scBFA is better informed by cell type markers than quantification models. Each latent factor learned from each method was evaluated based on how much influence established cell type markers exerted on its embeddings, as measured by the area under the curve (AUROC) metric. Each boxplot represents the AUROC of all latent factors for a given method and benchmark. ZINB-WaVE is represented twice, once for the latent dimensions inferred by their gene detection pattern (ZINB-WaVEdropout) and once for the latent dimensions inferred from the gene counts (ZINB-WaVEexpr). a PBMC benchmark. b HSCs benchmark. c Pancreatic benchmark.