What Line of Evidence is Strongest for Evolution?

[swamidass (S. Joshua Swamidass)]

@Mung

So they are indeed such prima facie evidence for common descent.

I’m not convinced. But hopefully we are working our way through it. I’m not arguing against common descent. I’m arguing for clarification on how we decide that a phylogenetic tree is evidence for common descent. It appears to have something to do with “statistical significance.”

This is a legitimate question. I just wish it could have arisen much earlier, perhaps here:

@Mung

Are not phylogenetic trees the very evidence for common descent that sequence data provides?

Not in my mind. Perhaps you could provide a brief description of how software creates a tree given a set of sequences and the conditions under which the software would fail to produce a tree.

[Mercer (John Mercer)]

@Mung

So if I am reading you correctly, i could not make up some imaginary taxa, allocate some sequence data to them, and get a phylogenetic tree? Or that I could do that but … what?

You could input random sequence data. You would almost certainly get several trees, all with nonsignificant (high) p values. Taxa have absolutely nothing to do with this; that suggests that you really aren’t grasping that this is an simply mathematical process.

@Mung

Could you briefly describe the role of the p value? I could create a tree in the above hypothetical scenario and it would come with a p value, and that p value would indicate … what?

It is basically the probability of getting the tree to which it is attached over a tree derived from non related sequences, the null hypothesis.

[T_aquaticus]

@Mung

Would I be too far off base if I thought that meant that something other than “chance” was at play? Not chance, therefore common descent? Or, we can rule out chance, and we can rule out gremlins , therefore common descent?

It’s a bit different. It is asking if chance could produce data that looks like common descent.

This is one of the basic concepts in statistics. For example, you may do an experiment where you try to determine if red heads are taller than with blond hair. You survey 100 people, 50 with red hair and 50 with blond hair. When you crunch the numbers and find that red heads are 1 inch taller on average than blonds. Have you proven that red heads are taller?

The question you have to ask is what are the chances of randomly selecting red heads that are slightly taller than the average red head, and/or randomly selecting blonds that are slightly shorter than the average blond? What you are asking is the probability of chance producing data where red heads are 1 inch taller on average. That is the basics of the famous Student’s T-test.

[swamidass (S. Joshua Swamidass)]

@T_aquaticus

It’s a bit different. It is asking if chance could produce data that looks like common descent.

@T_aquaticus, I think the situation is far more complex than this.

1. There are several ways of computing a p-value, and each one might use a different null hypothesis.
2. There are several ways of constructing a tree, some of which do not look for nested clades.
3. There are several ways of assessing the tree-likeness of a dataset.
4. There are several ways the tree-likeness can be violated by known processes.

To exposit this fully, we probably need another conversation. Perhaps we should start that thread. The point, however, is that using a the right software, we can test directly whether the data falls into nested clades. We can also test specific theories of special creation. The only theories of special creation that survive are those that are evidentially equivalent common descent.

(NOTE: “mimic CD” is a biased description of these models)

[Mung]

@Mercer

Taxa have absolutely nothing to do with this; that suggests that you really aren’t grasping that this is an simply mathematical process.

It was an example. The “taxon” would just be a label that is associated with a particular set of sequences. It’s the set of sequences that matter and i understand that.

[Mercer (John Mercer)]

@AllenWitmerMiller

I have a question. If the p-values weren’t so impressive, wouldn’t those resulting graphical structures be more like confused, tangled networks (as in M:M networks)—and not anything much like trees at all?

No, I’m pretty sure you’d only get lots of trees with unacceptably high p-values.

@AllenWitmerMiller

Indeed, wouldn’t Mung’s imagined “trees” actually be tangled spaghetti? (They’d have various random clumping, no doubt, but still just a mess. Right?)

No, I’m pretty sure that’s not the case. If you want to see the reticulation at the nodes of a tree caused by hybridization and other phenomena, here’s at least one program for that–I haven’t used it.

From a phylogenetic tree to a reticulated network - PubMed

[Mercer (John Mercer)]

@swamidass

@T_aquaticus, I think the situation is far more complex than this.

In his/her defense, we’re trying to keep it simple as Mung is working from very basic misconceptions.

[T_aquaticus]

@Mercer

In his/her defense, we’re trying to keep it simple as Mung is working from very basic misconceptions.

I only have a simple understanding of the methods used for phylogenetics, so any correction is much appreciated.

[Mercer (John Mercer)]

@Mung

It was an example. The “taxon” would just be a label that is associated with a particular set of sequences. It’s the set of sequences that matter and i understand that.

Then please don’t bring it up.

[Mung]

@Mercer

It is basically the probability of getting the tree to which it is attached over a tree derived from non related sequences, the null hypothesis.

ok, I’m struggling with this. The null hypothesis is that the sequences are not related. Are randomly generated trees used for that? Randomly generated sequences? In order to compare the probability some assumption has to be made about which sequences are related and which are not, would that be correct?

I might need to download some code and have a look at it to see what is gong on.

@Mercer

Then please don’t bring it up.

I work better with examples. I beg your indulgence.

[swamidass (S. Joshua Swamidass)]

@Mercer

@T_aquaticus, I think the situation is far more complex than this.

In his/her defense, we’re trying to keep it simple as Mung is working from very basic misconceptions.

I have found that ultimately works against understanding. By simplifying out the exceptions, we are vulnerable to @Pnelson claiming that homoplasy demonstrates common descent is false. It is better not to simplify, or these false objections will continue to confuse people no end.

One of the best pieces of advice about learning BioLogy was from a Computer Science professor who began doing computational biology in the 1990s. Contrasting this with computer science, he explained there are rules in biology, but there are exceptions to every rule, and every exception is important. We should teach students the rules, but upfront acknowledge that there are exceptions, and the exceptions are important. We are not ignorant of them, but are only cutting them out to aid their learning.

@Mung

Then please don’t bring it up.

I work better with examples. I beg your indulgence.

Then let us start with teaching you how to run the programs yourself. Can you find one online that you feel you can work with?

[Mung]

@swamidass

Then let us start with teaching you how to run the programs yourself. Can you find one online that you feel you can work with?

I doubt any are written in Ruby, which would be my preference, but I’ll spend some time looking. Barring that either Python or R would be my next choices. I’m certain i can find something for either of those.

I wonder if there are any book/software packages that come together? Anyone?

ETA:

http://nbisweden.github.io/MrBayes/

But we may not want to go there right away.

[Mung]

I’d prefer a simple program without a lot of bells and whistles and one in which the source code is available. If anyone has any suggestions.

[swamidass (S. Joshua Swamidass)]

@Mung

I’d prefer a simple program without a lot of bells and whistles and one in which the source code is available. If anyone has any suggestions.

MrBayes is great. Download it, go through the tutorials, and report back what you find out.

[John_Harshman]

@Mung

Something more is required, whether consideration of p values or something else. That is the ONLY point I have been trying to make.

Next time, try actually saying that from the start. Saves time and effort.

[John_Harshman]

@T_aquaticus

The method does something similar to what you are asking. It randomizes the data and compares it to the hypothesized tree. If the tree you are testing isn’t significantly different from random data then it casts doubt on the tree being the product of common descent.

Well, that’s not exactly what a bootstrap does. Instead it randomly resamples sites or characters, with replacement, to get a data set of the same size but of somewhat different composition from the original — i.e. missing some sites but having several copies of others — and analyzes that data set. It resamples many times, say 100. The bootstrap percentage of a node is the percentage of resampled data sets in which that node appears in the analysis. It’s a test, in other words, of the self-consistency of the data: do different samples of sites agree on the same tree, or do they display nested hierarchical structure? We would not expect random data to be consistent in that way.

[John_Harshman]

@Mung

I’d prefer a simple program without a lot of bells and whistles and one in which the source code is available. If anyone has any suggestions.

Joe Felsenstein’s PHYLIP package would fit your needs there. Free! Online!

[John_Harshman]

Let me present a very simplified explanation of how most phylogeny programs work: First, create a tree, possibly with reference to the data, possibly not. Second, evaluate the fit of the data to the tree using some optimality criterion, i.e. some algorithm that produces a number that describes the fit. The two most common criteria are minimum inferred length (the minimum number of changes that have to take place over the tree in order to explain the data) and likelihood (the probability of observing the data given the tree and a particular model of evolution). Rinse and repeat for a whole bunch of trees, keeping the tree that fares best under the optimality criterion. OK, that’s the “best” tree. But, especially under likelihood criteria, even random data are probably going to give you a single best tree. The question is whether that tree is really better than the others. That’s when statistical testing comes in. The bootstrap, which I’ve already described, is one such test. Likelihood ratio tests are another. If, given some test, the tree we have turns out to be a significantly better fit to the data than other trees, that’s evidence that the tree is real, and so is common descent. We do not expect data to display this hierarchical structure without common descent. Did that help?

Incidentally, Mr. Bayes doesn’t work that way, though the way it does work is analogous: it too picks a tree, evaluates it, and wanders around the space of all trees making these optimality (in this case likelihood) evaluations. But what it does with those evaluations is rather more complicated than I’ve described.

[swamidass (S. Joshua Swamidass)]

Don’t forget to explain discordant mutations.

[John_Harshman]

@swamidass

Don’t forget to explain discordant mutations.

Who are you talking to?