The unbiased and extremely high-throughput nature of modern scRNA-seq approaches has proved invaluable for describing heterogeneous cell populations^1–3. Prior to the use of single-cell genomics, detailed definitions of cellular states were routinely obtained via carefully curated panels of fluorescently labeled antibodies directed at cell surface proteins, which are often reliable indicators of cellular activity and function⁴. Recent studies^5,6 have demonstrated the potential for coupling ‘index-sorting’ measurements with single-cell transcriptomics, enabling the mapping of immunophenotypes onto transcriptomically-derived clusters. However, massively parallel approaches based on droplet microfluidics^1–3, microwells^7,8 or combinatorial indexing^9,10 do not utilize cytometry for cellular isolation, and therefore cannot couple protein information with cellular transcriptomes, representing a significant limitation for these approaches. Targeted methods to simultaneously measure transcripts and proteins in single cells are limited in scale and/or can only profile a few genes and proteins in parallel^11–15 (Supplementary Table 1).

Here we describe Cellular Indexing of Transcriptomes and Epitopes by sequencing (CITE-seq), a method that combines highly multiplexed antibody-based detection of protein markers together with unbiased transcriptome profiling for thousands of single cells in parallel. We demonstrate that the method is readily adaptable to two different high-throughput single-cell RNA sequencing applications and show by example that it can achieve a more detailed characterization of cellular phenotypes than scRNA-seq alone.

We hypothesized that a DNA oligonucleotide conjugated to an antibody could be measured by sequencing as a digital readout of protein abundance. We conjugated antibodies to oligonucleotides designed to 1) be captured by oligo dT-based RNA sequencing library preparations, 2) contain a barcode sequence for identification the antibody and 3) allow subsequent specific amplification by PCR (Fig 1a). We adopted a commonly used streptavidin-biotin interaction to link the 5’ end of oligos to antibodies, and included a disulfide link, which allows the oligo to be released from the antibody in reducing conditions (Supplementary Fig. 1a). The antibody-oligo complexes are incubated with single-cell suspensions using conditions comparable to staining protocols used in flow cytometry, after which cells are washed to remove unbound antibodies and processed for scRNA-seq. In this example, we encapsulated single cells into nanoliter-sized aqueous droplets in a microfluidic apparatus designed to perform Drop-seq¹ (Fig. 1b). After cell lysis in droplets, both cellular mRNAs and antibody-derived oligos anneal to polyT-containing Drop-seq microparticles (Supplementary Fig. 1b,c) via their 3’ polyA tails. A unique barcode sequence on the oligos attached to the Drop-seq microparticle indexes the cDNA of mRNAs and antibody-oligos of each co-encapsulated cell in the reverse transcription reaction. The amplified antibody-derived tags (hereafter referred to as ADTs) and cDNA molecules can be separated by size and converted into Illumina-sequencing-ready libraries independently (Supplementary Fig. 1c,d). Importantly, the two library types are designed to be sequenced together, but because they are generated separately, their relative proportions can be adjusted in a pooled single lane to ensure that the appropriate sequencing depth is obtained for each library.

To assess the ability of our method to distinguish single cells based on surface protein expression, we designed a proof-of-principle ‘species-mixing’ experiment, leveraging a species-specific and highly expressed protein marker CD29 (Integrin beta-1). A mixed suspension of human (HeLa) and mouse (4T1) cells was incubated with a mixture of DNA-barcoded anti-mouse and anti-human CD29 antibodies. After washing to remove unbound antibodies, we performed Drop-seq¹ to investigate the concordance between species of origin of the transcripts and ADTs (Fig. 1c–e, Supplementary Fig. 2a). We deliberately used a high cell concentration to obtain high rates of multiplets (droplets containing two or more cells), to correlate mixed-species transcriptome data with mixed-species ADT signals from individual droplets. 97.2% of droplets that were identified as having contained either human, mouse, or mixed cells by transcriptome (Fig. 1c), had the same species classification by ADT counts (Fig. 1d). Cell counts based on RNA or ADT are highly correlated between both methods (Fig. 1e), demonstrating low drop-out rate of ADT signals. We performed the same experiment using a commercially available system from 10× Genomics and obtained comparable results (Supplementary Fig. 2b–e).

We next wanted to characterize the quantitative nature of our CITE-seq ADT protein readout. Flow cytometry is the gold-standard for identification and enumeration of cell subsets based on quantitative differences in surface markers^16,17. We therefore aimed to benchmark the sensitivity of CITE-seq protein detection to flow cytometry using immune cells as a model system. We prepared a set of CITE-seq antibodies directed against markers commonly used in flow cytometry to identify and discriminate relevant immune sub-populations. We performed multiparameter flow cytometry (Fig. 2a) and CITE-seq (Fig. 2b) experiments using the same set of antibodies on aliquots of the same pool of peripheral blood mononuclear cells. Using ADT levels we were able to construct cytometry-like ‘bi-axial’ gating plots (Fig. 2b) and compare these qualitatively and quantitatively to the flow cytometry data (Fig. 2a). Cell distribution profiles based on expression of marker proteins associated with various T cell subsets, B cells, plasmacytoid, myeloid dendritic cells and monocytes were remarkably similar (Fig. 2a,b; Supplemental Fig. 3a,b).

Next, we asked whether relative quantitative differences in expression levels observed by flow cytometry can be observed by CITE-seq. For this, we focused on the marker CD8a, since it exhibits a wide quantitative range of levels across immune cell populations. We incubated cord blood mononuclear cells (CBMCs) with CITE-seq antibody conjugates and fluorophore-conjugated antibodies, so that some CD8a epitopes on each cell would be labeled by fluorophore and some by oligo. Cells were sorted (by fluorescence-activated cell sorting, FACS) into separate pools based on CD8a fluorescence (CD8a very high (+++), CD8a high (++), CD8a intermediate(+) and CD8a low(+/−); Fig. 2c,d, Supplementary Fig. 3c). Each pool was then split and separately reanalyzed by flow cytometry and CITE-seq. For each pool defined by FACS, similar relative expression levels of CD8a were observed by both methods (Fig. 2e,f; Suppementary Fig. 3d,e). We conclude that CITE-seq ADT levels are consistent with gold-standard flow cytometry and can therefore enable high-resolution immunophenotyping in concert with transcriptomics.

We next aimed to perform a broad immunophenotypic and transcriptomic characterization of a complex immune cell population using CITE-seq. The immune system has been extensively profiled by cell surface markers¹⁶ and scRNAseq^3,6,18, with both methods reliably identifying the same cell types at appropriate proportions. It is therefore an ideal system to validate the multimodal readout of CITE-seq. We prepared a CITE-seq panel of 13 well-characterized monoclonal antibodies that recognize cell-surface proteins routinely used as markers for immune-cell classification (Supplementary Table 2). Measuring protein abundance by antibodies can be complicated by non-specific binding, resulting in higher background and/or false positive results. To estimate background of antibody binding within experiments, we developed a low-level ‘spike-in’ control. We reasoned that a rare ‘spiked-in’ population of murine cells could be easily distinguished transcriptomically but should not cross-react with our anti-human antibodies, enabling us to define background ADT levels directly from the data. We therefore spiked a low number of mouse fibroblasts (~4% 3T3) into our CBMCs, incubated the cell pool with our CITE-seq antibody panel and ran the 10× Genomics single cell workflow to measure cDNA and ADT profiles from 8,005 cells in total. Unsupervised graph-based clustering based on RNA expression revealed recognizable cell types indicated by expression of select marker genes (Fig. 3a, Supplementary Fig. 4). Murine cells clustered separately, and also exhibited low ADT counts for each marker, allowing us to set a baseline for signal vs noise to more clearly delineate positive from negative cell populations (Supplementary Fig. 5a,b). Through this thresholding step, we identified three antibody-oligo conjugates with no specific binding (i.e., no signal over background threshold) and excluded these from further analysis (Supplementary Fig. 5b).

We detected strong ADT enrichment for different markers in the correct immune populations: We observe CD3e within the T cell cluster (Fig. 3b), and CD4 and CD8a in largely non-overlapping T cell subpopulations (Fig. 3b). We observe CD19 ADTs almost exclusively in the B cell cluster (Fig. 3b), CD56 and CD8a ADTs in the NK cluster, and CD11c, CD14 in the monocyte and dendritic cell cluster, together with CD16, also observed in the NK cells (Fig. 3b). We could also correctly identify a rare precursor cell population present at less than 2% in cord blood (CD34+ cells; Fig 3b). We observed that the ADT levels per-cell exhibited higher counts than mRNA-levels of the same genes and were less prone to ‘drop-out’ events. Consistent with this, we find low correlations between mRNA and ADT on a single cell basis and higher correlation when averaging expression within clusters (Supplementary Fig. 6). We used the ADT levels and clustering information based on transcriptome to construct multimodal CITE-seq ‘bi-axial’ gating plots, revealing similar profiles that are well-established by flow cytometry (Fig. 3c, Supplementary Fig. 5c). For example, we could resolve strong anti-correlation of CD4 and CD8a ADT levels in T cells and quantitative differences in marker expression between subsets, such as expression differences of CD8a between NK and T cells (blue and red cells; Fig. 3c) or CD4 between monocytes and T cells (yellow and turquoise cells; Fig. 3c). In addition to clustering cells based on their transcriptome, we performed clustering based on ADT levels, resulting in clear and consistent cell type separation (Supplementary Fig. 7).

We next asked whether CITE-seq could enhance our characterization of immune cell phenotypes, compared to scRNA-seq data alone. We noted an opposing gradient of CD56 and CD16 ADT levels within our transcriptomically-derived NK cell cluster, potentially corresponding to CD56^bright and CD56^dim subsets^19,20 (Fig. 3b, Supplementary Fig. 6a), and therefore sub-divided our NK cell cluster based on CD56 ADT levels (Fig. 3d). When comparing the molecular profiles of these groups, we observed protein and RNA changes that were highly consistent with literature^19,20. We observed an apparent complementarity between levels of CD16 (Fig. 3e), and to a lesser extent of CD8a ADTs (Supplementary Fig. 6b), compared to CD56 ADTs within these two subsets. For 11 genes that have been previously characterized as differentially expressed within these subtypes^19–21, we detected up or downregulation consistent with the literature in 10 cases, including GZMB, GZMK and PRF1. This illustrates the potential for integrated and multimodal analyses to enhance discovery and description of cellular phenotypes, particularly when differentiating between cell populations with subtle transcriptomic differences.

The ability to layer additional molecular measurements on top of scRNA-seq data represents an exciting advance and growing challenge for the single cell community. CITE-seq enables multimodal analysis of single cells at the scale afforded by droplet-based single-cell sequencing approaches. We demonstrate the value of multimodal analysis to reveal phenotypes that could not be discovered by using scRNA-seq alone, and also envision CITE-seq enabling new studies of post-transcriptional gene regulation at the single cell level. In contrast to flow and mass cytometry, detection of oligo-barcoded antibodies is not limited by signal collision; a 10 nucleotide sequence can easily encode more barcodes than there are human proteins, enabling the potential for large-scale immunophenotyping with panels of tens to hundreds of antibodies. In addition, we envision that mild cell permeabilization and fixation procedures that are used for intracellular cytometry assays will also be compatible with CITE-seq, which may significantly expand the number of markers and biological questions that can be interrogated. A modified version of CITE-seq in which only ADTs are analyzed on a massively parallel scale without capturing cellular mRNAs (cytometry by sequencing) can also be envisaged. A conceptually similar approach, Abseq, has recently been described²² which, in contrast to CITE-seq, focuses on the detection of single cell protein levels using DNA barcodes using highly advanced custom microfluidics.

Finally, we have shown that the CITE-Seq is fully compatible with a commercially-available single-cell instrument (10× Genomics) and should be readily adaptable to other droplet, microwell and combinatorial indexing-based high-throughput single-cell sequencing technologies^2,7–10 with no, or minor customizations. We believe that CITE-seq has the potential to advance single-cell biology by layering an extra dimension on top of single-cell transcriptome data.