CAMPrep: Data preprocessing for CAM

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

Description

This function perform preprocessing for CAM, including norm-based filtering, dimension deduction, perspective projection, local outlier removal and aggregation of gene expression vectors by clustering.

Usage

CAMPrep(data, 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)

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.
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 function. 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 function. 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. 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.

Details

This function is used internally by CAM function to preprocess data, or used when you want to perform CAM step by step.

Low/high-expressed genes are filtered by their L2-norm ranks. Dimension reduction is slightly different from PCA. The first loading vector is forced to be c(1,1,...,1) with unit norm normalization. The remaining are eigenvectors from PCA in the space orthogonal to the first vector. Perspective projection is to project dimension-reduced gene expression vectors to the hyperplane orthogonal to c(1,0,...,0), i.e., the first axis in the new coordinate system. local outlier removal is optional to exclude outliers in simplex formed after perspective projection. Finally, gene expression vectors are aggregated by clustering to further reduce the impact of noise/outlier and help improve the efficiency of simplex corner detection.

Value

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

Valid	logical vector to indicate the genes left after filtering.
Xprep	Preprocessed data matrix.
Xproj	Preprocessed data matrix after perspective projection.
W	The matrix whose rows are loading vectors.
SW	Sample weights.
cluster	cluster results including two vectors. The first indicates the cluster to which each gene is allocated. The second is the number of genes in each cluster.
c.outlier	The clusters with the gene number smaller than MG.num.thres.
centers	The centers of candidate corner clusters (candidate clusters containing marker genes).

Examples

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

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

#preprocess data
rPrep <- CAMPrep(data, dim.rdc = 3, thres.low = 0.30, thres.high = 0.95)

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