Glioblastoma is characterized by interpatient and intratumoral heterogeneity, making it difficult to assess treatment response and predict disease progression from standard clinical imaging. Integrating mechanistic tumor growth models with machine learning offers a path to address this challenge. By combining reaction–diffusion models with MRI data and image-localized biopsies, this approach links imaging features to underlying tumor biology. Incorporating mechanistic priors improves prediction of tumor cell density and spatial structure compared to purely data-driven methods. The framework further identifies distinct tumor ecosystem states (proliferative, immune-inflamed, and diffusely invasive) that are not captured by routine imaging. These results demonstrate how hybrid modeling can infer latent tumor biology and detect clinically meaningful shifts, supporting personalized treatment strategies.