I am trying to use Arlequin to detect outliers from a data set of ~20,000 SNPs. I can see that a large number of them are above the 1% quantile line when plotted but I want to know if it is possible to get a list of these as there are too many to look at simply using the graph. I have seen papers with large numbers of markers using arlequin to detect outliers so I am assuming it can be done, I just don't know how?
I assume that you are trying to detect loci under selection and for that you need a set of populations with enough sampling size to be compared. I wonder if the large number of outliers is connected to the marker you are using.To detect loci under selection or outliers you have to assume that the markers you use are neutral and that departures from neutrality are detected by a Bayesian analysis. SNPs, in most cases, do not meet that condition of neutrality and maybe biased to selective conditions.
In the results directory (name of your input file.res), you should have a file named fdist2_simOut.txt which contains the p-values for each of your SNPs. It should allow you to precisely identify the outliers.
Alfonso: Thanks. I am aware that SNPs may be biased and that it is likely that I have a large number of false positives however this is only part of the analysis I am doing to determine which of my SNPs are under selection (only those which show up in a range of analysis will be considered real).
Yann: Is the p value not the p value for the Fst value? I was under the impression that outliers where the ones which plot above the 1% quantile line? I have a large number of SNPs which are below this line but show a signifincant Fst p value.
The p-values show whether or not the observed Fst values of your markers are significantly different of the simulated Fst values (so here significant does not mean different from zero, but different from your neutral simulated distribution).
If you have SNPs with significant p-values, but that are below the 1% line in your graphical representation, it may be because the 1% line is plotted using a limited number of heterozygosity values (20 when I tested this with the example file), so it is not very precise. I would better trust the p-values for each SNP, since they are computed using the He and Fst values of the marker.
To answer your other question (how calculate the quantile lines), you can find the mathematical details in: Excoffier et al. (2009), Heredity, 103: 285–298. As I understand it and very roughly (I hope no statistician is reading that), they use a kernel density function to extrapolate a joint distribution for He and Fst from the simulated values, and then compute the p-values from this distribution. With the equations in the paper, I guess it is possible to compute quantile lines (i.e. calculate Fst values knowing the p-values and He), but once again, I think it would be much simpler to just look directly at the p-values for your markers.
I hope my answers are clear enough and will be of some help!
Ok, thanks, I think I was looking at the p values wrong. Someone else suggested that the reason I have signifincant points below the line is because the line is not made up of enough points so I added the points to look at this. The problem is that I have some significant points almost directly below some of the points used to make up the line (see attached- red points are signifincant to p
I am not sure about what could cause that. The only thing I can think of is the possibility that individual p-values are not calculated very accurately in this area, since there is only a low number of simulated values at the edge of the distribution to compute the joint density. However, if it was the case, I think the 1% quantile line would be more “jagged”, so I am really not sure it could be a problem. One solution to check if this is an issue would be to do the analysis again and see if you get similar p-values. Sorry I can’t give you a more precise answer.
Just out of curiosity: the black dots are observed values, right? Do you know why you have some gaps in the distribution (bottom left corner of the graph)?
Thanks. I had an email converstation with someone else yesterday and he suggested that it was becasue there was not enough simulations however I've run it again with 150,000 simulations and the although the line changes, I still get points plotting below the line. Having run the analysis again, the p values for those under the line are changing.
The person I was speaking to yesterday said that gaps like mine sometimes occur when you have low sample numbers but I'm not sure what counts as low sample numbers (I have 53 and 56 individuals in my two populations).
I'd say that more than 50 individuals per population is fine, but maybe the low number of populations is a problem.
Anyway, if I were you, I would not focus on the points close to the 1% quantile line, it is likely the simulated Fst distribution is too narrow compared to the one for your real populations. And since you are going to do other analyses, you will see which SNPs are detected as under selection by the different approaches.
Regarding the significant points below the 99% quantile, one possibility is the quantile is printed after reducing the original data (i.e., spline). I think otherwise it should be more ragged (high and low peaks).
Regarding the changing p-values, I wouldn't worry too much (well, depending on the magnitude of the chage). That is what you expect when you run simulations from different starting points.
Thanks Yann and Pablo. We have decided to pay more attention to the P values than the line for the time being. As I am doing a range of different analysis only those consistently showing as outliers will be considered. Thanks for all your help.
- Badges
- Science topic