CAM: Convex Analysis of Mixtures

In Lululuella/CAMTHC: Convex Analysis of Mixtures for Tissue Heterogeneity Characterization

Description

This function performs a fully unsupervised computational deconvolution to identify marker genes that define each of the multiple subpopulations, and estimate the proportions of these subpopulations in the mixture tissues as well as their respective expression profiles.

Usage

CAM(data, K = NULL, corner.strategy = 2, dim.rdc = 10, thres.low = 0.05,
  thres.high = 0.95, cluster.method = c("K-Means", "apcluster"),
  cluster.num = 50, MG.num.thres = 20, lof.thres = 0.02,
  quick.select = NULL, sample.weight = NULL, generalNMF = FALSE,
  cores = NULL)

Arguments

data	Matrix of mixture expression profiles. Data frame, SummarizedExperiment or ExpressionSet object will be internally coerced into a matrix. Each row is a gene and each column is a sample. Data should be in non-log linear space with non-negative numerical values (i.e. >= 0). Missing values are not supported. All-zero rows will be removed internally.
K	The candidate subpopulation number(s), e.g. K = 2:8.
corner.strategy	The method to find corners of convex hull. 1: minimum sum of margin-of-errors; 2: minimum sum of reconstruction errors. The default is 2.
dim.rdc	Reduced data dimension; should be not less than maximum candidate K.
thres.low	The lower bound of percentage of genes to keep for CAM with ranked norm. The value should be between 0 and 1. The default is 0.05.
thres.high	The higher bound of percentage of genes to keep for CAM with ranked norm. The value should be between 0 and 1. The default is 0.95.
cluster.method	The method to do clustering. The default "K-Means" will use kmeans. The alternative "apcluster" will use apclusterK-methods.
cluster.num	The number of clusters; should be much larger than K. The default is 50.
MG.num.thres	The clusters with the gene number smaller than MG.num.thres will be treated as outliers. The default is 20.
lof.thres	Remove local outlier using lofactor. MG.num.thres is used as the number of neighbors in the calculation of the local outlier factors. The default value 0.02 will remove top 2% local outliers. Zero value will disable lof.
quick.select	The number of candidate corners kept after quickhull and SFFS greedy search. If Null, only quickhull is applied. The default is 20. If this value is larger than the number of candidate corners after quickhull, greedy search will also not be applied.
sample.weight	Vector of sample weights. If NULL, all samples have the same weights. The length should be the same as sample numbers. All values should be positive.
generalNMF	If TRUE, the decomposed proportion matrix has no sum-to-one constraint for each row. The default is FALSE. TRUE value brings two changes: (1) Without assuming samples are normalized, the first principal component will not forced to be along c(1,1,..,1) but a standard PCA will be applied during preprocessing. (2) Without sum-to-one constraint for each row, the scale ambiguity of each column vector in proportion matrix will not be removed.
cores	The number of system cores for parallel computing. If not provided, one core for each element in K will be invoked. Zero value will disable parallel computing.

Details

This function includes three necessary steps to decompose a matrix of mixture expression profiles: data preprocessing, marker gene cluster search, and matrix decomposition. They are implemented in CAMPrep, CAMMGCluster and CAMASest, separately. More details can be found in the help document of each function.

For this function, you needs to specify the range of possible subpopulation numbers and the percentage of low/high-expressed genes to be removed. Typically, 30% ~ 50% low-expressed genes can be removed from gene expression data. The removal of high-expressed genes has much less impact on results, and usually set to be 0% ~ 10%.

This function can also analyze other molecular expression data, such as proteomics data. Much less low-expressed proteins need to be removed, e.g. 0% ~ 10%, due to a limited number of proteins without missing values.

Value

An object of class "CAMObj" containing the following components:

PrepResult	An object of class "CAMPrepObj" containing data preprocessing results from CAMPrep function.
MGResult	A list of "CAMMGObj" objects containing marker gene detection results from CAMMGCluster function for each K value.
ASestResult	A list of "CAMASObj" objects containing estimated proportions, subpopulation-specific expressions and mdl values from CAMASest function for each K value.

Examples

#obtain data
data(ratMix3)
data <- ratMix3$X

#set seed to generate reproducible results
set.seed(111)

#CAM with known subpopulation number
rCAM <- CAM(data, K = 3, dim.rdc = 3, thres.low = 0.30, thres.high = 0.95)
#Larger dim.rdc can improve performance but increase time complexity

## Not run: 
#CAM with a range of subpopulation number
rCAM <- CAM(data, K = 2:5, dim.rdc = 10, thres.low = 0.30, thres.high = 0.95)

#Use "apcluster" to aggregate gene vectors in CAM
rCAM <- CAM(data, K = 2:5, dim.rdc = 10, thres.low = 0.30, thres.high = 0.95,
            cluster.method = 'apcluster')

#CAM with quick selection to reduce time complexity
rCAM <- CAM(data, K = 3, dim.rdc = 10, thres.low = 0.30, thres.high = 0.95,
        quick.select = 20)

#CAM with different sample weights (e.g. adjusted based on sample quality)
rCAM <- CAM(data, K = 3, dim.rdc = 5, thres.low = 0.30, thres.high = 0.95,
        sample.weight = c(rep(10,11),rep(1,10)))

#CAM for general NMF (no sum-to-one contraint for proportion matrix)
rCAM <- CAM(data, K = 3, dim.rdc = 5, thres.low = 0.30, thres.high = 0.95,
       generalNMF = TRUE)

## End(Not run)

Lululuella/CAMTHC documentation built on May 5, 2019, 2:39 a.m.