OmicVerse Microbiome × Metabolomics — Paired Integration

Goal

Take two AnnData objects sharing the same obs_names in the same order — one with microbe / ASV / OTU counts, one with metabolite intensities — and find microbe-metabolite pairs that co-vary across samples. Three methods, three different statistical assumptions:

1. Spearman pairwise correlation (paired_spearman) — fast; per-pair r + p + BH-FDR; symmetric / non-parametric.
2. CCA (paired_cca) — linear canonical components; finds shared latent modes; small number of components, each explaining a covariance pair.
3. MMvec (MMvec) — PyTorch low-rank model logits = U·V.T + β over per-microbe metabolite distributions; gives an embedding biplot + conditional probabilities P(metabolite | microbe).

Plus a synthetic-recovery validation harness (simulate_paired) and a method-comparison plot (plot_paired_method_comparison).

This is a two-modality integration skill; the upstream microbe table is presumed to come from the 16S amplicon skill, and the metabolite table from any of the metabolomics skills. The Franzosa 2019 IBD dataset is the bundled real-data demo (fetch_franzosa_ibd_2019).

Quick Workflow