Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
BP: Fachverband Biologische Physik
BP 15: Computational Biophysics III
BP 15.8: Talk
Wednesday, March 11, 2026, 11:45–12:00, BAR/SCHÖ
Gradient-Estimating Gillespie Simulators for Parameter Inference in Stochastic Models — •Ludwig Burger1, Annalena Kofler2, Lukas Heinrich3, and Ulrich Gerland1 — 1Physics of Complex Biosystems, School of Natural Sciences, Technical University Munich — 2Max Planck Institute for Intelligent Systems, Tübingen — 3Data Science in Physics, School of Natural Sciences, Technical University Munich
Stochastic models are ubiquitous in (biological) physics, yet fitting such parameterized models to experimental data remains challenging. While gradients in deterministic systems can be obtained efficiently through numerical or automatic differentiation, these tools cannot be directly applied to stochastic simulation algorithms such as the Gillespie algorithm, where sampling from a discrete set of reactions introduces non-differentiable operations. In this work, we adapt three gradient estimators from machine learning for use in Gillespie simulations: the Gumbel-Softmax Straight-Through estimator, the Score Function estimator, and the Alternative Path estimator. We extend all three estimators to address the specific requirements of gradient estimation in Gillespie simulations. We analyze the statistical properties of the estimators and highlight practical advantages and limitations in two representative systems: a minimal bimolecular association-dissociation model and the repressilator. Our results demonstrate that gradient estimators can be effectively integrated into the Gillespie algorithm, providing a systematic approach for gradient-based parameter inference in stochastic models.
Keywords: Stochastic Models; Gillespie Stochastic Simulation Algorithm; Gradient Estimation; Parameter Inference