Stoeckius et al. 2018 — Cell Hashing (Genome Biology 19:224)

1. Citation

Stoeckius M, Zheng S, Houck-Loomis B, Hao S, Yeung BZ, Mauck WM III, Smibert P, Satija R. “Cell Hashing with barcoded antibodies enables multiplexing and doublet detection for single cell genomics.” Genome Biology 2018;19:224. doi:10.1186/s13059-018-1603-1. PMID 30567574; PMCID PMC6300015.

• Received 2018 Jun 5; Accepted 2018 Dec 4; published 2018 Dec 19. Collection date 2018.
• Open access: https://pmc.ncbi.nlm.nih.gov/articles/PMC6300015/ (accessed 2026-07-13).
• Data: Gene Expression Omnibus accession GSE108313. Software: Seurat (https://github.com/satijalab/seurat). Protocols: www.cite-seq.com.
• Marlon Stoeckius and Shiwei Zheng contributed equally; Smibert and Satija are corresponding authors.

Author-list discrepancy [summarizer flag]. The tasking citation listed additional authors (Yeung D, Hafemeister C, Papalexi E, Mimitou EP, Jain J, Srivastava A, Stuart T, Fleming LM, Rosenthal AJ, Smibert CA). The article as published at PMC lists eight authors, given above. The list above is the source’s own.

License (verbatim, from the article’s Copyright and License information):

“Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.”

Competing interests (verbatim, load-bearing for provenance):

“MS, PS and BHL have filed a patent application based on this work (US provisional patent application 62/515-180). BZY is an employee at BioLegend Inc., which is the exclusive licensee of the New York Genome Center patent application related to this work. All other authors declare that they have no competing interests.”

2. Access note

Full text (Abstract, Introduction, Results, Discussion, Methods, Computational methods, data availability, competing interests) read in full from PMC. Not read: the three additional files — Additional file 1 (supplementary figures S1–S5, PDF), Additional file 2 (Table S1, µg antibody per condition), Additional file 3 (Table S2, sample composition). Several numbers below are cited by the main text to those supplementary figures; the figures themselves were not inspected. Figure panels (Fig. 1–3) were not viewed as images; their content is taken from the legends and body text.

3. Thesis

Oligo-tagged antibodies against ubiquitously expressed surface proteins can uniquely label cells from distinct samples so those samples can be pooled into a single droplet scRNA-seq run, then computationally demultiplexed — simultaneously enabling sample assignment, cross-sample multiplet detection, and cost-reducing “super-loading.”

4. Problem & context

• Per-cell library-prep cost remains high; profiling tens to hundreds of thousands of cells is costly.
• Multiplets are an unsolved problem: multiplets give higher-complexity libraries than singlets, but the source states this signal “is not sufficient for unambiguous identification” (citing ref. 11).
• Technical/batch effects have been shown to mask biological signal in integrated scRNA-seq analysis (citing ref. 12), motivating experimental rather than purely computational solutions.
• Prior art the paper builds on: demuxlet (Kang et al., ref. 13) pools genotypically distinct donors and uses genetic polymorphisms as the sample fingerprint; and CITE-seq (Stoeckius et al., ref. 18) from the same group, which converts surface-protein detection into a sequenceable readout.
• Stated constraint of the genetic route, in the Introduction: “this elegant approach requires pooled samples to originate from previously genotyped individuals” (verbatim).

5. Method / approach

5.1 The reagent — what is attached, to what, when

• First-generation (used in the main PBMC experiment): monoclonal antibodies against four ubiquitously and highly expressed immune surface markers — CD45 [clone HI30], CD98 [clone MEM-108], CD44 [clone BJ18], CD11a [clone HI111] — combined into eight identical pools (pool A through H), each pool conjugated to a distinct Hashtag oligonucleotide (HTO).
• Second-generation / “universal” reagent: a pool of antibodies against B2M [clone 2M2] (MHC class I / beta-2-microglobulin) and CD298 [clone LNH-94] (sodium-potassium ATPase subunit), purchased from BioLegend. Rationale: these are “among the most broadly expressed surface proteins in human tissues” (citing ref. 21, Human Protein Atlas). Using a pool rather than a single antibody “mitigates the possibility that stochastic or cell type variation in the expression of any one marker would introduce bias in HTO recovery.”
• Each HTO carries a unique 12-bp barcode, sequenced alongside the transcriptome.
• Conjugation: iEDDA click chemistry (methyltetrazine-PEG4-NHS ester on antibody at 30-fold excess, 30 min RT; 5′-amine HTO reacted with 20-fold excess trans-cyclooctene-PEG4-NHS in 1X borate buffered saline + 20% DMSO, 30 min), covalent and irreversible; conjugates formed by mixing and incubating ≥ 1 h at RT; cleanup by Amicon centrifugal filtration (30 kDa MWCO to concentrate to 1 mg/ml; 50 kDa MWCO after conjugation).
• Workflow step: cells from each sample are stained separately with their HTO-conjugated antibody pool, then pooled and loaded into the droplet instrument. Staining is done as for CITE-seq (ref. 18).

Oligonucleotide sequences (functional strings, verbatim):

Hashtag oligo: GTGACTGGAGTTCAGACGTGTGCTCTTCCGATCTxxxxxxxxxxxxB AAAAAAAAAAAAAAAAAAAAAAAAAAAAAA*A A
HTO additive:  GTGACTGGAGTTCAGACGTGTGCTC
ADT additive:  CCTTGGCACCCGAGAATTCC
SI-PCR:        AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGC*T*C
RPI-x:         CAAGCAGAAGACGGCATACGAGATxxxxxxxxGTGACTGGAGTTCCTTGGCACCCGAGAATTCCA
D7xx_s:        CAAGCAGAAGACGGCATACGAGATxxxxxxxxGTGACTGGAGTTCAGACGTGTGC
x: Barcode or index sequence;  B: T,G,C, not A;  *: Phosphorothioate bond

(Note: whitespace within the Hashtag oligo line reflects markup in the PMC rendering of the polyA tail; the intended sequence is the 12-bp barcode x…x + B + a polyA tract ending in a phosphorothioate bond.)

5.2 Library generation — the extra library and its cost

• HTOs use a different amplification handle than CITE-seq antibody-derived tags (ADTs), so HTO, ADT, and scRNA-seq libraries are independently amplified and pooled at desired quantities.
• 10x Genomics Single Cell 3′ v2 workflow, standard through cDNA amplification. 2 pmol each of the HTO and ADT additive oligos are spiked into the cDNA amplification PCR.
• 0.6X SPRI separates the large cDNA fraction (on beads → transcriptome library) from the ADT/HTO fraction (in supernatant). An additional 1.4X volume of SPRI beads is added to the ADT/HTO fraction to bring the ratio to 2.0X; wash in 80% ethanol, elute in water, then a second 2.0X SPRI to remove excess single-stranded oligos.
• Separate PCRs: ADT library (SI-PCR + RPI-x primers); HTO library (SI-PCR + D7xx_s primers).
• Sequenced on Illumina HiSeq2500 (two rapid run flowcells), aiming for a 90% : 5% : 5% contribution of transcriptome : HTO : ADT libraries in the sequencing data.
• Stated added sequencing cost of Cell Hashing: “approximately 2–5% of transcriptome sequencing costs” (Discussion).

5.3 The main experiment

• PBMCs from eight human donors (A–H), each stained with one HTO-conjugated pool, while simultaneously running a CITE-seq titration of seven immunophenotypic markers (CD8 [RPA-T8], CD45RA [HI100], CD4 [RPA-T4], HLA-DR [L243], CD3 [UCHT1], CCR7 [G043H7], PD-1 [EH12.2H7]; BioLegend, 1–2 conjugated oligos per antibody on average).
• Pooled in equal proportion with an equal number of unstained HEK293T cells and 3% mouse NIH-3T3 cells as negative controls (both unlabeled by HTOs).
• Run in a single lane of 10x Genomics Chromium Single Cell 3′ v2.
• Donors A–G were stained with CITE-seq antibodies; donor H was not (unstained control) — this is later used as an independent test of donor calls.
• Genotyping of all 8 PBMC samples and HEK293T with the Illumina Infinium CoreExome 24 array, so HTOs and genotypes could be used as independent demultiplexing approaches.

5.4 “Super-loading” (definition + numbers)

Verbatim (Results): “Following the approach in Kang et al. [13], we ‘super-loaded’ the 10x Genomics instrument, loading cells at a significantly higher concentration with an expected yield of 20,000 single cells and 5000 multiplets.”

• Expected: 20,000 singlets + 5000 multiplets.
• By Poisson statistics, 4365 of those multiplets should be combinations from distinct samples and can be discarded → unresolved (within-sample) multiplet rate of 3.1%.
• Achieving a similar multiplet rate without multiplexing would yield ~4000 singlets.
• Since the cost of commercial droplet systems is fixed per run for sample prep, multiplexing allows profiling ~400% more cells for the same cost.
• Discussion adds: the per-cell library-prep saving approaches an order of magnitude as the number of multiplexed samples increases; and Cell Hashing “enables even a single sample to be highly multiplexed, as cells can be split into an arbitrary number of pools.”
• Offsetting cost, credited to Kang et al.: “savings in library prep are partially offset by reads originating from multiplets, which must be sequenced and discarded.”

5.5 Computational pipeline (named software + versions)

Species-control removal: a low-resolution pre-clustering (PCA on the 500 most highly expressed genes, then k-medoids on a distance matrix from the first 2 PCs) identified 160 NIH-3T3 and 2233 HEK293T cells out of 20,854 barcodes with ≥ 200 UMI; all were excluded.

5.6 HTO classification procedure — end to end (the core method)

1. Barcode set. “We classified all barcodes where we detected at least 200 RNA UMI, regardless of HTO signal” (20,854 barcodes in the main experiment).

2. Normalization — centered log ratio (CLR). Raw HTO counts are divided by the geometric mean of that HTO across cells and log-transformed (natural log):

   x_i' = log( x_i / ( Π_{i=1..n} x_i )^(1/n) )

   where x_i = count for a given HTO in cell i, n = total cell number.

3. Clustering to seed the background. k-medoids clustering of all cells on the normalized HTO data with k = 9 (8 HTOs + 1). Eight clusters were each highly enriched for one HTO; the ninth was enriched for cells with low expression of all HTOs. This is “an initial solution to the demultiplexing problem.”

4. Per-HTO background fit. For each of the 8 HTOs independently: identify the k-medoids cluster with the highest average expression of that HTO and exclude those cells; then fit a negative binomial distribution to the remaining HTO values, after further excluding the highest 0.5% of values as potential outliers.

5. Threshold (the exact cutoff). Compute the q = 0.99 quantile of the fitted negative binomial and threshold every cell against this HTO-specific value.

   Verbatim (Results): “Barcodes with HTO signals above the 99% quantile for this distribution were labeled as ‘positive,’ and barcodes that were ‘positive’ for more than one HTO were labeled as multiplets.”

6. Calls from the positives (Methods, verbatim): “Barcodes that were positive for only one HTO were classified as singlets. Barcodes that were positive for two or more HTOs were classified as multiplets and assigned sample IDs based on their two most highly expressed HTO. Barcodes that were negative for all eight HTOs were classified as ‘negative.’”

7. Visualization. Euclidean distances on the normalized HTO data → tSNE (Fig. 1d).

Downstream transcriptome clustering: PCA on the 1000 most highly variable genes (variance/mean ratio), distance matrix from the first 10 PCs → tSNE + graph-based clustering in Seurat; seven hematopoietic clusters annotated by canonical markers.

5.7 Ancillary experiments

• CITE-seq antibody titration (multiplexed via hashing): concentrations 5, 3, 1, 0.5, 0.25, 0.06, and 0 µg per test, staggered across batches so total antibody + oligo stays constant. Staining index computed as SI = (Pos_0.5 − Neg_0.5) / (2 × mad(Neg)) — difference of positive and negative peak medians over twice the mean absolute deviation of the negative peak. Negative peak approximated using unstained control cells (donor H); positive peak from the maximally ADT-enriched Seurat cluster in donors A–G.
• Cell-line cross-check of background binding: four cell lines (HEK293T, THP1, K562, KG1), each labeled with three distinct Cell Hashing oligos, mixed; demultiplexed, then RNA-clustered into four transcriptomic clusters.
• Dilution / oligo-competition experiment: Jurkat cells stained with decreasing concentrations (1:100, 1:500, 1:1000) of the B2M/CD298 hashing antibodies, ~2000 cells per 10x lane, vs. a non-hashed control lane; subsampled to the same cells and reads per cell to compare UMI and gene counts.
• Low-quality-cell rescue: re-applying the previously determined HTO thresholds to barcodes with 50–200 UMI.

6. Key claims / findings

1. Pairwise HTO counts look like species-mixing plots — mutually exclusive between singlets (Fig. 1b; axes clipped at 99.9% quantiles to exclude visual outliers).
2. Main experiment yield: after removing HEK293T and mouse cells, 14,002 singlets and 2974 identifiable multiplets — “in line with expectations.”
3. All 28 possible cross-sample doublet combinations were visible as distinct groups in an HTO-only tSNE, alongside the 8 singlet clusters (Fig. 1d).
4. “Negative” barcodes (negligible signal for every HTO) consisted primarily (86.5%) of the unlabeled HEK293T and mouse cells.
5. Multiplets show a positive shift in RNA UMI, as expected (Fig. 1e) — but the large dynamic range of UMI in multiplets illustrates that UMI count alone cannot unambiguously call multiplets; the same challenge holds for total HTO signal (Additional file 1: Fig. S1A).
6. Transcriptome clustering of hashed singlets recovered seven hematopoietic subpopulations interspersed across all 8 donors (Fig. 1f).
7. Concordance with demuxlet is strong overall, “even when considering the precise sample mixture in called doublets” (row-normalized confusion matrix, Fig. 2a). 13,421 concordant singlet/doublet classifications.
8. 871 barcodes were HTO-singlets but demuxlet-”ambiguous.” Their HTO classification strength (reads on the top HTO) was identical to fully concordant singlets (Fig. 2b), but they had reduced RNA UMI counts (Fig. 2c). Authors’ conclusion: these could not be genetically classified at their sequencing depth.
9. Sequencing depth of this experiment: ~24,115 reads per cell, which the authors state is below the recommended depth for using demuxlet.
10. 2528 barcodes received discordant singlet/doublet calls between the two techniques. Only barcodes called doublets by both showed a positive shift in transcriptomic complexity (Fig. 2d) → the authors read the discordant calls as conservative false positives from both methods, “perhaps due to ambient RNA or HTO signal.”
11. Restricting to demuxlet doublet calls with > 95% probability produced a 75% drop in the number of discordant calls (Fig. 2e).
12. 216 / 11,464 (1.9%) of barcodes called singlets by both methods received discordant donor assignments. Using donor H (the only donor not stained with CITE-seq antibodies) as an independent arbiter: in 40 instances where demuxlet — but not Cell Hashing — called a cell donor H, 37 had robust (> 1000) ADT counts. The authors call these misclassification errors from demuxlet, “in line with demuxlet’s estimated error rate of 1–2% [13].”
13. Cell-line experiment: demultiplexing vs. transcriptomic cluster identity showed nearly perfect concordance (99.7%), i.e., a low misassignment rate for that experiment.
14. Cell Hashing false-negative rate = 0.9%: 117 barcodes classified “negative” by HTO but whose transcriptomes clustered within PBMC singlet subtypes; stated to have “negligible effects on cell type proportion estimates” (Additional file 1: Fig. S4).
15. Recovering low-RNA cells: among 4344 barcodes with 50–200 UMI, HTO classification recovered 1110 additional singlets and called 3108 negatives. Predicted singlets projected cleanly onto B / NK / T / myeloid populations on tSNE; “negative” barcodes did not separate by their forced classification, consistent with ambient-RNA mixtures. Hashing therefore gives a sample-identity readout independent of the transcriptome, allowing the UMI cutoff (previously 200) to be relaxed.
16. Titration result: all antibodies showed only background signal in negative-control conditions and very weak signal-to-noise at 0.06 µg/test; signal-to-noise for most antibodies began to saturate at 0.5–1 µg/test, comparable to recommended flow-cytometry concentrations.
17. No transcriptomic-complexity penalty from hashing: Jurkat cells hashed with the B2M/CD298 pool at 1:100–1:1000 had read-vs-UMI relationships indistinguishable from non-hashed cells at all tested concentrations (Additional file 1: Fig. S5), addressing the concern that polyA-containing HTOs might compete with cellular mRNA.
18. HEK293T caveat found via demuxlet: 225 barcodes were called HEK by demuxlet but clustered transcriptomically with PBMCs; they expressed tenfold fewer UMI than transcriptomically classified HEK293T cells and lacked HEK293T-specific transcripts (e.g., NGFRAP1) — excluded from all further analysis.
19. Of cells transcriptomically annotated as HEK293T/NIH-3T3 (unlabeled), 60.1% were HTO-annotated “negative” and 32.1% were HTO-annotated singlets — “in agreement with our expected ratios in our ‘super-loaded’ 10x Genomics experiment,” i.e., these singlet calls are PBMC+HEK/3T3 doublets.

7. Load-bearing statements (VERBATIM — license-aware mode; CC BY 4.0 permits reproduction)

1. The threshold (Results, “Hashtag-enabled demultiplexing…”):

   “Briefly, we modeled the “background” signal for each HTO independently as a negative binomial distribution, estimating background cells based on the results of an initial k-medoids clustering of all HTO reads (“Methods” section). Barcodes with HTO signals above the 99% quantile for this distribution were labeled as “positive,” and barcodes that were “positive” for more than one HTO were labeled as multiplets. We classified all barcodes where we detected at least 200 RNA UMI, regardless of HTO signal.”

2. The two-column trade-off (Discussion) — the passage for route selection:

   “Our approach is complementary to pioneering genetic multiplexing strategies, with each having unique advantages. Genetic multiplexing does not utilize exogenous barcodes and therefore does not require alterations to existing workflows prior to or after sample pooling. In contrast, Cell Hashing requires incubation with antibodies against ubiquitously expressed surface proteins but can multiplex samples with the same genotype. Both methods do slightly increase downstream sequencing costs, due to the increased depth or read length needed to identify SNPs (genetic approaches) or sequencing of HTO libraries (Cell Hashing; approximately 2–5% of transcriptome sequencing costs). We believe that researchers will benefit from both approaches, enabling multiplexing for a broad range of experimental designs. In particular, we envision that our method will be most useful when processing genetically identical samples subjected to diverse perturbations (or experimental conditions/optimizations, as in our titration experiment) or to reduce the doublet rate when running cells from a single sample.”

3. The genotype route’s stated requirement (Introduction):

   “While this elegant approach requires pooled samples to originate from previously genotyped individuals, in principle, any approach assigning sample fingerprints that can be measured alongside scRNA-seq would enable a similar strategy.”

4. The disagreement explanation + depth (Results, “Genotype-based demultiplexing validates Cell Hashing”):

   “We conclude that these barcodes likely could not be genetically classified at our relatively shallow sequencing depth (~ 24,115 reads per cell), which is below the recommended depth for using demuxlet, but likely represent true single cells based on our HTO classifications.”

   and, on the doublet disagreements:

   “We note that this does reflect a minority of barcodes (compared to 13,421 concordant classifications) and that in these discordant cases, it is difficult to be certain which of these methods is correct.”

5. The undetectable-doublet limitation + the epitope limitation (Discussion):

   “Informatic detection of multiplets based on transcriptomic data also remains an important challenge for the field, for example, to identify doublets originating from two cells within the same sample.”

   “We however caution that there can be instances when a cell type of interest does not express these virtually ubiquitous surface proteins, which would result in failure to successfully label and demultiplex these cells.”

8. Stated scope, assumptions, limitations (the source’s own caveats)

• Epitope dependence. The universal reagent assumes B2M and CD298 are expressed; the authors explicitly caution that a cell type of interest may not express them, which “would result in failure to successfully label and demultiplex these cells.”
• Nuclei not covered. For single-nucleus sequencing, “an additional set of hashing reagents directed against nuclear proteins would further generalize this approach” — i.e., the reagents presented do not cover nuclei.
• Human-only demonstration. Reagents are described as applicable “to virtually any human sample with readily available commercial reagents.” Mouse NIH-3T3 cells appear only as unlabeled negative controls. Generalization “regardless of species, tissue, or technology” is framed as a future improvement, not a demonstrated result.
• Within-sample doublets are not detected. Hashing detects cross-sample multiplets; the residual is the 3.1% unresolved multiplet rate in the super-loaded design, and the Discussion names same-sample doublet detection as still “an important challenge for the field.”
• A hashed “singlet” can still be a doublet with an unlabeled cell. Methods: singlet calls “could also represent doublets of a PBMC with a HEK293T or NIH-3T3 cell, as the latter two populations were unlabeled.”
• Titration experiment is proof-of-concept only: “an ideal titration experiment would use cells from the same donor for all conditions and a larger range of concentrations.”
• Random-forest caveat (low-quality-cell classification): “we note that this classifier is guaranteed to return a result for each barcode.”
• Both methods slightly increase sequencing cost; library-prep savings are partially offset by multiplet reads that must be sequenced and discarded (credited to Kang et al.).
• Conflict of interest: patent application filed on this work; a co-author is a BioLegend employee and BioLegend is the exclusive licensee.

9. Failure modes / invalidity patterns

10. What the source does NOT address (confident silences)

• No decision rule mapping a study design to a route. The Discussion gives use-cases where hashing is “most useful” and lists each route’s advantage/requirement, but it states no if/then rule, no decision tree, and no threshold for choosing genotype vs. hashing. It calls the approaches “complementary” and says “researchers will benefit from both approaches.” See §12.
• No stated recommended sequencing depth for demuxlet. The paper says its own ~24,115 reads/cell is “below the recommended depth for using demuxlet” but never gives the recommended number.
• No head-to-head accuracy metric against a third ground truth. Neither method is scored against an independent gold standard; where they disagree the paper says outright it “is difficult to be certain which of these methods is correct” and then argues from UMI-complexity shifts and ADT signal.
• Batch effects are asserted, not measured. The abstract names “sample-specific batch effects” as a motivating challenge, and the Discussion says multiplexing should help “as experimental designs seek to minimize technology-driven batch effects” — but no experiment in this paper quantifies a reduction in batch effect. The batch-effect claim is a motivating/design argument, backed by citations to prior work (refs. 11, 12, 14), not by evidence generated here. There is no hashed-vs-unhashed batch-effect comparison, no variance-decomposition, no metric.
• No cost model in currency. Costs are given as percentages/ratios (2–5% added sequencing; ~400% more cells; “approaching an order of magnitude”), never as dollars.
• No performance guidance for scaling. No statement about how many HTOs can be multiplexed at once beyond the 8 used (and 3-per-cell-line in the cell-line test); no statement about how the 0.99 quantile behaves as HTO count grows, or how to pick k for k-medoids other than “k = 9” for 8 HTOs.
• No non-human / non-mammalian / tissue (vs. dissociated cell) demonstration.
• No treatment of unequal pooling proportions. All samples were pooled “at equal concentration.”
• No mention of Seurat’s HTODemux by name or of any specific function for HTO classification — only that the tools are “freely available as part of the open-source R package Seurat.”

11. Open questions / ambiguities the source leaves unresolved

• Which method is right where they disagree — stated explicitly as unresolved.
• Whether the discordant doublet calls are ambient RNA, ambient HTO, or something else (“perhaps”).
• Whether hashing generalizes across species, tissues, and technologies (framed as future work).
• Whether a nuclear-protein hashing reagent works (proposed, not tested).
• How to detect within-sample doublets (“remains an important challenge for the field”).
• The 216 discordant-donor singlets are only partly adjudicated: donor H gave an arbiter for 40 of them; the rest are not resolved.

12. Guidance answers

The method

• What is attached, against what, at what step: oligo-tagged (HTO) monoclonal antibodies against ubiquitous surface proteins — 1st gen CD45/CD98/CD44/CD11a; 2nd-gen “universal” pool B2M + CD298. Each HTO carries a unique 12-bp barcode. Staining is done per sample, before pooling (standard CITE-seq staining), then samples are pooled and loaded.
• Extra library + cost: a separate HTO library (distinct amplification handle from CITE-seq ADTs; amplified with SI-PCR + D7xx_s), separated from cDNA by 0.6X SPRI and cleaned with 2× 2.0X SPRI, sequenced alongside the transcriptome (the run targeted 90:5:5 transcriptome:HTO:ADT). Stated cost: “approximately 2–5% of transcriptome sequencing costs.”
• Classification procedure end to end: see §5.6. Normalization = centered log ratio (CLR) (divide by the geometric mean of the HTO across cells, natural log). Clustering = k-medoids on normalized HTO data with k = 9. Background distribution = negative binomial, fit per HTO after excluding (a) the k-medoids cluster with the highest average expression of that HTO and (b) the highest 0.5% of the remaining values. Cutoff = the q = 0.99 quantile of that fitted distribution, HTO-specific. Threshold quoted verbatim in §7.1.
• Singlet / multiplet / negative: verbatim in §7 and §5.6 step 6 — positive for exactly one HTO = singlet; positive for ≥ 2 HTOs = multiplet (sample IDs from the two most highly expressed HTOs); negative for all 8 = negative. Only barcodes with ≥ 200 RNA UMI were classified (later relaxed to 50–200 UMI for the low-quality-cell rescue).
• Super-loading + doublet rates: loading at significantly higher concentration than standard, here targeting 20,000 singlets and 5000 multiplets in one lane. With multiplexing: 4365 of the 5000 multiplets are cross-sample and discardable → 3.1% unresolved multiplet rate while keeping ~20,000 cells. Without multiplexing, that same 3.1% rate would be achieved only at ~4000 singlets → multiplexing gives ~400% more cells for the same cost. Observed in practice: 14,002 singlets, 2974 identifiable multiplets after removing control cells.
• Software / versions / artifacts: Drop-seq tools v1.0 → hg19-mm10 aligned BAM + DGE (top 50,000 barcodes); CITE-seq-Count (Hoohm/CITE-seq-Count, default params, max hamming distance 1) → HTO/ADT count matrices; PLINK v1.07 → VCF (GT) from the Infinium CoreExome 24 array; demuxlet (VCF + tagged BAM → *.best with BEST ∈ {SNG, DBL, AMB} and PRB.DBL); Seurat v3.0 for normalization, classification, clustering; ranger for the random-forest cell-type classifier. Data: GSE108313.

The comparison — provenance

• Yes, there is a head-to-head comparison (“Genotype-based demultiplexing validates Cell Hashing”).

• Design of that comparison — who owns what:

  • The data are the hashing authors’ own. One 10x lane, their 8 PBMC donors, their pooling design, their super-loading, their sequencing (~24,115 reads/cell).
  • The genotypes are the hashing authors’ own: they genotyped all 8 PBMC samples and HEK293T on the Illumina Infinium CoreExome 24 array, and built the VCF with PLINK v1.07.
  • The hashing authors ran the demuxlet arm themselves, with default parameters, on their own VCF + their own Drop-seq-tagged BAM. The demuxlet authors (Kang et al.) are not authors here and there is no indication they were involved.
  • Both arms were therefore executed and interpreted by the group that owns the hashing method. Two of them have a patent application on this work, and a co-author is a BioLegend employee (BioLegend being the exclusive licensee). Every comparative claim below is the hashing authors’ assessment of the other method, not a neutral third-party benchmark.
  • Directionality of the framing is visible in the section title itself: demuxlet is cast as the validator of Cell Hashing.

• Where they disagreed, and the authors’ explanation:

  • 871 HTO-singlets called AMB by demuxlet → authors attribute this to their own shallow sequencing depth: ~24,115 reads/cell, which they state is below the recommended depth for using demuxlet (the recommended depth is not stated). Supporting evidence they give: discordant barcodes had identical HTO classification strength but lower RNA UMI, and (Additional file 1: Fig. S2A-D) lower read and SNP counts. This is the hashing authors explaining away a failure of the arm they ran, on data whose depth they chose.
  • 2528 discordant singlet/doublet calls (vs. 13,421 concordant) → attributed to conservative false positives from both methods, “perhaps due to ambient RNA or HTO signal.” They note it “is difficult to be certain which of these methods is correct.” Filtering demuxlet to > 95% doublet posterior removed 75% of the discordance.
  • 216/11,464 (1.9%) discordant donor calls among mutual singlets → adjudicated using the unstained donor H: of 40 cells demuxlet (but not hashing) called donor H, 37 had > 1000 ADT counts, so the authors call these demuxlet misclassifications, “in line with demuxlet’s estimated error rate of 1–2%” (attributed to Kang et al. [13]).

• Both columns of the trade-off, in the source’s words: the Discussion paragraph is reproduced verbatim at §7.2, and the Introduction’s statement of the genotype requirement at §7.3. Summarized:

  • Genetic multiplexing advantage: no exogenous barcodes; no alteration to existing workflows before or after pooling.
  • Genetic multiplexing requirement: pooled samples must come from previously genotyped individuals (and, from the Results, sufficient reads and SNPs / adequate depth).
  • Cell Hashing requirement: incubation with antibodies against ubiquitously expressed surface proteins (i.e., a wet-lab change) — and, from the Discussion caution, the cells must express those epitopes.
  • Cell Hashing advantage: can multiplex samples with the same genotype; can split a single sample into arbitrarily many pools; gives a transcriptome-independent identity signal usable to rescue low-RNA cells.
  • Cost, both: both slightly raise sequencing cost — genetic via increased depth or read length to identify SNPs; hashing via the HTO library (~2–5% of transcriptome sequencing costs).

• DOES THE SOURCE STATE A RULE FOR WHEN TO USE WHICH ROUTE? No — it states no decision rule. It calls the approaches “complementary … with each having unique advantages” and says “We believe that researchers will benefit from both approaches.” The closest thing to guidance is a one-sided use-case preference for its own method, verbatim:

  “In particular, we envision that our method will be most useful when processing genetically identical samples subjected to diverse perturbations (or experimental conditions/optimizations, as in our titration experiment) or to reduce the doublet rate when running cells from a single sample.”

  That is a statement of where hashing is most useful (a conditional favoring hashing, phrased as “we envision”), not a rule for when to prefer genotype-based demultiplexing, and not a decision procedure. The genotype route’s requirement (pre-existing genotypes; adequate depth/SNPs) and advantage (no workflow change) are stated, but the source never composes them into an if/then rule. Plainly: the paper offers complementarity plus a preference for its own method’s niche; a selection rule must be constructed by the reader, not quoted from this source.

Limitations / failure modes

• Cells/samples where hashing cannot work: cell types not expressing the (virtually ubiquitous) target surface proteins — verbatim quote at §7.5. Nuclei are not covered by the presented reagents. Demonstrations are human; species generalization is future work.

• Ambient/background HTO signal: named as a possible cause of conservative false-positive doublet calls (the 2528 discordant calls). Stated mitigations: (a) the classification procedure itself excludes the top-expressing cluster and the highest 0.5% of values before fitting the negative binomial background; (b) unlabeled negative-control cells (HEK293T, NIH-3T3) are spiked in; (c) a dedicated four-cell-line experiment was run specifically “to further ensure that background binding levels did not lead to incorrectly demultiplexed samples,” yielding 99.7% concordance with transcriptomic identity. Note: the > 95%-posterior filter that removes 75% of discordance is a demuxlet-side knob, not a hashing-side one — the source names no tunable HTO-side stringency parameter beyond the fixed q = 0.99 quantile.

• Doublets the method CANNOT detect (verbatim):

  “Informatic detection of multiplets based on transcriptomic data also remains an important challenge for the field, for example, to identify doublets originating from two cells within the same sample.”

  Quantified for this design as the unresolved multiplet rate of 3.1%; also, doublets formed with an unlabeled cell are called singlets (Methods).

• Batch/technical effects reduced by pooling — exact wording and what backs it. The two relevant statements are:

  (Abstract) “Despite rapid developments in single cell sequencing, sample-specific batch effects, detection of cell multiplets, and experimental costs remain outstanding challenges.”

  (Discussion) “Still, as sequencing costs continue to drop, and experimental designs seek to minimize technology-driven batch effects, multiplexing should facilitate the generation of large scRNA-seq and CITE-seq datasets.”

  and, in the Introduction, the motivating premise:

  “Similarly, technical and “batch” effects have been demonstrated to mask biological signal in the integrated analysis of scRNA-seq experiments [12], necessitating experimental solutions to mitigate these challenges.”

  Evidence backing it in this paper: none. The batch-effect framing is motivational and is supported only by citations to prior work (refs. 11, 12; ref. 14 Tung et al. on batch effects and design). The paper runs no experiment comparing batch effects with and without pooling, reports no batch-effect metric, and the Discussion’s phrasing is prospective (“should facilitate”). Per the guidance’s instruction not to accept a bare claim as an evidenced one: treat “hashing reduces batch effects” as an unevidenced design argument in this source. (The one adjacent measured result is negative-control-shaped, not batch-shaped: hashing does not degrade transcriptomic complexity vs. non-hashed cells, Additional file 1: Fig. S5.)

Pin

• Journal / volume / article: Genome Biology, volume 19, article 224.
• DOI: 10.1186/s13059-018-1603-1. PMID 30567574. PMCID PMC6300015.
• Dates: Received 2018-06-05; Accepted 2018-12-04; Published 2018-12-19; collection date 2018.
• License text: reproduced verbatim in §1 (CC BY 4.0; CC0 waiver applies to the data unless otherwise stated).
• Access date: 2026-07-13 (read from PMC).
• Data accession: GEO GSE108313. No SRA accession is given in the article text. Code: Seurat (github.com/satijalab/seurat); CITE-seq-Count (github.com/Hoohm/CITE-seq-Count); protocols at www.cite-seq.com.