Cell2State: Learning Cell State Representations From Barcoded Single-Cell Gene-Expression Transitions
Genetic barcoding coupled with single-cell sequencing technology enables direct measurement of cell-to-cell transitions and gene-expression evolution over a long timespan. This new type of data reveals explicit state transitions of cell dynamics. Motivated by dimension reduction methods for dynamical systems, we develop a *cell-to-state* (cell2state) learning method that, through learning from such multi-modal data, maps single-cell gene expression profiles to low-dimensional state vectors that are predictive of cell dynamics. We evaluate the cell2state method using barcoded stem cell dataset (Biddy et al. (2018)) and simulation studies, compared with baseline approaches using features that are not dynamic-aware. We demonstrate the merits of cell2state in challenging downstream tasks including cell state prediction and finding dynamically stable clusters. Further, our method reveals potentiallatent meta-states of the underlying evolution process. For each of the meta-states, we identify a set of marker genes and development pathways that are biologically meaningful and potentially expand existing knowledge.
PDF Abstract
