starmie: Basic Usage

knitr::opts_chunk$set(tidy = TRUE, 
                      fig.align = 'center', fig.width = 8, fig.height = 6)

Starmie: making population structure analyses easier

A very common part of modern population genetics analysis is inferring underlying population structure from genetic markers such as single nucleotide polymorphisms (SNPs) or microsatellites. The two main methods for this task are the Bayesian STRUCTURE algorithm or the frequentist ADMIXTURE. We have found that processing the output of these programs and performing meaningful inference and visualization of the results is far more difficult than it should be. This is why we wrote starmie.

Some key features:

This vignette outlines how to use starmie to do basic tasks after running STRUCTURE at the command line.

A basic STRUCTURE pipeline in R.

To use all the options in starmie for STRUCTURE output, we require for each run the 'out_f' file produced by the program and the logging information so we can produce MCMC diagnostics. To get the latter the output of STRUCTURE must be redirected to a file. Below we present an example of running STRUCTURE in multiple runs for each K in parallel. We assume that the mainparams and extraparams files are correctly specified and that the user has access to the path of the STRUCTURE binary. Also make sure RANDOMIZE option is turned off, so independent seeds can be set in each run.

input_file <- system.file("inst/extdata/microsat_testfiles", "locprior.str", package = "starmie")
main_params <- system.file("inst/extdata/microsat_testfiles", "mainparams", package = "starmie")
extra_params <-  system.file("inst/extdata/microsat_testfiles", "extraparams", package = "starmie")
runStructure("path/to/structure", input_file, main_params, extra_params, "run", 5, 2, 2)

Parsing STRUCTURE files, the 'struct' object

The basic unit of analysis for starmie is the struct object, which contains the model information in the STRUCTURE out_f file and optionally the logging information for the MCMC diagnostics. As an example, we have run STRUCTURE on simulated microsatellite data from the STRUCTURE example data and save the out_f. To create a struct object from a run we use the following:

library(starmie)
# path to file name
k6_file <- system.file("extdata/microsat_testfiles/", "locprior_K6.out_f", 
                       package = "starmie")
# create struct object
k6_msat <- loadStructure(k6_file)

k6_msat

The STRUCTURE object contains the following information about a single run:

list_names <- names(k6_msat)
list_description <- c("K parameter supplied to STRUCTURE",
                      "Input parameters",
                      "Assigned cluster membership proportions",
                      "Pairwise Fst values between inferred clusters",
                      "Average nucleotide distance within clusters",
                      "Within cluster average Fst values",
                      "Model fit diagnositcs", 
                      "Individual ancestral probability of membership to cluster",
                      "Estimated ancestral allele frequencies for each cluster",
                      "MCMC burn-in diagnositcs",
                      "MCMC post burn-in diagnostics")

knitr::kable(data.frame(attributes = list_names, 
                        description = list_description))

Of most interest to users would be the ancest_df which is the Q-matrix of individual cluster membership probabilities. To extract that information for inspection use the helper function getQ.

Q_hat <- getQ(k6_msat)

Why you're here: the infamous bar plot...

To make the bar-plot simply type:

plotBar(k6_msat, facet = FALSE)

This will group the known sample labels into population labels if they were supplied to the STRUCTURE run. Alternatively, you can facet the inferred cluster labels to make it easier to see outliers and geographical groupings.

plotBar(k6_msat)

If you have not given population labels to your samples you can also add them using the populations argument.

Loading multiple 'struct' objects the 'structList'

The structList is a container for manipulating multiple struct objects. Some potential use-cases are:

On our example microsatellite data to add the second run for the results of running STRUCTURE $K$ = 6 ,we first load the output file, and then pass both struct objects to the structList function.

k6_file_run2 <- system.file("extdata/microsat_testfiles/", 
                            "run2_locprior_K6.out_f", 
                       package = "starmie")

k6_run2 <- loadStructure(k6_file_run2)

k6_all <- structList(k6_msat, k6_run2)

k6_all

We can also compare the cluster labeling by using plotMultiK (and see that label-switching over different MCMC runs is a problem!)

plotMultiK(k6_all)

Diagnostics: or checking out your chains

A very simple approach to determining whether you need to rerun a STRUCTURE a model is to plot the estimated log-likelihood over each iteration over the post burn-in MCMC phase. If the chains have converged the log-likelihood should stabilize towards the final iterations and the variance within a run should be relatively low. The plotMCMC can plot the log-likelihood or admixture coefficient against the iteration over different runs and different $K$ values. Note this requires the logging file to be read by loadStructure.

Here we show an example when $K$ = 10 and the number of runs is also 10.

multiple_runs_k10 <- exampleStructure("mcmc_diagnostics")

mcmc_out <-plotMCMC(multiple_runs_k10, facet = FALSE)

head(mcmc_out$mcmc_info)

Inference on K is hard

Usually you would run STRUCTURE multiple times for multiple values of $K$ and then use estimates of the log-likelihood to determine the 'best' choice of $K$ that explains the population structure in your data. There are two choices for model selection - either use the maximum mean log-posterior probability estimated by STRUCTURE or use the Evanno method. The bestK function returns the value of $K$ that is estimated by these methods and also produces diagnostic plots.

multi_K <- exampleStructure("multiple_runs")
bestK(multi_K)
bestK(multi_K, "structure")

CLUMPPING together - avoiding label switching

We have written R implementations of the popular CLUMMP and CLUMPAK algorithms for combining Q-matrices over different runs of STRUCTURE. Usually, this step is performed after choosing a value for $K$, when the analyst would like to refine their estimates of cluster memberships. To perform CLUMPPING create a structList consisting of the same value of $K$ for multiple runs. In each case the Q-matrices and a matrix of column permutations for each run are returned.

We return to the example of our two runs of $K$ = 6, stored in k6_all defined above.

Q_list <- lapply(k6_all, getQ)
clumpak_results <- clumpak(Q_list)

clumppy <- clumpp(Q_list, method = "greedy")

# plot the results
plotMultiK(clumppy$Q_list)

Several other algorithms to correct label switching are available, including fast implementations for large values of K.

Other visualisations

As the STRUCTURE model outputs other information, we have implemented some multidimensional scaling plots to visualize some of the neglected features of the STRUCTURE model.

For example, we can plot the net nucleotide distance between clusters:

plotMDS(k6_msat)

Bugs, feature requests and miscellana

Please submit any bugs or feature requests as an issue to https://github.com/sa-lee/starmie/issues

References

Kopelman, N. M., Mayzel, J., Jakobsson, M., Rosenberg, N. A. & Mayrose, I. Clumpak: a program for identifying clustering modes and packaging population structure inferences across K. Mol. Ecol. Resour. 15, 1179–1191 (2015).

Stephens, M. Dealing with label switching in mixture models. J. R. Stat. Soc. Series B Stat. Methodol. 62, 795–809 (2000).

Jakobsson, M. & Rosenberg, N. A. CLUMPP: a cluster matching and permutation program for dealing with label switching and multimodality in analysis of population structure. Bioinformatics 23, 1801–1806 (2007).

Hubisz, M. J., Falush, D., Stephens, M. & Pritchard, J. K. Inferring weak population structure with the assistance of sample group information. Mol. Ecol. Resour. 9, 1322–1332 (2009).

Falush, D., Stephens, M. & Pritchard, J. K. Inference of population structure using multilocus genotype data: dominant markers and null alleles. Mol. Ecol. Notes 7, 574–578 (2007).

Falush, D., Stephens, M. & Pritchard, J. K. Inference of population structure using multilocus genotype data: linked loci and correlated allele frequencies. Genetics 164, 1567–1587 (2003).

Pritchard, J. K., Stephens, M. & Donnelly, P. Inference of population structure using multilocus genotype data. Genetics 155, 945–959 (2000).

Verity, R. & Nichols, R. A. Estimating the Number of Subpopulations (K) in Structured Populations. Genetics (2016).

Appendix

sessionInfo()


Try the starmie package in your browser

Any scripts or data that you put into this service are public.

starmie documentation built on May 1, 2019, 8:01 p.m.