mofaplus-multi-omics

What it does

The skill guides an agent to set up and run a MOFA+ model using mofapy2 to jointly decompose multiple omics views (RNA, ATAC, etc.) into latent factors. It covers loading and aligning AnnData views, configuring data, building and training an EM-based model, saving the trained HDF5 model, and post-hoc analyses such as inspecting variance explained per factor per view, extracting factor scores, correlating with metadata, visualizing factor relationships, and deriving top-weighted features for enrichment. It also includes downstream steps like clustering samples by factor scores and preparing inputs for enrichment analyses.

How it works

• Load and prepare data: represent each omics layer as an AnnData object, align cell indices across modalities, and normalize as appropriate.
• Create the MOFA+ model object: instantiate entry_point, configure data options (e.g., scale groups, scale views), provide data as a list-of-lists [groups][views], and name views and groups. Set data options before data and register data matrices, with per-view likelihoods and per-group samples.
• Set model and training options: specify number of factors, spike-slab prior, automatic relevance determination for factors and weights, and training hyperparameters (iterations, convergence, ELBO tracking, pruning thresholds, seed).
• Build and train: execute ent.build() then ent.run() and save the model to an HDF5 file via ent.save(output_path).
• Inspect variance explained: load the MOFA+ HDF5 to compute variance explained per factor per view, averaging across groups, and display as a table/heatmap.
• Extract factor scores and correlate with metadata: load Z matrices (factor scores) per sample, join with sample metadata, and perform simple ANOVA-style comparisons between groups.
• Visualize factors: create scatter plots of factor scores colored by metadata; generate heatmaps of top feature weights per factor for a given view.
• Downstream clustering and enrichment: cluster samples by standardized factor scores (e.g., KMeans), and prepare top-weighted genes per factor for downstream enrichment analyses.
• Provide key parameters: list defaults and options for factors, likelihoods, and scale_views.

When to use it

• When integrating two or more omics layers from the same cells/samples (e.g., RNA with ATAC, RNA with proteomics, methylation with RNA).
• To identify shared and view-specific sources of variation without supervised labels.
• To compare latent factors across patient groups, treatments, or timepoints in multi-group analyses.
• To reduce multi-omics dimensionality before clustering, trajectory inference, or survival modeling.
• To highlight which features drive each factor via weight loadings and to annotate factors via correlation with metadata.

What it can touch

• Data inputs: AnnData objects for RNA and ATAC views, with aligned cell/sample indices.
• Model components: HDF5 MOFA+ model file saved as an output, loaded for downstream inspection.
• Outputs: variance explained data (per factor per view), factor scores (Z), feature weights (W) per view, clustering labels, and CSV/plots.

Caveats

• Requires Python 3.9+ and specific packages: mofapy2, anndata, muon, matplotlib, seaborn, numpy, pandas, scipy. Model storage is HDF5.
• The approach relies on proper data preprocessing and alignment across views; mismatches in sample indices can invalidate results.
• The heuristic for pruning and the interpretation of factors depend on model options (e.g., factors count, ARD settings) and may vary with data complexity.
• The workflow includes optional steps for enrichment that require external tools or libraries (e.g., gseapy) installed separately.