mbpca: Extension of PCA to analyze multiple data sets

In mengchen18/mogsa: Multiple omics data integrative clustering and gene set analysis

Description

Three approaches are supplied in this function, consensus PCA (CPCA), generalized CCA (GCCA) and multiple co-inertia analsyis (MCIA).

Usage

mbpca(
  x,
  ncomp,
  method,
  k = "all",
  center = TRUE,
  scale = FALSE,
  option = "uniform",
  maxiter = 1000,
  moa = TRUE,
  verbose = TRUE,
  k.obs = "all",
  w = NA,
  w.obs = NA,
  unit.p = FALSE,
  unit.obs = FALSE,
  pos = FALSE
)

Arguments

x	A list of matrix or data.frame, where rows are variables and columns are samples. The columns among the matrices need to be match but the variables do not need to be.
ncomp	An integer; the number of components to calculate. To calculate more components requires longer computational time.
method	A character string could be one of c("globalScore", "blockScore", "blockLoading"). The "globalScore" approach equals consensus PCA; The "blockScore" approach equals generalized canonical correlation analysis (GCCA); The "blockLoading" approach equals multiple co-inertia anaysis (MCIA);
k	The absolute number (if k >= 1) or the proportion (if 0<k<1) of non-zero coefficients for the variable loading vectors. It could be a single value or a vector has the same length as x so the sparsity of individual matrix could be different.
center	Logical; if the variables should be centered
scale	Logical; if the variables should be scaled
option	A charater string could be one of c("lambda1", "inertia", "uniform") to indicate how the different matrices should be normalized. If "lambda1", the matrix is divided by its the first singular value, if "inertia", the matrix is divided by its total inertia (sum of square), if "uniform", none of them would be done.
maxiter	Integer; Maximum number of iterations in the algorithm
moa	Logical; whether the output should be converted to an object of class moa-class
verbose	Logical; whether the process (# of PC) should be printed
k.obs	The absolute number (if k >= 1) or the proportion (if 0<k<1) of non-zero coefficients for the observations. Sparse factor scores for observation are used by sparse concordance analysis. (New arguments from v1.12)
w	The weight of variables. It could be given in the following format: 1) NA or a numeric value: all variables have the same weight; 2) A vector of numeric values, the vecter has the same length as x: variables in each block shares the same weight; 3) A list of vector, each vector in the list has the same length as the number of row in the corresponding table/block, then each variable use a different weight. See detail how to select weight. (New arguments from v1.12)
w.obs	The weight of observations, see w. (New arguments from v1.12)
unit.p	A logical value, whether the loading vectors (for variables) for each table/block should be unit length.
unit.obs	A logical value, whether the score vectors (for observations) for each table/block should be unit length. (New arguments from v1.12)
pos	A logical value, whether only retain non-negative coefficients in loading and score vectors. (New arguments from v1.12)

Details

Select of weight for variables: In omics data, it is often true that low intensity variables suffers more noise. Therefore, The variables with higher intensities are more reliable. If we consider this, we can use the total sum intensity of a variable (or a tranform of it) as weight, the model would prefer to select high intensity variables.

Value

An object of class moa-class (if moa=TRUE) or an list object contains the following elements:

tb - the block scores

pb - the block loadings

t - the global scores

w - the wegihts of block scores to construct the global scor

Note

no note

Author(s)

Chen Meng

References

For clustering problem: Meng et al. 2015 moCluster: Identifying Joint Patterns Across Multiple Omics Data Sets. Journal of proteome research.

See Also

see moa for non-iterative algorithms for multi-block PCA.

Examples

data("NCI60_4arrays")
tumorType <- sapply(strsplit(colnames(NCI60_4arrays$agilent), split="\\."), "[", 1)
colcode <- as.factor(tumorType)
levels(colcode) <- c("red", "green", "blue", "cyan", "orange", 
                     "gray25", "brown", "gray75", "pink")
colcode <- as.character(colcode)



moa <- mbpca(NCI60_4arrays, ncomp = 10, k = "all", method = "globalScore", option = "lambda1", 
             center=TRUE, scale=FALSE)
plot(moa, value="eig", type=2)
r <- bootMbpca(moa, mc.cores = 1, B=6, replace = FALSE, resample = "sample")

moas <- mbpca(NCI60_4arrays, ncomp = 3, k = 0.1, method = "globalScore", option = "lambda1", 
              center=TRUE, scale=FALSE)


scr <- moaScore(moa)
scrs <- moaScore(moas)
diag(cor(scr[, 1:3], scrs))

layout(matrix(1:2, 1, 2))
plot(scrs[, 1:2], col=colcode, pch=20)
legend("topright", legend = unique(tumorType), col=unique(colcode), pch=20)
plot(scrs[, 2:3], col=colcode, pch=20)

gap <- moGap(moas, K.max = 12, cluster = "hcl")
gap$nClust


hcl <- hclust(dist(scrs))
cls <- cutree(hcl, k=4)
clsColor <- as.factor(cls)
levels(clsColor) <- c("red", "blue", "orange", "pink")
clsColor <- as.character((clsColor))

heatmap(t(scrs[hcl$order, ]), ColSideColors = colcode[hcl$order], Rowv = NA, Colv=NA)
heatmap(t(scrs[hcl$order, ]), ColSideColors = clsColor[hcl$order], Rowv = NA, Colv=NA)

genes <- moaCoef(moas)
genes$nonZeroCoef$agilent.V1.neg

mengchen18/mogsa documentation built on June 7, 2020, 6:05 p.m.