Speaker
Description
Mechanistic ordinary differential equation models are central to systems biology, pharmacology and emerging dynamical digital twins, but they are still commonly built by hand through extensive literature review, manual specification of reaction mechanisms, and repeated parameter fitting. As biological datasets become larger and more diverse, this process becomes difficult to scale, motivating automated approaches to mechanistic model discovery.
We present a Bayesian framework for discovering biochemical reaction networks directly from time-resolved concentration data under mass-action kinetics. Starting from a library of candidate reactions, we infer parsimonious networks and rate parameters with full posterior uncertainty, using a projection-predictive search strategy in trajectory space that favors compact reaction sets while preserving predictive dynamics. We demonstrate the approach on synthetic benchmarks and real biological time-course data, where it recovers interpretable reaction networks, while quantifying structural uncertainty.
This work contributes a biologically grounded route toward automated, uncertainty-aware inference of reaction networks from limited experimental data.
