particular gene families.
$\begingroup$
$\endgroup$
I would start with metaplots for a visual/exploratory analysis to guide your statistical testing. Cut the genome into short segments e.g. centred on the centre of each nucleosome peak, overlay all the segments on top of each other, and plot the abundance of secondary structure features at each base count as you step away from the centres. This should give an easy view on whether there is any clear structure in the relative positions of the different objects.
Do the plots both ways round, and try different subsets of the objects (e.g. ones overlapping genes or not, different stringency thresholds for peak calling). This should hopefully give you some insights on what is worth testing for.
