Ali Essam Ghareeb, Benjamin Chang, and colleagues from the independent AI lab FutureHouse, the University of Oxford, and Fordham University published Robin, an open‑source agent that proposes repurposing existing drugs for specified diseases. The system designs experiments, recommends candidate drugs, and analyzes lab results in an iterative human‑in‑the‑loop workflow; humans performed the wet‑lab tests on isolated human cells.
How Robin works
Robin follows an iterative cycle: (i) identify possible biological mechanisms underlying a disease, (ii) design experiments to test those mechanisms, (iii) find existing drugs that modulate the mechanisms, (iv) have humans run the experiments in a lab, and (v) analyze the results and update hypotheses. For natural‑language tasks Robin primarily uses OpenAI’s GPT‑4‑mini.
Given a disease name, Robin generates questions about it. The Crow literature‑search agent produces concise summaries of relevant medical literature, which Robin uses to select ten candidate mechanisms. For each mechanism, Crow searches for relevant studies and proposes experimental designs; Claude 3.7 Sonnet evaluates these designs pairwise to produce a ranking.
From the top experimental design, Robin searches the web for 30 drugs that act on the mechanism, prioritizing agents that are commercially available (and thus safety‑tested) and have not previously been used to treat the disease. The Falcon literature‑search agent creates a report for each drug summarizing evidence and limitations, and Claude 3.7 Sonnet ranks these reports in a tournament.
Human researchers review the ranked list and test the top candidates in the lab using the previously generated experimental designs. After the experiments, researchers upload the data and instruct Robin which analysis to run. The Finch data‑analysis agent performs the requested analyses and returns a summary. Robin then generates follow‑up experiments and the cycle continues until humans judge a candidate satisfactory.
Software components and licensing
Robin itself is available under the Apache 2.0 license. It depends on earlier agents: Crow and Falcon (bundled as “Literature”) are free for research use only, while Finch is available under Apache 2.0.
Results for dry age‑related macular degeneration (dAMD)
The authors ran the pipeline for dry age‑related macular degeneration (dAMD). Robin hypothesized that increasing RPE (retinal pigment epithelium) phagocytosis — the process by which RPE cells clear debris and pathogens — could be therapeutic. The system proposed candidate drugs to boost RPE phagocytosis; two proved effective in cell‑based assays:
- Y‑27632 (a research compound): in the initial run, treated eye cells showed nearly a 2× increase in RPE phagocytosis compared with the same test without the drug; a human follow‑up analysis reproduced the result.
- Ripasudil (a drug approved in Japan for glaucoma): in the second run, experiments showed a 1.89× increase in RPE phagocytosis, and a human follow‑up analysis reported a 1.75× increase.
The authors note that automating interpretation of biological data is difficult because the inherently ambiguous nature of such data can lead humans and AI to reach different conclusions across runs.
Limitations
All tests reported were performed on isolated human eye cells; no clinical trials in patients were conducted. Therefore it remains unknown whether the observed in vitro effects translate to therapeutic benefit in living patients.
Context and significance
This work builds on earlier efforts to accelerate scientific discovery with agentic systems. Prior systems have generated machine‑learning research proposals at human‑level quality, produced and executed code to test hypotheses, and in some cases generated papers describing the experiments. Google’s AI Co‑Scientist produced biomedicine research proposals that were later validated in the lab.
Robin’s combination of hypothesis generation, experimental design, result analysis, and iterative updates suggests AI agents could streamline parts of medical research. If combined with robotic laboratory automation — as demonstrated for example by RoboChem, where an AI model controls automated lab instruments — similar pipelines might further reduce the time and human effort needed to test hypotheses.
Why this matters: developing a new drug in the United States typically takes over a decade and can cost more than $1 billion, while regulators approve on the order of 50 new drugs per year. Repurposing already‑approved drugs for new indications, or exploiting shared underlying mechanisms across diseases, is a cost‑efficient strategy that could accelerate the discovery of treatments.
Conclusion
Robin demonstrates that an agentic, largely open‑source pipeline can propose repurposed drug candidates, design and prioritize experiments, and analyze in vitro results in an iterative human‑in‑the‑loop process. The initial cell‑based findings for dAMD are promising, but clinical validation is required before any conclusions about patient benefit can be drawn.
