Speaker
Description
Physics-Informed Neural Networks (PINNs) have developed into a flexible framework for embedding mechanistic knowledge within data-driven models. Our work provides two complementary studies using PC9 lung cancer cell microscopy data.
First, we apply a Biologically-Informed Neural Network (BINN) to spatiotemporal (2D+t) data under a reaction–diffusion model, where diffusion and growth are treated as unknown functions of cell density. We augment this architecture with symbolic regression (SR) to recover interpretable analytical expressions for the diffusion and growth functions.
Second, we explore Bayesian PINNs for inverse problems with noisy, sparse biological data. We use an ordinary differential equation model of population dynamics and compare fully Bayesian inference using Hamiltonian Monte Carlo with approximate Bayesian approaches. We further examine the universal PINN (UPINN) framework, highlighting trade-offs between flexibility and mechanistic constraints, and propose an iterative UPINN–SR strategy to infer governing structure when prior knowledge is limited.
Together, these contributions offer practical, interpretable, and user-friendly workflows for applying physics-informed learning to biological systems.
