What data can I run through the single-cell pipeline?

10x Genomics single-cell expression data delivered as raw paired-end FASTQs. The organism for each sample is set in the samplesheet, and alignment runs against human, mouse, monkey, rat, zebrafish or yeast references, so one run can mix species. The library chemistry is detected from the barcodes rather than typed in, so there is nothing to look up before launching.

How are my cell types labelled?

Two independent ways. Clusters are labelled from canonical marker genes for the major cell types: T, NK, B, plasma, monocyte, macrophage, DC, pDC, neutrophil, mast, endothelial, fibroblast, epithelial and erythroid cells, with T-cell clusters split further into CD4, CD8, Treg or mixed. Alongside that, reference-based prediction runs with Azimuth for human, mouse and monkey, or SingleR for rat and zebrafish. Where the two disagree, that cluster is the one worth a closer look.

Which annotation reference should I choose?

For human, mouse and monkey you pick the Azimuth reference that matches your tissue: PBMC or whole blood, lung, bone marrow, kidney, heart, pancreas, tonsil, adipose, fetal development, or motor cortex. Rat and zebrafish samples are annotated with SingleR instead. ProjecTILs then maps CD4 and CD8 T cells onto reference atlases to add T-cell states, using human maps for human and monkey and mouse maps for mouse.

What comes back from a single-cell run?

An annotated Seurat object as RDS, plus an H5AD export so the same object opens in Python tooling such as Scanpy without conversion work. You also get cell-type labels, T-cell states and marker gene tables, differential expression tables per cluster and per treatment, QC gate and integration check tables, and MultiQC and HTML reports for the experiment.

How do I read the report?

Start at the QC gate table, where every sample is marked PASS, WARN or FAIL, then the integration check table, which says whether the samples shared enough cells to be integrated. From there the UMAP shows the clusters with their labels, the marker tables and DotPlots show the gene evidence behind each label, and the differential expression tables show what changes between treatments inside each cell population.

Why is one of my samples missing from the annotated object?

The QC gate filters cells on gene count and mitochondrial percentage, using the mitochondrial gene pattern for that organism, and flags each sample PASS, WARN or FAIL. A sample that fails stops there and appears only in the MultiQC report, so it cannot quietly skew normalization, clustering or the cell-type labels. The report shows the numbers behind that decision.

Is this the right pipeline for my data?

It is built for 10x Genomics single-cell expression data. If you are working with bulk RNA-seq, that pipeline aligns or quasi-aligns your reads and returns count tables and a MultiQC report, and TCR libraries are handled by the TCR detection pipeline.

In Horizon10x Genomics expression analysis

Single-cell RNA-seq

From 10x Genomics reads to an annotated, analysis-ready Seurat object: alignment, QC gating, integration, clustering, automated cell-type labels, and differential expression in one run.

Input

Raw paired-end FASTQ reads from 10x Genomics, with the organism for each sample set in the samplesheet.

Output

An annotated Seurat object (RDS) with automated cell-type labels and visualizations, an H5AD export, DGE tables, QC and integration check tables, and MultiQC and HTML reports.

New to this method?

Read the single-cell RNA-seq guide

What scRNA-seq measures, how a Seurat analysis runs from QC to annotation, and common questions about clustering and integration.

Read the guide

At a glance

From reads to labeled cells

Fifteen steps turn 10x Genomics reads into an annotated Seurat object with cell-type labels, differential expression, and reports for review.

PreprocessingProcessingPost-processing

InputPaired-end 10x FASTQOrganism set per sample in the samplesheet

01 / PreprocessingChemistry detectionDetect the chemistry

02 / PreprocessingFastQCCheck the raw reads

03 / ProcessingSTARsoloAlign and count cells

04 / ProcessingSeurat QCGate sample quality

05 / ProcessingSCTransformNormalize

06 / ProcessingSeurat integrationIntegrate samples

07 / ProcessingPCA, UMAP and clusteringMap and cluster cells

08 / ProcessingscdsRemove doublets

09 / Post-processingMarker-based labelingLabel cell types

10 / Post-processingProjecTILsResolve T-cell states

11 / Post-processingAzimuth and SingleRPredict with references

12 / Post-processingMarkers and DotPlotsShow the evidence

13 / Post-processingDifferential expressionCompare expression

14 / Post-processingDensity plot and H5ADExport for Python

15 / Post-processingMultiQC and HTML reportBring results together

OutputAnnotated Seurat objectRDS, H5AD, DGE tables & reports

InputPaired-end 10x FASTQOrganism set per sample in the samplesheet

01 / PreprocessingChemistry detectionDetect the chemistry

02 / PreprocessingFastQCCheck the raw reads

03 / ProcessingSTARsoloAlign and count cells

04 / ProcessingSeurat QCGate sample quality

05 / ProcessingSCTransformNormalize

06 / ProcessingSeurat integrationIntegrate samples

07 / ProcessingPCA, UMAP and clusteringMap and cluster cells

08 / ProcessingscdsRemove doublets

09 / Post-processingMarker-based labelingLabel cell types

10 / Post-processingProjecTILsResolve T-cell states

11 / Post-processingAzimuth and SingleRPredict with references

12 / Post-processingMarkers and DotPlotsShow the evidence

13 / Post-processingDifferential expressionCompare expression

14 / Post-processingDensity plot and H5ADExport for Python

15 / Post-processingMultiQC and HTML reportBring results together

OutputAnnotated Seurat objectRDS, H5AD, DGE tables & reports

Samples that fail the QC gate stop there and appear only in the MultiQC report; passing samples continue through normalization and integration.

Inside the pipeline

Why each step is here

01 / preprocessing

Chemistry detection

Detect the chemistry

What it does: Subsamples Read 1 of each sample's paired-end FASTQs (100,000 reads by default) and checks the barcodes against the 10x whitelists.
Why it is included: The library chemistry determines how barcodes and UMIs are read, so it has to be right before alignment.
What it produces: A detected 10x chemistry for every sample.
Why it matters: No chemistry settings to look up or guess; the pipeline works it out from the data.

02 / preprocessing

FastQC

Check the raw reads

What it does: Runs quality checks on the raw reads.
Why it is included: Read quality needs to be visible before interpreting cell-level results.
What it produces: Per-sample quality results that feed the final MultiQC report.
Why it matters: Spot read-quality concerns before relying on expression results.

03 / processing

STARsolo

Align and count cells

What it does: Aligns the reads to the reference genome for each sample's organism (human, mouse, monkey, rat, zebrafish or yeast, set per sample in the samplesheet), identifies real cells, and builds a gene-by-cell count matrix.
Why it is included: Counts per cell are the foundation every later step builds on.
What it produces: Per-sample gene-by-cell count matrices.
Why it matters: Six organisms supported in one run, each sample aligned to its own reference.

04 / processing

Seurat QC

Gate sample quality

What it does: Filters cells on gene count and mitochondrial percentage, using the right mitochondrial gene pattern for each organism, and a QC gate flags each sample as PASS, WARN or FAIL. Failing samples stop here and only appear in the MultiQC report.
Why it is included: Low-quality cells and samples distort normalization, clustering and labels downstream.
What it produces: QC gate tables and filtered count matrices for passing samples.
Why it matters: Weak samples are flagged automatically instead of silently skewing the experiment.

05 / processing

SCTransform

Normalize

What it does: Normalizes passing samples with SCTransform.
Why it is included: Cells differ in sequencing depth, and normalization makes their expression values comparable.
What it produces: Normalized expression values per sample.
Why it matters: Clusters and markers reflect biology rather than sequencing depth.

06 / processing

Seurat integration

Integrate samples

What it does: Integrates samples in the same experiment with Seurat (CCA by default), and an integration check reports whether the samples shared enough cells to integrate.
Why it is included: Batch effects between samples can masquerade as biological differences.
What it produces: An integrated object and an integration check table.
Why it matters: Samples line up for comparison, with a clear report when they cannot.

07 / processing

PCA, UMAP and clustering

Map and cluster cells

What it does: Runs PCA, UMAP and clustering at several resolutions.
Why it is included: Dimensionality reduction and clustering turn millions of measurements into readable cell groups.
What it produces: UMAP embeddings and cluster assignments at multiple resolutions.
Why it matters: See the structure of the experiment at a glance, at coarse or fine resolution.

08 / processing

scds

Remove doublets

What it does: Scores and removes likely doublets.
Why it is included: Two cells captured in one droplet look like a cell with an impossible mixed identity.
What it produces: A count matrix with likely doublets removed.
Why it matters: Clusters and labels are not distorted by artificial hybrid cells.

09 / post-processing

Marker-based labeling

Label cell types

What it does: Labels each cluster from the percentage of its cells expressing canonical markers for the major cell types: T, NK, B, plasma, monocyte, macrophage, DC, pDC, neutrophil, mast, endothelial, fibroblast, epithelial and erythroid cells. T-cell clusters are further split into CD4, CD8, Treg or mixed, using marker genes matched to the organism.
Why it is included: Cluster numbers mean little to a biologist until they carry cell-type names.
What it produces: Cell-type labels for every cluster.
Why it matters: Results read in biological language from the first look.

10 / post-processing

ProjecTILs

Resolve T-cell states

What it does: Maps the CD4 and CD8 T cells onto reference atlases to assign T-cell states (human maps for human and monkey, mouse maps for mouse).
Why it is included: T-cell state, not just type, is often where the biology of interest sits.
What it produces: T-cell state annotations.
Why it matters: T cells come back with atlas-level state detail without manual reference mapping.

11 / post-processing

Azimuth and SingleR

Predict with references

What it does: Predicts cell types with Azimuth for human, mouse and monkey, where the user picks the reference that matches the tissue: PBMC/whole blood, lung, bone marrow, kidney, heart, pancreas, tonsil, adipose, fetal development or motor cortex. SingleR is used instead for rat and zebrafish.
Why it is included: A second, reference-based opinion catches labeling mistakes and adds finer subtypes.
What it produces: Reference-based cell-type predictions alongside the marker labels.
Why it matters: Two independent annotations to trust, or to reconcile where they differ.

12 / post-processing

Markers and DotPlots

Show the evidence

What it does: Generates top marker genes per cluster and DotPlots of user-defined gene panels.
Why it is included: Labels need visible evidence that reviewers and collaborators can check.
What it produces: Marker tables and DotPlot figures.
Why it matters: Every label arrives with the gene evidence behind it.

13 / post-processing

Differential expression

Compare expression

What it does: Runs differential expression per cluster against all other cells, and for integrated experiments also between treatments within each cell population or Seurat cluster.
Why it is included: The question most experiments ask is what changes, and where.
What it produces: DGE tables per cluster and per treatment comparison.
Why it matters: Treatment effects are visible inside each cell type, not just across the whole sample.

14 / post-processing

Density plot and H5AD

Export for Python

What it does: Generates a UMAP density plot and exports the final object to H5AD for use in Python.
Why it is included: Many teams continue analysis in Python tooling such as Scanpy.
What it produces: A density plot and an H5AD export.
Why it matters: The same annotated object moves between R and Python without conversion work.

15 / post-processing

MultiQC and HTML report

Bring results together

What it does: Builds a MultiQC report and an HTML report per experiment.
Why it is included: Quality metrics and results are most useful reviewed together, by everyone involved.
What it produces: Consolidated MultiQC and HTML reports.
Why it matters: Share results, and context, as a link, not a transfer: one readable overview researchers, cores and collaborators can open directly.

Your deliverables

What comes back

Annotated Seurat object (RDS) and H5AD export for Python
Cell-type labels, T-cell states, and marker gene tables
Differential expression tables per cluster and per treatment
QC gate and integration check tables
MultiQC and HTML reports per experiment

Common questions

Single-cell RNA-seq FAQ

What the pipeline accepts, how cell types are decided, what comes back, and how to read the report once it opens.

Still deciding what your data needs? Email the team.

Take the next step

Run it in Horizon

Single-cell RNA-seq is available in Horizon. Sign in to work with your data and launch a run.

Open Horizon

Need a tailored pipeline?

Talk with our team about building or adapting a workflow for your research, including deployment in your AWS environment.

Contact the team