By EVOBYTE Your partner in bioinformatics
Introduction
If bulk genomics is a city skyline, single‑cell omics is walking the streets. Instead of one average signal, we see thousands to millions of individual cells, each with its own transcriptome, chromatin accessibility, or surface proteins. That granularity is changing how we study tumors, immune responses, and development. Yet for many data scientists, single‑cell datasets feel unfamiliar. The matrices are sparse, the preprocessing is opinionated, and the jargon—scRNA‑seq, scATAC‑seq, CITE‑seq, UMI, UMAP—arrives fast.
This guide explains what “single‑cell omics” actually includes, why it’s different from bulk data, and how to build a credible analysis pipeline. We’ll touch on key acronyms, point out common pitfalls, and share tiny code examples you can adapt immediately.
What “single‑cell omics” really means
Single‑cell omics refers to molecular profiling at cellular resolution. In practice, the workhorse is single‑cell RNA sequencing, or scRNA‑seq, which measures gene expression with unique molecular identifiers (UMIs) to count transcripts per cell. Alongside it, single‑cell ATAC‑seq (scATAC‑seq) captures chromatin accessibility, revealing which regulatory regions are open. Techniques like CITE‑seq add antibody‑derived tags to quantify surface proteins together with RNA, while emerging “multiome” assays jointly profile RNA and ATAC in the same cell. Spatial transcriptomics maps expression back onto tissue coordinates, trading some throughput for location context.
These modalities complement one another. RNA tells you what a cell is doing, ATAC hints at how it’s regulated, and proteins anchor phenotype. Together, they sharpen clustering, stabilize cell‑type annotation, and expose trajectories that RNA alone can blur.
How single‑cell data changes the analysis game
Single‑cell matrices are sparse by design. Each droplet or well captures only a fraction of the transcripts, so zero inflation is common. Batch effects become more visible because small technical differences shift clusters. Doublets—two cells captured as one—distort expression and must be filtered. And because every cell is a sample, naïve statistical tests inflate significance unless you aggregate or use mixed models.
Preprocessing reflects these realities. Instead of library‑size normalization alone, we typically use per‑cell QC (mitochondrial percentage, total counts, detected genes), followed by count‑aware normalization and variance stabilization. Dimensionality reduction (PCA first, then UMAP or t‑SNE) and graph‑based clustering replace classic distance‑based methods. For interpretability, marker‑gene enrichment and reference‑based label transfer help assign biological names to clusters.
Here’s a minimal Python walkthrough to ground those steps.
import scanpy as sc
adata = sc.read_10x_mtx("path/to/sample/") # counts: cells x genes
adata.var["mt"] = adata.var_names.str.startswith("MT-") # human mt genes
sc.pp.calculate_qc_metrics(adata, qc_vars=["mt"], inplace=True)
adata = adata[ (adata.obs.n_genes_by_counts > 200) &
(adata.obs.pct_counts_mt < 10) ].copy()
sc.pp.normalize_total(adata, target_sum=1e4)
sc.pp.log1p(adata)
sc.pp.highly_variable_genes(adata, n_top_genes=3000, flavor="seurat_v3")
adata = adata[:, adata.var.highly_variable].copy()
sc.pp.scale(adata, max_value=10)
sc.tl.pca(adata, n_comps=50)
sc.pp.neighbors(adata, n_neighbors=15, n_pcs=30)
sc.tl.umap(adata)
sc.tl.leiden(adata, resolution=0.6)
With a few lines, you’ve moved from raw counts to clusters on a UMAP, ready for marker discovery and annotation.
Putting modalities together: integration and multi‑omics
Most projects span multiple runs, donors, or conditions. Integration methods correct batch effects while preserving biology. On the RNA side, canonical correlation analysis and mutual nearest neighbors remain reliable, while modern graph integration methods scale to millions of cells. For scATAC‑seq, peak‑by‑cell matrices are even sparser, so we often aggregate peaks into gene activity scores before co‑embedding with RNA. In true multiome datasets, we align modalities within the same cell, linking open chromatin to its putative target genes. That lets you move from correlation to plausible regulation.
If you work in R, this tiny snippet shows the rhythm of RNA integration with a popular toolkit:
library(Seurat)
objs <- lapply(list("sample1","sample2"), function(p) Read10X(p) |> CreateSeuratObject())
objs <- lapply(objs, SCTransform, verbose=FALSE)
features <- SelectIntegrationFeatures(objs, nfeatures=3000)
objs <- PrepSCTIntegration(objs, anchor.features=features)
anchors <- FindIntegrationAnchors(objs, normalization.method="SCT",
anchor.features=features)
combined <- IntegrateData(anchors, normalization.method="SCT")
combined <- RunPCA(combined) |> RunUMAP(dims=1:30) |> FindNeighbors(dims=1:30) |> FindClusters()
Once modalities share a common space, you can transfer labels from a curated reference, stabilize annotations across studies, and quantify differential abundance without folding technical noise into your effect sizes.
A quick story: finding hidden resistance in a tumor biopsy
Imagine a lung tumor biopsy with ten thousand cells. Bulk RNA‑seq reports a strong interferon signature and suggests an inflamed microenvironment. After scRNA‑seq, the landscape looks different. Most malignant cells form a dominant cluster with druggable EGFR signaling. A smaller pocket shows high epithelial‑mesenchymal transition scores and chromatin opening around TGF‑β enhancers in the matched scATAC‑seq. When treatment starts, that minority expands. You didn’t just detect resistance after the fact; you saw its regulatory wiring at baseline and flagged a combination strategy. This is why multi‑omics matters: expression points to state, and accessibility points to levers.
Practical tips that save projects
Start with honest QC and don’t fear dropping cells. A clean forty thousand beats a noisy hundred thousand. Use cell‑level covariates sparingly in downstream models; over‑regressing can erase biology. Keep raw counts and full metadata in versioned files, and pin package versions in your environment to ensure figures are reproducible months later. When annotating clusters, mix automated label transfer with manual marker checks, and write down your rationale. For scalability, store matrices in chunked formats and compute on the cloud only what you can’t compute locally. Finally, plan your question before your pipeline. Pseudotime, trajectory inference, and RNA velocity shine in developmental contexts, while differential abundance and cell–cell interaction models suit tumor microenvironments.
Summary / Takeaways
Single‑cell omics turns averages into distributions, revealing cell states, rare populations, and regulatory programs that bulk assays miss. The core ideas are straightforward once you accept sparsity and batches as first‑class citizens. With scRNA‑seq for expression, scATAC‑seq for regulation, and multi‑omics to connect the two, you can build analyses that hold up under replication and guide experiments rather than chase noise. Start with clean QC, integrate carefully, annotate transparently, and keep your code and metadata reproducible. What single‑cell question would have changed your last project’s conclusion? Start there next.