Nikhat, A.; Mandal, T.; Veerasubramanian, N.; Chakrabarti, S.

Cell-state discovery at single-cell resolution is currently a major endeavor of modern biology. While immense effort has gone into dealing with associated technical noise, there is little appreciation of the fundamental limits imposed by intrinsic biological stochasticity. Using the circadian clock as an example where true cell-states can be precisely defined (the oscillator phase), we study how bursty transcription limits the achievable cellular and spatial resolution of circadian-phase inference. Combining multiplexed smFISH to measure endogenous gene-expression and a novel supervised learning algorithm, we demonstrate how accurate inference is possible using just 3 genes but only after averaging over 20-70 cells. Commonly used algorithms generate single-cell clusters completely different from the true states, successfully recovering them only after averaging. Further, by decoding all states within a population of asynchronized cells, we demonstrate how coarse-graining provides a principled approach to spatially-resolved phase inference. Our results argue that coarse-graining is likely essential for meaningful cell-state assignment across biological systems, even when technical noise is minimized.