Mathematical oncology increasingly draws on complementary modeling strategies to address the complexity of tumor dynamics and treatment response. Data-driven AI models excel at extracting patterns from complex data such as imaging, clinical, and molecular characterization, offering rapid and scalable predictions, whereas mechanistic models based on biological and physical principles provide interpretability and explanatory power even in the context of limited data. Recent developments in mechanistic learning, the integration of prior biological knowledge into (deep) learning architectures, illustrate how these two paradigms can be combined to overcome their respective limitations. In this presentation, I will review recent strategies in mechanistic learning and illustrate their application in the domain of radiotherapy. Examples include spatio-temporal modeling of brain tumor growth from longitudinal imaging and prediction of treatment response dynamics using hybrid approaches combining generative computer vision and mechanistic tumor growth models, aiming for counterfactual simulations of alternative radiotherapy schedules based on probability maps of tumor progression. The value of additional data types, including histopathology and multi-omics characterization, will be discussed in the context of biology-adaptive treatment strategies. By combining the scalability of AI with the interpretability of mechanistic models, mechanistic learning offers a pathway toward predictive and clinically useful tools.