consensus-nmf-multirun

Run cNMF once per variant (subset × QC combination) using run_cnmf_subset.py, manually pick K from each k_selection plot, then transfer usage scores onto the full dataset, merge correlated programs, annotate via g:Profiler, and test against a factor with ANOVA.

Inspect the k_selection.png plot showing reproducibility and stability across K, and pick the largest K on the high plateau before the curves fall. State your expected K before viewing the plot, then confirm against the inflection point.

Use consensus cNMF when you have one dataset with factorial conditions and want programs robust to QC and subsetting. Use per-sample NMF with cross-donor consensus (genenmf-metaprogram-discovery) when you have multiple donors and want recurrent programs across them.

cNMF requires raw integer counts, but most scanpy pipelines normalize .X and discard the raw values. Store raw counts with adata.layers['counts'] = adata.X.copy() before any normalization so the cNMF input file contains integer counts.

The default r > 0.7 works for most projects, cutting the hierarchical clustering dendrogram at distance 0.3. Use r > 0.9 to preserve subtle distinctions or r > 0.5 only when you want a heavily aggregated overview, since loose thresholds collapse distinct biology.

Plain f_oneway cannot include covariates, so pass --covariate to condition_anova.py to fit program ~ C(factor) + C(batch) via statsmodels OLS. Alternatively restrict comparisons to within-batch subsets and document the confound explicitly.