Cell-DRL Reconstructs Unseen Cellular Paths in Health and Disease

Diseases can be conceived as deviations from healthy cell state manifolds. Numerous methods leverage the power of single-cell RNA sequencing for reconstructing cell state trajectories. Such approaches generally rely on sufficient sampling of cell states covering the entire trajectory. Since patients typically undergo treatment only after symptoms have already manifested, clinical samples covering intermediate disease states are generally unavailable, limiting our ability to understand the underlying intermediate paths of disease evolution. Reconstruction of missing states on disease trajectories is a major open challenge. To overcome the limitations of current approaches, we developed Cell-DRL, a deep reinforcement learning agent capable of geometric reasoning on biological manifolds, generating actions in gene expression space, and learning stochastic policies to reconstruct trajectories connecting two distant anchoring cellular states. We validated Cell-DRL on ground truths scenarios by hiding intermediate states and demonstrated the capacity to reconstruct multipotent hematopoietic stem cell states from distinct lineage-specific progenitors. We showcased the power of Cell-DRL to recover unseen cellular states in healthy as well as disease scenarios at the individual patient level. Finally, Cell-DRL predicted a novel human cardiac fibroblast-to-cardiomyocyte trans-differentiation path, which we validated in vitro. We expect that Cell-DRL will be crucial to gain valuable mechanistic insights into the development and progression of diseases at high temporal resolution.