This software implements a method for computing posterior association probabilities of SNPs (and other quantities) in genome-wide association studies using Bayesian variable selection and model averaging (Peltola et al., 2012).

References:

• Peltola T, Marttinen P, Jula A, Salomaa V, Perola M, Vehtari A (2012) Bayesian Variable Selection in Searching for Additive and Dominant Effects in Genome-Wide Data. PLoS ONE 7(1): e29115. http://www.plosone.org/article/info:doi/10.1371/journal.pone.0029115

The source code is available at the homepage of the software or from their GitHub repositories. See below for compilation. There are no ready-made binary executables at the moment, but if you need one, send me an e-mail.

The software has been developed (and tested) on 64bit Linux.

Requirements: C, C++ and Fortran compilers (use GNU compiler collection) and BLAS, LAPACK (e.g., ATLAS or OpenBLAS) and Boost libraries.

Compilation: use the makefile provided (i.e., run make on command line in the source directory). You may have to adjust the paths of the libraries in the makefile.

The software have been developed only on 64 bit Linux. Complitation under Windows (using Cygwin) should be possible, albeit minor modification may be neccessary.

The software is used from command line. On Linux use terminal client and on Windows run cmd to access the command line. Then, change to the directory of the binary and execute it with the name of your configuration file as an attribute:

cd /directory/of/the/binary/
./bmagwa config.ini

; Example configuration file
; (lines starting with ; are comments)

[datafiles]
; Data files are in PLINK format (use only standard formats; not long or
; transposed etc.; see http://pngu.mgh.harvard.edu/~purcell/plink/).
;
; fam contains information on individuals (only phenotype is used).
file_fam = testdata/testdata.fam
; Genotypes in binary format (SNP-major mode; this is default in PLINK).
file_g = testdata/testdata.bed
; Covariates (excluding constant, which is added automatically).
file_e = testdata/testdata.e
; Optional phenotype file (overruns phenotypes of fam if provided).
file_y = testdata/testdata.y
; Set following to 1 to force coding of genotypes to 0,1,2 according to the
; number of minor alleles (i.e. if computing allele freq with current coding
; gives > 0.5, 0s are swapped to 2s and 2s to 0s). Usually you probably use 0.
recode_g_to_minor_allele_count = 1

[sizes]
; Number of individuals.
n = 1000
; Number of SNPs.
m_g = 10000
; Number of covariates (excluding constant).
m_e = 2

[sampler]
; Sampler types: PMV / NK / KSC / G.
; The latter three have limited implementation (no effect types other than A
; can be used).
type = PMV
; Number of iterations.
do_n_iter = 1000000
; Skip (in iterations) between computing rao-blackwellization (RB).
n_rao = 500
; Number of RB-steps to discard as burnin (affects _rao.dat outputfiles)
; relative to n_rao (so these settings start computing RB posterior estimates
; after 500*1000 iterations).
n_rao_burnin = 1000
; Keep at 0 (experimental; set to 1 to continue adaptations after burnin).
adaptation = 0
; Skip between printing status to _log.txt file.
verbosity = 100000
; Skip between saving MCMC samples.
thin = 10
; Skip between updating tau2 and missing genotypes.
n_sample_tau2_and_missing = 10
; Delayed rejection: 0 off, > 0 gives the threshold for restricting
; DR moves to move sizes less than or equal to the value.
delay_rejection = 10
; Initial value of the move size proposal distribution parameter (geom. dist).
p_move_size = 0.2
; The move size proposal distribution parameter for state changes of near-by
; snps.
p_move_size_nbc = 0.25
; The move size proposal distribution parameter for swaps of near-by snps.
p_move_size_nbs = 0.7
; SNP neighborhood size / 2 (i.e., to one direction as difference in the
; indices of snps) for state changes and swaps of near-by SNPs.
max_SNP_neighborhood_size = 20
; Adapt move size distribtion: 0 or 1.
adapt_p_move_size = 1
; Set to 0 to use expected jump distance optimization.
; Set to (0,1] to use acceptance rate coercion with the specified goal value
; for the rate.
; Has not effect is adapt_p_move_size = 0.
p_move_size_acpt_goal = 0
; Maximum move size: [1,255] (not tested with values > 20).
max_move_size = 20
; Don't adapt the proposal distribution for additions/removals (i.e., use
; uniform dist.): 0 or 1.
flat_proposal_dist = 0

[thread]
; Number of chains to run simultanously.
n_threads = 2
; Prefix for result files (directory should exist; in this case
; 'testdata/results/').
basename = testdata/results/chain
; Seed for random number generator for each chain (comma separated list).
seeds = 1234,2345

[model]
; Allowed effect types (possible values are A,H,D,R,AH; comma separated list).
types = A

[prior]
; Use individual tau2 parameters: 0 or 1.
use_individual_tau2 = 1
; Prior weights on different effect types (t parameter prior).
type_A = 1
type_H = 1
type_D = 1
type_R = 1
type_AH = 1
; Model size prior (mean and variance) used to set prior for gamma parameter.
e_qg = 20
var_qg = 300
; Residual variance (sigma2 parameter) prior.
; Give either R2mode_sigma2 (expected mode for proportion of variance
; explained) or s2_sigma2 (directly the parameter of invchi2-prior).
R2mode_sigma2 = 0.2
;s2_sigma2 = 1.0
nu_sigma2 = 1
; Effect size (tau2_t parameters) priors.
; _A is for additive effect. _H, _R, _D for others.
nu_tau2_A = 5
s2_tau2_A = 0.05
nu_tau2_H = 5
s2_tau2_H = 0.05
; Alpha prior mean value.
mu_alpha = 1
; Constant term effect size prior parameter (1/sigma^2_alpha_0).
inv_tau2_e_const_val = 0
; Covariate term effect size prior parameter (1/sigma^2_alpha_l).
inv_tau2_e_val = 0.001

Most of the output files are in binary format. A Python script/module to convert the results to text file format is included (run python postprocess.py for help). Note that the binary format may not be portable across computers with different architectures.

The output file names are prefixed with the basename-setting and the number of the chain.

Note: thinning affects the output. If do_n_iter = 100 and thin = 10 only 10 numbers will be in the (MCMC trace) output files.

• basename_prior.txt

  Contains the values of prior parameters.

• basenameX_log.txt

  The state of the sampler is printed to this file every verbosity iterations. You can check this file to monitor the progress.

• basenameX_samplerstats.txt

  Some statistics from the run (sampling times etc.).

Format of a single value is given after filename.

• basenameX_alpha.dat [double ('d')]

  Contains the value of alpha for each MCMC iteration.

• basenameX_jumpdistance.dat [unsigned char ('B')]

  Contains the realized number of changes made to the variable inclusion vector for each MCMC iteration (not thinned!).

• basenameX_loci.dat [unsigned int32 ('I')]

  Contains the SNPs included in the model for each MCMC iteration. You need basenameX_modelsize.dat to parse this file.

• basenameX_log_likelihood.dat [double ('d')]

  Contains the marginal log likelihood (conditional on beta variance parameters) of the linear model for each MCMC iteration.

• basenameX_log_prior.dat [double ('d')]

  Contains the log of model prior (model size and effect type prior) for each MCMC iteration. Does not include prior contribution of the beta variance parameters.

• basenameX_modelsize.dat [unsigned int32 ('I')]

  Contains the number of SNPs in the model for each MCMC iteration.

• basenameX_move_size.dat [unsigned char ('B')]

  Contains the size of move proposed for each MCMC iteration. Not thinned.

• basenameX_move_type.dat [unsigned char ('B')]

  Contains the type of move proposed for each MCMC iteration. Not thinned.

• basenameX_pve.dat [double ('d')]

  Contains the total proportion of variance explained and the proportions of variance explained by genetic effects and covariates for each MCMC iteration.

• basenameX_rao.dat [double ('d')]

  Contains the Rao-Blackwellised posterior association probability for each SNP.

• basenameX_rao_types.dat [double ('d')]

  Contains the Rao-Blackwellised posterior probabilities of effect types for each SNP (conditional on this SNP being included in the model). Note the ordering of types. Not created if only one possible type is specified in types.

• basenameX_sigma2.dat [double ('d')]

  Contains the residual variance for each MCMC iteration.

• basenameX_types.dat [unsigned char ('B')]

  Contains the effect type of each SNP included in the model for each MCMC iteration. Use basenameX_loci.dat to find the corresponding SNPs. Note the ordering of types. Not created if only one possible type is specified in types.

Ordering of types: Effect types are in order (with corresponding numeric value in basenameX_types.dat) A=0, H=1, D=2, R=3, AH=4 (excluding those not defined in types; however, the order in types is insignificant).

A simple dataset generated with PLINK is included in 'testdata' directory. It has 10000 SNPs and 1000 individuals. The last 20 SNPs are causal with heritability contribution of 0.02 each. To run the BMAGWA software on the dataset, on the command line run:

cd /directory/of/the/binary/
./bmagwa testdata/testdata.ini

Running the sampler for the dataset took under 10 minutes on 2.8GHz Intel Core2 Quad CPU. You can monitor the progress by looking into testdata/results/chain0_verbose.dat (you can tweak the sampling options in the configuration file if it seems to take too long...).

Print out from the program should look something like this:

-------------------------------------------------------------
BMAGWA software version 2.0
http://becs.aalto.fi/en/research/bayes/bmagwa/
-------------------------------------------------------------

Warning: using default value of 0 for option sampler.save_beta
Warning: using default value of 0 for option prior.s2_sigma2
Warning: using default value of 0 for option prior.eh_tau2_A
var y = 0.990225
var x = 0.353579
mean x = 0.530252
Precomputing SNP covariances
Initializing sampler 0
Initializing sampler 1
Creating thread 0
Creating thread 1
Completed join with thread 0 having a status of 0
Completed join with thread 1 having a status of 0

You may then use the provided Python script to convert the binary files to text files and to compute posterior association probabilities from the MCMC chain. On command line:

python postprocess.py convert testdata/results
python postprocess.py mcmcpos testdata/results/chain 10000 50000 1

The .txt files are probably easy to open in your preferred numerical software for further analysis. For example, in Matlab:

p_rao1 = load('testdata/results/chain0_rao.dat.txt');
p_rao2 = load('testdata/results/chain1_rao.dat.txt');
p_mcmc = load('testdata/results/chain_mcmcpos.txt');

% manhattan plot
figure(1); clf;
plot(p_mcmc, '.');

% compare rao-blackwellized and mcmc frequency estimates
figure(2); clf;
plot(p_mcmc, 0.5 * (p_rao1 + p_rao2), '.');

% modelsize histogram for the first chain
figure(3); clf;
hist(load('testdata/results/chain0_modelsize.dat.txt'));

% see which SNPs have posterior assocation probability > 0.5
find(p_mcmc > 0.5)

• Other sampler types than PMV do not allow effect types other than A.
• Near-by SNP state change proposal is not available when there are more than one effect type in use.
• Acceptance rate coercion is experimental (not tested / will not work in all cases).

• E-mail: tomi.peltola at aalto.fi
• WWW: http://becs.aalto.fi/en/research/bayes/bmagwa/
• GitHub: https://github.com/to-mi

Acknowledgements: The software uses inih, Boost, BLAS, LAPACK and LINPACK libraries.

Unit tests are located in the src/tests-subdirectory.

Additional dependencies to compile the tests include googletest and Cuba.

Tests in SamplerTestsF cases will fail as the threshold is set to too stringent. The differences in the posterior inclusion probabilities between the two computation methods are then printed and may be examined. Note that neither methods is exact.

GPL v3. See licenses/gpl-3.0.txt.

Licenses for the utilized libraries are also included in the licenses directory.