Cellular identity is controlled by cell type specific expression of transcription factors (TFs), and it is reflected in the cell’s epigenetic landscape maintained by epigenetic regulator proteins (ERs). Functional dissection of cellular identity has focused mainly on a small number of lineage-defining master regulators, yet there is increasing evidence that multiple TFs and ERs work together to establish and retain the vast number of different cell types and cell states in the human body. For a more quantitative understanding of cellular identity, and of the complexities of its regulation, I propose to develop a machine-learning approach for in silico prediction of TF/ER cocktails that can transdifferentiate any human cell type into any other cell type, thus defining an operational rulebook of cellular transdifferentiation. To this end, I will train a machine-learning model called generative adversarial networks (GANs) on large-scale CRISPR single-cell sequencing (CROP-seq) datasets generated in the host lab. Exploiting unique features of the deep-learning generative approach, the resulting model will be able to generalize the learned genetic perturbations across cell types in silico. I will experimentally validate several of these predicted TF/ER transdifferentiation cocktails in the context of the human hematopoietic system. Importantly, the proposed approach is hypothesis-free and data-driven, exploiting recent advances in machine learning to infer fundamental aspects of the regulation of cellular identity from high-throughput functional CRISPR single-cell sequencing data.