Differences between in vivo and ex vivo models drive divergent perturbation outcomes

Ex vivo experimental models are extensively used as reductionist models, yet the impact of model choice on molecular states and perturbation responses has not been systematically evaluated. Here, we compared the outcomes of genetic perturbations in hematopoietic stem cells between in vivo and ex vivo model systems. This revealed strong baseline differences characterized by a reduced interferon response and elevated growth and metabolism signatures in ex vivo models, a pattern recapitulated in various orthogonal human and mouse systems. We further found substantial differences between perturbation effects between in vivo and ex vivo models, with some perturbations showing opposite effects. Finally, we evaluated whether AI-based perturbation prediction models can predict in vivo perturbation effects from ex vivo perturbation effects, which proved challenging. Together, our findings reveal systematic biases in ex vivo models, demonstrate how these biases influence perturbation outcomes, suggest approaches to enhance ex vivo systems, and provide a test case for computational prediction of perturbation effects.