aba-recipe-pack

Run a pagoda2.1 single-dataset scRNA-seq workflow from raw count import through QC, filtering, variance QC, PCA, UMAP, Leiden, marker genes, marker plots, optional annotation, and RDS/h5ad export. Also covers multimodal data (CITE-seq RNA+ADT, ATAC/multiome) via facets, with WNN/CCA integration.

CITE-seq (paired RNA + ADT surface-protein) analysis in R/Seurat v5 — load a 10x multimodal bundle, build a Seurat object with an RNA assay AND an ADT assay, CLR-normalize ADT per-feature (margin=2), run the standard RNA QC/PCA/UMAP/clustering pipeline, and characterize clusters with biaxial protein plots (FeatureScatter) + cross-modality FeaturePlots (`adt_CD3` vs `rna_CD3E`). For joint clustering driven by BOTH modalities, step up to seurat-wnn-multimodal.

Multi-sample scRNA-seq integration with R/Seurat v5 — load N samples, merge, split the RNA assay into per-sample layers, preprocess (Normalize→HVG→Scale→PCA) on the layered object, then IntegrateLayers (CCAIntegration / RPCAIntegration / HarmonyIntegration / JointPCAIntegration / scVIIntegration / FastMNNIntegration) into a corrected reduction. Cluster + UMAP on the integrated reduction; JoinLayers before DE. The Seurat v5 layer-based replacement for v4's FindIntegrationAnchors / IntegrateData anchor flow.

Map a scRNA-seq (or CITE-seq) query onto a pre-built multimodal reference in R/Seurat v5 — find SCT-flavoured anchors against the reference's supervised PCA (spca) with FindTransferAnchors(), then MapQuery() to transfer cell-type labels at MULTIPLE granularities (l1/l2/l3) AND impute predicted protein levels, all projected onto the reference's pre-computed WNN UMAP. Produces a query Seurat object with `predicted.celltype.l*` columns, per-cell `.score` confidences, a `predicted_ADT` assay, and a `ref.umap` reduction.

Map a scRNA-seq query onto a labeled Seurat reference (R/Seurat v5) — find transfer anchors between reference and query, transfer cell-type labels and metadata via TransferData, and project query cells into the reference's PCA + UMAP space via MapQuery. The reference's UMAP model must be saved with return.model = TRUE. Covers the generic case (user-supplied reference Seurat object) and points at Azimuth for pre-built tissue references.

Joint scRNA + scATAC analysis on a 10x Multiome dataset with R/Seurat (v5) + Signac — build ONE Seurat object holding an RNA assay AND a Signac ChromatinAssay, preprocess each modality separately (RNA→SCTransform→PCA; ATAC→TF-IDF→SVD/LSI), fuse them with `FindMultiModalNeighbors` (WNN), cluster + UMAP on the weighted graph, and add a gene-activity matrix + CoveragePlot diagnostics. The RNA+ATAC specialization of the general WNN pattern.

Single-sample scRNA-seq QC + clustering + markers with R/Seurat v5 — Read10X/Read10X_h5 → CreateSeuratObject → percent.mt + nFeature/nCount QC subset → NormalizeData/FindVariableFeatures/ScaleData → RunPCA + ElbowPlot → FindNeighbors/FindClusters (Louvain) → RunUMAP → FindAllMarkers (Wilcoxon). Default canonical Seurat path; SCTransform sibling recipe handles low-depth / regularized-NB normalization.

Single-sample scRNA-seq workflow with SCTransform regularized-NB normalization (R/Seurat v5) — replaces NormalizeData + FindVariableFeatures + ScaleData with one SCTransform() call, downstream PCA/UMAP/clustering/markers run on the SCT assay, and PrepSCTFindMarkers gates DE on merged multi-model objects. Preferred when sequencing depth is low or varies a lot across cell types.

Weighted Nearest Neighbor (WNN) multimodal clustering in R/Seurat v5 — process each modality independently (RNA + ADT, or RNA + ATAC) to get its own dimensional reduction, then learn cell-specific modality weights with FindMultiModalNeighbors() and cluster on the joint wsnn graph. Produces one shared UMAP where the modalities' contributions are weighted per cell, plus a per-cell `RNA.weight` metadata column revealing which cells are RNA-driven vs protein/chromatin-driven.

Decision guide for differential expression in scRNA-seq — picks between per-cell tests (Wilcoxon / MAST / LR / ROC) and pseudobulk DESeq2/edgeR depending on whether the question is cluster markers, within-sample two-group, or cross-sample condition effects. Settles the per-cell-vs-pseudobulk question via Squair 2021. Use when the user asks "Wilcoxon or MAST or DESeq2 for my scRNA DE?" or "per-cell or pseudobulk?".

Survey the experimental design of a study's sample metadata — surface the design structure, factor-to-factor associations (clustered Cramér's V heatmap), and CONFOUNDING / aliasing between biological and technical factors, before any downstream analysis. Outputs 2-3 scientist-facing figures plus a RED/AMBER/GREEN confounding verdict.

Decision guide for distinguishing batch effect from real biological variation when scRNA-seq samples separate on the UMAP. Compares per-axis diagnostics (cluster composition by sample, marker conservation, QC-confound check, designed-control test, cell-type composition). Use when the user is staring at a UMAP that splits by sample and is choosing between "integrate it away" vs "this is the signal I want to study."

Decision guide for picking a scRNA-seq normalization — LogNormalize (shifted-log), SCTransform v2, analytic Pearson residuals, and where CLR fits (ADT only). Use when the user asks "which normalization should I use" or is choosing between scanpy's `normalize_total`+`log1p`, Seurat's `NormalizeData`, `SCTransform`, or `sc.experimental.pp.normalize_pearson_residuals`.

Differential expression testing in R/Seurat (v5) — the practical method-choice recipe. Three paths — Path A FindAllMarkers (cluster markers, Wilcoxon), Path B FindMarkers (two-group test, optional MAST/LR with covariates), Path C AggregateExpression + DESeq2 (PSEUDOBULK for multi-sample condition effects). Includes the explicit ban on running bulk DE tools (DESeq2, edgeR, limma, pydeseq2) directly on per-cell scRNA matrices — those are bulk + pseudobulk tools only.