Dissecting the cellular heterogeneity embedded in single-cell transcriptomic data is challenging. Although many methods and approaches exist, identifying cell states and their underlying topology is still a major challenge. Here, researchers from the MIT Computer Science and Artificial Intelligence Laboratory and the Broad Institute introduce the concept of multiresolution cell-state decomposition as a practical approach to simultaneously capture both fine- and coarse-grain patterns of variability. The researchers implement this concept in ACTIONet, a comprehensive framework that combines archetypal analysis and manifold learning to provide a ready-to-use analytical approach for multiresolution single-cell state characterization. ACTIONet provides a robust, reproducible, and highly interpretable single-cell analysis platform that couples dominant pattern discovery with a corresponding structural representation of the cell state landscape. Using multiple synthetic and real data sets, the researchers demonstrate ACTIONet’s superior performance relative to existing alternatives. They use ACTIONet to integrate and annotate cells across three human cortex data sets. Through integrative comparative analysis, they define a consensus vocabulary and a consistent set of gene signatures discriminating against the transcriptomic cell types and subtypes of the human prefrontal cortex.
ACTIONet framework overview
a Main steps in ACTIONet. b ACTION-based matrix decomposition. Dimensionality reduction for feature selection (reduction) is coupled with AA to perform individual-level decomposition and identify k latent cell-state patterns (archetypes). A column c of the cell influence matrix C encodes the influence of cells on the patterns. A column h of the cell-state encoding matrix H encodes the relative contribution of each pattern to the transcriptome of each cell. c Multilevel ACTION decomposition with an increment in the number of archetypes per level. Concatenation of individual-level matrices defines multilevel encoding (H*), cell influence (C*), and profile matrices (W*). d, e Metric cell space defined by measuring distances on multilevel cell-state encodings. f Construction of a sparse network representation of the cell space.
Availability – ACTIONet is implemented in an easy-to-use and freely available computational environment for both R and Python: ACTIONet (https://github.com/shmohammadi86/ACTIONet).