mofa-r

Create a model with create_mofa() from a named list of feature-by-sample matrices, configure it with prepare_mofa() using the default data, model, and training options, then call run_mofa() which invokes the mofapy2 Python engine and writes an HDF5 output file.

Use the MOFA2 R bindings when your pipeline is in R/Bioconductor, your inputs are SingleCellExperiment or MultiAssayExperiment objects, or you want ggplot2-native plots. Use mofapy2 directly when the rest of the workflow is Python-based; both share the same engine and HDF5 format.

MOFA2 calls the Python mofapy2 engine, so a Python environment is required. Either let basilisk provision its conda env automatically on the first run_mofa() call, or point reticulate::use_python() at a Python with mofapy2 installed before loading MOFA2.

No. MOFA2 has no biplot function because Z and W are scaled independently with no shared singular values. Use the L1 supplementary-projection approach instead, projecting covariate correlations or group barycenters onto the factor map.

Per-factor R² in calculate_variance_explained() is computed against the full data matrix, not against residuals after previous factors, so it is not a partition. Treat it as a per-factor diagnostic rather than an additive variance budget.

The bilinear model Y ≈ ZW^T is sign-flip equivariant per factor, and factors are only sorted post-hoc by total R². For cross-run comparisons, impose a canonical sign, for example flipping so the largest absolute loading is positive.