Abstract
<jats:p>Circulating cytokines encode immune state, yet their pleiotropy and cell-type specificity make constructing a unified atlas of immune cell responses to them challenging. Here, I transformed a single-cell atlas of approximately 10 million human peripheral blood mononuclear cells from 12 donors exposed to 90 cytokines into a multiscale model of cytokine response. A GPU-accelerated implementation of dimension-scalable single-cell perturbation integration network (D-SPIN) allowed for the creation of a signed, directed model of 9.6 million cells, 1,634 immune regulatory genes, and 40 cellular programs. The gene networks and cellular programs span canonical cytokine pathways and lineage relationships and delineated cytokine-specific activation and repression across immune states. Beyond established circuitry, the model nominated candidate regulatory interactions and identified the mitochondrial antioxidant SOD2 as a prominent hub of innate immune cell signaling. Next, I developed CytoCarto, a web application that projects cytokine profiles onto these networks to prioritize dysregulated programs, candidate effector genes, cellular contexts, and disease-associated signatures. In a proof-of-concept analysis I input the cytokine profile of a patient with mitochondrial encephalopathy, lactic acidosis, and stroke-like episodes (MELAS) undergoing an episode of sterile inflammation and found CytoCarto prioritized metabolically reprogrammed monocytes and SOD2, consistent with a role for mitochondrial redox signaling in innate immunity.</jats:p>