AIPhylo

This is the beginning of an attempt to train an AI to infer phylogenies.

Dependencies

PyCogent

Simulating alignments

We follow the scheme set out in Philippe et al 2005: http://bmcevolbiol.biomedcentral.com/articles/10.1186/1471-2148-5-50

from simulator import simulate_alignment

# simulate a 100bp alignment
aln, tree = simulate_alignment(0.1, 0.5, 0.1, 100, 1.0, 'random')
print aln
print tree.asciiArt()

Should give you something like this:

>a
TATCATTCTATACGTTGACCGCTTTTAGCGCAAGTTAAGCGGTTCTGTAATATCTGGTCTTGTAGTACTGTTACGAGGAGCTAGCTACCCTTTAGGGCTA
>c
GGCCATGGTAAAAAAGATGTTGCCAACCCCGGCGCTACCATGATAAATGGGTTGAGACTCGTCATTTGCCCGGATCAGGTTGTCGTCGCTGCCGACGACG
>b
GGCCATGGTAACAAAGATGTGGCCACGCCCGGCGCTACCATGATAAATGCTATGAGATGTCAGTCTCGAGCGGAAAGCGAGTCGCTACTTGCCCAGGCTT
>d
TATCATTCTATACGTTGACTGCTTTTCGCGCACGTTAAAGGGTTATGTAATCTCCGGTCTGGTGGGATTCTGAGGTCGGCAAGCCAAGCTTTTAGCGTCG


                    /-a
          /edge.0--|
         |          \-d
-root----|
         |          /-b
          \edge.1--|
                    \-c

The tree seems to be arbitrarily midpoint rooted. I think that's just a PyCogent thing for viewing.

Is the ML tree the true tree?

We can answer this with the ml4 function, which finds the ML tree by getting the likelihood (under the JC69 model) of all 3 unrooted 4-taxon trees, and returns True if the true tree is the ML tree, an d False otherwise:

from simulator import *
from ml import *

aln, tree = simulate_alignment(0.675, 0.15, 0.16, 10000, 0.9, 'random')
ml4(aln, tree) 

A quick test

In Phillipe et al figure 3, the suggestion is that when p = 0.675, q = 0.15, and t = 0.9, the BL50 is ~0.20. We can roughly test this by simulating a set of datasets with r = 0.2 and seeing if we get roughly 50% of datatsets where the ML tree is the true tree, and 50% where it isn't.

input = [simulate_alignment(0.675, 0.15, 0.2, 10000, 0.9, 'random') for _ in range(20)]
results = [ml4(i[0], i[1]) for i in input]

# percent True
float(sum(results))/float(len(results))

When I do this, I get 0.65, which is close enough to suggest our code works as expected.