Sony AI researchers published Hakken, a system that predicts undocumented links between scientific concepts. It reads sequences of concept maps built from large volumes of research papers. A transformer model combines these with a language model's semantic knowledge. It predicts both whether a relationship exists and what type it is. A separate explanation component supplies reasoning for each prediction so scientists can judge it. The authors say the design is domain-agnostic but demonstrate it on biomedicine. They scored 1.5 million aging-related hypotheses above a confidence threshold. Biologists reviewed batches of these, and three went to laboratory testing. Two were confirmed, naming new interactions between TP53 and BAMBI and between RAF1 and TNF. The team reports a new benchmark result for time-aware relation prediction.

• 1.5 million hypotheses scored
• 3 predictions taken to wet-lab
• 2 predictions confirmed
• 66-page paper