Multi-resolution single-cell state characterization via joint archetypal/network analysis

Abstract

Dissecting the cellular heterogeneity embedded in single-cell transcriptomic data is challenging. Although a large number of methods and approaches exist, robustly identifying underlying cell states and their associations is still a major challenge; given the nonexclusive and dynamic influence of multiple unknown sources of variability, the existence of state continuum at the time-scale of observation, and the inevitable snapshot nature of experiments. As a way to address some of these challenges, here we introduce ACTIONet, a comprehensive framework that combines archetypal analysis and network theory to provide a ready-to-use analytical approach for multiresolution single-cell state characterization. ACTIONet uses multilevel matrix decomposition and network reconstruction to simultaneously learn cell state patterns, quantify single-cell states, and reconstruct a reproducible structural representation of the transcriptional state space that is geometrically mapped to a color space. A color-enhanced quantitative view of cell states enables novel visualization, prediction, and annotation approaches. Using data from multiple tissues, organisms, and developmental conditions, we illustrate how ACTIONet facilitates the reconstruction and exploration of single-cell state landscapes.