Validation

Checked against the tools you already trust.

Reviewers ask “would Seurat, Scanpy or the original R package have given the same answer?” For each method below, scView was run beside that package on public data, and the table says how closely the two agree. “Identical” means the same numbers; “to 1e-13” means they agree to 13 decimal places.

Each row: the reference implementation and version, the agreement metric, the public dataset, and the CHANGELOG entry or module docstring on main that recorded it. The bars count matching decimal places out of 16 (double precision). Where scView deliberately differs from the reference, the row says so.

Quality control

Step	Agreement	Reference	Data	Source
Barcode ranksDid the run capture cells well?	to 1e-10	DropletUtils barcodeRanks (knee, inflection)	pbmc_1k_v3; every metric equals the Cell Ranger summary	CHANGELOG 0.32.0
Ambient RNABackground RNA in every droplet	to 1e-16	R SoupX 1.6.2: soup profile; adjustCounts to 5e-11	10x pbmc, heart and neuron 1k	ambient.py
Doublet scoresDroplets that hold two cells	Identical	sc.pp.scrublet (scanpy 1.11), bitwise	Kang 2018, paul15	CHANGELOG 0.33.0
Doublet classifierA second, independent doublet call	0.948	AUC; R scDblFinder scores 0.951	10x human + mouse mix	CHANGELOG 0.35.0
Sample hashtagsWhich sample each cell came from	Every cell	same class and tag as Seurat 5.1.0 HTODemux and DropletUtils 1.24 hashedDrops	GSE108313, 20,238 cells	hashing.py

Clusters and cell types

Step	Agreement	Reference	Data	Source
Marker genesThe genes that define each group	Identical	scanpy rank_genes_groups (Wilcoxon, t-test)	100k cells in 15 s instead of 2.5 min	CHANGELOG 0.34.0
Copy-number callsTumour cells from normal cells	0.9999996	Pearson vs infercnvpy 0.6.1 per-cell scores	maynard2020_3k	CHANGELOG 0.39.0

Differences between conditions

Step	Agreement	Reference	Data	Source
Pseudobulk limma-voomGene changes with donors as replicates	to 1e-9	TMM, voom and moderated t vs edgeR 4.10 / limma 3.68	Kang 2018	CHANGELOG 0.41.0
Cell-type proportionsDid a population really grow?	Exact	propeller-style test vs R limma, “to numerical precision”	robust eBayes	CHANGELOG 0.18.0
Neighbourhood abundanceShifts too small to be a cluster	to 5e-7	Milo vs pertpy (pydeseq2 solver); edgeR logFC r = 0.989	Kang, 779 neighbourhoods	milo.py

Pathways and signals

Step	Agreement	Reference	Data	Source
TF and pathway activityWhich regulators are switched on	to 1e-5	decoupler univariate linear model, “to 5 decimals”	CollecTRI, PROGENy	CHANGELOG 0.15.0
Gene-set scoresHow active a set of genes is in each cell	Identical	decoupler aucell	9 gene sets on Kang and paul15	CHANGELOG 0.31.0
Cytokine responseWhich signals drive a response	0.9997	Spearman of z-scores vs CytoSig 0.0.3	Kang: 13 cell types × 43 cytokines	cytokine_activity.py
Cell–cell signallingWhich cells talk to which	Identical	liana-py 1.10 method scores, CellPhoneDB p, specificity rank	Kang 2018, paul15	CHANGELOG 0.26.0

Datasets together

Step	Agreement	Reference	Data	Source
BBKNN integrationLining up batches	Identical	bbknn 1.6.0 (max difference 0); Scanorama within 1.8e-6 of 1.7.4	Kang 2018	batch_methods.py
“Not in reference” flagCells the reference has never seen	36% vs 4%	B cells held out of the reference: B cluster flagged 36%, elsewhere at most 4%	Kang 2018	CHANGELOG 0.36.0

Immune repertoire and screens

Step	Agreement	Reference	Data	Source
ClonotypesWhich T or B cells share an ancestor	One-to-one	scirpy 0.22 CDR3 amino-acid definition	10k-cell 10x PBMC run	CHANGELOG 0.32.0
CRISPR screensWhich knockouts actually worked	99.3%	same class as pertpy 1.0.3 Mixscape (99.65% with exact neighbours)	Papalexi 2021, 20,729 cells, 25 targets	CHANGELOG 0.41.0

The co-pilot

Step	Agreement	Reference	Data	Source
Claims checkIs each sentence backed by a fact?	0 refuted	126 of 194 claims confirmed by rule; the rest left for the judge	23 answers on Kang, pbmc_1k_v3, ovary	CHANGELOG 0.19.0

Checks are rerun when a method changes; the CHANGELOG entry of that release records the new number. Found a disagreement? Open an issue.