A 2025 Science Advances study found that low-dose metformin's glucose-reducing effect in mice depended on a brain Rap1 pathway. The evidence comes from experimental animals, and the parallel for medical AI is limited: a plausible explanation cannot replace external and clinical validation.
Taiwan-based Repurgenesis said in late July 2026 that it had signed a memorandum of understanding with Switzerland-based Medvisis to pursue AI-assisted drug repurposing. The public announcement names a platform and a commercial network, but not a candidate, dataset, validation plan or binding development milestone.
AI-designed drug candidates have moved into mid- and late-stage trials, but APPI News found no publicly identified US approval for a novel drug originating from an AI-led discovery program as of August 4, 2026. Early data point to faster candidate selection and stronger Phase 1 results, while later trials still impose the usual tests of safety and efficacy.
A 2025 systematic review found that non-benzodiazepine muscle relaxants improved several short-term outcomes in acute low back pain compared with placebo. Adverse events were more frequent, while a broader review found that evidence for most back-pain treatments remained uncertain.
A 2024 systematic review found that explainable AI increased clinicians' trust in five of 10 empirical studies, while two reported mixed effects. A COVID-19 drug-repurposing project shows how an inspectable workflow can support expert review, but explanations do not establish clinical validity.
GPT-5 Pro generated a mechanism for unexplained T-cell data and predicted the direction of an unpublished CAR-T experiment, according to a November 2025 case study. The result shows how a model can help experts form testable hypotheses, but it does not establish autonomous discovery or prove the proposed mechanism.
OpenAI and Molecule.one reported on June 17 that a GPT-5.4 workflow helped identify conditions that improved a difficult Chan-Lam coupling across 10,080 microscale reactions. An automated laboratory supplied the testing capacity, while chemists selected proposals, corrected plans and repeated key experiments by hand.
OpenAI reported in June 2026 that GPT-Rosalind passed 36.1 percent of the 750 tasks in its new LifeSciBench evaluation. Performance fell on tasks involving figures, files or web sources, while the benchmark's design and publication by OpenAI complicate interpretation of its model rankings.
