BulkLMM.jl

BulkLMM is a Julia package for performing genome scans for multiple traits (in "Bulk" sizes) using linear mixed models (LMMs). It is suitable for eQTL mapping with thousands of traits and markers. BulkLMM also performs permutation testing for LMMs taking into account the relatedness of individuals. We use multi-threading and matrix operations to speed up computations.

The current implementation is for genome scans with one-degree of freedom tests with choices of adding additional covariates. Future releases will cover the scenario of more-than-one degrees of freedom tests.

We consider the case when a univariate trait of interest is measured in a population of related individuals with kinship matrix $K$. Let the trait vector, $y$ follow the following linear model.

$$ y = X\beta + \epsilon,$$

where

$$V(\epsilon) = \sigma^2_g K+\sigma^2_e I.$$

where $X$ is a matrix of covariates which would include the intercept, candidate genetic markers of interest, and (optionally) any background covariates. The variance components $\sigma^2_g$ and $\sigma^2_e$ denote the genetic and random error variance components respectively.

For a single trait and candidate marker, we use a likelihood ratio test to compare a model with and without the candidate genetic marker (and including the intercept and all background covariates). This process is repeated for each marker to generate the genome scan. The result is reported in LOD (log 10 of likelihood ratio) units.

Users can specify if the variance components should be estimated using ML (maximum likelihood) or REML (restricted maximum likelihood). The scans can be performed with the variance components estimated once under the null, or separately for each marker. The latter approach is slower, but more accurate.

Under the null hypothesis that no individual genetic marker is associated with the trait, traits are correlated according if the kinship matrix is not identity, and the genetic variance component is non-zero. Thus, a standard permutation test where we shuffle the trait data randomly, is not appropriate. Instead, we rotate the data using the eigen decomposition of the kinship matrix, which de-correlates the data, and then shuffle the data after rescaling them by their standard deviations.

Scans for multiple traits are performed by running univariate LMMs for each combination of trait and marker. We are exploring algorithms for optimizing this process by judicious use of approximations.

This package uses multi-threading to speed up some operations. You will have to start Julia with mutliple threads to take advantage of this. You should use as many threads as your computer is capable of. Further speedups may be obtained by spreading (distributing) the computation across mutliple computers.

The package BulkLMM.jl can be installed by running

using Pkg
Pkg.add("BulkLMM")

To install from the Julia REPL, first press ] to enter the Pkg mode and then use:

add BulkLMM

The most recent release of the package can be obtained by running

using Pkg
Pkg.add(url = "https://github.com/senresearch/BulkLMM.jl", rev="main")

We demonstrate basic usage of BulkLMM.jl through an example applying the package on the BXD mouse strains data.

First, after successfully installed the package, load it to the current Julia session by

using BulkLMM
using CSV, DelimitedFiles, DataFrames, Statistics

The BXD data are accessible through our published github repo of the BulkLMM.jl package as .csv files under the data/bxdData directory.

The original data for BXD spleen traits BXDtraits_with_missing.csvcontains missing values. We saved the data after removed any missings to the file named "spleen-pheno_nomissing.csv" under the same directory.

bulklmmdir = dirname(pathof(BulkLMM));
pheno_file = joinpath(bulklmmdir,"..","data/bxdData/spleen-pheno-nomissing.csv");
pheno = readdlm(pheno_file, ',', header = false);
pheno_processed = pheno[2:end, 2:(end-1)].*1.0; # exclude the header, the first (transcript ID)and the last columns (sex)

Required data format for traits should be .csv or .txt files with values separated by ',', with each column being the observations of $n$ BXD strains on a particular trait and each row being the observations on all $m$ traits of a particular mouse strain.

Also load the BXD genotypes data. The raw BXD genotypes file BXDgeno_prob.csv contains even columns that each contains the complement genotype probabilities of the column immediately preceded (odd columns). Calling the function readBXDgeno will read the BXD genotype file excluding the even columns.

geno_file = joinpath(bulklmmdir,"..","data/bxdData/spleen-bxd-genoprob.csv");
geno = readdlm(geno_file, ',', header = false);
geno_processed = geno[2:end, 1:2:end] .* 1.0;

Required data format for genotypes should be .csv or .txt files with values separated by ',', with each column being the observations of genotype probabilities of $n$ BXD strains on a particular marker place and each row being the observations on all $p$ marker places of a particular mouse strain.

For the BXD data,

size(pheno_processed) # (number of strains, number of traits)

(79, 35554)

size(geno_processed) # (number of strains, number of markers)

(79, 7321)

Compute the kinship matrix $K$ from the genotype probabilities using the function calcKinship

kinship = calcKinship(geno_processed); # calculate K

kinship = round.(kinship; digits = 12);

For example, to conduct genome-wide associations mapping on the 1112-th trait, we can run the function scan() with inputs of the trait (as a 2D-array of one column), geno matrix, and the kinship matrix. Type ?scan() for more detailed description of the function.

traitID = 1112;
pheno_y = reshape(pheno_processed[:, traitID], :, 1);

@time single_results = scan(pheno_y, geno_processed, kinship);

  0.059480 seconds (80.86 k allocations: 47.266 MiB)

The output single_results is an object containing model results about the variance components (residual variance and the heritability parameter) estimated under the null baseline model, and the lod scores, as the fields named respectively as "sigma2_e", "h2_null", and "lod". By default, variance components are estimated from maximum-likelihood (ML). The user may choose REML for estimating by specifying in the input "reml = true".

# VCs: residual variance, heritability which is the proportion of genetic variance to total variance
(single_results.sigma2_e, single_results.h2_null)

(0.0942525841453798, 0.850587848871709)

# LOD scores calculated for a single trait under VCs estimated under the null (intercept model)
single_results.lod; 

BulkLMM.jl supports permutation testing for a single trait GWAS. Simply run the function scan() and set the optional keyword argument permutation_test = true with the required number of permutations as nperms = # of permutations. For example, to ask the package to do a permutation testing of 1000 permutations, do

@time single_results_perms = scan(pheno_y, geno_processed, kinship; permutation_test = true, nperms = 1000);

  0.079464 seconds (94.02 k allocations: 207.022 MiB)

Similarly to the results of the single-trait scan with no permutation, single_results_perms contains the fields sigma2_e, h2_null, and lod for the original trait. Additionally, we report the results of permutation tests as the raw LOD scores computed for each permuted copies, which are stored in a matrix named as L_perms of dimension $p \times nperms$, where each column contains the LOD scores corresponding to $p$ markers on one permuted copy, and each row are the LOD scores for a particular marker fitted on all 1000 permuted copies.

size(single_results_perms.L_perms)

(7321, 1000)

Based on the results of the permutation test, we can use the function get_thresholds() to obtain the LOD thresholds according to the quantile probabilities, based on the significance levels requested.

For example, if we would like to see the significant LOD scores with significance levels of 0.10 and 0.05, we can run the function get_thresholds() and give raw results of LOD scores from permutation testing and the desired significance (0.10, 0.05). The user can ask for results of as many significance levels as they want. In this case, the function reports the 90th and the 95th quantiles among LOD scores testing all 1000 permuted copies. The results are as follows:

lod_thresholds = get_thresholds(single_results_perms.L_perms, [0.10, 0.05]);
round.(lod_thresholds, digits = 4)

3.3644  
3.6504

Let's plot the BulkLMM LOD scores of the 1112-th trait and compare with the results from running GEMMA:

Note: to get results from GEMMA, one would need to run GEMMA on a Linux machine with input files of the same trait (here the 1112-th trait, X10339113), genetic markers and the kinship matrix, and finally convert the LRT p-values into corresponding LOD scores. Alternatively, you may simply load the results we obtained by following the procedures mentioned above. The resulting LOD scores from GEMMA are a .txt file in data/bxdData/GEMMA_BXDTrait1112/gemma_lod_1112.txt.

For reproducing this figure, we need to do the following steps:

First, read in the gmap.csv and the phenocovar.csv under data/bxdData/ directory as

gmap_file = joinpath(bulklmmdir,"..","data/bxdData/gmap.csv");
gInfo = CSV.read(gmap_file, DataFrame);
phenocovar_file = joinpath(bulklmmdir,"..","data/bxdData/phenocovar.csv");
pInfo = CSV.read(phenocovar_file, DataFrame);

Next, load the results preprocessed from GEMMA:

gemma_results_path = joinpath(bulklmmdir,"..","data/bxdData/GEMMA_BXDTrait1112/gemma_lod_1112.txt")
Lod_gemma = readdlm(gemma_results_path, '\t'); # load gemma LOD scores results available in the package

Finally, we use the QTL plotting function from the package BigRiverQTLPlots.jl:

using BigRiverQTLPlots
traitName = pInfo[traitID, 1] # get the trait name of the 1112-th trait

plot_QTL(
	single_results_perms, 
	gInfo, 
	significance= [0.10, 0.05],
	legend = true,
	label = "BulkLMM.jl",
	title = "Single trait $traitName LOD scores"
)
plot_QTL!(
	vec(Lod_gemma), 
	gInfo, 
	linecolor = :purple, 
	label = "GEMMA", 
	legend = :topright
)

To get LODs for multiple traits, for better runtime performance, first start julia with multiple threads following Instructions for starting Julia REPL with multi-threads or switch to a multi-threaded julia kernel if using Jupyter notebooks.

Then, run the function bulkscan() with the matrices of the traits of interest, genome markers, and the kinship. Type ?bulkscan() for more detailed description of the function.

Here, we started a 16-threaded julia session in julia version 1.9.2. Specific session info is as follows:

versioninfo()

Julia Version 1.9.2
Commit e4ee485e909 (2023-07-05 09:39 UTC)
Platform Info:
	OS: Linux (x86_64-linux-gnu)
	CPU: 48 × Intel(R) Xeon(R) Silver 4214 CPU @ 2.20GHz
	WORD_SIZE: 64
	LIBM: libopenlibm
	LLVM: libLLVM-14.0.6 (ORCJIT, cascadelake)
	Threads: 17 on 48 virtual cores
Environment:
	JULIA_NUM_THREADS = 16

@time multiple_results_allTraits = bulkscan(pheno_processed, geno_processed, kinship);

2.112011 seconds (107.94 k allocations: 5.053 GiB, 2.59% gc time)

Please Note: the default method and modeling options for bulkscan() takes an approximated approach for the best runtime performance. The user may choose to use other methods and options provided for more precision but longer runtime, following the detailed instructions in ?bulkscan().

The output multiple_results_allTraits is an object containing our model results:

• the matrix of LOD scores $L_{p \times m}$, where $p$ is the number of markers and $m$ is number of traits; each column corresponds to the LOD scores resulting from performing GWAS on each given trait.
• variance components (heritability) results will be returned in various formats depending on the specific method and other options by the user. For more details, enter ?bulkscan().

size(multiple_results_allTraits.L)

(7321, 35554)

To visualize the multiple-trait scan results, we can use the plotting function plot_eQTL from BigRiverQTLPlots.jl to generate the eQTL plot. In the following example, we only plot the LOD scores that are above 5.0 by calling the function and specifying in the optional argument threshold = 5.0:

plot_eQTL(multiple_results_allTraits.L, pheno, gInfo, pInfo; threshold = 5.0)

If you find any bugs, please post an issue on GitHub or contact the maintainer (Zifan Yu) directly. You may also fork the repository and send us a pull request with any contributions you wish to make.

Check out NEWS.md to see what's new in each BulkLMM.jl release.