Abstract: Cell types and states can be reprogrammed by activating combinations of transcription factors (TFs).However, the TF sets that reprogram cells from one state to another are unknown in the general case.There are $>10^{16}$ plausible TF sets in the human genome, making experimental search intractable and motivating *in silico* approaches to search this hypothesis space.Here, we describe a probabilistic model to design reprogramming interventions trained on a large corpus of single cell reprogramming data.Our model achieves strong performance on cell state and function prediction tasks and performance exhibits a data scaling law.Using our model in a simulated lab-in-the-loop, we were able to design successful reprogramming interventions significantly faster than pure experimental approaches.