NK Cell Multimodal Single-Cell Analysis

Analysis of NK cell differentiation using CITE-seq data from Foltz et al. 2024 (Science Immunology).

Data Sources

• GSE264696: CITE-seq (NK cells, 28-protein panel)

Target Cell Types

• CD56bright NK cells
• CD56dim NK cells
• eML-1 (enriched memory-like, from CD56bright)
• eML-2 (enriched memory-like, from CD56dim)

Environment

conda env create -f envs/environment.yml
conda activate nk_analysis

Reproducibility

• Random seed: 42
• All dependencies pinned in environment.yml

Notebooks

01_preprocessing_qc.ipynb

• Loaded CITE-seq data from two donors (21,167 cells, 28-protein panel)
• Removed 31,053 mouse spike-in genes and applied QC filters (≥200 genes, <15% mito)
• Ran TOTALVI to jointly model 4,000 HVGs and 28 surface proteins
• Batch-corrected across donors, extracted 20-dimensional latent embeddings
• Annotated NK subsets via hierarchical protein gating: CD56bright (CD117+), CD56dim (CD57+CD16+), eML (NKG2A+ only)
• Final dataset: 19,443 cells with denoised protein expression

02_classifier_training.ipynb

• Trained classifiers on TOTALVI latent embeddings to predict NK subsets from RNA
• Compared BalancedBaggingClassifier and XGBoost on 3-class problem (excluding Unassigned)
• XGBoost achieved best performance: macro F1 = 0.789 (5-fold CV)
• eML classification challenging (F1 = 0.56) due to small sample size (4%) and phenotypic overlap
• Saved trained models for downstream inference

References

Foltz JA, Tran J, Wong P, Fan C, Schmidt E, Fisk B, Becker-Hapak M, Russler-Germain DA, Johnson J, Marin ND, Cubitt CC, Pence P, Rueve J, Pureti S, Hwang K, Gao F, Zhou AY, Foster M, Schappe T, Marsala L, Berrien-Elliott MM, Cashen AF, Bednarski JJ, Fertig E, Griffith OL, Griffith M, Wang T, Petti AA, Fehniger TA. Cytokines drive the formation of memory-like NK cell subsets via epigenetic rewiring and transcriptional regulation. Science Immunology. 2024 Jun 28;9(96):eadk4893. doi: 10.1126/sciimmunol.adk4893

Data availability:

• CITE-seq: GSE264696