Normalization choice drives biological interpretation in single-cell RNA-seq cancer studies: A systematic benchmarking of 465 computational pipelines.

Single-cell RNA sequencing (scRNA-seq) has transformed our understanding of tumor heterogeneity, yet the impact of computational pipeline choices on biological conclusions remains poorly characterized. Here, we systematically benchmark 465 combinations of 5 normalization methods and 4 clustering algorithms across 3 cancer datasets encompassing over 434,000 cells. We introduce the Biological Discordance Score (BDS), a metric that quantifies how marker gene interpretation changes across pipelines. We find that normalization method choice has a greater impact on biological interpretation than clustering algorithm selection. Strikingly, pipelines with similar clustering agreement (ARI) can identify up to 86% different marker genes, a discrepancy invisible to standard evaluation metrics. Log-normalization consistently achieves the best balance of performance and interpretive stability across cancer types. Using a Borda count meta-ranking framework with bootstrap confidence intervals, we provide unified recommendations for pipeline selection. Our results demonstrate that computational choices have profound but underappreciated consequences for biological discovery in cancer single-cell studies.