Disentangling Transcription Factor Programs for Direct Reprogramming with Generative Modeling
Abstract
Transcription factor-induced reprogramming enables the direct transition between differentiated cell types without passing through a pluripotent state. This avoids risks associated with stem cell-based approaches, such as tumorigenicity and uncontrolled differentiation, while offering a comparatively rapid way to generate immune cells suitable for therapeutic use. However, identifying efficient transcription factor combinations remains a major challenge due to the intractable space of possible factors and their concentrations. Only very few cocktails have been identified so far, often with limited efficiency, and their discovery is highly resource-intensive, leaving many immune lineages directly inaccessible (Rosa et al. 2018). In this work, an adapted version of CellFlow (Klein et al., 2025), a flow-matching-based generative modeling framework for single cells, is trained on a single-cell RNA-seq dataset of human embryonic fibroblasts subjected to highly multiplexed transcription factor perturbations. Our model disentangles the underlying cellular transitions of the reprogramming landscape by learning how transcription factor treatments shape gene expression, and generates predicted gene expression profiles for arbitrary TF combinations. We were able to validate the performance of the model on previously established and experimentally confirmed TF combinations that reprogram fibroblasts into dendritic cells (Rosa et al., 2018). Building on this, we optimized previously established transcription factor combinations and identified new ones for reprogramming fibroblasts into various dendritic cell types. By analyzing the gene expression profiles of the generated cells, we demonstrate that these correspond to medically relevant phenotypes and cell states. Overall, this framework establishes a systematic and computationally feasible strategy to uncover transcription factor programs for direct lineage reprogramming. If experimentally validated, these predictions would provide access to previously unattainable immune states and have the potential to accelerate the development of next-generation immunotherapies.