omicverse.micro.MMvec.fit

Contents

omicverse.micro.MMvec.fit#

MMvec.fit(adata_microbe, adata_metabolite, verbose=False)[source]#

Fit MMvec on paired microbe / metabolite count tables and return self.

Trains a low-rank latent-space model logits[i, j] = U[i] · V[j] + β[j], where rows of U are microbe embeddings, rows of V are metabolite embeddings, and the loss is per-microbe negative log-likelihood of the sample-weighted metabolite distribution. The cooccurrence weights W[i, j] = Σ_s X_mb[s, i] · X_mt[s, j] / Σ_s X_mt[s, :] capture which microbes covary with which metabolites across samples.

Both AnnDatas must share obs_names in the same order (_check_paired enforces this). Training reports an early-stop validation loss every epoch when val_frac > 0; patience epochs without validation improvement stops training and restores the best checkpoint.

Parameters:

  • adata_microbe (AnnData) – Sample-aligned count tables.

  • adata_metabolite (AnnData) – Sample-aligned count tables.

  • verbose (bool, default False) – Print loss every epochs // 10 epochs.

Returns:

  • self — fitted; access microbe_embeddings_,

  • metabolite_embeddings_, cooccurrence(),

  • conditional_probabilities(), top_pairs(n).

Contents