Understanding tumor heterogeneity and therapy resistance remains a major challenge in cancer research and treatment development. While data-driven methods excel at extracting patterns from high-throughput biological data, mechanistic models provide critical insights into dynamical system behaviors. In this talk, I will present novel integrative frameworks bridging machine learning and mathematical modeling to gain deeper insights into the complex dynamics underlying tumor evolution and response to therapy. By leveraging single-cell RNA-seq, spatial transcriptomics data and clinical data, our approaches enable the identification of critical subclones and biomarkers driving resistance, while mapping tumor-immune ecosystem dynamics with spatiotemporal resolution. Our findings demonstrate that bridging data-driven methods with theoretical modeling provides a comprehensive approach to deciphering tumor heterogeneity and overcoming therapy resistance, ultimately guiding the development of more effective, personalized treatment strategies.