A variational framework for inferring population structure from SNP genotype data.

Project maintained by Anil Raj Hosted on GitHub Pages — Theme by mattgraham


fastStructure is an algorithm for inferring population structure from large SNP genotype data. It is based on a variational Bayesian framework for posterior inference and is written in Python2.x. Here, we summarize how to setup this software package, compile the C and Cython scripts and run the algorithm on a test simulated genotype dataset.


Anil Raj, Matthew Stephens, and Jonathan K. Pritchard. fastSTRUCTURE: Variational Inference of Population Structure in Large SNP Data Sets , (Genetics) June 2014 197:573-589 [Genetics, Biorxiv]



fastStructure depends on

A number of python distributions already have the first three modules packaged in them. It is also straightforward to install all these dependencies (1) using package managers for MACOSX and several Linux distributions, (2) from platform-specific binary packages, and (3) directly from source

Detailed instructions for installing these dependencies are provided in the README file in the fastStructure software bundle.

Getting the source code

To obtain the source code from github, let us assume you want to clone this repo into a directory named proj:

mkdir ~/proj
cd ~/proj
git clone

To retrieve the latest code updates, you can do the following:

cd ~/proj/fastStructure
git fetch
git merge origin/master

You can also retrieve the code using wget by doing the following:

wget --no-check-certificate

Building Python extensions

Before building python extensions, it is important to identify the path to the library files and, and header file gsl/gsl_sf_psi.h that are part of your GSL installation. For a default installation of GSL, the libraries (.so files) are usually found in /usr/local/lib and the header files (.h files) in /usr/local/include. In this case, you can add these lines to your .bashrc file on your home directory.

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib
export CFLAGS="-I/usr/local/include"
export LDFLAGS="-L/usr/local/lib"

Then, run source ~/.bashrc to set these environment variables.

To build library extensions, you can do the following:

cd ~/proj/fastStructure/vars
python build_ext --inplace

To compile the main cython scripts, you can do the following:

cd ~/proj/fastStructure
python build_ext --inplace

Each setup will create some .c and .so (shared object) files. This setup may give some warnings, which are OK. If you get errors that indicate the build failed, this might be because the wrong compiler is being used or environment variables (like LD_LIBRARY_PATH) are set incorrectly. To use a specific gcc compiler, you can do the following:

CC=/path/to/compiler python build_ext --inplace

Executing the code

The main script you will need to execute is To see command-line options that need to be passed to the script, you can do the following:

$ python

Here is how you can use this script

Usage: python
     -K <int>   (number of populations)
     --input=<file>   (/path/to/input/file)
     --output=<file>   (/path/to/output/file)
     --tol=<float>   (convergence criterion; default: 10e-6)
     --prior={simple,logistic}   (choice of prior; default: simple)
     --cv=<int>   (number of test sets for cross-validation, 0 implies no CV step; default: 0)
     --format={bed,str} (format of input file; default: bed)
     --full   (to output all variational parameters; optional)
     --seed=<int>   (manually specify seed for random number generator; optional)

fastStructure performs inference for the simplest, independent-loci, admixture model, with two choices of priors that can be specified using the `--prior` flag. Thus, unlike Structure, fastStructure does not require the `mainparams' and `extraparam' files. The inference algorithm used by fastStructure is fundamentally different from that of Structure and requires the setting of far fewer options. All options for fastStructure can be passed via the flags listed above.

Main options

The key options to pass to the scripts are the input file, the output file and the number of populations. Assuming the input file is named genotypes.bed (with corresponding genotypes.fam and genotypes.bim), the output file is named genotypes_output and the number of populations you would like is 3, you can run the algorithm as follows:

python -K 3 --input=genotypes --output=genotypes_output

This generates a genotypes_output.3.log file that tracks how the algorithm proceeds, and files genotypes_output.3.meanQ and genotypes_output.3.meanP containing the posterior mean of admixture proportions and allele frequencies, respectively. The orders of samples and SNPs in the output files match those in the `.fam` file and `.bim` file, respectively. Note that input file names should not include suffixes (e.g., .bed) and are relative to the main project directory (unless a full path is provided).

Input data format

The current implementation can import data from plink bed format and the original Structure format. If the data are in plink format, ensure that bed, bim and fam files for the dataset are all present in the same path.

While the original Structure program allowed for a more flexible input format, fastStructure expects a more specific Structure-like input format. Specifically, rows in the data file correspond to samples, with two rows per sample (note that only diploids are handled by this software), and columns correspond to SNPs. The first 6 columns of the file will be ignored; these typically would include IDs, metadata, etc. This software only handles bi-allelic loci. The two alleles at each locus can be encoded as desired; however, missing data should be encoded as `-9'.

Running on test data

A test simulated dataset is provided in test/testdata.bed with genotypes sampled for 200 individuals at 500 SNP loci. The output files in test/ were generated as follows:

$ python -K 3 --input=test/testdata --output=test/testoutput_simple --full --seed=100
$ ls test/testoutput_simple*
test/testoutput_simple.3.log  test/testoutput_simple.3.meanP  test/testoutput_simple.3.meanQ  
test/testoutput_simple.3.varP  test/testoutput_simple.3.varQ

$ python -K 3 --input=test/testdata --output=test/testoutput_logistic --full --seed=100 --prior=logistic
$ ls test/testoutput_logistic*
test/testoutput_logistic.3.log    test/testoutput_logistic.3.meanQ  test/testoutput_logistic.3.varQ
test/testoutput_logistic.3.meanP  test/testoutput_logistic.3.varP

$ tail -n 3 test/testoutput_simple.3.log
Marginal Likelihood (avg over genotypes) = -0.9777044544
Total time = 4.7611 seconds
Total iterations = 160

Executing the code with the provided test data should generate a log file identical to the ones in test/, as a final check that the source code has been downloaded and compiled correctly. The algorithm scales linearly with number of samples, number of loci and value of K; the expected runtime for a new dataset can be computed from the runtime in the above log file.

Choosing model complexity

In order to choose the appropriate number of model components that explain structure in the dataset, we recommend running the algorithm for multiple choices of K. We have provided a utility tool to parse through the output of these runs and provide a reasonable range of values for the model complexity appropriate for this dataset.

Assuming the algorithm was run on the test dataset for choices of K ranging from 1 to 10, and the output flag was --output=test/testoutput_simple, you can obtain the model complexity by doing the following:

$ python --input=test/testoutput_simple
Model complexity that maximizes marginal likelihood = 2
Model components used to explain structure in data = 4

Visualizing admixture proportions

In order to visualize the expected admixture proportions inferred by fastStructure, we have provided a simple tool to generate Distruct plots using the mean of the variational posterior distribution over admixture proportions. The samples in the plot will be grouped according to population labels inferred by fastStructure. However, if the user would like to group the samples according to some other categorical label (e.g., geographic location), these labels can be provided as a separate file using the flag --popfile. The order of labels in this file (one label per row) should match the order of samples in the input data files.

$ python

Here is how you can use this script

Usage: python
     -K <int>  (number of populations)
     --input=<file>  (/path/to/input/file; same as output flag passed to
     --output=<file>   (/path/to/output/file)
     --popfile=<file>  (file with known categorical labels; optional)
     --title=<figure title>  (a title for the figure; optional)
Assuming the algorithm was run on the test dataset for K=5, and the output flag was --output=test/testoutput_simple, you can generate a Distruct plot by doing the following:

$ python -K 5 --input=test/testoutput_simple --output=test/testoutput_simple_distruct.svg