Runtime Estimates and Sensitivity

[MPR]

Thanks again for providing a tool like this. We are hoping to apply
it to a rather large DNA sequence dataset. Are there are any general
guidelines for runtime for popABC?

It would be helpful just to know the largest dataset this program has
been used for and how long it took to run. We are prepared to run
popABC in parallel, but initial trial runs of only 10 iterations seem
to indicate that the method is not tractable for our combination of
data, settings and priors. We are now trying to figure out what the
runtime is most sensitive to—e.g. the number of loci, the number of
alleles per loci, the range of values for specific priors, number of
populations, etc.

Because we are new to this, we thought it would be best to check in
first and find out what general expectations are before we do any more
complex analysis. Any guidance you have on this would be very
helpful.

Best
Megan

[Joao Sollari Lopes]

Dear Megan,

Thanks for your email.
In terms of datasets I have used up to 50 sequence loci with 3 pop and
small sampled populations (15-25 individuals per population).
I also did some analysis with 4, 5 and 6 populations but without too
many iterations (around 1,000,000).

Now for some considerations:

no. loci: for each locus popABC will run a coalescent model and
calculate the respective summary statistics (except for STRs if you
specify a prior for recombination, i.e. linked loci). So you need to
count on that computation time when adding more loci

no.samples: under the coalescence, there is no need for too much
samples per population. And these can really affect the time taken for
the coalescence to reach the MRCA.

mutation rate: the mutation rate does not affect at all the time for
the coalescence, but it can increase the computation time required for
the calculation of the summary statistics.

recombination rate: recombination will increase immensely the
computation time since it has a great effect in the coalescence. If
you need to add it, constrain its prior as much as you can.

mig priors: increasing the migration rate does not affect much the
computation time, but it will add quite a lot of noise to the problem-
space, so you might need more simulations to cover it.

time priors: increasing the range of these does not affect computation
time too much.

ne prios: this parameter affects the coalescence significantly. The
rate of coalescence is calculated straight from the ne's.

As for estimating the computation time needed, just make a simple
extrapolation from how much it takes to run 100 or 1000 simulations.

Choosing the tolerance interval and size of the reference table (i.e.
total number of simulations) is quite a problem in itself. I have
written some stuff regarding these in "10 years of Software
development for approximate Bayesian computation" in Clary TS (ed.)
"Horizons in Computer science Research". Take a look.

Hope all this rambling helps,
Joao