Single-Cell and Single-Nucleus RNA Sequencing: Sample Preparation, Workflow, and Experimental Design for Frozen and Fresh Tissues

scRNA-Seq vs snRNA-Seq — Choosing the Right Approach

The choice between single-cell and single-nucleus RNA sequencing is the first fork in the sample preparation road, and it is dictated primarily by sample type — not by which method produces "better" data.

scRNA-seq captures cytoplasmic mRNA from intact, viable whole cells. It detects more genes per cell, recovers full-length mature transcripts, and provides higher overall transcriptomic sensitivity. A 2025 head-to-head comparison of 11 matched bone marrow samples across four independent laboratories confirmed that 10x Genomics-based droplet scRNA-seq consistently yields higher library complexity, greater gene detection, and finer clustering resolution than snRNA-seq from the same donors (1). The trade-off is that scRNA-seq requires fresh or freshly cryopreserved tissue. Enzymatic dissociation at 37°C — still the standard for most solid tissues — induces an immediate transcriptional stress response: immediate-early genes such as FOS and JUN are upregulated within minutes, and fragile cell types such as hepatocytes, neurons, and adipocytes can be selectively lost during processing. These dissociation artifacts become part of the dataset before the first cell enters a droplet.

snRNA-seq captures nuclear transcripts — predominantly unspliced pre-mRNA — from isolated nuclei. It detects fewer genes per nucleus and carries an inherently higher proportion of intronic reads, but it avoids the dissociation-induced transcriptional stress that plagues whole-cell methods. Nuclei can be isolated from frozen tissue, archived specimens, and tissues that resist enzymatic dissociation: adult brain, cardiac muscle, adipose, and fibrotic tumor samples. When intronic reads are included during preprocessing, the concordance between scRNA-seq and single-nucleus RNA sequencing improves substantially, and a gene-length-based correction — repurposed from bulk RNA-seq gene set enrichment analysis — can partially mitigate the systematic length bias between the two methods.

snRNA-seq also enables paired epigenomic readout. Because nuclei are isolated with intact chromatin, the 10x Genomics Single-Cell Multiome ATAC plus Gene Expression assay simultaneously measures chromatin accessibility and the transcriptome from the same nucleus. This is not possible with whole-cell scRNA-seq, where cytoplasmic membrane integrity must be maintained and the transposase cannot access nuclear chromatin. Conversely, scRNA-seq is the prerequisite for CITE-seq — antibody-based protein detection — because oligonucleotide-conjugated antibodies bind surface proteins on intact cell membranes.

The practical decision framework reduces to sample state. When fresh tissue is available and can be processed within hours of collection, scRNA-seq provides the highest-sensitivity transcriptomic readout. When tissue is frozen, archived, or inherently difficult to dissociate into viable single cells, snRNA-seq is the appropriate choice — not a compromise, but the method that matches the sample reality. A growing body of evidence supports a complementary strategy: profile a subset of fresh samples with scRNA-seq for maximum transcriptomic depth, and use snRNA-seq for the larger cohort — including frozen archival specimens — to maximize sample numbers while controlling for gene-length bias during integrated analysis.

Figure 1: Decision framework for selecting between scRNA-seq and snRNA-seq based on sample type, research goals, and platform compatibility. Fresh tissue processed within hours of collection yields the highest transcriptomic sensitivity with scRNA-seq and is compatible with CITE-seq for surface protein detection, while frozen or archived specimens are routed to snRNA-seq which avoids dissociation-induced transcriptional stress, preserves chromatin for paired ATAC-seq readout via the 10x Multiome assay, and captures predominantly unspliced pre-mRNA that requires intronic read inclusion for optimal gene detection.

Sample Dissociation — The Make-or-Break Step

If the sc-versus-sn decision is the first fork, the dissociation protocol is where the road surface matters. A poorly optimized dissociation protocol does not just reduce cell yield — it selectively eliminates certain populations and induces transcriptional programs that masquerade as biological signal.

Fresh Tissue: Enzymatic Dissociation

The standard approach uses tissue-specific enzyme cocktails — collagenase, trypsin, dispase, or papain — at 37°C with gentle mechanical trituration. Collagenase-based protocols work well for most epithelial and mesenchymal tissues. Trypsin is more aggressive and can strip surface proteins, making it incompatible with CITE-seq unless carefully titrated. Papain is preferred for neural tissue. The critical variable is incubation time: under-digestion leaves clumps that clog microfluidic channels; over-digestion reduces viability and enriches for dissociation-resistant cell types.

An alternative that has gained substantial traction is cold-active protease dissociation at 6°C. By performing enzymatic digestion at low temperature, metabolic activity and stress-response gene induction are minimized — FOS and JUN expression, which spike within minutes of 37°C processing, remain near baseline. The trade-off is somewhat lower cell yield for highly fibrotic or matrix-dense tissues.

Fresh-Frozen Tissue: The ACME HS Breakthrough

The most significant recent advance in sample preparation is the ACME HS (ACetic acid-MEthanol High Salt) method, published in early 2025 and validated across 41 human endocrine tissue samples (2). This protocol simultaneously dissociates and fixes intact single cells directly from fresh-frozen tissue — something previously considered infeasible. The key innovation is replacing standard PBS wash buffer with 3× saline-sodium citrate (SSC) at high ionic strength. The high-salt environment stabilizes RNA secondary structure, prevents RNase reactivation during tissue rehydration, and alters protein solubility to impede ribonuclease function. Methanol and acetic acid provide simultaneous fixation, while N-acetyl cysteine reduces mucus viscosity and oxidative damage. Dissociated cells can be cryopreserved in DMSO after processing and stored for at least 28 days without loss of mRNA integrity.

The significance of ACME HS for study design is practical: it means biobanked frozen specimens — the vast majority of archived clinical research samples — can now be analyzed by scRNA-seq rather than being restricted to snRNA-seq. For researchers designing prospective studies, it also means tissue can be frozen at the collection site and processed in batches later, removing the logistical constraint of same-day dissociation.

Nuclei Isolation for snRNA-seq

When whole-cell dissociation fails — or when chromatin accessibility data is needed — nuclei isolation for snRNA-seq is the fallback. The standard protocol uses detergent-based lysis (NP-40 or Triton X-100) in a Dounce homogenizer, followed by density gradient centrifugation to separate nuclei from debris. A 2026 optimized protocol for brain tissue emphasizes minimizing centrifugation speeds — excessive g-force shears nuclear membranes — and including RNase inhibitors at every step (3). For frozen tissue, nuclei isolation can be performed directly on pulverized frozen powder without prior thawing, preserving RNA integrity.

Common Pitfalls Across All Methods

Three mistakes recur: using RBC lysis buffers on precious samples (immunomagnetic depletion preserves more cells); failing to pre-wet filters (dry filters trap cells); and processing too many samples in parallel (dissociation time drifts across the batch). For any unfamiliar tissue type, a pilot dissociation with cell counting and viability assessment before committing 10x reagents pays for itself many times over.

Figure 2: Comparison of four tissue dissociation strategies for single-cell and single-nucleus RNA sequencing. Enzymatic dissociation at 37°C (left) offers high cell yield but induces stress-gene activation — FOS and JUN upregulation begins within minutes — and selectively loses fragile cell types including hepatocytes and neurons. Cold-active protease digestion at 6°C minimizes transcriptional artifacts but reduces yield from fibrotic and matrix-dense tissues. The ACME HS method (center) enables direct scRNA-seq from frozen tissue by replacing PBS with 3× SSC at high ionic strength, which stabilizes RNA secondary structure and prevents RNase reactivation, while methanol-acetic acid simultaneously fixes intact cells. Detergent-based nuclei isolation (right) provides the most robust option for tissues that resist whole-cell dissociation and preserves chromatin for single-cell ATAC-seq co-assay.

10x Genomics Workflow from Sample to Library

Once a viable single-cell or single-nucleus suspension is obtained, the 10x Genomics Chromium workflow proceeds through a standardized series of steps that have been refined across successive chemistry generations.

The prepared suspension is loaded onto a GEM-X microfluidic chip, where cells are co-encapsulated with barcoded gel beads in oil droplets — gel beads in emulsion, or GEMs. The current GEM-X v4 chemistry, running on the Chromium X platform, generates approximately 40-micrometer droplets that are more uniform than previous generations, reducing ambient RNA background and improving signal-to-noise separation. Within each GEM, cells are lysed and polyadenylated mRNA is captured by oligo-dT primers on the bead surface. Reverse transcription incorporates a cell-specific 10x barcode and a random unique molecular identifier (UMI) into each cDNA molecule, enabling both cell-of-origin identification and transcript-level counting. The entire encapsulation and barcoding process completes in roughly 4 minutes per chip.

After droplet breaking, the pooled barcoded cDNA is amplified by PCR — typically 11 to 14 cycles depending on input cell number — and then undergoes enzymatic fragmentation, end repair, A-tailing, and adapter ligation to produce Illumina-compatible sequencing libraries. The standard read configuration is 28 bp for Read 1 (cell barcode + UMI) and 90 to 150 bp for Read 2 (transcript insert). A typical project targeting 5,000 to 10,000 cells allocates 20,000 to 50,000 read pairs per cell, producing 250 to 500 million read pairs total. For researchers who need dedicated bioinformatics support for single-cell data, the analysis pipeline typically includes read alignment, UMI counting, quality filtering, clustering, and cell-type annotation.

For projects where fresh tissue processing is logistically impossible, the 10x Flex fixed-RNA workflow uses probe-based chemistry that hybridizes to target transcripts in fixed, permeabilized cells. This is a fundamentally different approach than poly(dT) capture — it does not require intact poly(A) tails or viable cells, making it suitable for FFPE specimens, but it targets a predefined gene panel rather than capturing the whole transcriptome. The choice between Flex and nuclei-based snRNA-seq for fixed specimens depends on whether whole-transcriptome discovery (snRNA-seq) or targeted gene expression from degraded RNA (Flex) better serves the research question.

For sample types where intact cells are consistently unattainable — frozen archived specimens, lipid-rich brain tissue, fibrotic tumors — single-nucleus RNA sequencing via nuclei isolation is the standard alternative. The isolated nuclei are loaded onto the same GEM-X chip and processed through the same library construction steps, with lysis conditions adjusted for the nuclear membrane rather than the plasma membrane.

Figure 3: The 10x Genomics GEM-X v4 workflow from sample to sequencing-ready library, accommodating four input paths — fresh single-cell suspension, ACME HS-dissociated frozen cells, isolated nuclei, and FFPE tissue sections via the Flex fixed-RNA chemistry. All paths converge on the GEM-X microfluidic chip where individual cells or nuclei are co-encapsulated with barcoded gel beads in ~40 μm oil droplets. Within each GEM, polyadenylated mRNA is captured by oligo-dT primers and a cell-specific 10x barcode plus UMI are incorporated during reverse transcription. After droplet breaking, pooled barcoded cDNA undergoes PCR amplification (11–14 cycles), enzymatic fragmentation, end repair, A-tailing, and adapter ligation to produce Illumina-compatible libraries with the standard 28 bp (Read 1) + 90–150 bp (Read 2) configuration. The Flex path (bottom) diverges by using probe-based hybridization rather than poly(dT) capture, targeting a predefined gene panel and enabling compatibility with degraded RNA from FFPE specimens at the cost of whole-transcriptome coverage.

Non-Model Organism Considerations

The dominance of human and mouse reference resources in single-cell genomics creates a genuine barrier for researchers working with non-model species, but it is not a methodological barrier — it is a computational and protocol-adaptation challenge.

The core sample preparation principles transfer across species: fresh tissue requires species-appropriate dissociation, frozen tissue follows the same ACME HS or nuclei isolation paths, and the 10x poly(dT) capture chemistry works universally — any organism with a polyadenylated transcriptome is compatible with the GEM-X platform, no species-specific reagents required.

Where species differences matter most is in tissue architecture and dissociation response. Zebrafish embryos require dechorionation before dissociation, and enzymatic protocols must be adjusted for developmental stage — blastula-stage embryos dissociate rapidly with mild trypsin treatment, while adult zebrafish brain requires papain-based protocols adapted from mouse brain methods. Plant tissues present a more fundamental challenge because the cell wall must be removed to generate protoplasts before single-cell encapsulation can proceed. Protoplast isolation uses cellulase and pectinase digestion in an osmotically stabilized buffer; the osmotic stabilizer concentration — typically 0.4 to 0.6 M mannitol or sorbitol — must be empirically optimized for each species to balance cell wall removal against protoplast bursting. Rice, poplar, and tomato have established protoplast protocols; for less-studied plant species, a pilot gradient of enzyme concentrations and incubation times is a necessary investment.

The computational side of non-model organism scRNA-seq is equally important. If a high-quality reference genome or transcriptome is not available, reads cannot be reliably aligned and genes cannot be quantified. For species without an annotated reference, a de novo transcriptome assembly from bulk RNA-Seq of the same tissue type serves as a custom alignment reference. This adds a preparatory step before the single-cell experiment begins but enables single-cell resolution in species where it would otherwise be impossible. For labs working with non-standard species, custom bioinformatics support can handle the reference assembly and annotation pipeline.

Figure 4: Sample preparation adaptations for non-model organisms in single-cell RNA sequencing. Top track — zebrafish: embryos require dechorionation before dissociation, with enzymatic protocol adjusted by developmental stage (mild trypsin for blastula-stage, papain-based protocols adapted from mouse brain methods for adult brain). Middle track — plant protoplasts: the cell wall must be removed via cellulase and pectinase digestion in osmotically stabilized buffer, with mannitol or sorbitol concentration (0.4–0.6 M) empirically optimized per species to balance cell wall removal against protoplast bursting; rice, poplar, and tomato have established protocols while less-studied species require pilot enzyme gradients. Bottom track — custom reference assembly: for species without an annotated reference genome, de novo transcriptome assembly from matched bulk RNA-seq of the same tissue type serves as an alignment reference, enabling single-cell resolution in any organism with a polyadenylated transcriptome.

Quality Control and Data Expectations

Quality control in single-cell experiments spans two distinct phases — wet-lab QC before the cells enter the microfluidic chip, and computational QC after sequencing — and decisions made in the first phase constrain what is possible in the second.

Wet-Lab QC: The Go/No-Go Checkpoint

Before loading a sample onto the 10x chip, three parameters determine success probability. Viability must exceed 80 percent for cultured cells and 70 percent for primary cells — lower viability means more ambient RNA from lysed cells, which contaminates the background and reduces effective sequencing depth. The cell-to-debris ratio should be at least roughly one-to-one; excessive debris competes for droplet occupancy and can be mistaken for cells during encapsulation. Clumps must be rare — a clump-to-single-cell ratio below roughly 0.05-to-one — because a clump entering a droplet produces an artifactual "cell" with a hybrid transcriptome from multiple real cells. Filtration through a 40-micrometer strainer immediately before loading removes most clumps, and DNase treatment during dissociation reduces DNA-mediated clumping from lysed nuclei.

Computational QC: From Counts to Clean Data

After sequencing, three per-cell metrics dominate the QC landscape. The number of unique genes detected per cell separates real cells from empty droplets and debris — typical thresholds range from 200 to 500 genes as a lower bound, but the 2025–2026 consensus has moved toward adaptive, median-absolute-deviation-based thresholds rather than fixed cutoffs. The total UMI count per cell reflects library complexity; cells with abnormally low counts may be damaged, and cells with abnormally high counts may be doublets. The mitochondrial read percentage reports cell stress — cytoplasmic RNA leaks from damaged or dying cells, and mitochondrial transcripts, being physically located in the cytoplasm, become overrepresented. For scRNA-seq, a mitochondrial percentage above 15 to 20 percent is typically a filtering criterion, though the exact threshold varies by tissue; metabolically active tissues such as heart and kidney naturally carry higher mitochondrial transcript fractions. For snRNA-seq, mitochondrial reads are substantially lower because nuclei contain minimal cytoplasm, and a threshold of 5 percent is more appropriate. The scRNA-seq analysis best-practices framework provides detailed guidance on parameter selection and iterative QC workflows, emphasizing that quality filtering should be visualized and justified rather than applied with default parameters (4).

Doublet Detection

The expected doublet rate for 10x Chromium is approximately 0.9 percent per 1,000 recovered cells, meaning a 10,000-cell dataset carries roughly an 8 to 9 percent doublet burden. DoubletFinder and scDblFinder — the two most widely adopted tools — simulate artificial doublets from the real data and score each cell barcode by its similarity to the simulated doublet profile. A newer tool, scUmaper (2026), adds a rule-based layer that flags cells co-expressing lineage-incompatible markers — for example, CD3E (T cells) and MS4A1 (B cells) — improving detection of heterotypic doublets that simulation-only approaches sometimes miss. Homotypic doublets — two cells of the same type — remain difficult to detect and are primarily identified by anomalously high UMI counts. For chromatin accessibility data, scIBD provides a self-supervised framework specifically designed for single-cell ATAC-seq doublet detection, addressing the unique error modes of epigenomic data that transcriptome-focused tools may miss (6).

Ambient RNA

RNA from lysed cells diffuses through the single-cell suspension and is captured in droplets alongside intact cells, creating a background signal that confounds differential expression analysis. CellBender (7), a deep-learning-based method, has become the standard ambient RNA removal tool in 2025–2026 workflows. Applied before doublet detection and normalization, it subtracts the estimated ambient profile from each cell's expression vector, improving the signal-to-noise ratio for lowly expressed genes and rare cell types. For projects requiring rigorous single-cell RNA sequencing data analysis, ambient RNA removal, doublet detection, and adaptive thresholding should be standardized across all samples in a study. Beyond these preprocessing steps, experimental design itself is a QC concern: pseudoreplication — treating cells from the same donor as independent observations — remains the single largest source of false positives in single-cell differential expression, and study designs must account for donor-level replication from the outset (5).

Figure 5: Quality control pipeline for single-cell RNA sequencing data, spanning wet-lab go/no-go checkpoints (viability >80% cultured cells, >70% primary cells; debris ratio ≥1:1; clump-to-single-cell ratio <0.05) through computational filtering. The 2025–2026 consensus has shifted from hard cutoffs to adaptive, median-absolute-deviation-based thresholds for genes detected per cell and UMI counts. Mitochondrial read percentage flags damaged cells (>15–20% for scRNA-seq; >5% for snRNA-seq due to minimal cytoplasmic carryover). Doublet detection proceeds via DoubletFinder/scDblFinder simulation-based scoring, supplemented by scUmaper for rule-based identification of heterotypic doublets co-expressing lineage-incompatible markers (e.g., CD3E and MS4A1). Ambient RNA removal by CellBender is applied before doublet detection and normalization to subtract the estimated background profile from each cell's expression vector, improving signal-to-noise for lowly expressed genes and rare cell types.

Integrating Sample Preparation with the Broader Experimental Design

Sample preparation decisions cascade into every downstream analysis choice. A tissue dissociation protocol that enriches for certain cell types — trypsin-sensitive cells lost, collagenase-resistant cells retained — produces a dataset whose cell-type proportions do not reflect the original tissue. A project that uses scRNA-seq for fresh samples and snRNA-seq for frozen archival samples must account for gene-length bias when integrating data across the two methods. A QC threshold set too stringently removes rare cell types that resemble damaged cells in their metric profile; a threshold set too leniently retains artifacts that distort clustering.

Consider a concrete example from tissue dissociation. In a multi-center liver atlas project, enzymatic dissociation at 37°C using collagenase-based protocols selectively depleted hepatocytes — the most metabolically active parenchymal cells — while enriching for dissociation-resistant hepatic stellate cells and Kupffer cells. When the same tissue blocks were processed by nuclei isolation for snRNA-seq, hepatocyte nuclei were recovered at proportions matching histological expectations. The discrepancy was not a biological finding; it was a sample preparation artifact introduced before the first droplet was formed. This is why pilot experiments comparing dissociation methods are not optional for any tissue type without an established protocol — and why single-cell studies must report dissociation conditions in their methods sections as rigorously as sequencing parameters.

These sample-prep-level decisions interact directly with the experimental design variables covered in our companion guide on Experimental Design for Single-Cell Studies: cell number, biological replicates, and sequencing depth trade-offs. A well-designed experiment with three to five biological replicates per condition provides the statistical power to detect true biological differences — but only if the sample preparation protocol is consistent across all replicates, and only if QC thresholds are applied uniformly rather than optimized per sample.

For projects combining transcriptomics with epigenomic or proteomic readouts, the sample preparation strategy must accommodate all modalities from the outset. Single-cell ATAC-seq requires nuclei isolation and cannot be performed on whole-cell suspensions optimized for scRNA-seq. Similarly, spatial transcriptomics requires intact tissue sections rather than dissociated cells, meaning the sample preparation workflow diverges at the very first step — a decision that must be made before tissue collection, not after.

For a broader view of how single-cell and single-nucleus transcriptomics fit into the larger single-cell and spatial biology landscape — including spatial transcriptomics and multi-omics integration — see our hub guide on Single-Cell and Spatial Biology Services.

FAQ

When should I choose scRNA-seq over snRNA-seq?

Choose scRNA-seq when you have fresh, viable tissue and need maximum transcriptomic sensitivity — detecting subtle cell states, lowly expressed genes, or fine clustering resolution. scRNA-seq is also required for CITE-seq protein detection. If your tissue is frozen, archived, or inherently difficult to dissociate (brain, heart, adipose), snRNA-seq is the appropriate choice.

Can I use frozen tissue for scRNA-seq, or am I limited to snRNA-seq?

The ACME HS protocol (2025) now enables scRNA-seq directly from fresh-frozen tissue by simultaneously dissociating and fixing intact cells in a high-salt buffer. This means frozen biobank specimens are no longer restricted to snRNA-seq. If ACME HS is not available at your facility, snRNA-seq remains the standard for frozen tissue.

What viability should my sample have before loading onto the 10x chip?

Aim for viability above 80 percent for cultured cells and above 70 percent for primary tissue-derived cells. If viability falls below these thresholds, consider live-cell enrichment by FACS or magnetic bead depletion of dead cells before loading. Below roughly 50 percent viability, ambient RNA from lysed cells dominates the background, and the dataset quality degrades sharply.

How many cells should I target for a typical scRNA-seq project?

For cell-type discovery in a heterogeneous tissue, targeting 5,000 to 10,000 cells per sample provides robust identification of populations present above roughly 1 percent frequency. For detailed guidance on cell number, replication, and budget planning, see our companion guide on Experimental Design for Single-Cell Studies.

Can CD Genomics process non-model organisms?

Yes. The 10x poly(dT) capture chemistry works with any species possessing a polyadenylated transcriptome. We have processed samples from zebrafish, axolotl, rice, poplar, tomato, pig, cow, and other non-standard species. For species without an annotated reference genome, we perform de novo transcriptome assembly from matched bulk RNA-seq as a preparatory step.

What sequencing depth do you recommend?

For cell-type classification and clustering, 20,000 to 25,000 read pairs per cell. For differential expression analysis within a cell type, 50,000 read pairs per cell. For rare transcript detection, 50,000 to 100,000 or more may be required. Sequencing saturation can be checked from a pilot sample before committing the full cohort.

How do I know if my dissociation protocol is working before committing to sequencing?

Run a pilot dissociation on a small portion of your tissue. Count cells, measure viability by trypan blue or fluorescent staining, and check for clumps under the microscope. If viability and cell numbers are in range, load a small number of cells onto a 10x chip and sequence shallowly (5,000 to 10,000 reads per cell) to check for expected cell-type recovery, mitochondrial fraction, and doublet rate before scaling to the full experiment.

What is the turnaround time for a single-cell project?

A standard project of 4 to 8 samples typically requires 3 to 4 weeks from sample receipt to sequencing completion, plus 2 to 4 weeks for full bioinformatic analysis including QC, clustering, cell-type annotation, and differential expression.

References:

1. Ghamsari L, et al. Comparative analysis of single-nucleus and single-cell RNA sequencing in human bone marrow mononuclear cells: methodological insights and trade-offs. bioRxiv. 2025. https://doi.org/10.1101/2025.09.08.675012
2. Utkina M, et al. Comparative evaluation of ACetic-MEthanol high salt dissociation approach for single-cell transcriptomics of frozen human tissues. Frontiers in Cell and Developmental Biology. 2025;12:1469955. https://doi.org/10.3389/fcell.2024.1469955
3. Zhang Y, et al. Optimized nuclei isolation and snRNA-seq reveal oligodendrocyte pathway dysregulation in MOGHE brain tissue from pediatric patients. Scientific Reports. 2026. https://doi.org/10.1038/s41598-026-54112-z
4. Luecken MD, Theis FJ. Current best practices in single-cell RNA-seq analysis: a tutorial. Molecular Systems Biology. 2019;15(6):e8746. https://doi.org/10.15252/msb.20188746
5. Squair JW, et al. Confronting false discoveries in single-cell differential expression. Nature Communications. 2021;12:5692. https://doi.org/10.1038/s41467-021-25960-2
6. Zhang R, et al. scIBD: a self-supervised iterative-optimizing method for heterotypic doublet detection in single-cell chromatin accessibility data. Genome Biology. 2023;24:250. https://doi.org/10.1186/s13059-023-03072-y
7. Fleming SJ, et al. Unsupervised removal of systematic background noise from droplet-based single-cell experiments using CellBender. Nature Methods. 2023;20:1323-1331. https://doi.org/10.1038/s41592-023-01943-7

For research purposes only, not intended for clinical diagnosis, treatment, or individual health assessments.