scPER: A Rigorous Computational Approach to Determine Cellular Subtypes in Tumors Aligned With Cancer Phenotypes From Total RNA Sequencing

Abstract Single‐cell RNA sequencing (scRNA‐seq) is a powerful technique for understanding cellular diversity, but processing large patient cohorts to identify phenotype‐associated cell populations remains challenging. Here, scPER ( E stimating cell P roportions using s ingle‐ c ell RNA‐seq R eference), a rigorous approach combining adversarial autoencoder and extreme gradient boosting to estimate tumor microenvironment cell compositions and identify phenotype‐associated subclusters for bulk RNA‐seq samples. Integrating scRNA‐seq datasets from diverse studies, scPER constructed comprehensive reference panels and disentangled confounders from true signals. scPER achieved superior accuracy in cellular proportion estimation compared to CIBERSORTx, BayesPrism, Scaden, MuSiC, SCDC, DeSide and ReCIDE. It showed high accuracy in predicting metastatic melanoma immunotherapy response and identified a critical T cell subcluster expressing FCRL3 and SLAMF7 . In metastatic urothelial cancer, scPER predicted TGFβ‐mediated inhibition of CD4 naïve T cells to diminish PD‐L1 checkpoint blockade efficacy. scPER enables robust integration of scRNA‐seq datasets to estimate cellular proportions across tumors and identify clinically relevant cell populations.