Single Cell Cell Cycle Scoring

Single cell cell cycle scoring is the computational assignment of cell cycle phases (G1, S, G2/M) to individual cells using transcriptomic signatures from single cell RNA sequencing (scRNA-seq) data. This guide is for researchers and bioinformaticians who need to account for cell cycle heterogeneity in their scRNA-seq analyses or to study proliferation dynamics. You will learn the core concepts, decision points, a practical workflow, common pitfalls, and the limits of interpretation. For authoritative technical references on biomedical data, see the NCBI Bookshelf.

At a Glance

Core Concepts

The cell cycle is divided into G1 (growth), S (DNA synthesis), G2 (preparation for mitosis), and M (mitosis). In single cell transcriptomics, cells from a population will be in different phases, introducing variability that can obscure biological signals of interest such as cell type differences or disease states. Scoring methods leverage sets of marker genes whose expression peaks during specific phases. For example, histones peak in S phase, while TOP2A and cyclin B peaking in G2/M. The official EMBL-EBI Training resources offer introductory modules on single cell data analysis that cover these concepts.

Two broad approaches exist. First, gene set scoring methods compute an aggregate expression score for each phase’s marker set. Seurat’s CellCycleScoring uses a list of known markers to produce G1/S and G2/M scores, then assigns the phase with the highest score. The scran package’s cyclone function applies a pre trained classifier using pairs of marker genes. Second, machine learning methods, including deep learning, train a model to predict phase from the full transcriptome or a selected feature set. A recent comprehensive study on deep learning models for cell cycle phase prediction from single cell RNA sequencing data demonstrates that neural network architectures can outperform simple scoring, especially when training data is representative.

Decision Criteria

When to Score and When to Regress

A critical decision is whether to regress out cell cycle effects before downstream analysis. If your biological question concerns cell type identity, differentiation, or treatment response not related to proliferation, regressing out phase scores (e.g., using Seurat’s ScaleData with vars.to.regress) can remove this confound. However, if you are studying proliferation, cancer stemness, or cell cycle regulation, retain the phase information and model it explicitly. The Galaxy Training Network provides workflow examples for scRNA-seq preprocessing that include optional cell cycle regression.

Choice of Marker Set

Most software uses curated human marker lists (e.g., from Tirosh et al. 2016). For mouse or other species, homology based conversion often works, but validation is advised. For disease specific contexts, such as acute myeloid leukemia, custom markers may be derived, a clinical study on randomized phase 2 trial of CPX 351 vs. CLAG M for medically unfit adults with acute myeloid leukemia illustrates how proliferation markers inform prognosis. Similarly, a study on a transient receptor potential related gene model for osteosarcoma used cell cycle genes for prognostic assessment. These examples show that marker selection must be tailored to the tissue and condition.

Method Selection

For most users, Seurat’s CellCycleScoring is simple and sufficient. If you observe batch specific phase biases or need to handle complex datasets, scran::cyclone offers robust classification. For large datasets or when you want to incorporate phase into a latent space, deep learning models (as in the deep learning study) can improve accuracy. However, deep learning requires training data and computational resources.

Practical Implementation Workflow

The following workflow assumes you have a normalized scRNA-seq expression matrix. Detailed code examples are available in the Bioconductor documentation for scran and in Seurat vignettes.

Step 1: Preprocessing and Quality Control

Load your data (e.g., from the NCBI Sequence Read Archive ). Filter low quality cells based on library size, number of genes, and mitochondrial percentage. Normalize (log transform or scale).

Step 2: Obtain Cell Cycle Marker Genes

Use a curated list. For human, the standard list includes:

• G1/S: MCM5, PCNA, TYMS, MCM2, etc.
• G2/M: TOP2A, MKI67, CCNB1, CCNB2, AURKA, etc.

You can also derive tissue specific markers from literature, such as the cancer associated fibroblast signatures linked to cell cycle in a single cell and machine learning framework for pancreatic cancer. This study identified CAFs associated genes that correlate with proliferation.

Step 3: Compute Phase Scores

Using Seurat:

library(Seurat)
s.genes <- c("MCM5", "PCNA", ...)  # G1/S markers
g2m.genes <- c("TOP2A", "MKI67", ...)  # G2/M markers
obj <- CellCycleScoring(obj, s.features = s.genes, g2m.features = g2m.genes)

The function returns G1.S and G2.M scores and a Phase column (G1, S, G2.M).

Using scran:

library(scran)
assignments <- cyclone(sce, pairs = mm.pairs)  # pre computed pairs

cyclone assigns G1, S, or G2.M based on pairs of genes.

Step 4: Quality Checks

Run PCA or UMAP and color by the assigned phase. You should see a continuous trajectory (especially if using a cell cycle sorted dataset) or at least a separation of S and G2/M from G1. If cells cluster by phase in a batch specific manner, a batch correction step may be needed before scoring.

Step 5: Downstream Use

If regressing, use the phase scores as covariates. If studying cycle, keep the phase labels or scores as a variable for differential expression or trajectory analysis. A causal analysis of breast cancer and coronary heart disease using Mendelian randomization and transcriptome data included cell cycle gene expression as a potential mediator, showing how phase information can be integrated into larger models.

Common Mistakes

1. Using default markers for divergent species. Human markers often work for mouse, but not for distant species. Always check homology or literature. For non model organisms, you may need to run a preliminary clustering and identify cycling cells by known proliferation markers.
2. Ignoring batch effects. Cell cycle scoring can be heavily influenced by sequencing batches. If you score each batch separately and then merge, you may get artificial phase clusters. Batch correct your data first or use a method robust to batch (e.g., scran’s cyclone with batch specific marker pairs).
3. Overinterpreting continuous scores as precise state. The G1/S score is not a true cell cycle coordinate. Cells can be in early or late G1, and the transition to S is gradual. Using discrete assignment loses this nuance.
4. Regressing out genuine biology. If your condition alters proliferation (e.g., drug treatment, oncogene activation), regressing out cell cycle can remove the very signal you want to study. Always check whether phase distribution changes across conditions before deciding to regress.

Limits of Interpretation

Cell cycle scoring from scRNA seq has several inherent uncertainties. First, the distinction between G0 (quiescence) and G1 is not reliable because quiescent cells show a similar transcriptome to early G1. Many methods assign all non cycling cells as G1, which conflates quiescence and active cycle entry. Second, the scoring is based on bulk phase markers, a cell in late G1 may express some S phase genes prematurely, leading to ambiguous calls. Third, the method assumes that phase markers are universally expressed across cell types, but in some lineages (e.g., post mitotic neurons) many markers are low or absent, causing misassignment to G1. Fourth, deep learning models, while more accurate, require a well annotated training set and may not generalize to new conditions or tissues. The deep learning paper notes that model performance degrades when applied to datasets with different sequencing depths.

Finally, cell cycle is a continuous process, not a discrete state. A single library captures a snapshot of RNA abundance that may not precisely reflect the phase because of transcript delays. For high resolution temporal studies, methods like FUCCI reporters or live imaging remain superior.

Frequently Asked Questions

Q1: Should I always regress out cell cycle effects?
No. Regress only if proliferation is not your biological focus and if phase related variation confounds your analysis (e.g., clustering by cell type). Always verify by checking whether phase distribution differs between groups. If it does, consider modeling rather than removing the effect.

Q2: Can I use cell cycle scoring on data from non human species?
Yes, but with caution. For mouse, the standard human marker lists are largely conserved. For zebrafish, Drosophila, or plants, you need to curate orthologs. The Galaxy Training Network provides instructions for using orthology mapping tools.

Q3: What if my cells are mostly in G1 and the scores show no variation?
This is typical for non proliferative tissues or differentiated cell populations. It may indicate that cell cycle effects are minimal. You can proceed without regression. However, if you expect cycling (e.g., in stem cells or tumors) and see no variation, check your marker list or data quality.

Q4: Can cell cycle scoring be performed on single nucleus RNA seq (snRNA seq)?
Yes, but nuclear RNA retains mostly unprocessed transcripts and may underrepresent short lived cell cycle mRNAs. The scores may be noisier. Some studies have used snRNA seq for cell cycle analysis with adjusted marker sets. Validate against expected biology.

References and Further Reading

• NCBI Bookshelf. Free biomedical books and technical references on cell biology and sequencing methods. link
• EMBL EBI Training. Single cell RNA seq analysis courses including cell cycle scoring. link
• Galaxy Training Network. Workflow based tutorials for scRNA seq preprocessing and cell cycle analysis. link
• Bioconductor. Software packages and vignettes for scran and cell cycle classification. link
• Deep learning models for cell cycle phase prediction from single cell RNA sequencing data. Brief Bioinform 2024. link
• A single cell and machine learning framework identifies CAFs associated signatures linking stromal heterogeneity to immune regulation in pancreatic cancer. J Gastrointest Oncol 2024. link
• Randomized phase 2 trial of CPX 351 vs. CLAG M for medically unfit adults with acute myeloid leukemia. Leukemia 2024. Demonstrates clinical relevance of proliferation markers. link
• A Transient Receptor Potential related gene model for prognostic assessment in osteosarcoma. Curr Med Chem 2024. Uses cell cycle genes in gene model. link
• Causal relationship and pathogenic mechanism analysis of breast cancer and coronary heart disease using transcriptome data. Immunogenetics 2024. Includes cell cycle expression analysis. link
• NCBI Sequence Read Archive. Repository for raw scRNA seq data. link

Related Articles

• Structural Variant Validation
• Repeat Expansion Sequencing
• Copy Number Variant Analysis
• Somatic Mosaic Variant Calling
• Germline Variant Filtering