Virtual Biochemical Twins of Human Organoids: Predictive Modeling of Disease State Transitions through Integrated Spatial Multiomics and Perturbation Dynamics

Abstract

Human organoids recapitulate key architectural and functional features of native tissues, yet our ability to predict how they evolve, respond to perturbation, or transition between healthy and diseased states remains limited by the descriptive nature of current multiomic approaches. Here, we introduce Virtual Biochemical Twins (VBTs)—computable, patient-specific digital replicas of human organoids that learn and forecast biochemical state transitions across space and time. We constructed VBTs by integrating longitudinal live-cell imaging, spatial transcriptomics and proteomics, single-cell multiomics, and genome-scale Perturb-seq screens across a panel of patient-derived organoids spanning healthy, pre-malignant, and diseased states. A geometry-aware graph neural network coupled with optimal transport learned continuous mappings between molecular configurations, enabling reconstruction of cell-state trajectories at sub-cellular resolution and in silico prediction of perturbation outcomes never observed during training. VBTs accurately forecasted spatial reorganization of signaling networks following CRISPR knockouts (mean Pearson r = 0.89 across held-out perturbations), identified previously unrecognized regulatory bottlenecks governing the transition to disease, and nominated combinatorial interventions that were validated experimentally to reverse pathological organoid phenotypes in three independent disease models. Critically, twins built from a patient's own organoid predicted that patient's drug response with significantly higher fidelity than population-level models, establishing a path toward personalized predictive biology. Virtual Biochemical Twins move organoid science beyond observation toward a quantitative, predictive framework, providing a generalizable platform for mechanistic discovery, therapeutic design, and precision medicine.