Single-Cell RNA-Seq: From 10x Data to Results Without Coding

You ran your 10x Chromium experiment. The core facility handed you a folder full of FASTQ files and a link to the Cell Ranger documentation. You opened it, saw "install Cell Ranger (requires 64 GB RAM and 8 CPU cores minimum)," and realized your MacBook isn't going to cut it.

Then you looked at the downstream analysis. Scanpy tutorials assume you know Python. Seurat tutorials assume you know R. Both assume you understand what "Leiden resolution 0.5" means and why it matters for your UMAP. You just want to know which cell types are in your tumor samples.

This guide explains what single-cell RNA-seq analysis actually involves, where the hard decisions are, and how to get from raw 10x data to annotated cell clusters without writing code.

What Single-Cell RNA-Seq Tells You

Bulk RNA-seq gives you the average gene expression across millions of cells. That's useful, but it's like measuring the average temperature of a hospital — it tells you nothing about who has a fever.

Single-cell RNA-seq (scRNA-seq) measures gene expression in individual cells. This lets you:

• Identify cell types — Distinguish T cells from macrophages from fibroblasts in a mixed tissue sample, based on their gene expression profiles
• Discover cell states — Find activated vs. exhausted T cells, or cycling vs. quiescent stem cells
• Find rare populations — Detect cell types that make up less than 1% of your sample, which bulk RNA-seq would completely miss
• Compare conditions — See how the cellular composition of a tumor changes after treatment, at single-cell resolution

If you're studying the tumor microenvironment, scRNA-seq can tell you not just that "immune genes are upregulated" but that there's a specific population of exhausted CD8+ T cells expressing PD-1 and LAG-3 that expands after immunotherapy.

The Traditional Pipeline

Here's the standard workflow that a computational biologist would run. Each step has tools, parameters, and judgment calls.

Step 1: Raw Read Processing (Cell Ranger)

Cell Ranger aligns reads to a reference genome, assigns reads to individual cells using barcodes, and counts transcripts per gene per cell. The output is a count matrix — rows are genes, columns are cells, values are counts. Computationally heavy (64+ GB RAM, several hours) but straightforward.

Step 2: Quality Control and Filtering

This is where it gets subjective. You need to remove empty droplets (barcodes with ambient RNA but no cell), dead cells (high mitochondrial gene percentage), doublets (two cells in one droplet), and low-quality cells.

The hard part: there are no universal thresholds. A mitochondrial cutoff of 10% might be right for PBMCs but too aggressive for cardiomyocytes (which are naturally mitochondria-rich). Filtering too aggressively removes real biology. Filtering too leniently keeps noise that distorts your clusters.

Step 3: Normalization and Feature Selection

Normalize counts so you can compare cells with different sequencing depths, then select highly variable genes — the 2,000-3,000 genes that vary most across cells. These drive the clustering.

Step 4: Dimensionality Reduction and Clustering

This produces the UMAP plots everyone puts in their papers: PCA to reduce dimensions → neighbor graph connecting similar cells → Leiden/Louvain clustering → UMAP projection into 2D.

The critical parameter is clustering resolution. Too low and you merge distinct cell types. Too high and you split one cell type into meaningless subclusters. There's no formula — you adjust and check whether clusters make biological sense.

Step 5: Cell Type Annotation

Clustering gives you "Cluster 0, 1, 2..." — not cell type names. You find marker genes per cluster, compare to known signatures, and assign labels. This is the most expertise-dependent step. Automated tools like CellTypist help for well-characterized tissues but struggle with unusual cell states.

Step 6: Differential Expression

With annotated clusters, you ask: which genes differ in T cells from treated vs. untreated tumors? Pseudobulk approaches are recommended for multi-sample designs.

Why This Is Hard for Non-Computational Researchers

The pipeline involves installing Python/R environments with dozens of dependencies, running code that needs 32-128 GB of RAM, making parameter choices that require experience, and debugging cryptic errors when something breaks. A computational postdoc might spend 1-3 weeks on a new dataset — not because they're slow, but because the iterative process of filtering, clustering, checking, and adjusting takes time.

How GeneChef Handles It

GeneChef runs the entire pipeline on cloud infrastructure through Galaxy, with the AI assistant helping you set up the right workflow for your specific experiment.

Step 1: Describe Your Experiment (~2 minutes)

Open the AI chat and describe what you have:

"I have 10x Chromium 3' v3 scRNA-seq data from 6 mouse lung samples — 3 bleomycin-treated and 3 PBS controls. I expect epithelial cells, immune cells, and fibroblasts. I need clustering, cell type annotation, and differential expression between conditions."

The AI uses this context to select appropriate tools and parameters. Mentioning "mouse lung" matters — it affects the reference genome, expected mitochondrial thresholds, and which cell type reference databases to use.

Step 2: AI Builds the Workflow (~1 minute)

The AI constructs a Galaxy workflow that chains together the appropriate tools: Cell Ranger (or STARsolo) for alignment and counting → quality filtering with tissue-appropriate thresholds → normalization → PCA → clustering → marker gene detection → automated cell type annotation → differential expression.

You see the full workflow as a visual diagram before running anything. Every parameter is visible and editable. If you know you want a specific mitochondrial threshold or clustering resolution, you can change it.

Step 3: Upload and Run (~5 minutes to start)

Upload your FASTQ files through the browser or provide S3/URL links. GeneChef runs the analysis on cloud infrastructure — no need to worry about RAM. The platform scales to handle datasets with 100,000+ cells.

For a typical 10x dataset (6 samples, ~30,000 cells total), expect 3-6 hours for the full pipeline.

Step 4: Interpret Your Results

GeneChef returns:

• UMAP plots — Colored by cluster, by sample, by condition, and by cell type annotation
• Cluster marker tables — Top differentially expressed genes per cluster with log fold changes and adjusted p-values
• Cell type proportions — Bar charts showing the fraction of each cell type per sample and per condition
• Differential expression results — Genes up- or down-regulated in each cell type between your conditions
• QC metrics — Violin plots of genes per cell, UMIs per cell, and mitochondrial percentage, so you can verify the filtering was appropriate

All results are downloadable for further analysis or figure preparation.

A Realistic Walkthrough

Here's what a typical analysis looks like end-to-end:

| Step | Time | What Happens | |------|------|-------------| | Describe experiment in AI chat | 2 min | AI builds Galaxy workflow | | Review and adjust workflow | 5 min | Check parameters, modify if needed | | Upload 6 FASTQ file pairs | 10 min | Browser upload or URL import | | Cell Ranger alignment | 2-3 hrs | Runs on cloud per sample | | QC, normalization, clustering | 30-60 min | Tissue-appropriate defaults | | Cell type annotation + DE | 30-60 min | Automated per cell type | | Total | ~4-5 hours | Mostly unattended compute |

Your active time is about 20 minutes. The rest is compute.

Cost and Time Comparison

| Approach | Setup Time | Analysis Time (6 samples) | Cost | |----------|-----------|--------------------------|------| | Hire a bioinformatician | 1-2 weeks | 1-3 weeks | $2,000-6,000 (consulting) | | Learn Scanpy/Seurat yourself | 1-3 months | 1-2 weeks | Free (but your time isn't) | | GeneChef | ~5 minutes | 4-5 hours compute | $299/month Professional |

Common Use Cases

Tumor microenvironment — Characterize immune infiltration, identify exhausted T cell populations, compare cellular composition between responders and non-responders to immunotherapy.

Developmental biology — Trace cell lineages during organogenesis, identify progenitor populations, map differentiation trajectories with RNA velocity.

Immune profiling — Characterize immune cell subsets in autoimmune disease, infection, or vaccination. PBMCs are the best-characterized tissue for automated annotation, so results tend to be reliable.

Drug response — Compare cellular responses to treatment at single-cell resolution. Find which cell types respond, which resist, and which new populations emerge.

When You Still Need a Human Expert

Automated pipelines handle the standard workflow well, but some analyses require expert judgment:

• Ambiguous cell types. If your clusters don't match known cell type signatures cleanly — maybe you're working with a poorly characterized tissue or a non-model organism — automated annotation will struggle. You'll need someone who knows the biology to interpret marker genes manually.
• Batch effects. Samples processed on different days or chips can show batch effects that dominate the biology. Integration methods (Harmony, scVI) help, but choosing the right one and validating it requires experience.
• Trajectory analysis. RNA velocity and pseudotime are powerful but finicky — sensitive to parameters and misleading if applied to data without a continuous trajectory.
• Multi-modal data. CITE-seq, Multiome, or spatial transcriptomics require substantially more complex and less standardized analysis.
• Very large datasets. GeneChef scales to 100K+ cells, but atlas-scale projects (500K+) may need specialized approaches and expert oversight.

Dropout, Sparsity, and Other Caveats

scRNA-seq data is inherently noisy:

• Dropout. Many expressed genes show zero counts because the mRNA wasn't captured. Absence of expression doesn't necessarily mean the gene is off.
• Ambient RNA contamination. Lysed cells release RNA into the suspension, adding background expression to every droplet. Tools like SoupX or CellBender can correct for this.
• Spatial information is lost. Standard scRNA-seq dissociates tissue into single cells — you know the cell types but not where they were. If spatial context matters, consider Visium or MERFISH as a complement.

Getting Started

If you have 10x data ready to analyze: check your data type with your core facility (3' v3? 5'? Multiome?), confirm the reference genome (GRCh38 for human, GRCm39 for mouse), and prepare a metadata spreadsheet with sample IDs, conditions, and batches. Start with the standard clustering pipeline — you can always add trajectory analysis or integration later.

GeneChef offers a free 14-day trial with full access to the Professional tier. Upload your 10x data, describe your experiment, and get annotated cell clusters back in an afternoon.

---

GeneChef is a managed bioinformatics platform built on Galaxy. It runs on AWS cloud infrastructure with AI-powered workflow building, so you can go from raw sequencing data to results without installing software or writing code.