Digital twins (DTs) in oncology aim to provide patient-specific models that integrate multi-source clinical, biological, and treatment data to support prediction, decision-making, and long-term risk monitoring. A key challenge in the development of such systems is to establish a rigorous methodological foundation capable of linking heterogeneous observational data to clinically meaningful predictions while preserving interpretability and causal coherence. In this work, we propose a multi-layer probabilistic framework for oncology DTs that explicitly connects three conceptual levels: a data level, where multimodal clinical, dosimetric, and molecular variables are collected; a causal level, where cause–effect relationships among selected variables are formalized; and a computational level, where these relationships are embedded into models suitable for prediction, simulation, and decision support. At the causal level, directed acyclic graphs are employed to disentangle exposure variables, outcomes, and confounders, allowing the investigation of mechanisms underlying treatment-related adverse events beyond purely predictive correlations. We introduce a probabilistic formalization of personalized risk assessment based on two main interpretable components: exposure to causal factors and individual propensity to develop adverse events driven by patient-specific susceptibility. This decomposition is naturally expressed through probabilistic graphical models, based on Bayesian Networks, which enable modular risk representation, accommodate heterogeneous and missing data, and support longitudinal updating. We provide a unified framework for predictive and causal inference within DT systems. The proposed methodology is demonstrated in the context of breast cancer radiotherapy, focusing on the prediction of severe late toxicity. Risk components are instantiated using retrospective and prospective data, including dosimetric, genomic, and clinical variables. The resulting models support individualized risk stratification and constitute the foundation for scalable, probabilistically grounded DT architectures in precision oncology. This work is in collaboration with the TETRIS project consortium: Claudio Fiorino, David Gibon, Sara Gutiérrez-Enríquez, Eva Onjukka, Sandrine Pereira, Ana Vega, and Tiziana Rancati.