AI SURROGATE MODELING FOR BIOLOGICAL SYSTEMS

Abstract

Biological systems exhibit complex, nonlinear dynamics governed by interactingmolecular and cellular processes, and are commonly modeled using differential equa- tions. While these mechanistic models provide a principled framework for under- standing biological behavior, their simulation becomes computationally prohibitive as system complexity and dimensionality increase. This limits large-scale exploration of parameter regimes, perturbations, and disease dynamics. To address this challenge, we develop a neural operator–based surrogate model for tau protein propagation on brain connectomes, enabling rapid approximation of high-dimensional spatiotemporal dynamics. The proposed Tau Brain Neural Operator (Tau-BNO) captures both local reaction kinetics and global transport processes, achieving improved predictive accu- racy over existing architectures while reducing simulation time from hours to seconds. However, the effectiveness of such surrogate models depends on access to large, high- quality training datasets generated from expensive numerical solvers. This creates a second bottleneck: data generation. To overcome this limitation, we introduce the IntelliAgent framework, a set of latent-space sampling strategies designed to select informative and diverse training trajectories. By combining diversity-based selection with prediction error signals, IntelliAgent reduces data requirements while preserv- ing predictive performance. Across multiple systems biology models, the proposed approach achieves target accuracy using only 40% of the training data required by conventional methods, significantly lowering computational cost. Together, Tau-BNO and IntelliAgent form a unified framework for scalable surrogate modeling, enabling efficient simulation and data-efficient learning for complex biological systems without requiring explicit knowledge of governing equations.