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.

1. Using dsb in an end-to-end CITE-seq workflow
2. Using dsb if empty drops are not available
3. How the dsb method works
4. Using the dsb method in Python
5. Frequently Asked Questions
   

See notes on upstream processing before dsb

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

The functions in this package return standard R matrix objects that can be added to any data container like a SingleCellExperiment, Seurat, or AnnData related python objects.

Our paper combined experiments and computational approaches to find ADT protein data from CITE-seq and related assays are affected by substantial background noise. We observed that ADT reads from empty droplets—often more than tenfold the number of cell-containing droplets—closely match levels in unstained spike-in cells, and can also serve as a readout of protein-specific ambient noise. We also remove cell-to-cell technical variation by estimating a conservative adjustment factor derived from isotype control levels and per cell background derived from a per cell mixture model. The 2.0 release of dsb includes faster compute times and functions for normalization on datasets without empty drops.

The default 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 - 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)

isotype.names = c("MouseIgG1kappaisotype_PROT", "MouseIgG2akappaisotype_PROT", 
                  "Mouse IgG2bkIsotype_PROT", "RatIgG2bkIsotype_PROT")

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 = isotype.names, 
  fast.km = TRUE # optional
  )

Not all datasets have empty droplets available, for example those downloaded from online repositories where only processed data are included. We provide a method to approximate the background distribution of proteins based on data from cells alone. Please see the vignette Normalizing ADTs if empty drops are not available for more details.

adt_norm = ModelNegativeADTnorm(
  cell_protein_matrix = cells_citeseq_mtx, 
  denoise.counts = TRUE, 
  use.isotype.control = TRUE, 
  isotype.control.name.vec = isotype.names, 
  fast.km = TRUE # optional
  )

To speed up the function 10-fold with minimal impact on the results from those in the default function set fast.km = TRUE with either the DSBNormalizeProtein or ModelNegativeADTnorm functions. See the new vignette on this topic.

See the simple visual guide below. Please search the resolved issues on github for questions or open a new issue if your use case has not been addressed.

Any alignment software can be used prior to normalization with dsb. To use the DSBNormalizeProtein function described in the manuscript, you need to define cells and empty droplets from the alignment files. Any alignment pipeline can be used. Some examples guides below:

See the “end to end” vignette for information on defining cells and background droplets from the output files created from Cell Ranger as in the schematic below.
Please note whether or not you use dsb, to define cells using the filtered_feature_bc_matrix file from Cell Ranger, you need to properly set the --expect-cells argument to roughly your estimated cell recovery per lane based on how many cells you loaded. see the note from 10X about this. The default value of 3000 is likely not suited to most modern experiments.

# Cell Ranger alignment
cellranger count --id=sampleid\
--transcriptome=transcriptome_path\
--fastqs=fastq_path\
--sample=mysample\
--expect-cells=10000\  

See end to end vignette for detailed information on using Cell Ranger output.

Important: set the -cells argument in CITE-seq-Count to ~ 200000. This aligns the top 200000 barcodes per lane by ADT library size.
CITE-seq-count documentation

# CITE-seq-Count alignment
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

I recommend following the comprehensive tutorials by Tommy Tang for using Alevin, DropletUtils and dsb for CITE-seq normalization.
ADT alignment with Alevin
DropletUtils and dsb from Alevin output
Alevin documentation

I recommend checking out the tutorials and example code below to understand how to use kallisto bustools outputs with dsb.
kallisto bustools tutorial by Sarah Ennis
dsb normalization using kallisto outputs by Terkild Brink Buus
kallisto bustools documentation

Example script

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

After alignment define cells and background droplets empirically with protein and mRNA based thresholding as outlined in the main tutorial.

From other groups
Singhaviranon Nature Immunology 2025
Yayo Nature 2024
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

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

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
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