Deep learning inference of universal dormancy pseudotime reveals the cellular targets of anti-cancer therapies

Abstract

Controlled exit from and re-entry into the cell cycle is essential for multi-cellular life, while aberrant quiescent and senescent cell states have been implicated in age-related diseases and cancer treatment evasion. Recent molecular and imaging studies suggest non-cycling cellular states exist along a continuum of deepening dormancy, whereby the probability of cell cycle re-entry decreases with distance from the restriction point. We trained a probabilistic deep-learning model that enables mapping of heterogeneous single cell transcriptomic datasets into an interpretable latent space that encodes a common "dormancy pseudotime". We demonstrate that our model enables robust inference of active cell cycle states, and validate in diverse biological contexts that the inferred location along dormancy pseudotime represents a continuum from quiescence to durably arrested states. Applying dormancy pseudotime inference to pre- and post-treatment time points from patients undergoing anti-cancer treatment, we uncover new insights into the distinct tumour cell dormancy states targeted by immune checkpoint inhibitors and platinum-taxane chemotherapy. Given the ubiquity of single cell transcriptomics, we anticipate that dormancy pseudotime analysis will be widely applied to shed new light on the complex interplay between cycling and non-cycling cellular states in health and disease.