Run sparse PCA using irlba SVD

Source: R/preprocessing_filtering_reduction.R

This function allows to run a PCA using IRLBA Singular Value Decomposition in a fast & memory efficient way. The increamental Lanczos bidiagonalisation algorithm allows to keep the matrix sparse as the "loci" centering is implicit. The function then multiplies by the approximate singular values (svd$d) in order to get more importance to the first PCs proportionnally to their singular values. This step is crucial for downstream approaches, e.g. UMAP or T-SNE.

pca_irlba_for_sparseMatrix(x, n_comp, work = 3 * n_comp)

Arguments

x

A sparse normalized matrix (features x cells)
n_comp

The number of principal components to keep
work

Working subspace dimension, larger values can speed convergence at the cost of more memory use.

Value

The rotated data, e.g. the cells x PC column in case of sc data.