Disease-associated NK-cell state remodeling reveals transcriptional profiles predictive of clinical phenotypes

Abstract

Natural killer (NK) cells adapt their cellular states to diverse tissue and inflammatory environments, yet how this plasticity manifests across pathological contexts and relates to patient-level disease phenotypes remains poorly defined. Here we established the Human NK-cell Atlas (HNKA) from over 1 million NK cells, including samples from 1,632 donors spanning 50 diseases. Within a shared NK-cell state architecture, analyses of state abundance and matched-reference exceedance distinguished disease-biased redistribution from transcriptional displacement. Fine-grained analyses localized disease-associated remodeling to specific NK-cell states, including CD56dim NK ISG15 in infection and autoimmunity and CD56bright NK ITM2C in tumors. These patterns were accompanied by recurrent interferon responses and state-specific regulatory activity. Cytokine perturbations recapitulated key transcriptional features of disease-associated NK-cell states. We developed NKceiver, a model based on state-specific NK-cell expression profiles for patient-level prediction. It distinguished 18 diseases and predicted COVID-19 severity and response to immune checkpoint inhibitors, with predictive performance retained in unseen patients and studies. Collectively, our results reveal that disease-associated remodeling within the diverse NK-cell state repertoire resolved by HNKA is biologically interpretable and clinically informative, supporting disease discrimination and prediction of clinical phenotypes.