LondonRobin scanned 551 papers in 30 minutes, surfacing drug candidates for leukemia and macular degeneration that researchers had missed.
2 Nature papers validate the approach; Google’s Co-Scientist and FutureHouse’s Robin both searched curated literature for cross-field connections rather than generating open-ended hypotheses.
Neither system has cleared animal trials; most cell-culture hits fail at that stage.
Sources: Ars Technica
WHAT REPOSITIONING IS
Drug repositioning means finding a new disease for an already-approved drug. The molecule has already passed toxicity trials, so a hit can reach patients in years rather than the decade-plus a novel compound needs.
THE LITERATURE PROBLEM
PubMed indexes over 36 million biomedical papers and adds roughly a million a year. No human can read across oncology, ophthalmology, neurology and cardiology at once — yet the connections that yield repositioning hits often sit precisely in those gaps between fields.
WHY RETRIEVAL, NOT GENERATION
Both Robin and Co-Scientist constrain the model to curated literature rather than letting it invent hypotheses freely. This is deliberate: open-ended generation hallucinates plausible-sounding mechanisms that don't exist. Retrieval over real papers keeps every claim traceable to a citation.
THE VALLEY OF DEATH
A drug candidate that works in a petri dish has cleared the easiest test in pharmacology. Roughly 90% of compounds that show activity in cell culture fail in animal models; of those that reach human trials, only about one in ten gets approved. The papers describe hits, not drugs.
THE INCENTIVE GAP
Repositioning an off-patent drug yields little exclusivity, so big pharma rarely funds the trials. Most repositioning candidates that survive cell culture die in this incentive gap — not from biology, from economics. AI surfacing more candidates does not by itself solve who pays to test them.
THE NATURE BAR
Nature publishes roughly 8% of submissions, and validation papers — claiming an AI system did something useful — face harder review than pure-method papers. Two Nature papers on the same week is the field signaling that retrieval-based scientific assistants have crossed from demo to publishable result.