Single-Cell RNA Sequencing · scRNA-seq Analysis

Resolve cell populations. Map transcriptional diversity.

Single-cell RNA sequencing data analysis to resolve transcriptomic diversity and cell population structure at single-cell resolution — essential for immunology, oncology, developmental biology, and tissue-heterogeneity studies.

24
Clusters resolved (demo atlas)
25+
Cell types annotated
Cell Ranger
+ Seurat pipeline
SingleR
Reference-based annotation

Single-cell resolution for tissue heterogeneity

Where bulk RNA-seq reports an average signal across a tissue, scRNA-seq lets you see which cell types are present, how they differ transcriptionally, and which genes define each population.

Our pipeline is built on Cell Ranger for primary processing and Seurat for downstream analysis, covering everything from raw FASTQ files through quality control, normalization, clustering, cell-type annotation, and differential expression — including fine-grained sub-clustering of specific lineages when biologically relevant.

  • Primary processing of raw scRNA-seq FASTQ files with Cell Ranger
  • Quality control and filtering of low-quality cells and genes
  • Normalization and highly variable gene (HVG) identification
  • Dimensionality reduction (PCA) and graph-based clustering (Louvain)
  • UMAP and t-SNE visualization of transcriptional landscapes
  • Marker gene identification and cell-type annotation (markers + SingleR)
  • Differential gene expression between clusters
  • Fine-grained sub-clustering of specific lineages (e.g., T/NK, myeloid)

Cell Ranger + Seurat

Industry-standard primary processing with Cell Ranger Count (10x Genomics), followed by Seurat-based QC, clustering, annotation and differential expression.

Cell Ranger Seurat SingleR Louvain

Layered Analysis Strategy

Broad cluster-level annotation followed by targeted sub-clustering of lineages of interest — substantially increasing the biological resolution of a single scRNA-seq run.

From first principles to expert-level analysis

New to single-cell work? Start with Foundations. Planning a study or reviewing results? Jump to Practitioner or Expert. Each level builds on the one before it, so you can read this page as a short course or dip in where you need to.

The big idea in plain language: a tissue is a mixture of many cell types. Single-cell RNA sequencing reads the active genes of each cell separately, so you can see who is in the mixture and what each cell is doing.

LEVEL 1

Bulk gives an average, single-cell gives a crowd

Bulk RNA-seq blends every cell into one number per gene. Single-cell sequencing keeps one profile per cell, like counting the people in a room instead of measuring its average height.

LEVEL 1

Each cell gets a barcode

In a droplet system, every cell is captured with a bead carrying its own barcode. All RNA from that cell is tagged with it, so reads can be sorted back to their cell after sequencing.

LEVEL 1

The result is a big table of counts

Rows are genes, columns are cells, and each entry is how many RNA molecules were seen. Most entries are zero, because each cell is sampled shallowly.

LEVEL 1

Quality control removes bad cells

Broken or dying cells, empty droplets and two cells captured together look wrong in the data. They are filtered out before any biology is read.

LEVEL 1

Similar cells are grouped

Cells with similar gene activity are placed near each other on a map (UMAP) and grouped into clusters. Each cluster is a candidate cell type or state.

LEVEL 1

Marker genes give clusters a name

Genes that are high in one cluster and low in others, such as CD3E in T cells, are used to label clusters. A reference tool such as SingleR provides a second opinion.

For scientists planning a project or reading a report: what each pipeline step does and how it can go wrong.

LEVEL 2

UMIs count molecules, not reads

Each captured RNA molecule gets a random tag (UMI). Reads sharing a cell barcode, gene and UMI are copies of one molecule from amplification, so they are counted once.

LEVEL 2

Three QC numbers matter most

UMIs per cell, genes per cell and the share of mitochondrial reads. Low genes with high mitochondrial share suggests a damaged cell; unusually high UMIs and genes suggests a doublet (Ilicic et al., 2016).

LEVEL 2

Normalization removes depth differences

Cells are sequenced to different depths by chance. Log-normalization rescales each cell to a common total so cells can be compared. SCTransform does this with a regression model.

LEVEL 2

Use the informative genes

Thousands of genes barely change between cells. Highly variable genes (HVGs) carry the cell-type signal, so clustering uses only them, which cuts noise and computing time.

LEVEL 2

PCA, then a neighbour graph

PCA compresses the HVGs into a few dozen axes. Each cell is linked to its nearest neighbours, and Louvain clustering cuts the graph into communities. The elbow plot guides how many PCs to use.

LEVEL 2

UMAP is a picture, not a measurement

UMAP and t-SNE squeeze many dimensions into two for viewing. Neighbours are reliable; the distances between far-apart islands are not. Conclusions should rest on the clusters and markers.

For method developers and reviewers: the statistical caveats that decide whether an apparent result holds up.

LEVEL 3

Zeros are mostly sampling, not "dropout"

For UMI data from droplets, counts of a gene in a homogeneous cell population fit a negative binomial without a need for extra zero-inflation (Svensson, 2020). Many zeros are simply low expression sampled shallowly.

LEVEL 3

Ambient RNA and doublets blur types

RNA floating in the suspension is captured with every droplet and can make a cell appear to express another type's genes. Tools such as SoupX and CellBender model it; Scrublet and DoubletFinder score doublets.

LEVEL 3

Cells are not independent replicates

Cells from one donor share that donor's biology. Testing cells as if they were independent samples inflates significance, so multi-donor comparisons are best made on pseudobulk profiles (Squair et al., 2021; Zimmerman et al., 2021).

LEVEL 3

Clustering depends on its settings

The resolution parameter decides whether a population is one cluster or several. Louvain can produce poorly connected communities, which Leiden was designed to avoid (Traag et al., 2019). Stability checks are advisable.

LEVEL 3

Batches need explicit handling

Samples processed apart differ technically. Integration methods such as Harmony or scVI align them, but over-correction can erase real biology, so integration is benchmarked on both mixing and biology (Luecken et al., 2022).

LEVEL 3

Annotation is a hypothesis

Reference-based labels transfer well only when the reference contains the cell type. Novel or rare states get the nearest wrong label. Marker evidence and reference scores should agree before a label is trusted.

Barcode
A short DNA tag that marks which cell (or droplet) a read came from.
UMI
Unique molecular identifier: a random tag that lets copies of one molecule be counted once.
GEM
Gel bead in emulsion: the droplet that holds one bead and, ideally, one cell.
Doublet
A droplet that captured two cells, which then look like one odd cell.
Ambient RNA
Free RNA from broken cells that is picked up in every droplet.
HVG
Highly variable gene: a gene whose level varies across cells more than expected by chance.
PCA
Principal component analysis: finds the main axes along which cells differ.
UMAP
A method that draws high-dimensional cells on a 2-D map, keeping neighbours together.
Louvain
A graph method that groups cells into densely connected communities (clusters).
Marker gene
A gene that is much higher in one cluster than in the others.
Pseudobulk
Counts summed across cells of one type within one sample, to be tested as a bulk profile.
Batch effect
A technical difference between samples processed separately, not due to biology.

How a cell becomes a row of counts

The path from a living cell to a count matrix, and the data properties that shape every later analysis step.

Figure 1 Droplet-based single-cell capture schematic
1 · Suspension + beads2 · Droplets (GEMs)3 · Barcoded RT4 · Pool, library, sequence cells (colour) and gel beads (grey) one cell + one bead per droplet (ideal) doublet cell lyses · mRNA captured on bead cDNA gets cell barcode + UMI pooled cDNA sequencing library FASTQ → Cell Ranger reads sorted back to cells by barcode
Figure 1 | From cell to library. Cells and barcoded beads meet in oil so that each droplet ideally holds one of each. Inside, the cell lyses and its mRNA is tagged with the bead's barcode and a molecular tag. Droplets that capture two cells give doublets (Zheng et al., 2017; Macosko et al., 2015).
Figure 2 Read structure and UMI counting schematic
What each captured molecule carries Cell barcode UMI poly(dT) cDNA of the gene Which cell · which molecule · captures mRNA tails · which gene Counting after amplification cell A · gene X · UMI ACGT… cell A · gene X · UMI ACGT… cell A · gene X · UMI ACGT… cell A · gene X · UMI TTGA… collapse Gene X in cell A = 2two molecules, four reads Three reads share one UMI, so they are PCR copies of a single molecule. Tag lengths depend on the chemistry version.
Figure 2 | Why UMIs matter. Amplification makes uneven numbers of copies of each molecule. Counting distinct UMIs per gene and cell, as Cell Ranger does, corrects this bias and turns reads into molecule counts.
Figure 3 Barcode rank plot: cells versus empty droplets simulated
Figure 3 | Calling cells. Barcodes are ranked by UMI count. Real cells form a plateau of high counts, then the curve drops sharply to the low-count ambient background of empty droplets. Cell Ranger uses this shape to decide which barcodes are cells.
Figure 4 Most zeros are sampling, not dropout simulated
Figure 4 | Mean-variance of counts. Each point is a gene in one simulated cell type. Variance follows the negative binomial curve (green), which sits above the Poisson line (grey) for abundant genes but needs no extra zero-inflation term for UMI data (Svensson, 2020).
Table 1 Choosing a single-cell platform
ApproachStrengthTrade-off
Droplet (10x Chromium)Thousands of cells per run; 3′ or 5′ countingShallow per cell; needs viable suspension
Plate-based (Smart-seq)Full-length transcripts; deep per cellHundreds of cells; costly per cell
Nuclei (snRNA-seq)Frozen or hard-to-dissociate tissueFewer transcripts; nuclear bias
Split-pool barcodingFixed cells; very large scaleMore complex design and analysis
This service. The page describes 10x Chromium data processed with Cell Ranger; other platforms can be discussed with our team.

Input checklist for scRNA-seq projects

Standard requirements — please confirm specifics with our team based on tissue type and platform.

ParameterRequirement
Sample typeSingle-cell suspensions from tissue, blood, or cultured cells (fresh or cryopreserved)
Recommended inputViable single-cell suspension; confirm target cell number with team
ViabilityHigh viability preferred (≥ 70–80% typical for quality libraries)
Cell concentrationPlatform-dependent (e.g. 10x Genomics recommended ranges)
Library platform10x Genomics Chromium or equivalent; confirm with team
Sequencing depthProject-dependent; typically ≥ 20,000–50,000 reads/cell for 3′ gene expression
Sample handlingMinimize time from dissociation to loading; keep cells on ice
MetadataSample IDs, tissue/condition labels, and any batch information

Contact us at info@nucleomeinfo.com for a tailored sample checklist based on tissue type and platform.

Aim for gentle dissociationHarsh or long dissociation damages cells and raises ambient RNA. Keep cells cold, filter clumps, and load as soon as possible after dissociation.
Include biological replicatesSeveral donors or animals per condition allow condition effects to be separated from individual variation, which matters for any comparison between groups.
Avoid confounding batch and conditionIf all treated samples are processed on one day and controls on another, the two cannot be told apart later. Spread conditions across batches and record them.
Tell us the plan for the tissueFresh, cryopreserved or nuclei-based input, and the cell types you most want to resolve, shape depth and design. Share these early.

From raw FASTQ to annotated clusters

Every scRNA-seq dataset moves through the same structured pipeline, from raw reads to annotated, biologically interpretable clusters.

Figure 5 Quality-gated single-cell workflow infographic
STAGE 01

Cell Ranger processing

FASTQ files are aligned and counted; gene-barcode matrices are generated, with cell calling.

raw matrixfiltered matrix
cells called · reads in cells
STAGE 02

QC & filtering

Cells and genes are filtered on gene count, UMI count and mitochondrial content.

clean matrix
before/after QC plots
STAGE 03

Normalization & HVG

LogNormalize and selection of highly variable genes to capture biological signal.

normalizedHVG list
variable-gene plot
STAGE 04

Dimensionality reduction

PCA on the HVG matrix; the elbow plot guides how many PCs to keep.

PCs
elbow plot
STAGE 05

Clustering & embedding

Shared-nearest-neighbour graph with Louvain clustering; UMAP and t-SNE for display.

clustersUMAPt-SNE
cluster sizes · stability
STAGE 06

Marker identification

Wilcoxon rank-sum test per cluster (adj. p < 0.05, log₂FC > 0.25).

marker tables
adj. p · log₂FC
STAGE 07

Cell-type annotation

Known markers plus SingleR reference scoring, combined into consensus labels.

labelled UMAPscore heatmap
markers agree with reference
STAGE 08

Differential expression

Population-level DEGs (adj. p < 0.05, log₂FC > 0.5).

DEG tables
design reviewed
STAGE 09

Sub-analysis & delivery

Subset, re-normalize (SCTransform) and re-cluster lineages of interest; results via secure drive.

sub-clustersviolin plots
access control
Figure 5 | Decision points. Amber gates mark where metrics are reviewed before the next stage. Thresholds are set from the distributions in your own data rather than copied from another study.
01

Data Processing

Raw FASTQ files processed with Cell Ranger Count; gene-barcode expression matrices generated.

02

QC & Filtering

Low-quality cells and genes removed based on gene count, UMI count, and mitochondrial content.

03

Normalization & HVG

LogNormalize and highly variable gene identification to capture biological signal.

04

Dimensionality reduction — PCA on HVG matrix; elbow plot guides PC selection.

05

Clustering & visualization — SNN graph + Louvain clustering; UMAP and t-SNE plots.

06

Marker identification — Wilcoxon rank-sum (adj. p < 0.05, log₂FC > 0.25) per cluster.

07

Cell-type annotation — Known markers + SingleR; annotation-score heatmaps.

08

Differential expression — Population-level DEGs (adj. p < 0.05, log₂FC > 0.5).

09

Fine-grained sub-analysis — Subset, re-normalize (SCTransform), re-cluster lineages of interest.

Keep the cells, drop the artefacts

Three simple numbers per cell, UMI count, gene count and mitochondrial share, separate healthy cells from damaged ones and from doublets. The thresholds below are examples; real cut-offs are set from your own distributions.

Figure 6 Cell-level QC with adjustable thresholds simulated
Figure 6 | Reading the QC scatter. Each point is a cell, coloured by mitochondrial share. Damaged cells sit at low gene counts with high mitochondrial reads; doublets sit at the top right. Drag the sliders to see which cells each threshold removes (Ilicic et al., 2016).
Figure 7 What each filter removes simulated
Figure 7 | Retention by reason. Counts of cells kept and removed for the thresholds in Figure 6. Too strict a filter can discard rare or metabolically active cell types, which naturally have high mitochondrial content.
Figure 8 Doublet scores simulated
Figure 8 | Finding doublets. Methods such as Scrublet create artificial doublets from the data and score every cell by its similarity to them (Wolock et al., 2019). Real doublets form a second peak at high scores. The doublet rate rises with the number of cells loaded.
Table 2 QC signals and what to do
SignalSuggestsTypical action
Low genes and UMIsEmpty droplet or debrisRemove
High mitochondrial %Damaged or dying cellRemove above a data-based cut-off
Very high genes and UMIsDoubletRemove or flag
Marker mixing across typesAmbient RNACorrect or interpret with care
Cluster driven by one sampleBatch effectIntegrate, or model batch
Beyond the stated service. Doublet scoring and ambient-RNA correction are described here as good practice; confirm with our team whether they are included in your project.

Find the genes that matter, then compress them

After filtering, counts are normalized, informative genes are selected, and PCA compresses them into a few dozen axes that feed the neighbour graph.

Figure 9 Highly variable genes simulated
Figure 9 | Selecting variable genes. Each point is a gene. Most follow the expected trend of variance against mean (line); genes well above it (orange) vary because cell types differ, and are kept for PCA. Choosing about 2,000 HVGs is a common default.
Figure 10 PCA elbow plot illustrative
Figure 10 | How many PCs? The first components capture real structure and then the curve flattens into noise. The elbow guides how many PCs to pass to the neighbour graph. The curve here is a stylised example, not computed from data.
Normalization options. LogNormalize divides each cell's counts by its total, scales and takes a log. SCTransform fits a regularized negative binomial model per gene and returns Pearson residuals (Hafemeister and Satija, 2019; Choudhary and Satija, 2022). Comparisons of transformations find that simple log-based methods remain competitive (Ahlmann-Eltze and Huber, 2023), so the right choice depends on the question.

Group the cells, then name the groups

Clusters are statistical groupings. Marker genes and reference comparison turn them into cell types.

Figure 11 Clusters on a UMAP and the resolution trade-off simulated
Figure 11 | Too few, just right, too many. Cells are simulated from 12 known types. Move the slider: too few clusters merge distinct types, too many split real ones. The agreement score (adjusted Rand index) rises towards the true number of 12 and falls again when types are over-split; because k-means is imperfect, its peak can sit a little above 12. A simple k-means stands in for Louvain here, whose resolution parameter plays the same role.
Figure 12 A marker gene on the map simulated
Figure 12 | Feature plot. Colour shows the normalized level of one marker (CD3E, a T-cell gene) on the same map. Expression is high in the T-cell islands and low elsewhere, with scattered zeros even inside T cells, which is normal for shallow sampling.
Figure 13 Marker dot plot simulated
Figure 13 | Evidence for each label. Dot size is the share of cells expressing a gene and colour is its average level. A clean diagonal means each type has its own markers; off-diagonal dots, such as NKG7 in CD8 T cells, show shared biology that makes labels harder.
Figure 14 Reference-based annotation scores illustrative
Figure 14 | A second opinion from a reference. SingleR compares each cluster with labelled reference profiles (Aran et al., 2019). Bright cells show the best match. Confusion between closely related types, such as monocytes and macrophages, is expected. Values are stylised.

Zoom in on a lineage, and handle batches

Fine-grained sub-clustering raises resolution inside a compartment. When samples were processed apart, batch handling keeps technical differences from masquerading as biology.

Figure 15 Sub-clustering the T/NK compartment simulated
Figure 15 | Resolution by re-analysis. Left: T and NK cells in the whole-dataset map, where closely related subtypes overlap. Right: the same cells after subsetting, re-normalizing (SCTransform) and re-clustering, where the subtypes separate. Colours show the simulated true subtype.
Figure 16 Batch effects and integration illustrative
Figure 16 | Mixing batches without merging types. Two batches of the same six cell types, plotted on a map. Without correction each type splits by batch. After correction batches mix inside each type while the types stay apart. The mixing score is the share of each cell's 15 nearest neighbours from the other batch. Correction here is a simple shift, for illustration; methods such as Harmony and scVI are far more general (Korsunsky et al., 2019; Lopez et al., 2018).
Figure 17 Cell-level tests versus pseudobulk simulated
Figure 17 | The pseudoreplication problem. Three donors per group with no true difference. Testing individual cells (orange) reports many false hits because cells from one donor are not independent. Testing donor-level sums (green) gives the expected uniform p-values (Squair et al., 2021; Zimmerman et al., 2021).

Markers and population DEGs

Cluster markers use adj. p < 0.05 and log₂FC > 0.25; population-level DEGs use adj. p < 0.05 and log₂FC > 0.5, both from Wilcoxon rank-sum tests.

Figure 18 Volcano plot with the page's thresholds simulated
Figure 18 | Two thresholds, two questions. Set 0.25 for cluster markers and 0.5 for population DEGs to reproduce the page's defaults. A higher fold-change cut-off favours large effects and drops subtle ones; lower p cut-offs trade sensitivity for fewer false positives.
Table 3 Choosing a test
QuestionSuitable approach
What marks cluster 4?Wilcoxon rank-sum, cluster versus rest
Cluster against clusterWilcoxon rank-sum, pairwise
Treated versus control, one sample eachCell-level test; interpret cautiously
Treated versus control, several donorsPseudobulk with a count model (e.g. edgeR, DESeq2)
Design note. Marker discovery is a descriptive task where cell-level tests are standard. For condition comparisons across several donors, ask our team about donor-aware pseudobulk testing.

Annotated landscapes, ready to interpret

At project close, clients receive complete results and high-resolution images via a secure drive link.

Expression Matrices

Cell Ranger gene-barcode count matrices (raw and filtered).

QC Reports

Cell/gene filtering metrics and before/after QC visualizations.

Dimensionality Reduction

PCA elbow plots and reduced-dimension embeddings.

Clustering & Visualization

UMAP and t-SNE plots of clusters.

Marker Gene Tables

Top marker genes per cluster, with log fold-change and adjusted p-values.

Cell-Type Annotation

Annotation-score heatmaps and consensus cell-type-labeled UMAP plots.

Differential Expression

Cluster-vs-cluster and population-level DEG tables.

Sub-Lineage Analysis

Subset clustering, marker tables, and violin plots for lineages of interest (if applicable).

Full Data & Figures

Complete results and high-resolution images via secure drive link.

Immune and non-immune cell atlas from a complex tissue

A single-cell RNA-seq dataset was processed end-to-end — from raw FASTQ through Cell Ranger quantification, QC, PCA/UMAP, Louvain clustering, and reference-based cell-type annotation.

Analysis Overview & Metrics

Complex tissue · Cell Ranger + Seurat · SingleR annotation

Clusters identified (Louvain)24
Cell types annotated25+
Marker gene testingWilcoxon rank-sum
Cluster markers (thresholds)adj. p < 0.05, log₂FC > 0.25
Population DEGsadj. p < 0.05, log₂FC > 0.5
Fine-grained sub-analysisT/NK + myeloid

Beyond the initial cluster-level annotation, fine-grained sub-clustering of the T/NK and myeloid compartments resolved additional immune subtypes not distinguishable at the whole-dataset resolution.

Cell Types Annotated

T cells NK cells B cells Monocytes Macrophages Neutrophils Erythrocytes Endothelial Fibroblasts Chondrocytes Epithelial HSC / MEP GMP / CMP

Sub-Lineage Resolution

T/NK: naive/memory T cells, NK variants
Myeloid: Monocyte, Neutrophil, Macrophage, Erythroblast, Platelet

A layered analysis strategy (broad annotation followed by targeted sub-clustering) substantially increased the biological resolution of a single scRNA-seq run.

Where single-cell analysis is heading

Single-cell RNA-seq is now one layer in a wider toolbox. Here is how the field adds more modalities, better statistics and learned models, and what recent benchmarks say about them.

Figure 19 Beyond RNA: layers measured per cell schematic
RNA Surface proteinsCITE-seq antibody tags ChromatinATAC with RNA (multiome) Spatial positionwhere the cell sits Sample hashingpool samples, flag doublets
Figure 19 | One cell, several readouts. Antibody tags add surface proteins (Stoeckius et al., 2017), multiome kits add chromatin accessibility, spatial methods add location, and sample hashing lets pooled samples be separated and doublets be flagged (Stoeckius et al., 2018).
Table 4 Where the options sit
OptionGainCost
Standard scRNA-seqCell types and states from RNAProtein and chromatin are inferred
CITE-seqProtein-based immune typingAntibody panel to design
MultiomeRegulatory regions per cellNuclei prep; extra analysis
SpatialTissue contextLower depth or resolution trade-offs
Pooled with hashingLess batch effect; cheaperDemultiplexing step
Layered analysis. The page's own approach, broad annotation followed by targeted sub-clustering, is a way of adding resolution without more experiments.
BENCHMARKS

Integration is benchmarked

A benchmark of 16 integration methods across 13 tasks scored both batch removal and conservation of biology, showing that no single method wins everywhere (Luecken et al., Nature Methods, 2022).

STATISTICS

Pseudobulk for condition effects

Methods that treat cells as replicates are biased towards highly expressed genes and produce false positives when samples differ; aggregating to the sample level corrects this (Squair et al., Nature Communications, 2021).

CLEAN-UP

Removing ambient RNA

CellBender uses a deep generative model to remove background RNA and call cells more accurately, improving marker specificity (Fleming et al., Nature Methods, 2023).

DYNAMICS

RNA velocity

The ratio of unspliced to spliced reads hints at the direction a cell is moving along a trajectory (La Manno et al., Nature, 2018); scVelo generalises the model (Bergen et al., Nature Biotechnology, 2020).

SCALE

Atlases of millions of cells

Seurat v5 uses sketching to analyse very large datasets and integrates across modalities (Hao et al., Nature Biotechnology, 2024), and reference mapping labels new data against atlases.

FOUNDATION MODELS

Large pre-trained models

Models such as scGPT are pre-trained on millions of cells (Cui et al., Nature Methods, 2024). Zero-shot evaluations found that they can fail to beat simple baselines for clustering and batch integration (Kedzierska et al., Genome Biology, 2025).

Scope note. Nucleome's single-cell service covers the workflow and deliverables described on this page: Cell Ranger processing, QC filtering, normalization and HVG selection, PCA, Louvain clustering, UMAP and t-SNE, Wilcoxon-based marker and differential expression analysis, marker plus SingleR annotation, and sub-clustering with SCTransform. Doublet and ambient-RNA handling, batch integration, pseudobulk testing, multimodal data, RNA velocity and learned models are shown to place the service in its scientific context; contact our team to discuss extensions for your project.

Service highlights

End-to-end single-cell analysis with rigorous QC and layered biological interpretation.

End-to-End Pipeline

From raw FASTQ to annotated, interpretable clusters — Cell Ranger + Seurat.

Rigorous QC Filtering

Gene count, UMI count, and mitochondrial content filters to remove low-quality/dead cells.

Dual Visualization

UMAP and t-SNE clustering visualization for robust interpretation.

Reference-Based Annotation

SingleR with reference atlases alongside marker-driven annotation for consensus labels.

Fine-Grained Sub-Analysis

Sub-lineage analysis available for immune and other compartments of interest.

Statistically Grounded

Wilcoxon rank-sum marker and DE testing with defined significance thresholds.

Frequently asked, in plain language

From the basics to the finer points of design and statistics.

basicWhat is single-cell RNA-seq?
It measures the gene activity of each cell separately, so you can identify the cell types in a tissue and the genes that define them, rather than reading one average for the whole sample.
basicWhat will I receive?
Raw and filtered expression matrices, QC reports, PCA elbow plots, UMAP and t-SNE plots, marker gene tables, annotation-score heatmaps and labelled UMAPs, differential expression tables, sub-lineage results where applicable, and full figures through a secure drive link.
basicHow many cells do I need?
It depends on how rare the cell types of interest are and how many conditions you compare. Our team can advise on target cell numbers for your tissue and platform.
intermediateWhy does viability matter so much?
Dying cells leak RNA, which raises background in every droplet, and they show high mitochondrial content that can form misleading clusters. High-viability suspensions give cleaner data.
intermediateHow are cell types named?
Clusters are labelled from known marker genes and checked against a reference with SingleR. Labels are reported as a consensus, and clusters where the two disagree are flagged for review.
intermediateWhy UMAP and t-SNE both?
They are two ways to draw the same data in two dimensions. Seeing both guards against over-reading a quirk of one layout. Neither should be used to measure distances between islands.
intermediateWhat if some of my cells are doublets?
A small share of droplets capture two cells and look like unusual hybrids. The mitochondrial, gene and UMI filters catch some of them, and dedicated doublet scoring can be discussed for dense loadings.
advancedCan I compare conditions between donors?
Yes, but because cells from one donor are not independent, the best practice for multi-donor designs is pseudobulk testing on summed counts. The page's standard population DEGs use Wilcoxon rank-sum tests; ask us how to handle replicated designs.
advancedDo you correct batch effects?
When samples were processed in different batches, integration can be applied so that clusters reflect biology. It must be chosen carefully, as over-correction can merge real cell types. Share batch information with your sample sheet.
advancedWhy a log fold-change threshold at all?
With thousands of cells, even tiny differences become statistically significant. A fold-change threshold, 0.25 for cluster markers and 0.5 for population DEGs on this page, keeps the lists focused on differences large enough to matter.
advancedDoes sub-clustering create new cell types?
It reveals structure that was hidden at whole-dataset resolution, but sub-clusters should be supported by markers and, where possible, independent evidence. Over-splitting is a risk, as Figure 11 shows.

Key literature & resources

Selected sources behind the methods and reference values on this page. Figures labelled "simulated", "schematic" or "illustrative" are generated for this page, not client data.

  1. Zheng GXY, et al. Massively parallel digital transcriptional profiling of single cells. Nature Communications 2017;8:14049.
  2. Macosko EZ, et al. Highly parallel genome-wide expression profiling of individual cells using nanoliter droplets. Cell 2015;161:1202–1214.
  3. Picelli S, et al. Smart-seq2 for sensitive full-length transcriptome profiling in single cells. Nature Methods 2013;10:1096–1098.
  4. Satija R, et al. Spatial reconstruction of single-cell gene expression data. Nature Biotechnology 2015;33:495–502.
  5. Stuart T, et al. Comprehensive integration of single-cell data. Cell 2019;177:1888–1902.
  6. Hao Y, et al. Dictionary learning for integrative, multimodal and scalable single-cell analysis. Nature Biotechnology 2024;42:293–304.
  7. Wolf FA, Angerer P, Theis FJ. SCANPY: large-scale single-cell gene expression data analysis. Genome Biology 2018;19:15.
  8. Luecken MD, Theis FJ. Current best practices in single-cell RNA-seq analysis: a tutorial. Molecular Systems Biology 2019;15:e8746.
  9. Heumos L, et al. Best practices for single-cell analysis across modalities. Nature Reviews Genetics 2023;24:550–572.
  10. Ilicic T, et al. Classification of low quality cells from single-cell RNA-seq data. Genome Biology 2016;17:29.
  11. Wolock SL, Lopez R, Klein AM. Scrublet: computational identification of cell doublets in single-cell transcriptomic data. Cell Systems 2019;8:281–291.
  12. McGinnis CS, Murrow LM, Gartner ZJ. DoubletFinder: doublet detection in single-cell RNA sequencing data using artificial nearest neighbors. Cell Systems 2019;8:329–337. PDF
  13. Young MD, Behjati S. SoupX removes ambient RNA contamination from droplet-based single-cell RNA sequencing data. GigaScience 2020;9:giaa151.
  14. Fleming SJ, et al. Unsupervised removal of systematic background noise from droplet-based single-cell experiments using CellBender. Nature Methods 2023;20:1323–1335.
  15. Svensson V. Droplet scRNA-seq is not zero-inflated. Nature Biotechnology 2020;38:147–150.
  16. Hafemeister C, Satija R. Normalization and variance stabilization of single-cell RNA-seq data using regularized negative binomial regression. Genome Biology 2019;20:296.
  17. Choudhary S, Satija R. Comparison and evaluation of statistical error models for scRNA-seq. Genome Biology 2022;23:27.
  18. Ahlmann-Eltze C, Huber W. Comparison of transformations for single-cell RNA-seq data. Nature Methods 2023;20:665–672.
  19. Blondel VD, et al. Fast unfolding of communities in large networks. Journal of Statistical Mechanics 2008;P10008.
  20. Traag VA, Waltman L, van Eck NJ. From Louvain to Leiden: guaranteeing well-connected communities. Scientific Reports 2019;9:5233.
  21. McInnes L, Healy J, Melville J. UMAP: uniform manifold approximation and projection for dimension reduction. arXiv:1802.03426, 2018.
  22. Chari T, Pachter L. The specious art of single-cell genomics. PLoS Computational Biology 2023;19:e1011288.
  23. Aran D, et al. Reference-based analysis of lung single-cell sequencing reveals a transitional profibrotic macrophage. Nature Immunology 2019;20:163–172.
  24. Soneson C, Robinson MD. Bias, robustness and scalability in single-cell differential expression analysis. Nature Methods 2018;15:255–261.
  25. Squair JW, et al. Confronting false discoveries in single-cell differential expression. Nature Communications 2021;12:5692.
  26. Zimmerman KD, Espeland MA, Langefeld CD. A practical solution to pseudoreplication bias in single-cell studies. Nature Communications 2021;12:738.
  27. Benchmarking integration of single-cell differential expression. PMC10030080
  28. Korsunsky I, et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nature Methods 2019;16:1289–1296.
  29. Lopez R, et al. Deep generative modeling for single-cell transcriptomics. Nature Methods 2018;15:1053–1058.
  30. Luecken MD, et al. Benchmarking atlas-level data integration in single-cell genomics. Nature Methods 2022;19:41–50.
  31. La Manno G, et al. RNA velocity of single cells. Nature 2018;560:494–498.
  32. Bergen V, et al. Generalizing RNA velocity to transient cell states through dynamical modeling. Nature Biotechnology 2020;38:1408–1414.
  33. Stoeckius M, et al. Simultaneous epitope and transcriptome measurement in single cells. Nature Methods 2017;14:865–868.
  34. Stoeckius M, et al. Cell Hashing with barcoded antibodies enables multiplexing and doublet detection for single cell genomics. Genome Biology 2018;19:224.
  35. Cui H, et al. scGPT: toward building a foundation model for single-cell multi-omics using generative AI. Nature Methods 2024;21:1470–1480.
  36. Kedzierska KZ, et al. Zero-shot evaluation reveals limitations of single-cell foundation models. Genome Biology 2025. Link
  37. Benjamini Y, Hochberg Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society B 1995;57:289–300.