Nazanin Ahmadi, Brown University. From PINNs to PIKANs: Physics-Informed AI for Systems Biology and Pharmacology

Physics-informed neural networks (PINNs) have opened exciting avenues for integrating mechanistic modeling with data-driven learning. However, their limitations—including difficulties with stiffness, sparse data, and highly complex loss landscapes that often lead to local minima—underscore the need for next-generation architectures and systematic optimization strategies. Recent advances bridge PINNs with Kolmogorov–Arnold Networks (PIKANs) to address some of the most challenging problems in biomedical modeling. One such development, AI-Aristotle [1], is a hybrid gray-box framework that combines PINNs with symbolic regression to estimate parameters and uncover missing physics in systems biology and pharmacology. Applications to pharmacokinetics model and ultradian glucose–insulin dynamics demonstrate the ability of AI-Aristotle to recover hidden mechanisms from noisy, limited data. Building on this, PIKANs, a physics-informed extension of Kolmogorov–Arnold Networks, are designed to overcome spectral bias and flexibly capture stiff pharmacological and physiological dynamics [2]. Together, these advances highlight how embedding physics into modern neural architectures—coupled with careful optimization—can enhance robustness, interpretability, and personalization, paving the way for next-generation AI-assisted discovery in systems pharmacology and biology.