Validation

Numerical checks

• Benjamini–Hochberg adjustment matches R p.adjust(method = 'BH')
• Welch t-test, one-sided one-sample t-test and hypergeometric tail match SciPy (ttest_ind equal_var=False, ttest_1samp, hypergeom.sf)
• OLS differential abundance with covariates and rank-deficiency detection match statsmodels
• Condition-adjusted (partial) Spearman correlation matches pingouin
• Cell-level counts → pseudobulk log2 CPM reproduces the stated formula

Numerical checks

• Quasi-binomial GLM for frequencies with event counts matches statsmodels (Binomial, var_weights = trials, scale = Pearson X²)
• PCA variances and scores match NumPy SVD
• Paired designs reproduce SciPy ttest_rel
• Tie-aware Spearman permutation p values match SciPy permutation_test (exact enumeration for small n)

Numerical checks

• Every per-capture Scrublet observed score and threshold flag matches independent direct package fits, tolerance 1e-12
• Every Harmony corrected coordinate matches independent direct package fit, tolerance 1e-10
• Known separated Leiden graph, independent exact weighted graph partition quality, tolerance 1e-8, and seeded repeatability
• Every TSNE coordinate matches direct sklearn, tolerance 1e-10
• Every diffusion coordinate and DPT pseudotime matches independent Scanpy Gaussian-neighbor/diffmap/DPT fit, tolerance 1e-7
• Specimen counts/fractions preserve technical-capture grouping and zero populations; all tested limma outputs match independent real-R reference fit
• All implemented Plot Studio statistical families match independent actual R stats calls and exported reproduction expressions
• Nested-unit aggregation, matched/repeated completeness, confounding, missingness, exact risk sets/tied event/censoring and descriptive Kaplan-Meier hand cases
• SVG finite geometry, physical width, panel labels and active-content rejection; actual PNG pHYs resolution and CRC replacement

Numerical checks

• Every tested CellTypist query score matches independent package calls at atol 1e-12
• Specimen-disjoint holdout, explicit Unassigned thresholds and zero-library rejection
• No model-download network access during reference training/prediction
• Every SoupX auto/manual corrected CSR pointer/index/value and rho matches independent calls
• All supplied called droplets excluded from ambient pool; mismatched counts rejected
• Real authenticated loopback HTTP count uploads execute CellTypist and SoupX
• Worker attachment, compact provenance, JSON.GZ/binary restore, tamper/stale guards and parent/copy independence
• DOM upload/mapping, consent invalidation and pairing gates
• Existing gzip, browser-runtime, large-CSV and RNA/toolkit regressions

Reference package versions: CellTypist 1.7.1, SoupX 1.6.2, scikit-learn 1.8.0, SciPy 1.17.0, NumPy 2.3.5

Not implemented: DecontX, Pretrained public-model discovery, RNA velocity, CellChat/NicheNet, Mixed-effects condition models, Unlimited native Seurat jobs

Numerical checks

• Independent actual-R rank(-values,ties.method=average) plus stated formula, 4 cells x 3 signatures x 4 rank cutoffs; absolute tolerance 1e-12
• Implicit zero ties, tied positive values, rank capping and missing-gene skip policy
• Per-cell score invariance after cell subsetting with unchanged feature universe
• Every-gene sparse subset values and original observation IDs match parent
• Specimen means and cell counts match source values

Numerical checks

• All scores in 21 independent raw/linear/logged GSVA/ssGSEA cases match direct calls at atol 1e-12
• ssGSEA alpha 0/.25/1 and normalized/unnormalized options
• All score-model effects/SE/t/P/BH match independent simple, batch and fixed-subject fits at atol 1e-12
• Reverse contrasts invert effects and retain P
• Actual authorized loopback expression upload, R calculation, archive and deletion
• Negative transformed values, exact sample/binary order, identifier/design guards
• Worker attachment preserves original expression and gene DE results
• JSON.GZ/binary persistence, output/input hash tamper and stale-library rejection
• Actual-R DOM workflow for consent, four plot types, display-only regrouping and exports
• 240 single cells become eight biological pseudobulk columns
• Bundled mouse annotation metadata survives native JSON roundtrip using only synthetic expression

Numerical checks

• Four tied/untied unequal-size rank-sum cases versus actual R wilcox.test(exact=FALSE,correct=TRUE); absolute P tolerance 1e-7
• MAD scaling versus actual R mad
• Sparse Poisson residual variances versus independent dense formula, including zeros
• All-cell PCA projection versus dense centered/scaled X times fitted loadings
• Separated clique Louvain labels and known modularity 0.5
• Independent Euclidean silhouette example
• BH across all marker comparisons and normalized-linear mean fold changes
• Every-gene pseudobulk raw sums by biological specimen

Not implemented: doublet detection, ambient correction, Harmony, Leiden, automatic reference annotation, differential abundance, advanced trajectory/communication, full Plot Studio

Numerical checks

• Independent R stats::lm matches all tested scores for counts, linear and signed logged input at absolute tolerance 1e-8
• Exported expression + record JSON + R script reproduce scores, including CSV-protected identifiers
• Network sign reversal negates scores; positive weight rescaling preserves scores
• Constant expression and exact linear fits yield null with explicit reasons
• All retained single cells scored; zero-only target genes excluded and reported
• Actual worker preserves raw expression and gene DE results
• JSON snapshot/binary portable roundtrips, forged coverage/hash rejection and atomic recovery rollback
• DOM workflow uses pasted and file networks, coverage review and explicit confirmation, four plot families, display-only regrouping and all exports
• HTML parsing confirms TF page is separate, with four cards; complete suite checks unique IDs and references

Numerical checks

• Every effect, moderated SE, t, P and BH value matches independently constructed R model.matrix / limma fits at absolute tolerance 1e-9
• Simple, batch plus continuous-covariate and fixed paired-subject designs
• Reverse contrasts invert effects and retain FDR
• Exported R contrast code executed with the exported score CSV and reproduces effects
• No residual variation, cells, confounding, duplicate subject/condition, missing/constant score and saturation guards