Mechanistic Learning for Computational Oncology: from Neural Model Order Reduction to Digital Twins
Computational oncology aims to provide quantitative tools to support personalized treatment and risk assessment by integrating biological knowledge, clinical data, and predictive modeling within a rigorous mathematical framework that preserves interpretability and causal coherence.
We present an overview of recent methodological advances within the emerging paradigm of mechanistic learning for personalized radiotherapy and computational oncology. We focus on the integration of physics-based models and data-driven approaches to overcome current computational and methodological limitations arising from the intrinsically multiscale and multiphysics nature of the tumor microenvironment.
From a physics-based modeling perspective, we combine scientific machine learning and reduced-order modeling (ROM) techniques to approximate high-dimensional parametric partial differential equations defined on complex geometries [1][2].
These components are integrated through domain decomposition and Schwarz-type coupling strategies (DD-ROM), enabling robust upscaled 3D-1D simulations suitable for whole–tumor mass modeling.
Building upon these advances, we then introduce a multi-layered, data-driven probabilistic framework for risk assessment that connects data acquisition, causal inference through directed acyclic graphs (DAGs), and computational modeling via probabilistic graphical models (PGMs). The resulting models enable individualized risk stratification and provide a scalable probabilistic foundation for digital twins in precision oncology.
[1] Vitullo, P., Colombo, A., Franco, N.R., Manzoni, A. and Zunino, P. Nonlinear model order reduction for problems with microstructure using mesh informed neural networks. Finite Elements in Analysis and Design, 229, 104068, (2024)
[2] Botta, P., Vitullo, P., Ventimiglia, T., Linninger, A. and Zunino, P. Physics-Informed Learning of Microvascular Flow Models using Graph Neural Networks. arXiv preprint, arXiv:2512.10792, (2025)
Contatto:
andrea1.manzoni@polimi.it