We developed a biology-informed model that uses real-world multimodal data to predict which metastatic breast cancer patients are most likely to benefit from trastuzumab deruxtecan (T-DXd), while also tracking how resistance emerges over time. Built on RNA-seq data and a molecular foundation model trained on more than one million transcriptomes, the approach identified distinct responder groups with longer treatment benefit and showed strong treatment-specific predictive power. Longitudinal analyses also uncovered consistent biological changes linked to acquired resistance, helping reveal new opportunities to improve patient selection and guide future therapeutic strategies.