Abstract
Immune checkpoint inhibitors (ICIs) can induce durable responses across cancers, yet T-cell biomarkers of response remain difficult to reproduce across single-cell RNA-seq studies. A major reason is that T-cell states are typically defined de novo within each cohort, making reported marker genes sensitive to cohort composition and analytic choices rather than stable cellular programs. Here we present sigNATURE (signature Normalization and Atlas-based T-cell Understanding for Reproducibility and Evaluation), a reference-guided framework that maps query cells onto large CD4+ and CD8+ T-cell atlases, evaluates published ICI-response markers in an atlas-aligned coordinate system, and quantifies the atlas support of mapped cells through a cell-level identifiability score. We applied sigNATURE to two independent ICI scRNA-seq cohorts comprising 36 non-small cell lung cancer patients and 15 skin cancer patients (11 basal cell carcinoma and 4 squamous cell carcinoma). Across cohorts, sigNATURE-derived features more robustly resolved response-associated T-cell structure than cohort-derived state definitions, yielding clearer unsupervised separation of responders and non-responders, enabling integrated analysis of independent studies in a shared atlas-aligned space, and improving mean response-prediction AUC from 0.469 to 0.746. Using identifiability score, we further identify terminally differentiated effector CD8+ T cells and regulatory CD4+ T cells as prominent response-associated states across studies, prioritizing published markers in terms of robust, atlas-resolvable cell states. Using this framework. Together, these results establish sigNATURE as a framework for improving the reproducibility, cross-cohort comparability, and mechanistic interpretability of single-cell ICI biomarkers.