Kan AI generere plausible videnskabelige hypoteser fra omfattende biomedicinsk litteratur på sekunder ?

Nye AI-systemer kan læse tusindvis af forskningsartikler og identificere nye forbindelser mellem studier. Disse modeller bruger transformer-arkitekturer, der er trænet på biomedicinske tekster, til at foreslå forskningsretninger. Farmaceutiske virksomheder tester dem for at fremskynde lægemiddeludviklingsprocesser. Hypoteserne kræver stadig streng eksperimentel validering, før de kan accepteres.

Background

Current systems can ingest millions of abstracts, rapidly surface statistically associated molecular or disease patterns, and even suggest mechanistic links that humans had missed—an approach sometimes called “robot scientist” or literature-based discovery. Pharmaceutical companies are testing them to accelerate drug discovery pipelines. However, the resulting hypotheses still require expert curation to distinguish plausible mechanistic narratives from statistical artifacts and to ensure biological feasibility. In controlled biomedical challenges, AI has produced testable drug–target or disease–pathway hypotheses that were later validated in lab experiments, showing promise but not yet matching the full rigor of hypothesis generation by seasoned investigators. Work continues on making these systems more explainable, reproducible, and aligned with experimental constraints so they can truly operate at “seconds” speed while maintaining scientific trustworthiness.

New AI systems use transformer architectures trained on biomedical texts to propose research directions. Current systems can already ingest millions of abstracts, rapidly surface statistically associated molecular or disease patterns, and even suggest mechanistic links that humans had missed—an approach sometimes called “robot scientist” or literature-based discovery. Pharmaceutical companies are testing them to accelerate drug discovery pipelines. These models use transformer architectures trained on biomedical texts to propose research directions.