MechGNN: A Graph Neural Network Framework for Learning Nanoparticle-Tumor Mechanical Interactions in Targeted Drug Delivery

janelia7_blocks-janelia7_biblio_header | block

2025 IEEE MIT Undergraduate Research Technology Conference (URTC). 2026 May 26;. doi: 10.1109/URTC68753.2025.11533032

Uppalapati K, Yimenicioglu B, Abdulkareem S, Eftekhari A

janelia7_blocks-janelia7_biblio_abstract | block

Abstract

Mechanical heterogeneity and collagen topology in tumor extracellular matrix (ECM) hinder nanoparticle (NP) transport and uptake, motivating models that couple NP mechanics with measured tumor mechanics. Non-invasive stiffness mapping (e.g., MR elastography) and histology-derived stiffness inference (STIFMap) enable learning on patientproximal mechanical maps. We introduce MechGNN, a dualgraph framework with cross-attention that represents NPs (core/shell/ligands) and ECM fiber networks as interacting graphs, and embeds physics-informed edge functions derived from Hertzian contact and receptor-ligand binding kinetics to regularize message passing. Using only public re-sources-STIFMap and open SHG collagen images quantified with CT-FIRE/CurveAlign/TWOMBLI-plus open FEM/BD simulations (FEniCS; Brownian dynamics) and open NP/ligand repositories (caNanoLab, BindingDB, IUPHAR/BPS Guide to Pharmacology), MechGNN achieves a 12−16%reduction in penetration-depth RMSE relative to the strongest baseline on synthetic-from-real ECM testbeds and sustains an 8−13% RMSE gain in leave one tumor type out tests; physics losses reduce constraint violation rates by ≈35%. The pipeline provides a reproducible bridge from measured ECM mechanics to mechanically designable NP properties.

Main Menu (Mobile)- Block

Main Menu - Block

MechGNN: A Graph Neural Network Framework for Learning Nanoparticle-Tumor Mechanical Interactions in Targeted Drug Delivery

Abstract

Janelia Authors