dsb: a method for normalizing and denoising antibody derived tag data from CITE-seq, ASAP-seq, TEA-seq and related assays.

The dsb R package is available on CRAN: latest dsb release
To install in R use install.packages('dsb')

Mulè, Martins, and Tsang, Nature Communications (2022) describes our deconvolution of ADT noise sources and development of dsb.

Vignettes:

1. Using dsb in an end to end CITE-seq workflow including multimodal clustering in Seurat
2. How the dsb method works
3. Normalizing ADTs if empty drops are not available
4. Python users: use dsb in Python with scverse software muon
5. FAQ etc.

Recent Publications Check out recent publications that used dsb for ADT normalization.

In the first “end to end” vignette, we demonstrate basic CITE-seq analysis starting from UMI count alignment output files from Cell Ranger though note that dsb is compatible with any alignment tool (see using other alignment tools). We load unfiltered UMI data containing cells and empty droplets, perform QC on cells and background droplets, normalize with dsb, and demonstrate protein-based clustering and multimodal RNA+Protein joint clustering using dsb normalized values with Seurat’s Weighted Nearest Neighbor method.

Background and motivation

Protein data derived from sequencing antibody derived tags (ADTs) in CITE-seq and other related assays has substantial background noise. Our paper outlines experiments and analysis designed to dissect sources of noise in ADT data we used to developed our method. We found all experiments measuring ADTs capture protein-specific background noise because ADT reads in empty / background drops (outnumbering cell-containing droplets > 10-fold in all experiments) were highly concordant with ADT levels in unstained spike-in cells. We therefore utilize background droplets which capture the ambient component of protein background noise to correct values in cells. We also remove technical cell-to-cell variations by defining each cell’s dsb “technical component”, a conservative adjustment factor derived by combining isotype control levels with each cell’s specific background level fitted with a single cell model.

Installation and quick overview

The method is carried out in a single step with a call to the DSBNormalizeProtein() function.
cells_citeseq_mtx - a raw ADT count matrix empty_drop_citeseq_mtx - a raw ADT count matrix from non-cell containing empty / background droplets.
denoise.counts = TRUE - implement step II to define and remove the ‘technical component’ of each cell’s protein library.
use.isotype.control = TRUE - include isotype controls in the modeled dsb technical component.

# install.packages('dsb')
library(dsb)

adt_norm = DSBNormalizeProtein(
  cell_protein_matrix = cells_citeseq_mtx, 
  empty_drop_matrix = empty_drop_citeseq_mtx, 
  denoise.counts = TRUE, 
  use.isotype.control = TRUE, 
  isotype.control.name.vec = rownames(cells_citeseq_mtx)[67:70]
  )

Please see the main vignette on CRAN for more details.

Selected publications using dsb

Publications from other investigators Izzo et al. Nature 2024 Arieta et al. Cell 2023 Magen et al. Nature Medicine 2023 COMBAT consortium Cell 2021 Jardine et al. Nature 2021 Mimitou et al. Nature Biotechnology 2021

Publications from the Tsang lab Mulè et al. Immunity 2024 Sparks et al. Nature 2023 Liu et al. Cell 2021 Kotliarov et al. Nature Medicine 2020

using other alignment algorithms

dsb was developed prior to 10X Genomics supporting CITE-seq or hashing data and we routinely use other alignment pipelines.

A note on alignment and how to use dsb with Cell Ranger is detailed in the main vignette. Cells and empty droplets are used by default by dsb.

To use dsb properly with CITE-seq-Count you need to align background. One way to do this is to set the -cells argument to ~ 200000. That will align the top 200000 barcodes in terms of ADT library size, making sure you capture the background. Please refer to CITE-seq-count documentation

CITE-seq-Count -R1 TAGS_R1.fastq.gz  -R2 TAGS_R2.fastq.gz \
 -t TAG_LIST.csv -cbf X1 -cbl X2 -umif Y1 -umil Y2 \
  -cells 200000 -o OUTFOLDER

If you already aligned your mRNA with Cell Ranger or something else but wish to use a different tool like kallisto or Cite-seq-count for ADT alignment, you can provide the latter with whitelist of cell barcodes to align. A simple way to do this is to extract all barcodes with at least k mRNA where we set k to a tiny number to retain cells and cells capturing ambient ADT reads:

library(Seurat)
umi = Read10X(data.dir = 'data/raw_feature_bc_matrix/')
k = 3 
barcode.whitelist = 
  rownames(
    CreateSeuratObject(counts = umi,
                       min.features = k,  # retain all barcodes with at least k raw mRNA
                       min.cells = 800, # this just speeds up the function by removing genes. 
                       )@meta.data 
    )

write.table(barcode.whitelist,
file =paste0(your_save_path,"barcode.whitelist.tsv"), 
sep = '\t', quote = FALSE, col.names = FALSE, row.names = FALSE)

With the example dataset in the vignette this retains about 150,000 barcodes.

Now you can provide that as an argument to -wl in CITE-seq-count to align the ADTs and then proceed with the dsb analysis example.

CITE-seq-Count -R1 TAGS_R1.fastq.gz  -R2 TAGS_R2.fastq.gz \
 -t TAG_LIST.csv -cbf X1 -cbl X2 -umif Y1 -umil Y2 \
  -wl path_to_barcode.whitelist.tsv -o OUTFOLDER

This whitelist can also be provided to Kallisto.
kallisto bustools documentation

kb count -i index_file -g gtf_file.t2g -x 10xv3 \
-t n_cores -w path_to_barcode.whitelist.tsv -o output_dir \
input.R1.fastq.gz input.R2.fastq.gz

Next one can similarly define cells and background droplets empirically with protein and mRNA based thresholding as outlined in the main tutorial.

A note on Cell Ranger –expect-cells

Note whether or not you use dsb, if you want to define cells using the filtered_feature_bc_matrix file, you should make sure to properly set the Cell Ranger --expect-cells argument roughly equal to the estimated cell recovery per lane based on number of cells you loaded in the experiment. see the note from 10X about this. The default value of 3000 is relatively low for modern experiments. Note cells and empty droplets can also be defined directly from the raw_feature_bc_matrix using any method, including simple protein and mRNA library size based thresholding because this contains all droplets.

Topics covered in other vignettes on CRAN: Integrating dsb with Bioconductor, integrating dsb with python/Scanpy, Using dsb with data lacking isotype controls, integrating dsb with sample multiplexing experiments, using dsb on data with multiple batches, advanced usage - using a different scale / standardization based on empty droplet levels, returning internal stats used by dsb, outlier clipping with the quantile.clipping argument, other FAQ.