Toward a Framework for Evaluating the Clinical Relevance of Cancer Cell Lines at Single-Cell Resolution

Abstract

Human cancer cell lines (CCLs) play a critical role in propagating our understanding of cancer biology and enabling preclinical testing of novel anti-cancer compounds. Previous studies have shown how myriad factors (e.g., media formulation, genetic and transcriptomic evolution) can compromise how faithfully CCLs represent in vivo phenotypes. Prior computational strategies have compared CCLs and TCGA bulk RNA-seq expression profiles, but the relationship between CCLs and patient tumors at single-cell resolution remains poorly characterized. To this end, we established a robust benchmarking framework that compares the effectiveness of diverse embedding approaches at relating CCLs to patient single-cell RNA-seq (scRNA-seq) profiles. To facilitate robust benchmarking, we designed a suite of control experiments, standardized preprocessing pipelines, and adapted metrics to evaluate mapping performance for each benchmarked method. We benchmarked seven methods: principal component analysis (PCA), canonical correlation analysis (CCA), Gene program, Celligner, Symphony, expiMap, and Geneformer. We observed that most methods excelled at tasks in simpler experiments, but showed reduced accuracy for certain lineages when mapping CCLs to their corresponding lineage in patient samples. We identified CCA as a promising candidate for integrating large-scale cross-system mapping of CCLs and tumor datasets. In sum, we have taken initial steps toward a framework to compare preclinical cancer models and patient tumor samples at single-cell resolution. Ultimately, with further development, this framework will guide the selection of preclinical models for specific clinical contexts.