omicverse-single-cell-cnmf-program-discovery

Consensus non-negative matrix factorization discovers gene expression programs by fitting cNMF across a grid of component ranks on a preprocessed AnnData object, then combining replicate factorizations into consensus program usages and cluster labels.

cNMF program discovery identifies consensus gene expression programs from single-cell AnnData inputs, converting transient notebook workflows into reproducible pipelines that output normalized usages and hard or RFC-derived cluster labels.

A GPU is not required to run cNMF for single-cell program discovery, as the workflow provides explicit CPU and GPU execution branches alongside single-worker and multi-worker configurations to fit available computing resources.

Consensus NMF gene program discovery requires a preprocessed single-cell RNA-seq AnnData object with normalized data and optional PCA, plus a writable output directory to persist intermediate factorization files and consensus combine step results.

To select the optimal K rank for consensus NMF on AnnData, fit factorizations across a user-provided grid of component ranks, then use the included selection guidance to evaluate K and density thresholds for stable consensus program generation.

Direct max-usage labeling assigns cells to programs based on highest usage values, while RFC-derived labels use a random forest classifier path to generate hard cluster labels, offering a classifier-based alternative within the cNMF workflow.