The biomedical research community has generated over 9 million publicly available datasets and samples spanning genomics, transcriptomics, proteomics, and clinical data—representing an unprecedented opportunity for scientific discovery. Yet these resources remain severely underutilized due to fundamental barriers: fragmented metadata across repositories, inconsistent terminologies, and the inability to semantically connect datasets based on biological meaning rather than exact keyword matches.

This talk will demonstrate how AI/ML approaches are transforming our ability to unlock these data treasures. I’ll present novel frameworks that combine large language models, graph neural networks, and retrieval-augmented generation to: (1) automatically infer standardized annotations of datasets from their unstructured, plain-text metadata by extracting biomedical context from publications and ontologies and (2) create cross-modal spaces where textual descriptions and molecular profiles become translatable to generate novel hypotheses.

Through concrete examples spanning rare disease research, drug repurposing, and high-content imaging, I’ll also show how these AI-powered approaches enable researchers to pose entirely novel biological questions to millions of existing samples—moving beyond pre-calculated results to generate custom analyses on-demand. The implications extend beyond biomedicine: as massive data collections proliferate across scientific domains, AI-driven semantic integration represents a paradigm shift from treating public data as static archives to dynamic discovery engines that accelerate scientific breakthroughs while maximizing return on prior research investments.