Dear All,

Many thanks for all of you who have replied to my post on Evoldir last
week. The responses were very useful and as many requested me to
forward them the replies, I have summarised them below. For privacy's
sake I have left out people's names.

Thank you once again,

Serinde

My original post:

I am running outlier analyses on a set of 25 markers but find
discrepancies between different methods, which I find hard to explain.
The dataset is comprised of both  putatively neutral microsatellites
and candidate SNP loci in genes of interest for divergently selected
populations over 6 generations (50 individuals per population).  Fdist
and Schlotterer's method are fairly consistent in the identification
of outlier loci, which also correspond to patterns of allelic
divergence and results obtained from exploratory analyses (AMOVA,
correspondence analysis). The output from BayeScan however is very
different: rather than indicating divergent selection at the loci
identified by Fdist and Schlotterer's methods (and others) other loci
are identified as under balancing selection. There therefore appears
to be a bias towards balancing selection rather than diversifying
selection at the expected loci, which doesn't make biological sense in
the case of my data. Analysing microsatellites and SNPs independently
does not radically change the findings.

I would like to know if anyone has observed similar discrepancies
between BayeScan and other outlier analyses or could think of ways how
such a shift (identification of different loci under balancing
selection rather than the same loci under divergent selection) could
be explained. Any suggestions would be extremely helpful. Thank you
very much in advance.

The responses:

-----

1)      The methods are really different, so you shouldn't be shocked
if they yield different results.

2)      The Beaumont & Nichols FDIST/Lositan test is very intuitive
and understandable, so if it does confirm likely divergent selection
at those candidate loci for which you already had biological
knowledge, then that is a strong indication!

3)      Consider that in most cases these tests are conducted on
samples from wild populations, with large Ne and very little knowledge
about them. You may have a lot of knowledge about your populations,
and over 6 generations. Also if your populations are small and there
is a strong genetic drift signal in your data maybe that affects the
results of Bayescan. I remember some experiments on Drosophila that
showed that when you have lab populations undergoing strong genetic
drift, the signal of selection may not be so obvious. Maybe FDIST
picks it up, and Bayescan doesn't.

4)      Do you have 2 or more populations? If you have more, you could
try to do tests per pairs.

5)      You could also try to use multivariate ordination techniques
(like PCA or FCA), separately for microsats and SNPs, and plot your
individuals alongside the alleles, to see if there is any decisive
pattern of association between certain alleles that you expect to be
under selection and those individuals that had been exposed to said
selective pressure (I suspect that if there is a real signal of
selection, you should see some striking scattering).

So, these were my two-cent - good luck with your research!

-----

I saw you message on evoldir. Sorry I canīt be of much use, but Iīd be
interested in the response you get so if you donīt post a summary on
evoldir, perhaps you could forward the replies to me? One thought did
occur. Perhaps your results are not as different from each other as
they appear, it just seems that the assessment of divergent selection
by BayeScan has shifted down for all loci. In that case outlier loci
look like loci under ordinary drift only, and loci under drift look to
be under stabilising selection. Perhaps a problem with your prior
settings?
Good luck in figuring out the difference and which program does the
better job in your case

-----

It is a general issue that BayeScan seems to be more conservative than
alternative tests (see attached if you haven't already), but be aware
of the major flaw in their discussion of the hierarchical model.
I don't know how large your Fst values are, but for marine fish
(approx. Fst=0.0001-0.04) we do also often get candidates for
balancing selection, but they're most likely false positives, and as
long as the neutral background Fst is relatively low, I don't think
it's possible to reliably detect true balancing selection in these
types of tests. Thus, I wouldn't worry too much about that result. The
lack of outliers in BayeScan may simply be due to low power in your
dataset, but I guess that if you couple these observations with time
series plots of Fst for the other outliers that could reinforce the
results (does it make sense).

-----

How large is your Fst values over time? Does it make sense that the
signal is simply too low for BayeScan to detect? A simulation trial
could verify that (like the Narum & Hess 2011 paper).

-----

I attach a paper that initiates some discussion for difference (Manel
et al 2008, BMC Evolutionary Biology 2009, 9:288). But not so easy to
explain. In other analysis, I also find very often difference.
Probably due to software assumption about demographic structure.

-----

We had a similar problem when applying BayesFst which uses the same
underlying model than Bayescan (although with different outputs and
explicit tests of positive selection on each locus) in the Eveno et al
paper (2008. Mol. Biol. Evol 25: 417-437). We interpreted this as the
fact that if mutation rates are really different between SSRs and SNPs
(which is the case a priori but might be more or less pronounced
depending on the species), the subdivided model used in Bayescan might
be affected and this could potentially lead to problem of convergence
and fitting of the different parameters. Did you test if your run had
actually converged correctly? what does it mean when you say that
using SNP only, this did not change the results radically? in this
case , a low number of loci might be insufficient also for Bayescan
which is fitted using the original data, while the Fdist2 uses
independently simulated data, and then compares neutral envelop to the
observed results. Be aware also that the fdist2 analysis might not be
completely adequate for SNP data (see in Eveno et al) due to their low
mutation rate and this is also visible in the Beaumont & Nichols paper
in one of the figures when they test the robustness of their model
(inflation of upper bound for low theta).

-----

Check out this paper on outliers: Nunes et al. 2011, Molecular Ecology
20: 193-205

-----

I would be careful regarding any conclusions I would make from few
markers, i.e. considering each of them separately or all together.
There is a lot of variation in terms of divergence (e.g. outliers)
even if the underlying demography is well know and informative (see
attached paper; Orozco-ter Wengel et al. 2011. Mol Ecol 20:1108-1121).
I would suggest to perform a bunch of simulations under a demographic
scenario that broadly fits your data and then look at the distribution
of simulated Fst values you get. The simulations I suggest should be
performed with software like ms (Dick Hudson) and then converted into
microsatellite data with ms2ms (Pidugu and Schloetterer) so they fit
the stepwise mutation model. Finally calculate the Fst for your couple
of thousands of simulated markers with MSA or any other such software
to get an idea of the distribution of Fst under the null of your
demography and  plain neutrality.

-----

I got completely different results from Bayescan and LOSITAN on the
same set of microsatellites and populations.  I am appending to my
response the set of responses that I received on a similar query to
evoldir.  The bottom line?  I think it may be the case that none of
these approaches are very reliable, but Bayescan was certainly much
more conservative, finding only one outlier (which wasn't one of the
six that LOSITAN identified).

What follows are the replies that I received to my evoldir query
regarding apparently non-neutral microsatellite loci.  For privacy's
sake, I have dropped people's names.  Thanks to all who provided
responses.

My original post: "Most analyses of population genetic variation and
structure in populations assume the neutrality of the markers used.
There are a number of software tools that test for selection on loci,
with LOSITAN (which uses the It uses the FST-outlier method (Vitalis
et al. 2001; Beaumont 2005) my preferred tool. In a data set of 31
microsatellite loci, as many of as six test as being significantly
under either balancing or positive selection across my samples. I
would just like to get a feel for whether the community at large,
faced with such a scenario, would 1) advocate dropping those loci from
the data set, 2) leaving them in or 3) presenting analyses both with
and without those loci included.  My question is asked from the
standpoint of presentation in publication - I intend to analyze both
the full and trimmed data set for my own interests in how non-neutral
loci affect population genetic analyses."

Responses:

I would be cautious with only using the classical  Fst method (for
example LOSITAN) for detecting loci under selection. The classical
methods for detection of loci under selections are based on simulating
a FST-null distribution across all loci and from this detect loci that
lie outside the credibility region, and therefore assumed to be under
selection. These methods apply a simple demographic model such as
coalescent-based approach assuming genetic drift to be the contributor
to differentiation among populations; outliers are therefore taken as
evidence of selection.
Novel methods (such as BAYESCAN ver. 2.01 (Foll and Gaggiotti, 2008))
extent the classical approach to include dynamic processes such as
gene flow are based on detection of LD among pairs of loci. The
demographic models are more advanced and more realistically describing
ecological scenarios, including migration among subpopulations. The
degree of differentiation (FST) decomposed into a locusŽ-specific
component (alpha), shared by all populations, and a
population-specific component (beta), shared by all loci. Selection is
assumed when alpha is necessary for explaining the observed pattern of
diversity.
For testing for loci under selection I have used both LOSITAN and
BAYESCAN, the latter seemed to be more realistic, when the population
structure and history was taken into consideration.

I guess I'd say your results are probably evidence that the method
used to identify "positive selection" is junk (like most such
methods). Almost certainly false positives...

I would look into how your loci are segregating your samples. You
could do PCAs or correspondence analyses of the allele distributions
among your samples and see how they segregate for each marker, then
you can compare what your presumably neutral and what your presumably
selected markers are doing. If segregation is the same among all
markers (i.e. throughout the genome), then I would say you do not have
selection, just some markers that are particularly good at picking the
biological signal, and thus, in my opinion, all markers should be
included into the estimation of differentiation. If you neutral
markers all segregate your samples one way, but your outlier loci do
it another (or several other ways), then you may have selection, and I
would report your neutral differentiation, and the differentiation due
to "selected" markers.

It depends on the impact on the results and the amount of data you
have.  I definitely prefer to check it. Often the impact for the final
conclusions is very limited and can be noted verbally.
(Well, dendrograms are notoriously instable, but one does not need
them anyway.) You might want to do some more realistic simulations to
explore other confounding factors such as sampling from a spatially
structured population, founding events and so on.

I would suggest excluding these 6 loci since you will have many left.
I would also check what each non neutral ones says.  I would also
check the repeat motive. Trinucleotides are more likely to be non
neutral than dinucleotidic loci. I would also, if possible try to know
where these loci come from in the genome (in a coding sequence
etc...).

I would strongly recommend to at least report the outlier results and
subsequent results with and without them.

I would definitely analyse data with and without. In my experience,
this will likely change things. But it might only affect only some of
the populations analysed (which is reasonable if any such locus is
indeed affected by some local selective force specific to some area of
your study).
If the loci "potentially under selection" do change the picture quite
a bit, you would need to check on BLAST if they are anywhere near a
genomic region that has some functional implication.
Please also note that you should consider also the fact that in
certain conditions, when a population is quickly expanding its range
at distribution margins, some rare alleles can suddenly increase in
frequency as a result of these peripheral founding at the front wave
of expansion (see recent papers by Excoffier & Co about the "allele
surfing" hypothesis). Reviewers will almost certainly ask you to
consider that.

The idea of comparing the analyses with and without the selected loci
seems interesting (just be aware of the smaller sample size, and thus
lower power, in the "neutral" fraction compared to the whole list of
loci). Also I have a feeling that outlier analyses such as LOSITAN are
quite prone to giving false positive hits.

serindevanwijk@gmail.com