ccrepe: Calculates compositionality-corrected p-values and q-values...

In AlmaasLab/micInt: Find microbial interactions

Calculates compositionality-corrected p-values and q-values for compositional data using an arbitrary distance metric.

Description

ccrepe calculates compositionality-corrected p-values and q-values for compositional data by generating first a null distribution of the distance metric generated by permutation and renormalization of the data, and then by generating an alternative distribution of the distance metric by bootstrap resampling of the data. For greater detail, see References
The two distributions are compared using a pooled-variance Z-test to give a compositionality-corrected p-value. The p-values can be calculated for all appropriate (passing certain quality-control measures) pairwise comparisons, or for a subset of user-specified ones.
Q-values are additionally calculated using the Benjamin-Hochberg-Yekutieli procedure (see References)

Usage

 ccrepe(
 x = NA,
 y = NA,
 sim.score = cor,
 sim.score.args = list(),
 min.subj = 20,
 iterations = 1000,
 subset.cols.x = NULL,
 subset.cols.y = NULL,
 errthresh  = 1e-04,
 verbose = FALSE,
 iterations.gap = 100,
 distributions = NA,
 compare.within.x = TRUE,
 renormalize = TRUE,
 memory.optimize = FALSE
 )

Arguments

x	First dataframe or matrix containing the relative or absolute abundances in cavity1 : columns are bugs, rows are samples. (Rows should therefore sum to a constant.) The subjectIDs, if present, are assumed to be the row names and NOT the first column of data.
y	Second dataframe or matrix (optional) containing the relative or absolute abundances in cavity2: columns are bugs, rows are samples. The subjectIDs, if present, are assumed to be the row names. If both x and y are specified, they will be merged by row names. If no row names are specified for either or both datasets, the default is to use the row numbers as subject IDs.
sim.score	A function defining a similarity measure, such as cor or nc.score. This similarity measure can be a pre-defined R function or user-defined. If the latter, certain properties should be satisfied as detailed below (also see examples). The default similarity measure is Spearman correlation. A user-defined similarity measure should: 1.Be able to take either two inputs which are vectors or one input which is either a matrix or a dataframe 2.In the case of two inputs, return a single number 3.In the case of one input, return a matrix in which the (i,j)th entry is the similarity score for column i and column j in the original matrix 4.Resulting matrix (in the case of one input) must be symmetric 5.The inputs must be named x and y
sim.score.args	A list of arguments for the measurement function. For example: In the case of cor, the following would be acceptable: sim.score.args = list(method='spearman',use='complete.obs' ).
min.subj	Minimum number of samples that must be non-missing in a bug/feature/column in order to apply the similarity measure to that bug/feature/column. This is to ensure that there are sufficient subjects to perform a bootstrap (default: 20).
iterations	The number of iterations of bootstrap and permutation (default: 1000).
subset.cols.x	A vector of column indices from x to indicate which features to compare
subset.cols.y	A vector of column indices from y to indicate which features to compare
errthresh	If feature has number of zeros greater than errthresh^(1/n) , that feature is excluded
verbose	Logical: an indicator whether the user requested verbose output, which prints periodic progress of the algorithm through the dataset(s), as well as including more detailed output. (default:FALSE)
iterations.gap	If output is verbose - number of iterations after issue a status message (Default=100 - displayed only if verbose=TRUE).
distributions	Output Distribution file (default:NA).
compare.within.x	A boolean value indicating whether to do comparisons given by taking all subsets of size 2 from subset.cols.x or to do comparisons given by taking all possible combinations of subset.cols.x or subset.cols.y. If TRUE but subset.cols.y=NA, returns all comparisons involving any features in subset.cols.x. This argument is only used when y=NA.
renormalize	Logical, should data be renormalized when preforming permutations? For relative data, it is strongly recommended to renormalize. However, when using absolute data, this argument must be set to FALSE.
memory.optimize	Logical, should the computations be run a memory-saving and highly vectorizing algorithm? Currently, this is only implemented in the case where y=NULL. The downside of this option is that printing of distributions is not supported. In addition, you should expect more numerical instability by enabling this option. However, this option is recommended with datasets with many OTUs

Value

Returns a list containing the calculation results and the parameters used.
Default parameters shown:

min.subj	Description above
errThresh	Description same as errthresh above
sim.score	A matrix of the simliarity scores for all the requested comparisons. The (i,j)th element of sim.score correponds to the similarity score of column i (or the ith column of subset.cols.1) and column j (or the jth column of subset.cols.1) in one dataset, or to the similarity score of column i (or the ith column of subset.cols.1) in dataset x and column j (or the jth column of subset.cols.2)in dataset y in the case of two datasets.
p.values	A matrix of the p-values for all the requested comparisons. The (i,j)th element of p.values corresponds to the p-value of the (i,j)th element of sim.score.
q.values	A matrix of the Benjamini-Hochberg-Yekutieli FDR corrected p-values. The (i,j)th element of q.values corresponds to the q-value fo the (i,j)th element of sim.score.
z.stat	A matrix of the z-statistics for all the requested comparisons. The (i,j)th element corresponds to the z-statistic which gave rise to the (i,j)th p-value.

Additional parameters if verbose=TRUE:

iterations	Description Above
subset.cols.x	Description Above
subset.cols.y	Description Above
iterations.gap	Description Above
sim.score.parameters	Description Above
compare.within.x	Description Above

Author(s)

Emma Schwager <[e-mail]>, Jakob Peder Pettersen <[e-mail]>

References

Emma Schwager and Colleagues. Detecting statistically significant associtations between sparse and high dimensional compositioanl data. In Progress.

Benjamini and Yekutieli (2001). "The control of the false discovery rate in multiple testing under dependency." The Annals of Statistics. Vol. 19, No. 4. pp. 1165-1188.

Examples

data <- matrix(rlnorm(40,meanlog=0,sdlog=1),nrow=10)
data.rowsum <- apply(data,1,sum)
data.norm <- data/data.rowsum
testdata <- data.norm
dimnames(testdata) <- list(paste("Sample",seq(1,10)),paste("Feature",seq(1,4)))
ccrepe.results  <-ccrepe  (x=testdata, iterations=20, min.subj=10)
ccrepe.results.nc.score <- ccrepe(x=testdata,iterations=20,min.subj=10,sim.score=nc.score)
ccrepe.results
ccrepe.results.nc.score

AlmaasLab/micInt documentation built on April 1, 2022, 10:37 a.m.