Virtual Knockout in the Era of Generative AI: From Intracellular Networks to Spatial and Cross-Modal Functional Simulations

Abstract

Single-cell perturbation data have enabled computational models that predict intervention-associated cellular responses. This narrative review organizes explicit gene-regulatory-network simulations, latent and compositional response models, graph, optimal-transport, and dynamical approaches, foundation-model representations, and emerging spatial or cross-modal settings through a virtual-knockout lens. We distinguish representation-level ablation from causal genetic intervention and separate held-out regimes for genes, cell types or contexts, perturbation combinations, chemical structures, and doses. We compare representative methods, experimental resources, metrics, and benchmark frameworks and operationally define a prospective holographic virtual cell by its modalities, regulatory and physical constraints, intervention-conditioned outputs with calibrated uncertainty, and prospective validation requirements. The holographic virtual cell and Design−Build−Test−Learn roadmap are presented as future-facing criteria rather than as an achieved clinical capability.