mgCov

Multi-group Shared Subspace Covariance Estimation

Reference: Shared Subspace Models for Multi-Group Covariance Estimation (JMLR, 2019)

Installation

Install mgCov using devtools:

# install.packages("devtools")
devtools::install_github("afranks86/mgCov")

Example

We’ll demonstrate the use of mgCov with a gene expression dataset from patients with multiple subtypes of acute lymphoblastic leukemia from Yeoh et al (2002). For simplicity we include a subset of 1000 genes. The data includes expression levels on 327 split across 7 leukemia subtypes.

library(mgCov)
data(leukemia)

sapply(data_list, function(x) nrow(x))
#>    BCR-ABL   E2A-PBX1 Hyperdip50        MLL     OTHERS      T-ALL 
#>         15         27         64         20         79         43 
#>   TEL-AML1 
#>         79

S <- getRank(data_list)

Vinit <- mgCov::subspaceInit(data_list, S)
EMFit <- subspaceEM(data_list, S=S)
#> [1] "Reached maximum iterations in line search."

Vfit <- EMFit$V ## inferred basis for the shared subspace

Now run (conditional) Bayesian covariance estimation using the inferred subspace.

samples <- fitBayesianSpike(V=Vfit, Ylist=data_list, 
                             niters=1000, nskip=10, verbose=FALSE)

Let’s compare the gene expression covariance matrices of E2A-PBX1 to MLL.

groups_to_plot = c(1, 2, 4)
names(groups_to_plot) <- names(data_list)[groups_to_plot]
create_plots(V=Vfit, samples, group1=2, group2=4, to_plot = groups_to_plot, view=c(1, 2))

We can compare the same groups on a different two dimensional subspace. By setting view to …

This is analogous to looking at the 3rd and 4th principal components in a standard PCA.

create_plots(V=Vfit, samples, group1=2, group2=4, to_plot = groups_to_plot, view=c(3, 4))
#> Warning: Removed 1 rows containing missing values (geom_label_repel).

We can compare different groups

create_plots(V=Vfit, samples, group1=2, group2=4, to_plot = groups_to_plot, view=c(3, 4))
#> Warning: Removed 1 rows containing missing values (geom_label_repel).

Advanced features

TO DO