scRL: Utilizing Reinforcement Learning to Evaluate Fate Decisions in Single-Cell Data

The rapid development of single-cell sequencing offers an unparalleled opportunity to delineate the heterogeneity of individual cells. However, current methods struggle to pinpoint the states of cell fate decisions. In this study, we introduce a novel approach called Single-cell Reinforcement Learning (scRL), which integrates reinforcement learning into single-cell data analysis using an actor-critic architecture. Among existing dimensionality reduction methods, we identified one with the best interpretability. Based on this latent space and combined with our scRL algorithm, we assessed the intensity of fate decisions at the single-cell level. Extensive evaluations demonstrate that scRL outperforms existing techniques, as well as their variants and alternative approaches, in assessing cell fate decisions. Moreover, scRL offers an alternative method for evaluating the intensity of cell lineage differentiation which shows competitive interpretability as well. The superiority of scRL in assessing fate decisions is confirmed across several types of single-cell datasets.