Dry age-related macular degeneration destroys the central vision of more than 8 million Americans. There is no approved treatment that slows its progression. It is one of the largest unmet needs in ophthalmology — a disease with known pathology, no shortage of research investment, and no cure. In May 2026, a multi-agent AI system called Robin read the literature, formed hypotheses, designed experiments, analysed results, and proposed a drug that no researcher had ever suggested for this condition. The discovery was published in Nature.
2.5 months, one drug candidate
Robin is an AI system built by FutureHouse, a nonprofit research organisation founded with a specific mission: automate scientific discovery. The system orchestrates three specialised AI agents — Crow for broad literature search, Falcon for deep evaluation of candidate molecules, and Finch for autonomous data analysis including writing and executing its own Python and R code. A supervising layer coordinates them through an iterative cycle of hypothesis, experiment, interpretation, and refinement.
The team applied Robin to dry age-related macular degeneration (dAMD). The disease is caused by the progressive degeneration of the retinal pigment epithelium (RPE) — a single layer of cells at the back of the eye that supports the photoreceptors above it. When RPE cells fail to clear cellular debris through a process called phagocytosis, that debris accumulates, and the photoreceptors they sustain begin to die. Central vision follows.
The discovery unfolded in three rounds:
Literature review → initial hypothesis. Crow surveyed the dAMD literature and Robin formed its first hypothesis: that enhancing RPE phagocytosis could slow disease progression. Falcon screened candidate molecules. The team tested ten of them in the lab. Robin then directed Finch to analyse the results — and identified that Y-27632, a Rho-kinase (ROCK) inhibitor, significantly increased phagocytosis in RPE cell culture.
Mechanism investigation. Robin proposed an RNA-sequencing experiment to understand why Y-27632 worked. The experiment was conducted by human researchers. Finch analysed the gene expression data and found that the drug upregulates ABCA1, a critical lipid efflux pump in RPE cells that helps clear the lipid deposits central to dAMD pathology.
Second drug candidate screen. Using its new mechanistic understanding, Robin proposed a second set of candidates. Among them was ripasudil — a ROCK inhibitor already approved for clinical use in glaucoma. When tested, ripasudil outperformed all other candidates. A drug already in clinical use for a different eye disease had been identified, for the first time, as a candidate for dAMD.
The entire process — from Robin's first hypothesis to paper submission — took 2.5 months. FutureHouse estimates it represented a 200-fold reduction in researcher time compared to a conventional experimental workflow. Human researchers executed the physical experiments; every intellectual step — the hypotheses, the experimental designs, the analyses, the main text figures — was generated by Robin.
to paper submission
in researcher time
currently untreatable
Co-Scientist arrives in the same issue
Robin's paper was published back-to-back in Nature on May 19, 2026, alongside a second paper describing Co-Scientist — a multi-agent AI system developed by Google DeepMind. The two teams worked independently. The coincidence of publication was not coordinated. That two separate research organisations arrived at the same architectural conclusion — that multi-agent AI systems can function as autonomous scientific collaborators — in the same week is itself a signal.
Co-Scientist is built on Gemini 2.0 and is designed as a general-purpose research partner, applicable across disciplines. Where Robin focused on a single disease, Co-Scientist's Nature paper documented results across three separate biomedical challenges simultaneously.
Acute myeloid leukaemia. Co-Scientist proposed novel repurposing candidates and combination therapies for AML, an aggressive blood cancer. Several suggestions were confirmed to inhibit tumour viability in multiple AML cell lines at clinically relevant concentrations.
Liver fibrosis. The system identified Vorinostat — an FDA-approved anti-cancer drug — as a candidate for treating liver fibrosis. In hepatic organoid experiments, Vorinostat reduced a key TGFβ-induced chromatin structural change by 91%.
Antimicrobial resistance. Co-Scientist generated hypotheses about the molecular mechanisms by which resistance genes transfer between bacteria — work relevant to understanding how AMR spreads at the evolutionary level.
Both papers carry the same caveat clearly: these are preclinical results. None of the drug candidates have entered human trials. The path from in-vitro cell line results or organoid experiments to clinical use is long and expensive, and the majority of candidates fail at some stage. The researchers are not claiming cures. They are claiming, more modestly but more significantly, that AI can now function as a genuine hypothesis-generating, experiment-designing, data-analysing partner in the early stages of scientific research.
"Both teams emphasize that these systems are designed to collaborate with researchers, and a scientist would always be in the loop."— Nature Press Release, May 19, 2026
Why ripasudil matters
The choice of ripasudil as Robin's lead candidate is worth examining closely, because it illustrates what AI brings to drug discovery that human researchers consistently underperform on: exhaustive, bias-free literature synthesis across disconnected fields.
Ripasudil has been approved in Japan since 2014 for glaucoma and ocular hypertension. It works by inhibiting Rho-kinase, an enzyme involved in cellular contractility and cytoskeletal regulation. Rho-kinase inhibition affects a wide range of cellular processes — including, it turns out, the phagocytic activity of RPE cells. The connection between glaucoma treatment and dAMD pathology runs through a mechanism that is well-documented in separate bodies of literature that rarely intersect in clinical practice.
Human researchers working within the ophthalmology field are deeply familiar with the dAMD literature. They are less likely to have systematically reviewed the biochemistry of ROCK inhibition across cell types, cross-referenced it with the RPE biology literature, and arrived at ripasudil as a candidate. That cross-domain synthesis is precisely where AI — which reads without disciplinary borders — has a structural advantage.
The path forward requires preclinical animal studies, then Phase I safety trials, then efficacy trials. If ripasudil works in human dAMD — and that is still an open question — it has a significant advantage over novel compounds: it already has a known human safety profile. Drug repurposing skips the years of early safety characterisation that most new drug candidates require. That alone could cut years from the development timeline.
A new kind of co-author
The most important claim in both papers is not that the AI made a specific discovery. It is that the AI performed the intellectual work of science — and did so reproducibly, transparently, and fast enough to matter. Every hypothesis, every experimental choice, every analysis in Robin's dAMD paper was generated by the system and documented in full agent trajectories made publicly available. The scientific reasoning is auditable in a way that human intuition-driven research often is not.
FutureHouse's mission is explicit: to automate scientific discovery. Not to assist scientists. To automate. The Robin system is a step toward a future where AI systems run large parts of the hypothesis-experiment-analysis cycle, with humans providing oversight, physical execution, and final judgement. The 200-fold reduction in researcher time is not a curiosity — it is an estimate of what happens to the pace of science if this approach scales.
That scaling is not guaranteed. Robin used a small, well-defined problem domain. The system's ability to operate in messier, less-well-characterised research areas — where the relevant literature is sparse, contradictory, or spread across languages and disciplines — remains untested. The fundamental limits of language-based AI systems, which both papers also document, matter here: scientific precision requires quantitative specificity that natural language can obscure.
But the dAMD result is a proof of concept with a named molecule, a validated mechanism, and a published paper in the most prestigious scientific journal in the world. For 8 million people with no treatment options, that is not an abstraction.
Buy me a coffee