Sony AI's Hakken Predicted 1.5M Aging Hypotheses and Two Held Up in Lab

Sony AI has built a system that proposes scientific facts nobody has published yet, then explains its reasoning well enough for a biologist to take it into the lab. The system, called Hakken (Japanese for "discovery"), generated 1.5 million candidate hypotheses about aging-related genes. Three were sent to a contract lab for wet-lab validation. Two came back confirmed as previously undocumented biological interactions.

Drowning in papers

Biomedical research produces over one million papers per year, a rate that keeps climbing. Existing AI tools mostly help researchers navigate what already exists: summarizing, retrieving, connecting dots already implicit in the literature. Hakken predicts relationships not yet documented anywhere, then hands the scientist a paper trail explaining why.

How it actually works

Hakken has two main components. The prediction model, THiGERLLM, takes a pair of entities (say, two genes) and predicts what kinds of relationships likely hold between them. The explanation module, PHELInE, then traces paths through the existing literature to justify each prediction.

THiGERLLM combines two views of the same problem:

• A temporal graph branch: The literature is converted into knowledge graphs that grow year by year, letting the model reason about how ideas evolve rather than treating knowledge as static. A graph neural network learns how evidence accumulates around each entity pair over time, and a hierarchical transformer compresses that history into a compact embedding.