Omicverse Single Cell Preprocessing

OmicVerse Single-Cell Preprocessing

Goal

Turn raw or already-QC'd single-cell AnnData into a reusable execution spine: QC, preprocessing, PCA, graph construction, embeddings, Leiden clustering, and optional marker discovery. Keep marker discovery in the same skill because it only becomes useful after the same cluster-ready object exists.

Quick Workflow

1. Inspect the input AnnData shape, layers, and whether a cluster key already exists.
2. If the data is raw, run ov.pp.qc(...) first; if it is already filtered, jump to preprocessing.
3. Choose one preprocessing branch with ov.pp.preprocess(...), then keep the HVG output and build scaled, PCA, neighbors, and an embedding.
4. Run ov.pp.leiden(...) after the graph exists.
5. If you need marker genes, choose a method that matches the input representation, then call ov.single.find_markers(...) and ov.single.get_markers(...) or ov.pl.markers_dotplot(...).
6. Validate the expected obs, var, obsm, obsp, and uns keys before writing output.

Interface Summary

• ov.pp.qc(adata, **kwargs) handles QC and doublet filtering.
• ov.pp.preprocess(adata, mode='shiftlog|pearson', target_sum=500000.0, n_HVGs=2000, organism='human', no_cc=False, batch_key=None, identify_robust=True) normalizes and selects HVGs.
• ov.pp.scale(adata, max_value=10, layers_add='scaled', to_sparse=True, **kwargs) creates the scaled layer.
• ov.pp.pca(adata, n_pcs=50, layer='scaled', inplace=True, **kwargs) stores PCA in scaled|original|X_pca.
• ov.pp.neighbors(adata, n_neighbors=15, n_pcs=None, use_rep=None, knn=True, random_state=0, n_jobs=None, method='umap', transformer=None, metric='euclidean', metric_kwds=mappingproxy({}), key_added=None, copy=False, **kwargs) builds the graph.
• ov.pp.umap(adata, **kwargs), ov.pp.mde(adata, ...), ov.pp.tsne(adata, ...), and ov.pp.sude(adata, ...) are alternative embedding branches.
• ov.pp.leiden(adata, resolution=1.0, random_state=0, key_added='leiden', local_iterations=100, max_levels=10, device='cpu', symmetrize=None, **kwargs) clusters the graph.
• ov.single.find_markers(adata, groupby, method='cosg', n_genes=50, key_added=None, use_raw=None, layer=None, groups='all', reference='rest', corr_method='benjamini-hochberg', rankby_abs=False, tie_correct=False, pts=True, **kwargs) finds markers.
• ov.single.get_markers(adata, n_genes=10, key='rank_genes_groups', groups=None, return_type='dataframe', min_logfoldchange=None, min_score=None, min_pval_adj=None) extracts marker tables.
• ov.pl.markers_dotplot(...) renders the marker summary after find_markers.

Read references/source-grounding.md before adding more interface-specific details.

Stage Selection

• Use this branch selection to avoid replaying the whole notebook when the object is already partially prepared.
• Use the full pipeline when the input is raw counts and no graph exists yet.
• Skip directly to marker discovery when the object is already normalized, clustered, and only marker extraction or dotplotting is missing.
• Use mode='shiftlog|pearson' when you want the notebook's default preprocessing path.
• Use method='wilcoxon' for statistically tested markers on log-normalized data.
• Use method='cosg' when you want fast marker discovery from raw counts.
• Use method='t-test', method='t-test_overestim_var', or method='logreg' only when their assumptions match the data.

Input Contract

• Start from AnnData.
• Preserve raw counts if you need a marker branch that expects them.
• Ensure the neighbor graph exists before ov.pp.leiden(...).
• Ensure a cluster key such as leiden exists before ov.single.find_markers(...) and ov.pl.markers_dotplot(...).
• Treat ov.settings.cpu_gpu_mixed_init() as a runtime toggle, not as a separate skill boundary.

Minimal Execution Patterns

import omicverse as ov

ov.settings.cpu_init()

adata = ov.pp.qc(adata, mode="mads", doublets=False, min_genes=0, min_cells=0)
adata = ov.pp.preprocess(
    adata,
    mode="shiftlog|pearson",
    n_HVGs=2000,
    target_sum=1e4,
    identify_robust=True,
)
adata.raw = adata
adata = adata[:, adata.var.highly_variable_features].copy()
ov.pp.scale(adata)
ov.pp.pca(adata, layer="scaled", n_pcs=50)
ov.pp.neighbors(adata, n_neighbors=15, n_pcs=50, use_rep="scaled|original|X_pca")
ov.pp.umap(adata)
ov.pp.leiden(adata, resolution=1.0)
ov.single.find_markers(adata, groupby="leiden", method="wilcoxon", key_added="rank_genes_groups")
markers = ov.single.get_markers(adata, key="rank_genes_groups", return_type="dict")

ov.single.find_markers(adata, groupby="leiden", method="cosg", key_added="cosg_markers", pts=True)
ov.pl.markers_dotplot(adata, groupby="leiden", key="cosg_markers", n_genes=5, show=False)

Constraints

• Do not hardcode notebook data paths or local environment names.
• Do not assume scrublet works on tiny fixtures; keep bounded smoke checks synthetic and lightweight.
• Keep cosg and the statistical methods separate: cosg expects raw counts, while wilcoxon, t-test, and logreg expect log-normalized data.
• Be explicit about mode, method, and use_rep; do not leave branch selection implicit.
• Keep the notebook's visualization polish cells optional unless the user explicitly wants the same figures.

Validation

• Check adata.layers["counts"] after preprocessing.
• Check adata.var["highly_variable_features"] before PCA.
• Check adata.obsm["scaled|original|X_pca"] after PCA.
• Check adata.obsm["X_umap"] or another embedding key after the embedding step.
• Check adata.obs["leiden"] before marker discovery.
• Check adata.uns["rank_genes_groups"] or the selected marker key before get_markers(...) or markers_dotplot(...).
• If you only validated a smoke path, say so; do not claim full-scale reproduction.

Resource Map

• Read references/branch-selection.md when choosing QC, preprocessing, graph, embedding, clustering, or marker branches.
• Read references/source-notebook-map.md to trace notebook cells into this reusable skill.
• Read references/source-grounding.md for inspected signatures, source-level branch behavior, and compatibility notes.
• Read references/compatibility.md when notebook prose, parameter names, or current runtime behavior diverge.