Same Data, Different Story: cell annotation in single-cell sequencing

Single-cell RNA-seq offers a more precise view than bulk RNA-seq because it measures transcripts in individual cells rather than averaging signals across a mixture. Sounds perfect, right?

However, a simple question still bothers researchers a lot: Which cell is which? How to label the correct the cell-type to the measured cells?

NO, WE WILL NEVER KNOW THE TRUTH, but we can for sure have some rational guesses.

TLDR;

If you don't have time reading to the end, the take home message is:

There is no single, fully automated pipeline that can definitively determine cell types. Instead, results from different approaches should be harmonized, with final annotation relying on careful evaluation and validation by human experts.

What is the general practice?

Fig. 1: Examples of GPT-4’s cell type annotation and comparisons with other methods. | Nature Methods

1. Expert approach: In some studies, investigators define target cell populations a priori and identify them according to a specific hypothesis and associated marker genes. This strategy demands substantial biological background and domain knowledge.
2. Computational approach: The most widely used approaches require less prior biological knowledge. Bioinformaticians can annotate cells using marker-gene scoring or reference-based mapping, often helped by machine-learning models. But…. the results can differ from infinite dataset * algorithm combinations. We’ll examine several options later in this article.
3. LLM approach: Can you believe that we can use large language model to annotate the cells? One of the newest publication form 2024 just use the GPT-4

Assessing GPT-4 for cell type annotation in single-cell RNA-seq analysis | Nature Methods

What I do?

I use both reference data and marker genes approaches under python-based framework scanpy to annotation my data

• Reference Data: CellTypist is a automated cell type annotation method. It has 53 models (until Aug 2025) including human and mouse in different tissues such as brain, heart, liver, lung.

CellTypist | automated cell type annotation for scRNA-seq datasets

• Marker Genes: python-based decopuler is used for import the dataset and apply the ML algorithm to annotate the cells

Dataset:

1. PanglaoDB: a database for the scientific community interested in exploration of single cell RNA sequencing experiments from mouse and human. It is from Karolinska Institutet from Sweden. Sadly, the project is not maintained nor funded since 2020.
2. Cellmarker 2.0: the updated release of the CellMarker database: a manually curated catalogue of experimentally supported marker genes for cell types across human and mouse tissues.

ML Algorithm:

1. AUCell: AUCell uses the “Area Under the Curve” (AUC) to calculate whether a critical subset of the input gene set is enriched within the expressed genes for each cell. The score is between 0 and 1.
2. ULM: This approach uses the molecular features from one observation as the population of samples and it fits a linear model with a single covariate. The score is the beta value (slope) of the linear regression.

How to assign the cell type from marker gene method?

Individual cell assignment

Our input data is a table with Cell * Gene, and we need the a clean public dataset with cell type as source and marker genes as target. Then we use ML method to generate a classification score table for each cells. This can be used for identify the cell type individually.

After that, we can use majority vote or statistical test to identify the high probably cell types for each cluster.

Cluster cell assignment

When we have the individual cell type classification, we would like to assign to the same cell type into cluster. Voting by the majority or rank based t-test.

All the analysis process can be found by the tutorial here:

Single-cell Enrichment Analysis — decoupler

Real Data Analysis

I use data from CYCLIN-B1/2 and -D1 act in opposition to coordinate cortical progenitor self-renewal and lineage commitment | Nature Communications

After cluster from Leiden resolution 1, I get the cluster as this

Results

I list the first 3 high probability cell types from each methods. As you can see, there is no general agreement across methods or datasets. The labels are not standardized, and even the level of annotation differs — in some cases cell types are defined at a broad level, while in others they are subdivided into more detailed subtypes.

Conclusion

Sadly, there is no conclusion here.

Cell annotation can be performed in countless ways — using large language models, manual curation, or computational methods. Even within computational approaches, there are numerous frameworks (R /Python), an ever-growing number of reference datasets, and many black- or white-box machine learning pipelines.

As Deng Xiaoping said

it doesn’t matter if a cat is black or white, if it catches mice it’s a good cat

Ultimately, regardless of the computational strategy used, expert evaluation is essential. A clear understanding of the hypothesis, the relevant biological context, the experimental design, and the characteristics of the data is key for cell type annotation.