Machine learning methods to reverse engineer dynamic gene regulatory networks governing cell state transitions

Deciphering the dynamic gene regulatory mechanisms driving cells to make fate decisions remains elusive. We present a novel unsupervised machine learning methodology that can be used to analyze a dataset of heterogeneous single-cell gene expressions profiles, determine the most probable number of states (major cellular phenotypes) represented and extract the corresponding cell sub-populations. Most importantly, for any transition of interest from a source to a destination state, our methodology can zoom in, identify the cells most specific for studying the dynamics of this transition, order them along a trajectory of biological progression in posterior probabilities space, determine the \"key-player\" genes governing the transition dynamics, partition the trajectory into consecutive phases (transition \"micro-states\"), and finally reconstruct causal gene regulatory networks for each phase. Application of the end-to-end methodology provides new insights on key-player genes and their dynamic interactions during the important HSC-to-LMPP cell state transition involved in hematopoiesis. Moreover, it allows us to reconstruct a probabilistic representation of the \"epigenetic landscape\" of transitions and identify correctly the major ones in the hematopoiesis hierarchy of states.