Learning cell-specific networks from dynamics and geometry of single cells

Cell dynamics and biological function are governed by intricate networks of molecular interactions. Inferring these interactions from data is a notoriously difficult inverse problem. Most existing network inference methods construct population-averaged representations of gene interaction networks, and they do not naturally allow us to infer differences in interaction activity across heterogeneous cell populations. We introduce locaTE, an information theoretic approach that leverages single-cell, dynamical information, together with geometry of the cell-state manifold, to infer cell-specific, causal gene interaction networks in a manner that is agnostic to the topology of the underlying biological trajectory. Through extensive simulation studies and applications to experimental datasets spanning mouse primitive endoderm formation, pancreatic development, and hematopoiesis, we demonstrate superior performance and the generation of additional insights, compared with standard population-averaged inference methods. We find that locaTE provides a powerful network inference method that allows us to distil cell-specific networks from single-cell data. A record of this paper's transparent peer review process is included in the supplemental information.