Modern biomedical research increasingly relies on integrating information across multiple heterogeneous datasets, yet real-world data are often fragmented, high-dimensional, and subject to substantial distributional differences across studies. In this talk, I present a unified framework for learning from such heterogeneous data that enables robust integration while avoiding misleading or spurious information transfer. I will introduce a sequence of methods that address key challenges in this setting, including robust estimation under heterogeneous sources, valid uncertainty quantification for reliable inference, and adaptive integration of multimodal data with missingness. These approaches leverage only summary-level information from external studies, improving practical feasibility while maintaining robustness and interpretability. We demonstrate their practical utility through applications to EHR-linked biobank data, the Alzheimer’s Disease Neuroimaging Initiative (ADNI), and Genotype-Tissue Expression (GTEx) datasets.