Why do so many in silico models fail when moved to the lab or clinic? Too often, they’re trained on incomplete, non-human, or non-representative datasets. This session tackles the “data gap” head-on: from interoperability bottlenecks and the black box problem to the limits of current virtual cell simulations (~50 million perturbations vs. the billions biology demands). Panelists will explore how to create “human-first” datasets that reflect real biology, unlock mechanistic interoperability, and close the discovery–development divide. The goal: build AI tools that can directly identify viable drug candidates instead of stalling in silico.