Single-cell preprocessing with omicverse

Overview

Follow this skill when a user needs to reproduce the preprocessing workflow from the omicverse notebooks t_preprocess.ipynb, t_preprocess_cpu.ipynb, and t_preprocess_gpu.ipynb. The tutorials operate on the 10x PBMC3k dataset and cover QC filtering, normalisation, highly variable gene (HVG) detection, dimensionality reduction, and downstream embeddings.

Instructions

1. Set up the environment
   • Import omicverse as ov and scanpy as sc, then call ov.plot_set(font_path='Arial') (or ov.ov_plot_set() in legacy notebooks) to standardise figure styling.
   • Encourage %load_ext autoreload and %autoreload 2 when iterating inside notebooks so code edits propagate without restarting the kernel.
2. Prepare input data
   • Download the PBMC3k filtered matrix from 10x Genomics (pbmc3k_filtered_gene_bc_matrices.tar.gz) and extract it under data/filtered_gene_bc_matrices/hg19/.
   • Load the matrix via sc.read_10x_mtx(..., var_names='gene_symbols', cache=True) and keep a writable folder like write/ for exports.
3. Perform quality control (QC)
   • Run ov.pp.qc(adata, tresh={'mito_perc': 0.2, 'nUMIs': 500, 'detected_genes': 250}, doublets_method='scrublet') for the CPU/CPU–GPU pipelines; omit doublets_method on pure GPU where Scrublet is not yet supported.
   • Review the returned AnnData summary to confirm doublet rates and QC thresholds; advise adjusting cut-offs for different species or sequencing depths.
4. Store raw counts before transformations
   • Call ov.utils.store_layers(adata, layers='counts') immediately after QC so the original counts remain accessible for later recovery and comparison.
5. Normalise and select HVGs
   • Use ov.pp.preprocess(adata, mode='shiftlog|pearson', n_HVGs=2000, target_sum=5e5) to apply shift-log normalisation followed by Pearson residual HVG detection (set target_sum=None on GPU, which keeps defaults).
   • For CPU–GPU mixed runs, demonstrate ov.pp.recover_counts(...) to invert normalisation and store reconstructed counts in adata.layers['recover_counts'].
6. Manage .raw and layer recovery
   • Snapshot normalised data to .raw with adata.raw = adata (or adata.raw = adata.copy()), and show ov.utils.retrieve_layers(adata_counts, layers='counts') to compare normalised vs. raw intensities.
7. Scale, reduce, and embed
   • Scale features using ov.pp.scale(adata) (layers hold scaled matrices) followed by ov.pp.pca(adata, layer='scaled', n_pcs=50).
   • Construct neighbourhood graphs with:
     • sc.pp.neighbors(adata, n_neighbors=15, n_pcs=50, use_rep='scaled|original|X_pca') for the baseline notebook.
     • ov.pp.neighbors(..., use_rep='scaled|original|X_pca') on CPU–GPU to leverage accelerated routines.
     • ov.pp.neighbors(..., method='cagra') on GPU to call RAPIDS graph primitives.
   • Generate embeddings via ov.utils.mde(...), ov.pp.umap(adata), ov.pp.mde(...), ov.pp.tsne(...), or ov.pp.sude(...) depending on the notebook variant.
8. Cluster and annotate
   • Run ov.pp.leiden(adata, resolution=1) or ov.single.leiden(adata, resolution=1.0) after neighbour graph construction; CPU–GPU pipelines also showcase ov.pp.score_genes_cell_cycle before clustering.
   • IMPORTANT - Defensive checks: When generating code that plots by clustering results (e.g., color='leiden'), always check if the clustering has been performed first:
     pythonCopy

     # Check if leiden clustering exists, if not, run it
     if 'leiden' not in adata.obs:
         if 'neighbors' not in adata.uns:
             ov.pp.neighbors(adata, n_neighbors=15, use_rep='X_pca')
         ov.single.leiden(adata, resolution=1.0)
     

   • Plot embeddings with ov.pl.embedding(...) or ov.utils.embedding(...), colouring by leiden clusters and marker genes. Always verify that the column specified in color= parameter exists in adata.obs before plotting.
9. Document outputs
   • Encourage saving intermediate AnnData objects (adata.write('write/pbmc3k_preprocessed.h5ad')) and figure exports using Matplotlib’s plt.savefig(...) to preserve QC summaries and embeddings.
10. Notebook-specific notes
    • Baseline (t_preprocess.ipynb): Focuses on CPU execution with Scanpy neighbours; emphasise storing counts before and after retrieve_layers demonstrations.
    • CPU–GPU mixed (t_preprocess_cpu.ipynb): Highlights Omicverse ≥1.7.0 mixed acceleration. Include timing magics (%%time) to showcase speedups and call out doublets_method='scrublet' support.
    • GPU (t_preprocess_gpu.ipynb): Requires a CUDA-capable GPU, RAPIDS 24.04 stack, and rapids-singlecell. Mention the ov.pp.anndata_to_GPU/ov.pp.anndata_to_CPU transfers and method='cagra' neighbours. Note the current warning that pure-GPU pipelines depend on RAPIDS updates.
11. Troubleshooting tips
    • If sc.read_10x_mtx fails, verify the extracted folder structure and ensure gene symbols are available via var_names='gene_symbols'.
    • Address GPU import errors by confirming the conda environment matches the RAPIDS version for the installed CUDA driver (nvidia-smi).
    • For ov.pp.preprocess dimension mismatches, ensure QC filtered out empty barcodes so HVG selection does not encounter zero-variance features.
    • When embeddings lack expected fields (e.g., scaled|original|X_pca missing), re-run ov.pp.scale and ov.pp.pca to rebuild the cached layers.
    • Pipeline dependency errors: When encountering errors like "Could not find 'leiden' in adata.obs or adata.var_names":
      • Always check if required preprocessing steps (neighbors, PCA) exist before dependent operations
      • Check if clustering results exist in adata.obs before trying to color plots by them
      • Use defensive checks in generated code to handle incomplete pipelines gracefully
    • Code generation best practice: Generate robust code with conditional checks for prerequisites rather than assuming perfect sequential execution. Users may run steps in separate sessions or skip intermediate steps.

Critical API Reference - Batch Column Handling

Batch Column Validation - REQUIRED Before Batch Operations

IMPORTANT: Always validate and prepare the batch column before any batch-aware operations (batch correction, integration, etc.). Missing or NaN values will cause errors.

CORRECT usage:

# Step 1: Check if batch column exists, create default if not
if 'batch' not in adata.obs.columns:
    adata.obs['batch'] = 'batch_1'  # Default single batch

# Step 2: Handle NaN/missing values - CRITICAL!
adata.obs['batch'] = adata.obs['batch'].fillna('unknown')

# Step 3: Convert to categorical for efficient memory usage
adata.obs['batch'] = adata.obs['batch'].astype('category')

# Now safe to use in batch-aware operations
ov.pp.combat(adata, batch='batch')  # or other batch correction methods

WRONG - DO NOT USE:

# WRONG! Using batch column without validation can cause NaN errors
# ov.pp.combat(adata, batch='batch')  # May fail if batch has NaN values!

# WRONG! Assuming batch column exists
# adata.obs['batch'].unique()  # KeyError if column doesn't exist!

Common Batch-Related Pitfalls

1. NaN values in batch column: Always use fillna() before batch operations
2. Missing batch column: Always check existence before use
3. Non-categorical batch: Convert to category for memory efficiency
4. Mixed data types: Ensure consistent string type before categorization

# Complete defensive batch preparation pattern:
def prepare_batch_column(adata, batch_key='batch', default_batch='batch_1'):
    """Prepare batch column for batch-aware operations."""
    if batch_key not in adata.obs.columns:
        adata.obs[batch_key] = default_batch
    adata.obs[batch_key] = adata.obs[batch_key].fillna('unknown')
    adata.obs[batch_key] = adata.obs[batch_key].astype(str).astype('category')
    return adata

Highly Variable Genes (HVG) - Small Dataset Handling

LOESS Failure with Small Batches

IMPORTANT: The seurat_v3 HVG flavor uses LOESS regression which fails on small datasets or small per-batch subsets (<500 cells per batch). This manifests as:

ValueError: Extrapolation not allowed with blending

CORRECT - Use try/except fallback pattern:

# Robust HVG selection for any dataset size
try:
    sc.pp.highly_variable_genes(
        adata,
        flavor='seurat_v3',
        n_top_genes=2000,
        batch_key='batch'  # if batch correction is needed
    )
except ValueError as e:
    if 'Extrapolation' in str(e) or 'LOESS' in str(e):
        # Fallback to simpler method for small datasets
        sc.pp.highly_variable_genes(
            adata,
            flavor='seurat',  # Works with any size
            n_top_genes=2000
        )
    else:
        raise

Alternative - Use cell_ranger flavor for batch-aware HVG:

# cell_ranger flavor is more robust for batched data
sc.pp.highly_variable_genes(
    adata,
    flavor='cell_ranger',  # No LOESS, works with batches
    n_top_genes=2000,
    batch_key='batch'
)

Best Practices for Batch-Aware HVG

1. Check batch sizes before HVG: Small batches (<500 cells) will cause LOESS to fail
2. Prefer seurat or cell_ranger when batch sizes vary significantly
3. Use seurat_v3 only when all batches have >500 cells
4. Always wrap in try/except when dataset size is unknown

# Safe batch-aware HVG pattern
def safe_highly_variable_genes(adata, batch_key='batch', n_top_genes=2000):
    """Select HVGs with automatic fallback for small batches."""
    try:
        sc.pp.highly_variable_genes(
            adata, flavor='seurat_v3', n_top_genes=n_top_genes, batch_key=batch_key
        )
    except ValueError:
        # Fallback for small batches
        sc.pp.highly_variable_genes(
            adata, flavor='seurat', n_top_genes=n_top_genes
        )

Examples

• "Download PBMC3k counts, run QC with Scrublet, normalise with shiftlog|pearson, and compute MDE + UMAP embeddings on CPU."
• "Set up the mixed CPU–GPU workflow in a fresh conda env, recover raw counts after normalisation, and score cell cycle phases before Leiden clustering."
• "Provision a RAPIDS environment, transfer AnnData to GPU, run method='cagra' neighbours, and return embeddings to CPU for plotting."

References

• Detailed walkthrough notebooks: t_preprocess.ipynb, t_preprocess_cpu.ipynb, t_preprocess_gpu.ipynb
• Quick copy/paste commands: reference.md