Single Cell Multi Omics Integration

Single-Cell Multi-Omics Tutorials Cheat Sheet

This skill walk-through summarizes the OmicVerse notebooks that cover paired and unpaired multi-omic integration, multi-batch embedding, reference transfer, and trajectory cartography.

MOFA on paired scRNA + scATAC (t_mofa.ipynb)

• Data preparation: Load preprocessed AnnData objects for RNA (rna_p_n_raw.h5ad) and ATAC (atac_p_n_raw.h5ad) with ov.utils.read, and initialise pyMOFA with matching omics and omics_name lists.
• Model training: Call mofa_preprocess() to select highly variable features and run the factor model with mofa_run(outfile=...), which exports the learned MOFA+ factors to an HDF5 model file.
• Result inspection: Reload downstream AnnData, append factor scores via ov.single.factor_exact, and explore factor–cluster associations using factor_correlation, get_weights, and the plotting helpers in pyMOFAART (plot_r2, plot_cor, plot_factor, plot_weights, etc.).
• Export workflow: Persist factors and weights through the MOFA HDF5 artifact and reuse them by instantiating pyMOFAART(model_path=...) for later annotation or visualisation sessions.
• Dependencies & hardware: Requires mofapy2; plots optionally rely on pymde/scvi-tools but run on CPU.

MOFA after GLUE pairing (t_mofa_glue.ipynb)

• Data preparation: Start from GLUE-derived embeddings (rna-emb.h5ad, atac.emb.h5ad), build a GLUE_pair object, and run correlation() to align unpaired cells before subsetting to highly variable features.
• Model training: Instantiate pyMOFA with the aligned AnnData objects, run mofa_preprocess(), and save the joint factors through mofa_run(outfile='models/chen_rna_atac.hdf5').
• Result inspection: Use pyMOFAART plus AnnData that now contains the GLUE embeddings to compute factors (get_factors) and visualise variance explained, factor–cluster correlations, and ranked feature weights.
• Export workflow: Reuse the saved MOFA HDF5 model for downstream inspection; GLUE embeddings can be embedded with scvi.model.utils.mde (GPU-accelerated MDE is optional, sc.tl.umap works on CPU).
• Dependencies & hardware: Requires both mofapy2 and the GLUE tooling (scglue, scvi-tools, pymde); GPU acceleration only affects optional MDE visualisation.

SIMBA batch integration (t_simba.ipynb)

• Data preparation: Fetch the concatenated AnnData (simba_adata_raw.h5ad) derived from multiple pancreas studies and pass it, alongside a results directory, to pySIMBA.
• Model training: Execute preprocess(...) to bin features and build a SIMBA-compatible graph, then call gen_graph() followed by train(num_workers=...) to launch PyTorch-BigGraph optimisation (can scale with CPU workers) and load(...) to resume trained checkpoints.
• Result inspection: Apply batch_correction() to obtain the harmonised AnnData with SIMBA embeddings (X_simba) and visualise using mde/sc.tl.umap coloured by cell type or batch.
• Export workflow: Training outputs reside in the workdir (e.g., result_human_pancreas/pbg/graph0); reuse them with simba_object.load(...) for later analyses.
• Dependencies & hardware: Requires installing simba and simba_pbg (PyTorch BigGraph backend). GPU is optional; make sure adequate CPU threads and memory are available for graph training.

TOSICA reference transfer (t_tosica.ipynb)

• Data preparation: Download demo AnnData references (demo_train.h5ad, demo_test.h5ad) and required gene-set GMT files via ov.utils.download_tosica_gmt(); confirm datasets are log-normalised before training.
• Model training: Create pyTOSICA with the reference AnnData, chosen pathway mask, label key, project directory, and batch size; train with train(epochs=...), then persist weights with save() and optionally reload via load().
• Result inspection: Generate predictions on query AnnData through predicted(pre_adata=...), embed with OmicVerse preprocessing and GPU-enabled mde (UMAP fallback available), and explore pathway attention to interpret transformer heads.
• Export workflow: Saved project folder keeps model checkpoints and attention summaries; reuse the exported assets to annotate future datasets without retraining from scratch.
• Dependencies & hardware: Needs TOSICA (PyTorch transformer) plus downloaded gene-set masks; avoid setting depth=2 if memory is constrained. GPU acceleration improves embedding (mde) but training runs on standard PyTorch (CPU/GPU depending on environment).

StaVIA trajectory cartography (t_stavia.ipynb)

• Data preparation: Load example dentate gyrus velocity data via scvelo.datasets.dentategyrus(), preprocess with OmicVerse (preprocess, scale, pca, neighbours, UMAP) to populate the AnnData matrices used by VIA.
• Model training: Configure VIA hyperparameters (components, neighbours, seeds, root selection) and instantiate/run VIA.core.VIA on the chosen representation (adata.obsm['scaled|original|X_pca']).
• Result inspection: Store outputs such as pseudotime (single_cell_pt_markov), cluster graph abstractions, trajectory curves, atlas views, and stream plots through VIA plotting helpers.
• Export workflow: Persist derived visualisations and animations (e.g., animate_streamplot_ov, animate_atlas) to files (.gif) for reporting; recompute edge bundles via make_edgebundle_milestone when needed.
• Dependencies & hardware: Relies on scvelo, pyVIA, and OmicVerse plotting; computations are CPU-bound though producing large stream/animation outputs benefits from ample memory.