Comparing AI versus Optimization Workflows for Simulation-Based Inference of Spatial-Stochastic Systems

Abstract

Model parameter inference is a universal problem across science. This challenge is particularly pronounced in developmental biology, where faithful mechanistic descriptions require spatial-stochastic models with numerous parameters, yet quantitative empirical data often lack sufficient granularity due to experimental limitations. Parameterizing such complex models thus necessitates methods that elaborate on classical Bayesian inference by incorporating notions of optimality and goal-orientation through low-dimensional objective functions that quantitatively capture the target behavior of the underlying system. In this study, we contrast two such inference workflows and apply them to biophysics-inspired spatial-stochastic models. Technically, both workflows are simulation-based inference (SBI) methods. The first method leverages a modern deep-learning technique known as sequential neural posterior estimation (SNPE), while the second is based on a classical optimization technique called simulated annealing (SA). We evaluate these workflows by inferring the parameters of two complementary models for the inner cell mass (ICM) lineage differentiation in the blastocyst-stage mouse embryo. This developmental biology system serves as a paradigmatic example of a highly robust and reproducible cell-fate proportioning process that self-organizes under strongly stochastic conditions, such as intrinsic biochemical noise and cell-cell signaling delays. Our results indicate that while both methods largely agree in their predictions, the modern SBI workflow provides substantially richer inferred distributions at an equivalent computational cost. We identify the computational scenarios that favor the modern SBI method over its classical counterpart. Finally, we propose a plausible approach to integrate these two methods, thereby synergistically exploiting their parameter space exploration capabilities.