Open matsen opened 6 years ago
I love the term rogue taxa. 😝
Yeah seems like an alternative to cft pruning for sure. Not quite sure from the Stamatakis paper, whether it'll improve things (he uses consensus tree improvement as validation), but I guess we just try different options and see if it leads to better ASR resolution empirically?
Er, I don't think if it as an alternative to pruning. I still like the pruning strategy, but I want to have some guidelines about when to stop adding taxa.
Shouldn't we run our seq list through their rogue taxa clipping service and then prune as usual? Their input = a file containing bootstrap trees. I think we'd need to tinker and do more than our original FastTree generation, right?
Perhaps! My feeling is that we care most about things wandering around in the Bayesian posterior, so we could start by running things with a generous taxon set then do some trimming based on that.
How do we generate a file of bootstrap trees? Here is the link to their program RogueNaRok
So maybe we do something to the effect of wide net prune -> beast -> rogue flagging -> edge plotting
? By seeing how many were flagged by the post-beast rogue trimming would tell us the quality of the prune step.
Our *.trees
files are NEXUS format. Perhaps we could just reformat to Newick and submit to rogue trimming.
This "rogue taxon" idea seems like a good hypothesis for what happened with 125, but has it actually been confirmed?
Seems like we should make sure that these sequences with long branches are actually invading the ancestral lineage.
Here's an idea that will easily see if the long branches could be invading the ancestral lineage.
Use our runs to build a consensus tree, with posterior support. If the long branches are moving around too much, then the smallest clades containing them will have low support.
Ascii art:
/--- long branch sister
-----|
\--------------- long branch
should have low support.
Forgot to respond, I like this idea.
Copying from Slack, this shows in our case that the extra 25 taxa are rogue and are invading the seed lineage.
With just 100 sequences, we have high sister confidence:
But with 125 sequences:
^why are the invaders clustered around BF520.1, specifically @matsen ? Are the rogue taxa attracted to our seed? Can we easily prune these rogues out by editing our current prune.py
script or will we need to use RogueNaRok?
No, I don't think they are especially.
My thought was that for this particular data set we will do fine with taking the 100 set, that we can justify by saying that the branches get too long for the additional taxa.
I the the main problem may actually be the clock model in BEAST. I'm getting started working on that.
In Bayesian phylogenetics, "rogue taxa" are taxa that move around the tree. I definitely think that this is what we were seeing with the long branches in the 125 set. See, e.g. https://scholar.google.com/scholar?q=%22rogue+taxa%22&btnG=&hl=en&as_sdt=0%2C48
How shall we proceed? We could try some existing solution, or directly look for this effect in terms of things moving on and off the lineage. If there's too much uncertainty and they have long branches, perhaps they get pruned?