Review of computational methods for estimating cell potency from single-cell RNA-seq data, with a detailed analysis of discrepancies between method description and code implementation

In single-cell RNA sequencing (scRNA-seq) data analysis, a critical challenge is to infer hidden dynamic cellular processes from measured static cell snapshots. To tackle this challenge, many computational methods have been developed from distinct perspectives. Besides the common perspectives of inferring trajectories (or pseudotime) and RNA velocity, another important perspective is to estimate the differentiation potential of cells, which is commonly referred to as "cell potency." In this review, we provide a comprehensive summary of 11 computational methods that estimate cell potency from scRNA-seq data under different assumptions, some of which are even conceptually contradictory. We divide these methods into three categories: mean-based, entropy-based, and correlation-based methods, depending on how a method summarizes gene expression levels of a cell or cell type into a potency measure. Our review focuses on the key similarities and differences of the methods within each category and between the categories, providing a high-level intuition of each method. Moreover, we use a unified set of mathematical notations to detail the 11 methods' methodologies and summarize their usage complexities, including the number of ad-hoc parameters, the number of required inputs, and the existence of discrepancies between the method description in publications and the method implementation in software packages. Realizing the conceptual contradictions of existing methods and the difficulty of fair benchmarking without single-cell-level ground truths, we conclude that accurate estimation of cell potency from scRNA-seq data remains an open challenge.