B3 - Physics-Informed Neural Network (PINN) Applications in Cellular Processes

Introduction: : Understanding complex cellular behaviors in dynamic biological systems remains a central challenge in systems biology and immunotherapy. In this study, we explore the application of Physics-Informed Neural Networks (PINNs) as a hybrid modeling approach to infer latent biological dynamics from sparse and noisy data. Physics-Informed Neural Networks are a class of neural networks that integrate known physical, biological, or mathematical laws—usually expressed as differential equations—directly into the model training process.

Materials and

Methods: : Leveraging a small cohort of individuals undergoing targeted immune modulation [1], we amalgamate several omics outputs at the single-cell level, collected across multiple timepoints. We developed PINN structures that consist of neural networks that map input features, such as time or treatment conditions, to cellular states while simultaneously enforcing differential equation constraints. In the PINNs, the total loss function includes both a data fidelity term and a physics-based residual, ensuring the outputs are consistent with biological laws as well as observed measurements.

Results, Conclusions, and Discussions:: Selected immune cell subsets, isolated and profiled were integrated into a PINN framework guided by mechanistic assumptions. Unlike conventional neural networks, PINNs embed prior biological knowledge directly into the learning process through differential equation constraints, leading to models that are both data-efficient and interpretable. This allows PINNs to generalize better from limited data and offer insights grounded in known biological mechanisms. Our results show improved interpretability of temporal cellular changes and identification of patient-specific response signatures, and a generalizable framework for combining experimental data with biological knowledge. This work highlights the potential of PINNs to bridge data-driven inference and mechanistic modeling in cellular biology.

Acknowledgements and/or References (Optional):: 1. Zhang, J. Y., Hamey, F., Trzupek, D., Mickunas, M., Lee, M., Godfrey, L., Yang, J. H. M., Pekalski, M. L., Kennet, J., Waldron-Lynch, F., Evans, M. L., Tree, T. I. M., Wicker, L. S., Todd, J. A., & Ferreira, R. C. (2022). Low-dose IL-2 reduces IL-21+ T cell frequency and induces anti-inflammatory gene expression in type 1 diabetes. Nature communications, 13(1), 7324. https://doi.org/10.1038/s41467-022-34162-3