OmicVerse Bulk RNA-seq — Cell-Type Deconvolution

Goal

Take a bulk RNA-seq cohort (AnnData, samples × genes) plus a single-cell reference (AnnData, cells × genes, with cell-type labels) and infer cell-type fractions per bulk sample. One unified class wraps four established methods under a method=... switch:

• TAPE (Cao 2022) — deep-learning autoencoder; default; CPU-friendly.
• Scaden (Menden 2020) — deep neural network; trains on pseudo-bulk mixtures; benefits from GPU.
• BayesPrism (Chu 2022) — full Bayesian model; multi-core CPU; produces posterior fractions.
• OmicsTweezer — joint reference-correction deconvolution.

Output is a pd.DataFrame indexed by bulk sample with one column per cell type — directly stackable as a per-sample bar chart, or summarisable per phenotype group via ov.pl.plot_grouped_fractions.

This skill is the opposite direction of bulk-to-single-deconvolution (Bulk2Single, which generates synthetic single cells from bulk). They share the word "deconvolution" but solve inverse problems — don't confuse them.

Quick Workflow