Reinforcement learning guides single-cell sequencing in decoding lineage and cell fate decisions

The rapid development of single-cell sequencing provides unparalleled opportunity for delineating the heterogeneous characteristics of individual cells, while current methodologies fall short in accurately deciphering the origins of cellular heterogeneity and pinpointing the states of cell fate decisions. In this study, we introduce a novel approach termed Single-cell Reinforcement Learning (scRL) by integrating a framework of reinforcement learning into single-cell data analysis with the assistance of an actor-critic architecture. scRL reveals that the intensity of fate decisions precedes the maximum potential of lineage differentiation and actual gene expression, which enables the forecasting and perturbation of cellular fate in an earlier stage. The superiority of scRL in assessing fate decisions is confirmed in several different patterns of single-cell datasets. By using scRL, we disclose a distinct role of Dapp1 in myeloid lineage decision during hematopoietic stem cells (HSCs) differentiation and uncover the priority of erythroid fate decision occurring post irradiation exposure.