ipyrad-analysis toolkit: treemix

The program TreeMix by Pickrell & Pritchard (2012) is used to infer population splits and admixture from allele frequency data. From the TreeMix documentation: “In the underlying model, the modern-day populations in a species are related to a common ancestor via a graph of ancestral populations. We use the allele frequencies in the modern populations to infer the structure of this graph.”

Required software

A minor detail of the treemix conda installation is that it installs openblas in place of MKL and will downgrade numpy and scipy to use this as a backend. This may interfere with other tools you have installed that require MKL, but it rarely does. To avoid any problems you could create a separate conda environment when installing treemix.

# conda install ipyrad -c bioconda
# conda install treemix -c bioconda
# conda install toytree -c eaton-lab

import ipyrad.analysis as ipa
import toytree
import toyplot

print('ipyrad', ipa.__version__)
print('toytree', toytree.__version__)
! treemix --version | grep 'TreeMix v. '

ipyrad 0.9.13
toytree 1.0.0
/usr/bin/sh: 1: treemix: not found

Required input data files

Your input data should be a .snps.hdf database file produced by ipyrad. If you do not have this you can generate it from any VCF file following the vcf2hdf5 tool tutorial. The database file contains the genotype calls information as well as linkage information that is used for subsampling unlinked SNPs and bootstrap resampling.

# the path to your HDF5 formatted snps file
data = "/home/deren/Downloads/ref_pop2.snps.hdf5"

Short Tutorial:

If you entered population information during data assembly then you may have already produced a .treemix.gz output file that can be used as input to the treemix command line program. Alternatively, you can run treemix using the ipyrad tool here which offers some additional flexibility for filtering SNP data, and for running treemix programatically over many parameter settings.

The key features offered by ipa.treemix include:

1. Filter unlinked SNPs (1 per locus) many times for replicate analyses.
2. Filter by sample or populations coverage.
3. Plotting functions.
4. Easy to write for-loops

# group individuals into populations
imap = {
    "virg": ["TXWV2", "LALC2", "SCCU3", "FLSF33", "FLBA140"],
    "mini": ["FLSF47", "FLMO62", "FLSA185", "FLCK216"],
    "gemi": ["FLCK18", "FLSF54", "FLWO6", "FLAB109"],
    "bran": ["BJSL25", "BJSB3", "BJVL19"],
    "fusi": ["MXED8", "MXGT4", "TXGR3", "TXMD3"],
    "sagr": ["CUVN10", "CUCA4", "CUSV6", "CUMM5"],
    "oleo": ["CRL0030", "HNDA09", "BZBB1", "MXSA3017"],
}

# minimum % of samples that must be present in each SNP from each group
minmap = {i: 0.5 for i in imap}

# init a treemix analysis object with some param arguments
tmx = ipa.treemix(
    data=data,
    imap=imap,
    minmap=minmap,
    seed=123456,
    root="bran,fusi",
    m=2,
)

Samples: 28
Sites before filtering: 349914
Filtered (indels): 0
Filtered (bi-allel): 13077
Filtered (mincov): 0
Filtered (minmap): 109383
Filtered (combined): 117718
Sites after filtering: 232196
Sites containing missing values: 221834 (95.54%)
Missing values in SNP matrix: 822320 (12.65%)
subsampled 29923 unlinked SNPs

# print the command string that will be called and run it
print(tmx.command)
tmx.run()

treemix -i /home/deren/Documents/ipyrad/newdocs/cookbook/analysis-treemix/test.treemix.in.gz -o /home/deren/Documents/ipyrad/newdocs/cookbook/analysis-treemix/test -m 2 -seed 123456 -root bran,fusi

# draw the resulting tree
tmx.draw_tree();

# draw the covariance matrix
tmx.draw_cov();

1. Finding the best value for m

As with structure plots there is no True best value, but you can use model selection methods to decide whether one is a statistically better fit to your data than another. Adding additional admixture edges will always improve the likelihood score, but with diminishing returns as you add additional edges that explain little variation in the data. You can look at the log likelihood score of each model fit by running a for-loop like below. You may want to run this within another for-loop that iterates over different subsampled SNPs.

# init a treemix analysis object with some param arguments
tmx = ipa.treemix(
    data=data,
    imap=imap,
    minmap=minmap,
    seed=1234,
    root="bran,fusi",
)

Samples: 28
Sites before filtering: 349914
Filtered (indels): 0
Filtered (bi-allel): 13077
Filtered (mincov): 0
Filtered (minmap): 109383
Filtered (combined): 117718
Sites after filtering: 232196
Sites containing missing values: 221834 (95.54%)
Missing values in SNP matrix: 822320 (12.65%)
subsampled 29923 unlinked SNPs

tests = {}
nadmix = [0, 1, 2, 3, 4, 5]

# iterate over n admixture edges and store results in a dictionary
for adm in nadmix:
    tmx.params.m = adm
    tmx.run()
    tests[adm] = tmx.results.llik

# plot the likelihood for different values of m
toyplot.plot(
    nadmix,
    [tests[i] for i in nadmix],
    width=350,
    height=275,
    stroke_width=3,
    xlabel="n admixture edges",
    ylabel="ln(likelihood)",
);

2. Iterate over different subsamples of SNPs

The treemix tool randomly subsamples 1 SNP per locus to reduce the effect of linkage on the results. However, depending on the size of your data set, and the strength of the signal, subsampling may yield slightly different results in different iterations. You can check over different subsampled iterations by re-initing the treemix tool with a different (or no) random seed. Below I plot the results of 9 iterations for m=2. I also use the global_=True option here which performs a more thorough (but slower) search.

# a gridded canvas to plot trees on
canvas = toyplot.Canvas(width=600, height=700)

# iterate over multiple set of SNPs
for i in range(9):

    # init a treemix analysis object with a random (no) seed
    tmx = ipa.treemix(
        data=data,
        imap=imap,
        minmap=minmap,
        root="bran,fusi",
        global_=True,
        m=2,
        quiet=True
    )

    # run model fit
    tmx.run()

    # select a plot grid axis and add tree to axes
    axes = canvas.cartesian(grid=(3, 3, i))
    tmx.draw_tree(axes)

# a gridded canvas to plot trees on
canvas = toyplot.Canvas(width=600, height=700)

# iterate over multiple set of SNPs
for i in range(9):

    # init a treemix analysis object with a random (no) seed
    tmx = ipa.treemix(
        data=data,
        imap=imap,
        minmap=minmap,
        root="bran,fusi",
        global_=True,
        m=3,
        quiet=True
    )

    # run model fit
    tmx.run()

    # create a grid axis and add tree to axes
    axes = canvas.cartesian(grid=(3, 3, i))
    tmx.draw_tree(axes)

3. Save the plot to pdf

import toyplot.pdf
toyplot.pdf.render(canvas, "treemix-m3.pdf")