taxon-specific sampling fractions?

[ptitle]

Hi,

Can Pesto accept taxon-specific sampling fractions? Or can you currently only provide a global sampling fraction? I'm not seeing mention of anything beyond global sampling fractions in the manuscript or manual, but wanted to ask to be sure.

thanks,
-Pascal

[kopperud]

Hi Pascal, thanks for your interest.

The simple answer is no. The method assumes that the species sampling probability is equal for all species.

The reason we did not implement it, is because we do not know how. If you are aware if there is a peer-reviewed publication where it is explained how to implement it (I am not sure where to look), then we could potentially add it as an option.

cheers
Bjørn

[ptitle]

Taxon-specific sampling fractions are implemented in a few diversification modeling approaches, including bamm and clads.
It is explained a bit in the second clads paper: https://academic.oup.com/sysbio/article/71/2/353/6316269

[kopperud]

an except from their paper

I suppose then we would have to rewrite the calculation of the sampling probability to be branch-specific, and then set the initial state of $E(t)$ for an internal node to depend on $E(t)$ for the left and right subtrees. If this is right, I think it should not be too much work to add. I might ask Sebastian for his opinion before I change anything in the software, though

[kopperud]

Out of curiosity, how important is this feature for empirical analyses, in your experience? In principle it makes sense that taxon sampling non-uniform, of course. However, are there many datasets where this kind of information is available?

Besides that, what would be a good way to provide this information in the program? Would it be okay to ask for a dictionary, for example

taxon_sampling = Dict(
    "Homo sapiens" => 1.0, 
    "Pan troglodytes" => 1.0,
    "Gorilla gorilla" => 0.5,
)

After the manuscript was accepted, I also wrote an R-package version of the same software ( https://github.com/kopperud/RPesto ). It behaves the same but has a bit fewer features. I could implement the taxon-sampling in both softwares, but it might be simpler to start with the R-package as it is not so big. Would you be more interested in using the R package or the Julia program?

[ptitle]

To answer your questions:

How important is it? I think it's fairly important to address the fact that different clades might have substantially different sampling in a phylogeny (especially larger phylogenies), and this should influence diversification rate parameters. There are a couple of ways to deal with it: (1) just have a phylogeny-wide sampling fraction, (2) implement taxon-specific sampling fractions, or (3) impute missing taxa in order to have a distribution of fully sampled trees, and then model diversification on those trees (with 100% sampling). In the last case, you would then drop the imputed taxa, and the idea is that the diversification rates attributed to the branches in the tree that were not imputed are more accurate.
I think the imputation approach is probably best, as the imputation can be done with the level of detail needed, but depending on the system, that can lead to very large trees that are computationally demanding for diversification modeling.
So if Pesto is fast, maybe this is just the recommended approach for accounting for variable sampling, rather than implementing taxon-specific sampling fractions.

As for how frequently the information is available, I think anytime we have known species diversity vs what is sampled in the phylogeny, this can be implemented. In vertebrates, we often at least know how many described species exist within families or genera, vs what has genetic data and was included in a phylogeny. Outside of vertebrates, this probably more of an unknown.

As for R vs julia, I personally barely use julia, but use R all the time, so R would be my preference as that is where I am comfortable. But don't implement this for my sake! I was just asking =)