Analyse single-cell RNA-seq data

What this involves

Single-cell RNA sequencing resolves transcriptional heterogeneity within tissues and cell cultures, but moving from cell-by-gene matrices to verified biology requires strict statistical discipline. Standard workflows quantify ambient contamination, identify cellular doublets, select highly variable genes, and partition cells into clusters. Because clusters are derived directly from the expression data, naive differential expression tests between clusters yield circular, invalid p-values, making external reference mapping or curated gene panels essential for authentic annotation.

Comparing conditions across samples introduces the most widespread point of failure in single-cell transcriptomics: treating individual cells as independent replicates. Doing so generates false statistical power and deflated p-values from minimal biological cohorts. A valid comparison requires aggregating counts per sample within each annotated cell type to construct pseudobulk profiles, which are subsequently modelled with negative binomial statistics while adjusting for batch effects.

The research agent executes these analyses on its own environment, handling alignments, filtering, dimensional reduction, and differential tests without local software installation. It queries reference databases such as CZ CELLxGENE and the Human Protein Atlas live to confirm cell-type identities, reads outputs back directly, and provides reproducible tabular summaries and figures.

How it goes

1. 1Quality control and doublet detectionFilter individual samples using sample-specific median absolute deviation (MAD) cutoffs for library size and gene counts, evaluate mitochondrial expression, and identify droplet doublets using nf-core/scrnaseq or nf-core/scdownstream.
2. 2Normalisation and dimensionality reductionApply log1p normalisation on counts-per-10k or pooled scran size factors, isolate approximately 2,000 highly variable genes, compute 30 to 50 principal components, and generate neighbourhood graphs alongside UMAP embeddings for visualisation.
3. 3Clustering and batch integrationEvaluate Leiden clustering across multiple resolutions (0.2 to 1.5) and introduce batch integration via Harmony or scVI only if shared cell populations visibly fracture by sample.
4. 4Cell-type annotationAssign labels using curated marker panels, Cell Ontology (cl) identifiers via search_ontology_terms, and reference comparisons against the CZ CELLxGENE atlas or the Human Protein Atlas, avoiding circular cluster-marker p-values.
5. 5Pseudobulk differential expressionSum raw integer counts per biological sample for each cell type and evaluate contrasts using PyDESeq2 with explicit design formulas, apeglm or ashr log-fold-change shrinkage, and false discovery rate control.

Start it here

Ask in plain language. The agent works on a Linux machine of its own — it installs and runs what the job needs, keeps every file, and reads the results back with the numbers explained. Nothing to set up, and a long run continues after you close the page.

What to watch for

• Treating single cells as independent biological replicates in differential expression tests deflates p-values and generates false-positive findings; counts must be summed per sample into pseudobulk profiles.
• Clustering resolution is a parameter choice rather than a biological result; sweep resolutions between 0.2 and 1.5 and reject clusters that appear at only a single setting.
• Marker p-values derived from testing clusters defined by those same expression data represent statistical double-dipping and should only be used to rank candidates for annotation.
• Applying batch correction algorithms such as ComBat prior to differential testing strips raw count distributions and invalidates negative binomial dispersion models; batch must be included in the design formula.
• UMAP projections distort distances and densities; trajectory lengths and biological distances must never be quantified directly from UMAP coordinates.
• Hard universal mitochondrial cutoffs inadvertently discard metabolically active or stressed cell populations; thresholds must be determined per sample using median absolute deviations.

What it runs

Questions