ForeSCITE Maps Research Trends into Biomedical Hypotheses

ForeSCITE’s compelling idea is to model scientific progress as directional movement rather than treating literature as a static knowledge base.

• –Clusters about three million biomedical conclusions into 407 thematic groups
• –Projects conclusion embeddings toward learned “future” directions, then decodes them into natural-language hypotheses
• –Retrospective tests outperformed prompt-based and retrieval-augmented baselines on semantic coverage and entity-pair novelty
• –Expert review found 37 CAR-T candidates both novel and biologically plausible
• –Publication trends can encode bias, and prospective experiments—not retrospective literature overlap—must determine whether these hypotheses are genuinely useful