PhD Candidate Biomedical Engineering, Brown University providence, Rhode Island, United States
Disclosure(s):
Nazanin Ahmadi Daryakenari, PhD: No financial relationships to disclose
Predictive modeling plays a crucial role in assessing the efficacy of antitumor drugs in a real-world setting. In this work, we leverage a pharmacokinetic-pharmacodynamic (PK-PD) model to establish a quantitative link between drug administration regimens and tumor growth dynamics. With precision medicine in oncology relying on accurate pharmacodynamic outcome predictions, enhancing model efficiency and generalization under limited data conditions is paramount. We investigate the integration of physics-informed learning into the state-space (Mamba) deep sequence model to improve predictive accuracy when training data is scarce. The study explores how incorporating physical information from ordinary differential equations (ODEs) can enhance model performance using a hybrid loss function that combines data-driven and physics-based components. We demonstrate that adding a physics-informed loss term significantly improves generalization. However, relying solely on the physics loss without labeled data leads to poor model performance, highlighting the complementary nature of data-driven and physics-informed learning. Our findings underscore the potential of physics-informed deep learning in drug discovery, offering early PD outcome prediction, uncertainty quantification, and insights into the impact of treatment schedules on tumor progression.
