ABS 003: Predictive Modeling and Discovery of Novel Therapeutics Using Physics-Informed Neural Networks to Overcome Drug and Surgical Resistance: Therapeutics in Glioblastoma Multiforme

Shivi Kumar ¹, Teryn Mitchell ², Osama Farany ³

¹ University of Pennsylvania
² Columbia University Medical Center
³ Moffitt Cancer Center

Van Wickle (2026) Volume 2, ABS003

Introduction: Tumor-associated arteries undergo complex molecular and physical remodeling that dictates therapeutic resistance and perfusion. This project combines multi-omic data (transcriptomic and epigenomic) with Physics-Informed Neural Networks (PINNs) to predict vascular remodeling and therapy response. The model enforces physical conservation laws of blood flow, diffusion, and pressure while being conditioned on patient-specific genetic and epigenetic features, yielding biologically interpretable and clinically relevant predictions.

Methods: Multimodal MRI data from 240 glioblastoma patients in The Cancer Imaging Archive, including T1-weighted contrast-enhanced, T2, FLAIR, and diffusion tensor imaging (DTI) sequences, were preprocessed using spatial normalization, skull stripping, voxel resampling, and intensity harmonization. Radiomic features describing tumor texture, morphology, and intensity heterogeneity were extracted from segmented tumor regions. DTI-derived diffusion tensors quantified white matter anisotropy to model directional tumor infiltration. Tumor evolution was represented using a reaction-diffusion equation. A three-dimensional Fourier Neural Operator integrated multimodal imaging and PDE constraints through a composite physics-informed loss function. Model performance was evaluated against U-Net, 3D CNN, and Vision Transformer baselines using Dice coefficient, longitudinal progression correlation, and neuroradiologist plausibility scoring.

Results: The physics-informed neural operator (PINO) framework achieved superior performance compared with baseline deep learning models across segmentation and progression prediction tasks. PINO produced a Dice similarity coefficient of 0.82 ± 0.03 versus 0.74 ± 0.05 for the U-Net model, demonstrating improved tumor boundary delineation. Longitudinal prediction of glioblastoma infiltration patterns showed stronger correlation with 12-month follow-up MRI scans (Pearson r = 0.71) compared with conventional CNN architectures (r = 0.46). Expert neuroradiologist review found that 87% of PINO-generated progression maps aligned with biologically plausible white matter tract invasion pathways, substantially exceeding baseline models. Performance remained stable across heterogeneous imaging protocols and tumor locations.

Discussion: The proposed physics-informed neural operator framework demonstrates that integrating mechanistic tumor growth equations with multimodal neuroimaging substantially improves prediction of glioblastoma evolution. By embedding diffusion-reaction dynamics and DTI-derived anisotropy into the learning process, the model generated biologically plausible infiltration trajectories while reducing unrealistic predictions common in purely data-driven systems. Integration of radiomic signatures further enabled characterization of intratumoral heterogeneity and progression risk. The framework’s robustness across imaging protocols and anatomical locations highlights its translational potential for precision neuro-oncology. Clinically, this approach may improve neurosurgical planning, adaptive therapy monitoring, and early stratification of high-risk glioblastoma patients.