Glossary#

all features#: Assay-wide all-true feature-selection artifact returned by DataStore.select_all_features and used by complete-universe operations. It is distinct from the physical feature metadata column I.
artifact#: A persisted analysis result such as a normalization, PCA, neighbourhood graph, or marker table. Scarf writes each one into the Zarr store next to a record of what produced it, so results outlive the session that computed them and can be inspected, compared, or reused later.
artifact lineage#: Directed record of the exact selections and upstream artifacts that produced a stored result. Granular methods pass returned ArtifactRef values explicitly; no result is selected implicitly for an assay.
ArtifactRef#: Reference to an artifact, returned by the method that wrote it. It identifies a stored result without loading it, so it can be passed to the next step or held for later comparison. Read the data with load_artifact, and the parameters and status with inspect_artifact.
assay#: Named modality inside a DataStore (for example RNA, ADT, ATAC).
batch correction#: Adjusting embeddings or graphs so technical sample batches mix while biological structure is preserved.
cell key#: Boolean column in cell metadata that can be captured as an analytical input. Default is I. Filtering snapshots it into an immutable selection artifact, combines thresholds with that selection, and leaves the column unchanged.
cLISI#: Cell-type LISI. Median label LISI inverted and scaled so higher values indicate better biological-label conservation. Computed with metric_clisi.
count matrix#: Sparse matrix of primary counts stored cell-major (n_cells × n_features), for features such as genes, peaks, or ADTs. Scarf stores that array as counts in the assay Zarr group. RNA assays also store countsT, the same values in gene-major order, so gene-wise stages can stream without scanning every cell.
DataStore#: Primary Scarf object that opens a Zarr store and exposes analysis methods.
DataStoreMerge#: Canonical class for merging DataStores into one Zarr file.
densMAP#: Density-preserving UMAP variant enabled by passing explicit graph and initialization refs to run_umap(..., use_density_map=True).
feature selection#: Immutable Boolean artifact aligned to the complete feature order of one assay. Producers such as select_hvgs return an ArtifactRef and leave feature metadata unchanged. Direct feature consumers require the exact ref through features=.
graph connectivity#: Mean fraction of cells from each label retained in its largest connected component on Scarf’s symmetrized graph. Computed with metric_graph_connectivity.
Harmony#: Batch correction applied to PCA embeddings with run_harmony before ANN construction.
highly variable genes#: Features selected with select_hvgs for neighbourhood-graph construction.
iLISI#: Integration LISI. Median batch LISI scaled so higher values indicate better batch mixing. Computed with metric_ilisi.
Leiden clustering#: Graph community detection via run_leiden_clustering. A manual call returns a cluster-label artifact without adding metadata columns. The RNA pipeline runs default resolutions 0.5, 0.75, 1.0, and 1.25, plus Paris unless disabled. Its artifact-backed silhouette stage scores Leiden resolutions in the graph’s coordinate space, and run["clusters"] is the selected Leiden candidate’s exact ref. Paris remains a diagnostic run output. This automatic choice is a reproducible baseline, not biological validation.
LISI#: Local Inverse Simpson Index. Per-cell measure of local label mixing in the KNN graph. Computed with metric_lisi.
LSI#: Latent Semantic Indexing. Linear dimension reduction used for scATAC-seq graphs.
mapping reference#: Immutable RNA mapping artifact built from a scaled PCA or Symphony neighbour chain with build_mapping_reference(neighbors). A writable query datastore uses it to create query-owned projections without changing the reference.
neighbourhood graph#: KNN graph of cells built by individual graph-construction methods or ds.pipeline.run. Embeddings, clustering, mapping, and multimodal integration reuse this graph.
Paris clustering#: Hierarchical graph clustering in Scarf (run_paris_clustering). Supports fixed cuts and branch-adaptive cuts guarded by configuration-null modularity, plus cluster-tree visualization.
partial PCA#: PCA trained on an immutable cell subset via pca_cell_selection. A lightweight batch-correction option when one sample is the reference.
PipelineRun#: Durable handle returned by ds.pipeline.run() and reopened through ds.pipeline.open(...). It maps stable output names to artifacts and exposes frozen cell and feature views plus reports. Plotting, marker loading, and export remain DataStore operations.
proportion-aware batch mixing#: Scarf summary that rescales mean batch LISI against the observed global batch proportions. Computed with metric_proportional_batch_mixing.
provenance#: Record stored with every artifact naming the operation that produced it, the scientific parameters it used, and the artifacts it consumed. It is what lets Scarf recognise that a new request describes a result the store already holds.
reuse#: Returning an existing artifact instead of recomputing it, when the requested operation, parameters, and inputs match its provenance. Changing a parameter produces a new artifact, and the steps that depended on the previous one are recomputed rather than reused.
SNN integration#: Shared-nearest-neighbor merge of explicit modality-specific connectivity-map refs via integrate_assays(sources, method="snn").
TopACeDo#: Manifold-preserving cell subsampling using the KNN graph (run_topacedo_sampler).
WNN integration#: Default Hao-inspired weighted nearest-neighbor merge for two or more explicit modality neighbour refs via integrate_assays(sources) or integrate_assays(sources, method="wnn"). Scarf scores the union of all existing, self-free KNN rows with affinity and the distance span from each modality’s nearest to its k-th neighbour as bandwidth. During scoring it L2-normalizes modality coordinate rows; the affinities themselves are not L2-normalized. Unlike Seurat defaults, it does not build a wider 200-neighbour candidate pool or use SNN-far bandwidth.