The development of predictive, scalable, and clinically actionable models remains a central challenge in computational oncology, due to the intrinsic multiscale and multiphysics nature of tumor growth, vascular remodeling, and therapy response. In particular, the accurate description of microvascular transport processes, including blood flow, oxygen transfer, and drug delivery, requires the solution of large-scale, high-dimensional parametric partial differential equations defined on complex mixed-dimensional geometries. These models, while mechanistically sound, are often computationally prohibitive for real-time analysis, uncertainty quantification, or patient-specific optimization. In this talk, we present recent advances in scientific machine learning approaches that tightly integrate reduced order modeling (ROM) with deep learning to overcome these limitations. We focus on neural reduced order models for parametric microvascular and tissue-scale problems, where physics-informed and data-driven components are combined to achieve both accuracy and efficiency. Particular emphasis is placed on microvascular flow and transport models formulated on vascular graphs and embedded tissue domains, which naturally arise in the modeling of tumor microenvironments and radiotherapy response. We introduce a hierarchy of methodologies that span three interconnected levels. First, we discuss neural representations of vascular potentials, where ROMs are used to approximate solution operators of elliptic and transport problems defined outside the vascular domain [1]. Second, we consider graph-based formulations of microvascular systems, where learning is performed directly on vascular graphs to capture flow redistribution, pressure fields, and perfusion heterogeneity across realistic anatomical configurations [2]. Third, we present upscaled ROM strategies, based on Schwarz-type iterative coupling, which allow for the efficient reconstruction of global three-dimensional tissue solutions from localized representations. Overall, the presented work highlights how the synergy between reduced order modeling, deep learning, and domain decomposition provides a mathematically consistent and computationally viable foundation for next-generation digital twins in oncology, bridging the gap between high-fidelity models and real-world clinical applicability.