An AI-Powered Scientist Proposes a Treatment for Blindness

Since the term “artificial intelligence” (AI) was first coined in the 1950s, what once belonged to the realm of science fiction has steadily become a reality. Each decade has brought its breakthroughs, from early chatbots to digital assistants like Siri and Alexa, to cutting-edge systems like AlphaFold for protein structure prediction, and ChatGPT for language generation.

As AI capabilities continue to grow, so does its impact—particularly in the field of scientific discovery. In recent years, researchers have begun exploring how to build AI scientists: tools designed to assist human researchers with hypothesis generation, literature review, experiment design, and data analysis.

FutureHouse, a non-profit AI research organization, is working to make this vision a reality. “When we started [a few years ago], people thought it was insane to build an AI scientist,” remarked Andrew White, the head of science and co-founder of FutureHouse.

In a recent arXiv preprint, the scientists introduced Robin, an AI-driven platform designed as a multi-agent system to automate various stages of scientific research.¹ Each agent performs an aspect of research: Crow performs a general search of the literature, Falcon conducts a deeper dive of literature reviews for in-depth analysis (both look at open-access resources), and Finch analyzes experimental data and refines hypotheses through iteration. Originally developed as individual tools, the team aimed to integrate them to generate a novel hypothesis, propose experimental assays, and identify therapeutic candidates for a target disease.

The team focused on age-related macular degeneration (AMD), the leading cause of severe vision loss in individuals aged 50 and above. An estimated 20 million Americans are affected by some form of AMD, which exists in two types: wet and dry.² Dry AMD (dAMD), which accounts for more than 80 percent of AMD cases, currently lacks an effective treatment. Using this condition as a case study, the researchers asked Robin to propose a potential therapeutic drug candidate, and it identified a mechanism of action involving phagocytosis.

“We just put these things together in a loop and found that we were able to get really exciting results,” said White. He emphasized that Robin carried out computational insights, while human scientists performed physical experimentation and manuscript writing.

After reviewing the literature with Crow, Robin identified 10 potential disease mechanisms. Using a tournament-style approach, it made pairwise comparison brackets and calculated rankings. Of these, the top candidate proposed enhancing retinal pigment epithelium (RPE) phagocytosis as a therapeutic strategy. Falcon then evaluated candidate drug molecules to achieve this goal, and Finch analyzed experimental data showing that the rho kinase (ROCK) inhibitor Y-27632 enhanced RPE phagocytosis in cell culture.

To investigate the mechanism, Robin proposed an RNA-sequencing experiment, which revealed that Y-27632 upregulated ABCA1, a lipid efflux pump which may play a critical role in RPE cell health. Building on this, Robin suggested a second round of drug candidates, leading to ripasudil—a clinically approved ROCK inhibitor for treating glaucoma in Japan, but not previously associated with dAMD.³ Follow-up experiments confirmed the drug’s effectiveness in enhancing phagocytosis of RPE cells.

Although ROCK inhibitors are used in other areas of ophthalmology, their role in promoting phagocytosis is less established. “One of those key functions was phagocytosis, which Robin picked up on in our paper as a disease mechanism,” said coauthor Ali Ghareeb, an ophthalmologist and scientist at FutureHouse.

Ghareeb explained that although wet and dry AMD share early pathologic changes in RPE cells, they have unique genetic risk factors and diverge substantially as the disease advances. This distinction is critically important as dAMD patients do not have effective treatment. He added, “The way that we created Robin was that you could take ideas from very diverse fields of medicine…and then synthesize them into something totally new. [Robin] is basically a cross-disciplinary expert in drug discovery.”

Image of coauthors Ali Ghareeb (left) and Andrew White (right) standing in the lab. Both men smile at the camera. — FutureHouse scientists Ali Ghareeb and Andrew White noted their excitement for the use of this AI agent system for future discoveries across all diseases.
FutureHouse

The study drew attention from the scientific community, including discussions of earlier work proposing ROCK inhibitors for AMD.⁴ After reading the paper, Konrad Kording, a neuroscientist at the University of Pennsylvania who was not involved in the study, said it prompted him to search online to see if the proposed strategy had already been explored in previous human-led research. “This is the big question with which the value proposition rises and falls. I think they readily admit that anyone in the field would have known about this being a meaningful hypothesis.”

“We all use tools, like Google Scholar and Grammarly, and AI is moving into that [space] now,” said Kording. Still, he noted, the paper raises a key question: Can humans eventually be removed from the research process to make science more efficient? “It’s a big dream, and I’m not sure if the time has come for it yet.” Kording expressed skepticism about the preprint's claims of fully automating scientific discovery, emphasizing the need for human oversight and validation. “That being said, they built a couple of tools, and those tools seem interesting.”

Since the preprint’s announcement, the FutureHouse team has stated that, to their knowledge, no other group has proposed using ROCK inhibitors to treat dry AMD and that it would have been very difficult to arrive at this hypothesis without their AI agents. They acknowledged, however, that while the discovery is promising, it is not yet a “move-37” moment—a bold, novel leap made by AI—and further validation is needed before it can be considered a viable treatment.

“[People] on Twitter are saying, ‘This is obvious in retrospect,’ which is great. It means that people believe it and it’s possible,” said White. “Everyone can debate how obvious it was, but at the end of the day it will be helpful for patients at the end of the pipeline.”