3063 - Enhanced Detection of Rare Cells in Hodgkin Lymphoma Using TabMap for Single-Cell RNA Sequencing Analysis
Presenter(s)
A. Heider1, S. Su2, R. Yan2, A. Subramanian2, L. Xing1, and M. S. Binkley3; 1Department of Radiation Oncology, Stanford University, Stanford, CA, 2Stanford University, Stanford, CA, 3Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA
Purpose/Objective(s): Rare immune cell populations including gamma delta T-cells (gdT-cells) have been implicated in predicting responses to radiotherapy and immunotherapy. We hypothesize that single cell and bulk RNA-seq aided by the machine-learning algorithm, TabMap, will allow for phenotyping of the tumor microenvironment with elucidation of rare, clinically relevant cells. We sought to apply TabMap to RNA-seq data from patients with Hodgkin lymphoma (HL) as a disease model containing rare cell types.
Materials/Methods: For 10 HL samples we performed either single-cell or single-nucleus RNA sequencing. We isolated 71921 cells and 18517 nuclei. A TabMap was then made through the process of transforming a 1D gene expression vector by multiplying it with a projection matrix composed of 0s and 1s. After this transformation, the data is reshaped into a 2D representation called a TabMap optimizing gene-gene interactions for cell clustering. To test performance, we generated pseudo-bulk samples (rare cell spike-in ranging from 0-20%) by aggregating single-cell RNA expression using bootstrapping, where cell-type-specific counts were iteratively sampled and averaged. The final pseudo-bulk expression matrix was normalized to counts per million. We applied CIBERSORTx for digital deconvolution.
Results: Using TabMap, we identified 16 distinct cell types including rare (abundances <5%) Reed-Sternberg and gdT-cells. When compared to conventional UMAP clustering, TabMap demonstrated superior performance in detecting our rare cell types in pseudo-bulk analyses. For gdT-cells, our pseudo-bulk down-sampling had a slope of 0.85, compared to 0.68 for UMAP (indicating ~17% higher accuracy for TabMap). For Reed-Sternberg cells, our pseudo-bulk down-sampling had a slope of 1.02, compared to 1.25 for UMAP (reducing overestimation by ~25%). For NK, Plasma, and B_Memory cells, pseudo-bulk down-sampling using TabMap had slopes of 0.79, 0.78, and 0.85, respectively, compared to 0.75, 1.75, and 0.5 for UMAP (~4% to ~97% improvement for TabMap).
Conclusion: Cell phenotyping from next generation gene expression assays is becoming increasingly important to identify biological correlates of radiation response. We demonstrate for HL that TabMap outperforms conventional cell clustering algorithms in identifying rare cell types including gdT-cells that are notoriously hard to resolve in single cell RNA-seq data. We anticipate this accuracy in cell phenotyping will ultimately allow for improved ability to identify biologic correlates of response to radiotherapy and immunotherapy informing translational research aimed to improve patient outcomes.