bctsne {bcTSNE} | R Documentation |
Calculate BC t-SNE by orthogonal gradient descent
bctsne(X, Z, k = 50, outDim = 2, perplexity = 30, maxIter = 1000)
X |
numeric matrix, input matrix |
Z |
numeric matrix, covariate matrix |
k |
integer of length 1, reduced dimension (number of eigenvectors) |
outDim |
integer of length 1, the output dimension |
perplexity |
numeric of length 1, the t-SNE perplexity |
maxIter |
integer of length 1, the maximum iterations for the BC t-SNE algorithm |
X
should be preprocessed (e.g. PCA, centered and scaled). Z
is the full model matrix, excluding the intercept.
list
wth the following items:
Xred
numeric matrix, the reduced dimension input to bctsne
Z
model matrix indicating batch membership
perplexity
perpelexity value used in computing t-SNE
Y
batch-corrected projection matrix
maxIter
maximum iterations used in training
## Create small simulated dataset, A, with embeded batch effects set.seed(2731) kRid <- 20 p <- 100 n <- 200 W <- matrix(rnorm(p*kRid), kRid) S <- matrix(rnorm(n*kRid), n) z <- sample(1:3, rep = TRUE, size = n) Z <- model.matrix( ~ -1 + as.factor(z)) l <- matrix(rnorm(kRid*NCOL(Z)), kRid) A <- (S - Z %*% t(l) ) %*% W ## Scale A to give input, X X <- scale(A) resUnadj <- Rtsne::Rtsne(X) ## Standard t-SNE resAdj <- bctsne(X = X, Z = Z, k = 10) ## Batch-corrected t-SNE ## Plot results, no true effects were included in the simulated data, so ## we expect all batches to overlap with bcTSNE; batch membership indicated ## by color plot(resUnadj$Y, col = z) plot(resAdj$Y, col = z)