Single-Cell Unified Polarization Assessment of Immune Cells

Information Science Computer Science Immunology Life Sciences Cell Biology
single-cell RNA sequencing immune cell polarization cytokine-driven computational method tumor infiltration transcriptome cell embeddings

Immune cells undergo cytokine-driven polarization in response to various stimuli, which alters their transcriptional profiles and functional states. This dynamic process plays a central role in immune responses in both health and disease. However, there has been a lack of systematic methods to assess cytokine-driven polarization in single-cell RNA sequencing (scRNA-seq) data. To address this gap, researchers developed the Single-Cell Unified Polarization Assessment (SCUPA), the first computational method for comprehensive evaluation of immune cell polarization.

Source of the Paper

The paper was co-authored by Wendao Liu and Zhongming Zhao, affiliated with The University of Texas MD Anderson Cancer Center UTHealth Houston Graduate School of Biomedical Sciences and Center for Precision Health, McWilliams School of Biomedical Informatics, The University of Texas Health Science Center at Houston, respectively. The paper was published on February 25, 2025, in the journal Bioinformatics.

Research Process

1. Data Collection and Preprocessing

The researchers first downloaded the Immune Dictionary scRNA-seq dataset from the Single Cell Portal, which contains single-cell transcriptomic data of immune cells in mouse lymph nodes treated with 86 cytokines. Additionally, other relevant datasets were downloaded from SeuratData and the Gene Expression Omnibus (GEO) database, including datasets on IFN-β-treated human peripheral blood mononuclear cells (PBMCs), cytokine-treated human macrophages, IL-2-treated mouse spleen, and pan-cancer infiltrating myeloid cells.

2. Generating Cell Embeddings and Dimensionality Reduction

The researchers used the single-cell foundation model Universal Cell Embeddings (UCE) to generate cell embeddings for all scRNA-seq datasets. To reduce dimensionality, they performed Principal Component Analysis (PCA) on the UCE cell embeddings and used the top 20 principal components as input features for machine learning models. Additionally, two-dimensional UMAP plots were generated for data visualization.

3. Identifying Fully Polarized Cells

The researchers identified fully polarized cells for each polarization state based on three criteria: (1) the cell is from a sample treated with a specific driving cytokine; (2) the mean expression of polarization marker genes in the cell is higher than in most other cells; and (3) the UCE cell embeddings of the cell are similar to those of other cells from samples treated with the driving cytokine. By calculating “embedding shifts” and cosine similarity, the researchers filtered out fully polarized cells, which were then used for training machine learning models.

4. Training and Testing Machine Learning Models

The researchers tested several machine learning models, including logistic regression, support vector machines (SVM), random forest, and semi-supervised learning. Ultimately, the SVM model was chosen due to its superior performance across all polarization states. During training, unpolarized cells were labeled as 0, while fully polarized cells were labeled as 1. The models were trained and tested 20 times, and the mean AUROC values were calculated.

5. Quantifying Statistical Uncertainty

Since immune cell polarization is a continuous process, the researchers used conformal prediction to quantify statistical uncertainty in polarization assessment. By calculating nonconformity scores, the model could predict each cell as polarized, unpolarized, intermediate, or uncertain.

6. Cross-Dataset Batch Effect Correction

To enhance SCUPA’s transferability across datasets, the researchers provided a simple and effective method for cross-dataset batch effect correction. By adjusting the UCE cell embeddings, the model could bypass complex data integration processes while preserving polarization information.

7. Benchmarking Single-Cell Foundation Models

The researchers compared the performance of three single-cell foundation models—UCE, scGPT, and scFoundation—in predicting immune cell polarization using cell embeddings. The results showed that UCE and scFoundation models performed similarly in predicting polarization states, while the scGPT model had lower performance.

Key Results

1. SCUPA Framework and Immune Cell Polarization States

SCUPA uses the immune cell polarization states from the Immune Dictionary as a reference to train machine learning models to distinguish between polarized cells from cytokine-treated samples and unpolarized cells. By leveraging UCE cell embeddings, SCUPA can effectively capture polarized cells across different species and experimental conditions.

2. SCUPA’s Performance in In Vitro Cytokine Stimulation Datasets

In the IFN-β-treated human PBMC dataset, SCUPA accurately classified stimulated and unstimulated cells, with AUROC values exceeding 0.99. In the cytokine-treated human macrophage dataset, SCUPA identified different cytokine-driven polarization states and revealed correlations between these states.

3. Application of SCUPA in In Vivo Cytokine Treatment Datasets

In the IL-2-treated mouse spleen dataset, SCUPA revealed IL-2-driven polarization states and found that IL-2 treatment significantly increased the proportion of polarized cells.

4. SCUPA’s Analysis of Pan-Cancer Infiltrating Myeloid Cell Dataset

In the pan-cancer infiltrating myeloid cell dataset, SCUPA revealed the polarization states and proinflammatory responses of myeloid cells across different cancer types. For example, myeloid cells in lymphoma exhibited the highest proinflammatory polarization scores, while those in pancreatic adenocarcinoma and kidney cancer had lower polarization scores.

Conclusion

SCUPA is the first computational method for comprehensive assessment of immune cell polarization, leveraging cell embeddings from the single-cell foundation model UCE to effectively capture transcriptional changes associated with different polarization states. The method has been validated in multiple independent datasets and has revealed polarization characteristics of tumor-infiltrating myeloid cells across various cancers. SCUPA provides a new tool for studying immune cell polarization, with significant potential for applications in cytokine-based therapies.

Highlights of the Study

  1. First Computational Method for Systematic Immune Cell Polarization Assessment: SCUPA fills the gap in assessing cytokine-driven polarization in scRNA-seq data.
  2. Leveraging the Single-Cell Foundation Model UCE: By using UCE cell embeddings, SCUPA can capture polarized cells across different species and experimental conditions.
  3. Validation in Multiple Independent Datasets: SCUPA performed excellently in both in vitro and in vivo experimental datasets, accurately classifying polarized cells.
  4. Revealing Polarization Characteristics of Tumor-Infiltrating Myeloid Cells: SCUPA revealed polarization states and proinflammatory responses in myeloid cells across different cancer types.

Additional Valuable Information

The code for SCUPA has been publicly released on GitHub (https://github.com/bsml320/scupa), allowing researchers to freely use and extend the method. Additionally, SCUPA is designed to integrate seamlessly with the widely used Seurat pipeline, facilitating comprehensive single-cell data analysis.