Likelihood-free parameter inference for spatiotemporal stochastic biological models using neural posterior estimation

Clinical Snapshot

Key Findings

• P Value: Not applicable — Bayesian inference framework; no frequentist hypothesis testing reported

• Effect Size: Not quantified with standard effect size metrics in the abstract; performance described qualitatively as superior to surrogate-based methods for the combined bias-proliferation model

• Primary Outcome: Neural posterior estimation successfully recovers biologically interpretable parameters (cell motility, directional bias, proliferation rates) across all four random walk models of barrier assay experiments, including the complex combined model where surrogate-based methods are unreliable

• Nnt Or Sensitivity: Not applicable in conventional clinical sense; methodological study — relevant metric is posterior calibration quality assessed via SBC rank statistics

• Confidence Interval: Posterior distributions reported rather than frequentist confidence intervals; SBC diagnostics used to assess calibration — specific coverage statistics not reported in abstract

Clinical Application

Open-source pipeline provided, which lowers the barrier to adoption. Requires familiarity with Python-based machine learning frameworks and simulation-based inference libraries (e.g., sbi package). Computational infrastructure for neural network training is needed but accessible via standard research computing environments. Not yet validated for routine laboratory use. Developed by Australian researchers at the University of Melbourne and Queensland University of Technology under the ARC Centre of Excellence for the Mathematical Analysis of Cellular Systems (MACSYS). Directly relevant to Australian mathematical biology and biomedical engineering research communities. No direct PBS, TGA, or RACGP clinical guideline implications at this stage. Potential long-term relevance to Australian cancer research (e.g., cancer metastasis modelling) and wound care research, but clinical translation requires substantial further validation steps. Mathematical biologists, computational cell biologists, and biomedical researchers developing stochastic models of cell migration, wound healing, tissue morphogenesis, and cancer invasion; not directly applicable to clinical patient populations at this stage

Abstract

Cell migration is a key biological process underlying wound healing, tissue development, and cancer metastasis, yet calibrating mathematical models of migration to experimental data remains a major challenge. Scratch and barrier assays are widely used to study collective cell spreading, and agent-based random walk models provide a natural stochastic description of these experiments. However, parameter inference for such models is hampered by intractable likelihoods, forcing researchers to rely on Approximate Bayesian Computation, which introduces biases and tuning difficulties, or surrogate models that require potentially erroneous noise model specifications. Here, we overcome these limitations using neural posterior estimation, a simulation-based inference framework that learns the full posterior distribution directly from stochastic simulations without surrogate approximations or explicit noise model specifications. We deploy this framework on four progressively complex random walk models of barrier assay experiments describing in vitro cell migration: an isotropic baseline, a model with directional bias (chemotaxis), a model with cell proliferation, and a combined model incorporating both bias and proliferation. For each model, we demonstrate inference in two settings: using one-dimensional summary statistics (column counts), and using a convolutional neural network that enables inference directly from raw two-dimensional spatial data. Neural posterior estimation performs well across all four models, recovering biologically interpretable parameters (e.g. cell motility, directional bias, proliferation rates) from cases where classical surrogate-based methods are adequate through to the combined model where the interplay of multiple mechanisms renders surrogate approximations unreliable. We validate all posteriors using simulation-based calibration diagnostics and provide an open-source implementation of our pipeline to facilitate its adoption and extension to more complex, spatially-structured biological models.

References

1. 1.Kimpson, T., Flegg, J., & Simpson, M. J. (2026). Likelihood-free parameter inference for spatiotemporal stochastic biological models using neural posterior estimation. Journal of Theoretical Biology. https://doi.org/10.1016/j.jtbi.2026.112475