Speaker
Description
In single-cell biology, stochastic reaction networks (SRNs) model molecular production, degradation, and interactions, and inferring their structure and parameters from data is central to understanding underlying biological mechanisms. Approximate Bayesian computation (ABC) offers a flexible Bayesian approach with posterior uncertainty quantification, but its reliance on extensive stochastic simulations results in high computational cost even for small systems, despite advances in sampling schemes. In this work, we propose a simulation-free approximate Bayesian computation (SFABC) framework that avoids explicit simulation of system dynamics. The method instead exploits theoretical constraints derived from the governing equations and uses convex optimization to evaluate whether a candidate parameter set is consistent with the observed data. Using synthetic benchmarks, we show that SFABC is compatible with different sampling schemes and achieves performance comparable to standard ABC algorithms, while substantially reducing computational cost that does not scale with sample size. Applied to steady-state scRNA-seq datasets, SFABC unravels transcriptional bursting kinetics through full posterior inference beyond point estimates. Using metabolic labelling data, we compare snapshot and time resolved models. Extending SFABC to model selection, we distinguish competing models of cell cycle–dependent transcriptional regulation using scRNA-seq data with cell cycle reporters.
