We updated how meta-programs (MPs) are calculated from individual programs. Instead of extracting gene sets for each program and then calculating a consensus, we keep the full vector of gene weights and calculate cosine similarities between the vectors. Consensus gene weights are then calculated as the average over all programs in a MP.
To impose sparsity in the decomposition, we include a 'specificity.weight' parameter, which is used to re-normalize NMF loadings based on how specific a gene is for a given program.
To determine the number of genes to be included in a MP, we calculate the cumulative distribution for the gene weights in a given MP. Only genes that cumulatively explain up to a fraction of the total weight ('weight.explained' parameter) are included in the MP gene set.
The definition and default of 'min.confidence' has changed. The confidence of a gene in a given MP is calculated as the fraction of programs in which the gene has been determined to be part of the invidual program (using 'weight.explained=0.8').
The parameter 'nprograms' in the function 'getMetaPrograms()' has been renamed to 'nMP', to avoid confusion
New defaults: expression matrices are now by default not scaled or centered (the behavior can be altered using the 'scale' and 'center' parameters)
First stable version