AetherCell: A generative engine for virtual cell perturbation and in vivo drug discovery

Abstract

Virtual cell modeling is currently hindered by a "data-utility paradox": biological information is fragmented between context-rich clinical RNA-seq and perturbation-dense experimental assays, leading to poor predictive generalization in human contexts. Here, we introduce AetherCell, a generative foundation model that unifies these disparate domains into a shared, platform-aligned transcriptomic manifold. By implementing a specificity-driven learning framework, AetherCell successfully recovers low-frequency, mechanism-specific signals often obscured by systematic noise. Across extensive benchmarks, AetherCell demonstrates robust generalization, accurately predicting responses to unseen compounds and genetic perturbations. We show that the model effectively translates signals from simple cell lines to complex 3D organoids, achieving high-fidelity, whole-transcriptome prediction. Building on this foundational manifold, we demonstrate precise drug response prediction across biological scales--including patient-derived organoids and clinical cohorts. We also implement a phenotype-knowledge mixture-of-experts strategy for precision drug repurposing. This approach is validated by the in vivo discovery of teriflunomide for dry eye disease and dabigatran for ulcerative colitis. Together, AetherCell establishes a scalable, human-centric virtual cell framework for translational biology and accelerated drug discovery.