Robin: The AI Scientist That Performs Scientific Discovery - And Discovered a New Drug Candidate for Blindness

 

Researchers from FutureHouse have developed Robin, a groundbreaking multi-agent artificial intelligence system capable of autonomously generating scientific hypotheses, proposing experiments, analyzing biological data, and refining therapeutic strategies through iterative experimentation. The study, published in Nature, represents one of the first demonstrations of an AI platform functioning as an active participant in the scientific discovery process rather than merely serving as a research assistant.

Robin was specifically designed to automate the major intellectual stages of biomedical research — literature review, hypothesis generation, experimental planning, data interpretation, and follow-up discovery. Using this system, the researchers identified ripasudil, a clinically approved glaucoma drug, as a potential new therapeutic candidate for dry age-related macular degeneration (dAMD), a leading cause of blindness worldwide. Importantly, this connection had never previously been proposed in the scientific literature.

The Growing Challenge of Scientific Knowledge Overload

Modern biomedical science generates enormous volumes of information through genomics, high-throughput screening, single-cell analysis, and large-scale imaging technologies. However, the ability of scientists to synthesize and connect this information has become an increasing bottleneck.

Drug discovery is particularly vulnerable to this problem because identifying new therapeutics requires integrating knowledge across molecular biology, pharmacology, genetics, toxicology, and clinical medicine. Many therapeutic opportunities remain hidden within disconnected scientific disciplines for years before researchers recognize them.

The FutureHouse team developed Robin to address this challenge directly. The system combines multiple specialized AI agents that collaborate together in a structured scientific workflow.

Robin incorporates:

  • Crow — a literature exploration agent designed for broad scientific review.
  • Falcon — a deep literature evaluation agent that analyzes therapeutic candidates.
  • Finch — an autonomous biological data-analysis agent capable of writing and executing original Python and R code.

 
Together, these agents form an integrated AI research system capable of performing semi-autonomous scientific discovery.

Robin Chose Retinal Phagocytosis as a Therapeutic Strategy

The researchers tested Robin using dry age-related macular degeneration, a disease affecting millions of individuals globally and currently lacking effective long-term therapies.

Beginning with only the phrase “dry age-related macular degeneration,” Robin reviewed hundreds of scientific papers and identified retinal pigment epithelium (RPE) phagocytosis enhancement as a promising therapeutic mechanism.

The retinal pigment epithelium plays an essential role in maintaining photoreceptor health by engulfing and clearing cellular debris from the retina. Dysfunction of this phagocytic process is strongly associated with AMD progression.

Robin then proposed a flow cytometry-based phagocytosis assay and generated thirty candidate therapeutic compounds for experimental testing.

Among the top-ranked candidates were:

  • Exendin-4
  • Fingolimod
  • MFGE8
  • Y-27632
  • AICAR/TUDCA combinations

 
One of Robin’s most important insights was the identification of ROCK inhibition as a potential strategy for enhancing retinal phagocytosis.

AI Performed Autonomous Experimental Data Analysis

Human scientists performed the proposed experiments using ARPE-19 retinal cells and fluorescent phagocytosis assays. The resulting raw flow cytometry datasets were uploaded back into Robin.

Robin then deployed Finch, its autonomous data-analysis agent.

Finch independently processed the flow cytometry data by:

  • excluding debris and dead cells,
  • performing gating analysis,
  • quantifying phagocytic activity,
  • conducting statistical testing,
  • generating publication-style visualizations.

 
The analysis confirmed Robin’s prediction that Y-27632, a ROCK inhibitor, significantly enhanced retinal phagocytosis.

This marked a major milestone because the AI system successfully completed an entire scientific reasoning loop:

  1. Literature synthesis
  2. Hypothesis generation
  3. Experimental proposal
  4. Biological data analysis
  5. Scientific interpretation

—all within one continuous workflow.

Robin Identified a Novel Molecular Mechanism

Following the initial success, Robin proposed a second experiment involving RNA sequencing of Y-27632-treated retinal cells.

Previous studies had primarily associated ROCK inhibition with cytoskeletal remodeling. However, Robin hypothesized that transcriptional mechanisms might also contribute to enhanced phagocytosis.

Finch analyzed the RNA-seq data and identified significant changes in genes associated with:

  • actin filament organization,
  • small GTPase signaling,
  • autophagy pathways,
  • lipid metabolism.

 
The most striking finding was the approximately three-fold upregulation of ABCA1, a lipid efflux transporter essential for retinal pigment epithelium homeostasis and strongly linked to macular degeneration biology.

Robin therefore not only identified a therapeutic candidate but also uncovered a potentially novel molecular mechanism linking ROCK inhibition to retinal health.

Robin Proposed Ripasudil as a Potential Therapy for Blindness

Using the insights from the first experimental cycle, Robin initiated a second round of therapeutic hypothesis generation.

This time, the system identified ripasudil, a clinically approved glaucoma drug used in Japan.

Ripasudil is a Rho kinase (ROCK) inhibitor with an established ocular safety profile, but it had never previously been proposed as a treatment for dry AMD.

Experimental validation showed that ripasudil:

  • increased retinal phagocytosis approximately 1.89-fold,
  • demonstrated greater potency than Y-27632,
  • showed minimal cytotoxicity,
  • remained effective in primary human retinal stem cell-derived RPE cultures.

 
Robin also identified KL001, a circadian clock modulator, as another novel enhancer of retinal phagocytosis.

Importantly, the researchers validated ABCA1 upregulation in primary human retinal cells treated with ripasudil, further strengthening Robin’s mechanistic findings.

Robin Dramatically Accelerated Scientific Discovery

One of the most remarkable aspects of the study was Robin’s efficiency.

The researchers estimated that Robin compressed a scientific discovery workflow requiring approximately 872–937 hours of human cognitive labor into less than two hours.

Robin analyzed roughly 825 scientific references in approximately thirty minutes — a task estimated to require over 800 hours of manual scientific reading.

Robin Outperformed General-Purpose AI Systems

The FutureHouse researchers also compared Robin against general-purpose AI research agents, including OpenAI’s Deep Research system.

When both systems were tasked with generating novel therapeutic candidates for retinal phagocytosis enhancement:

  • Robin identified multiple experimentally validated hits.
  • Deep Research produced no successful therapeutic candidates.
  • Deep Research failed to identify ROCK inhibition as a therapeutic mechanism.

 
These findings suggest that specialized, literature-grounded multi-agent scientific AI systems may substantially outperform broader general-purpose AI systems in biomedical discovery applications.

A New Era of AI-Driven Science

The authors emphasize that Robin remains a “lab-in-the-loop” system requiring human oversight and experimental validation. The system cannot yet autonomously perform laboratory experiments or generate fully executable protocols.

Nevertheless, Robin represents a major conceptual shift in the relationship between AI and scientific research.

Rather than simply summarizing information, Robin demonstrated the ability to connect non-obvious biological insights across multiple scientific disciplines and generate experimentally validated therapeutic hypotheses.

The implications extend far beyond ophthalmology. Similar AI systems could eventually accelerate drug discovery across cancer biology, neurodegeneration, immunology, metabolic disease, and regenerative medicine.

As the authors conclude, Robin establishes a new paradigm for AI-driven scientific discovery — one in which AI systems participate directly in the intellectual core of science itself.

Reference

Ghareeb, A.E., Chang, B., Mitchener, L. et al. A multi-agent system for automating scientific discovery. Nature (2026). https://doi.org/10

Image credit: Portions of the figure in this article were generated using ChatGPT (OpenAI) or Google Gemini.

Disclaimer: This blog post is intended solely for educational and scientific informational purposes. Any mention of therapeutic drug names, including FDA-approved medications, is for the purpose of accurate reporting and discussion of biomedical research and does not constitute medical advice, endorsement, or promotion. Readers should not interpret the content as a recommendation for any specific treatment. Always consult a qualified healthcare professional for medical advice or treatment decisions.

About Marin Biologic Laboratories

Our Recent Publication/Meeting Presentation on Gene Therapy

1. Development of VNX-101, an Adeno-Associated Virus with Less Immunogenicity and Efficient Long-Term Expression of a CD19 T-Cell Engager. Molecular Therapy Methods & Clinical Development, published online July 24, 2025.

3. Cell-Based Potency Assay for Anti-CD3-Anti-CD19 Diabody. Journal of Immunological Methods. 2025. 545-114004.

3. Development of a Pharmacokinetic (PK) Mouse Serum GLP ELISA for an Anti–CD19–AntiCD3 Diabody
bioRxiv 2025.03.19.644217; doi: https://doi.org/10.1101/2025.03.19.644217.

4. American Society of Hematology (ASH) Annual Meeting 2024.
Abstract link: Using Gene Therapy to Solve Challenges with CAR-T Cell Immunotherapy: Lead Selection and Preclinical Development of an Adeno-Associated Virus with Reduced Immunogenicity Exhibiting Efficient and Long-Term Expression of an Anti-CD19 T-Cell Engager.

Comprehensive Assay Solutions for In Vitro and Cell Based Potency Assays and Pharmacokinetics Studies- Our Expertise

With 30 years of expertise in cell culture, cell-based assays, and preclinical/clinical PK/PD analysis, we specialize in offering assay services essential for a wide variety of therapeutic drug development programs, preclinical studies, IND/BLA applications, and commercialization. Our comprehensive services include both preclinical non-GLP and GLP assays, as well as non-GMP and GMP assays, providing critical support throughout the entire development pipeline.

Watch the following video and explore our latest presentation on the development and validation of potency and pharmacokinetic (PK) assays for AAV vectors, highlighting innovative methodologies and industry-leading expertise.

 

 

Download the full presentation: Development of Custom Cell Based and In vitro Potency and Pharmacokinetics (PK) Assays for AAV vectors- Marin biologic Laboratories

 

Development of Cell-Based Potency Assays: Case Studies and Blogs from Marin Biologic Laboratories (MarinBio)

 

Drug Discovery & Development Assays Offered by Marin Biologic Laboratories (MarinBio)