baSeq docs › CITE-seq

CITE-seq

A CITE-seq analysis takes single-cell RNA and surface proteins measured in the same cells (antibody-derived tags, ADTs), normalises the proteins against their background, and clusters the cells on both together.

Start a CITE-seq analysis

In a project, click New session and choose CITE-seq as the kind of analysis. The sidebar marks CITE-seq with a cell and an antibody.

What to upload

Anything the single-cell upload reads, with the antibodies in it or beside it:

Cell Ranger feature barcoding output: the filtered_feature_bc_matrix folder or its .h5 file, with its Antibody Capture features (TotalSeq, BioLegend or BD AbSeq panels);
a MuData file (.h5mu) with an RNA and a protein modality;
an AnnData file (.h5ad) with the protein counts in obsm, as scvi-tools keeps them;
a Seurat object with an ADT assay, or a SingleCellExperiment with an antibody altExp;
a separate ADT count table beside each sample's RNA, as GEO often has them. It is paired with the RNA whose cells it shares.

Upload the raw_feature_bc_matrix too when you have it. Its empty droplets show how much antibody floats free in each lane, and dsb normalises against them. Without it, dsb takes the background from the cells' own negative population. Cell hashtags (HTOs) are kept apart from the proteins.

RNA and protein in one analysis

The single-cell steps run as they do in a single-cell RNA-seq analysis: quality control, doublets, normalisation, variable genes, PCA, integration, annotation, markers, pseudobulk and the rest. The protein steps sit between them. After integration, Multimodal Neighbors joins the RNA and the proteins into one graph, and clustering and the UMAP use it. Cells the RNA cannot tell apart, such as CD4 and CD8 T cells or naive and memory T cells (CD45RA and CD45RO come from one gene), fall into their own clusters, because their proteins differ. The weight of each modality is shown per cell and per cluster.

Reading dsb values

A dsb value is how many standard deviations a protein sits above its background in that cell. Values near 0 are background; values above about 3.5 are clearly expressed. Protein Gating and Protein Markers draw each protein's line between its negative and positive cells where the two populations meet, and fall back to 3.5 when they do not separate.

Step reference

Step	Stage	What it does	You get	Key inputs
Initial Plot	Setup	Reads the RNA and the antibody counts of every cell, recognises isotype controls and the gene behind each antibody, and keeps the empty droplets of an unfiltered matrix as the antibodies' background.	Cells per sample, every cell's RNA against its antibody counts, the antibodies with the largest share, and tables of the samples and the panel.	None.
Protein QC	QC	Removes cells whose antibody counts are far below or above their sample's (aggregates) or whose isotype controls take a high share (dead or sticky cells), and shows which antibodies stained cells and which sit at background.	The cells kept and removed, antibody counts per sample, and each antibody's signal over its background.	Outlier distance: 3, 4 or 5 (default) median absolute deviations, within each sample.
Normalize Proteins	QC	dsb (Mulè et al. 2022): each protein against its level in empty droplets (or the cells' negative population), each cell's technical noise removed. 0 is background; above about 3.5 is expressed. CLR, as in Seurat, on request.	A ridgeline of the most variable proteins, the noise removed per cell, and each protein's threshold and share of positive cells.	Normalization: dsb (default) or CLR.
Multimodal Neighbors	QC	One neighbour graph from RNA and protein. WNN (Hao et al. 2021) weighs each cell's RNA and proteins by how much each says about it; totalVI (scvi-tools) learns a joint model of both, on a GPU.	The RNA, protein and joint UMAPs side by side, and which modality defines each cell.	Method: WNN (default) or totalVI.
Perform Clustering	Analysis	Leiden clustering of the joint graph at resolutions 0.50, 0.75 and 1.00.	The clusters on the joint UMAP, which modality defines each cluster, and the samples in each.	None.
Protein Markers	Analysis	Every cluster against all other cells, every protein: difference, share of positive cells, AUROC and a Wilcoxon test, Benjamini-Hochberg adjusted.	A dot heatmap of the best markers (colour: mean, size: share positive), ordered by similarity, and a table.	Clustering resolution; markers per cluster (2, 3 or 5).
Explore Proteins	Analysis	Each protein on the UMAP beside its gene's RNA, and the clusters' mean protein against mean RNA.	Three panels per protein, and how well protein and RNA agree.	Clustering resolution; proteins (empty: the six that differ most between clusters).
Protein Gating	Analysis	Two proteins against each other as in flow cytometry, per condition, with quadrant gates where each protein's positive cells begin.	A density dot plot per condition with the share of cells in each quadrant, the same coloured by cluster, and the shares per condition and cluster.	The two proteins; clustering resolution.