Model Fitting

Setup

Load the ExpressionSet data objects containing protein expressions and genomic copy number, which was generated in the previous step.

suppressPackageStartupMessages(library(Biobase))
cnaset <- readRDS("data/cna-expression-set.rds") 
cna <- t(Biobase::exprs(cnaset))
protset <- readRDS("data/prot-expression-set.rds")
pexp <- t(Biobase::exprs(protset))
#standardize
pexp <- scale(pexp); cna <- scale(cna);

Load the cross validation test sets, which have balanced molecular subtypes across the test sets.

testSetIds <- readRDS(file = "data/prot_cv_10k_test_sets77.rds")
#create training sets.
allSampleIds <- seq_len(nrow(pexp))
trainSetIds <- lapply(testSetIds, function(testSetId) setdiff(allSampleIds, testSetId))

Load the last iteration of the tuning grid used in the CV.Vote step.

tmap <- readRDS(file = "data/prot_last_tune_grid.rds")

Establish a parallel backend to evaluate model fits efficiently.

library(doParallel)
library(parallel)
ncores <- detectCores()  - 1
cl <- makeCluster(ncores)
registerDoParallel(cl)

Create a directory to save all the model fits from cvVote.

rp <- "~/scratch-data/neta-bcpls/mfits/"
system(paste("mkdir -p ", rp))

CV.Vote

The CV.Vote step will estimate optimal tuning parameters for the BCPLS application, which are denoted as \(\hat \lambda^*_1, \hat \lambda^*_2, \hat \lambda^*_3\).

library(spacemap)
spmapcv <- cvVote(Y = pexp, X = cna, 
                  trainIds = trainSetIds, testIds = testSetIds, 
                  method = "spacemap", tuneGrid = tmap,
                  resPath = rp, aszero = 1e-4,
                  tol = 1e-4, cd_iter = 60e7)

The estimated optimal tuning parameters (previously computed) are:

##         lam1    lam2     lam3
## 976 86.76054 18.8506 56.80362

Boot.Vote

Fit the spaceMap model on bootstrapped data replicates subject to \(\hat \lambda^*_1, \hat \lambda^*_2, \hat \lambda^*_3\), which will create a bootstrap ensemble of networks. Then, construct a final network through majority vote on the bootstrap ensemble of networks.

ens <- bootEnsemble(Y = pexp, X = cna, tune = spmapcv$minTune, method = "spacemap", B = 1000, 
                    aszero = 1e-4, tol = 1e-4, seed = 55139L)
ensbv <- bootVote(ens)

Next Step in Analysis

Please see the Annotation article for the next step in the analysis.