omicverse.metabol.run_mofa

Contents

omicverse.metabol.run_mofa#

omicverse.metabol.run_mofa(views, *, n_factors=10, outfile='mofa_model.hdf5', scale_views=True, center_groups=True, max_iter=500, convergence_mode='fast', gpu_mode=False, seed=0)[source]#

Train MOFA+ on sample-aligned metabolomics + other-omics views.

Parameters:

  • views (dict) – Mapping {view_name: AnnData}. Each AnnData must have the same ``obs_names`` in the same order — MOFA+ concatenates samples across views by position. Mismatched indices raise a ValueError with the first few offending sample IDs.

  • n_factors (int (default: 10)) – Number of latent factors. Default 10 — MOFA+ drops factors with variance explained below dropR2=0.001 automatically.

  • outfile (str | Path (default: 'mofa_model.hdf5')) – Path to the HDF5 the trainer writes. Kept on disk so the user can reload it with ov.single.pyMOFAART if they want to drive the sc-oriented downstream helpers on their own.

  • scale_views (bool (default: True)) – Scale each view to unit total variance before training. The MOFA+ recommendation when view dimensions / dynamic ranges differ by >10x (metabolomics vs RNA-seq definitely qualifies).

  • center_groups (bool (default: True)) – Mean-centre each (view, sample-group) block before training.

  • max_iter (int (default: 500)) – Passed to pyMOFA.mofa_run.

  • convergence_mode (str (default: 'fast')) – Passed to pyMOFA.mofa_run.

  • gpu_mode (bool (default: False)) – Pass through to MOFA’s GPU path. Requires CuPy.

  • seed (int (default: 0)) – Deterministic seed.

Returns:

(n_samples, n_factors_retained) factor matrix indexed by the shared sample IDs, columns F1, F2, .... To plug it into a downstream step:

>>> factors = ov.metabol.run_mofa({'metabol': adata, 'rna': rna})
>>> adata.obsm['X_mofa'] = factors.reindex(adata.obs_names).to_numpy()

Return type:

pd.DataFrame

Contents