Surrogate-Driven Uncertainty Quantification for Agent-Based Models in Biology

Agent-based models (ABMs) capture how individual interactions give rise to system-level behavior, making them a natural choice for studying complex biological phenomena from tumor growth to disease outbreaks. Yet their stochasticity, computational cost, and high-dimensional parameter spaces, combined with data that are sparse, coarse-grained, and noisy, make it hard to know how far their predictions can be trusted. In this talk, I will present our SMoRe framework for uncertainty quantification in ABMs. This is built on explicitly formulated, mechanistic surrogate models whose interpretable parameters act as interlocutors between ABM inputs and real-world data. SMoRe ParS uses this bidirectional mapping to calibrate ABM parameters and determine which of them the data can actually constrain, accounting for both model stochasticity and measurement noise. SMoRe GloS extends the approach to global sensitivity analysis, recovering variance-based results in minutes rather than weeks of computation, including for spatial processes the surrogate does not explicitly represent. I will close with early results from SMoRe BayeS, a Bayesian formulation that carries uncertainty from data through the surrogate to full posterior distributions over ABM parameters. Examples will be drawn from cancer biology and infectious diseases.