For the modern physician, keeping pace with the exponential growth of medical literature and high-dimensional genomic data presents a significant “breadth and depth conundrum”: the need for specific expertise vs. the need for trans-disciplinary insight. However, a new generation of multi-agent AI systems—including Co-Scientist, Robin, and Empirical Research Assistance (ERA)—is transforming from simple search tools into structured “thinking engines” capable of generating novel medical hypotheses and analyzing experimental data.
Accelerating Drug Repurposing: AML and Beyond
Drug development is traditionally a multi-year, multi-billion-dollar process, but AI is now identifying therapeutic shortcuts through drug repurposing. The Co-Scientist system, built on the Gemini architecture, uses a “generate, debate, and evolve” paradigm to simulate the scientific method.
In a landmark validation for hematologic oncology, Co-Scientist independently proposed several single-agent and combination therapies for Acute Myeloid Leukemia (AML). The system identified Binimetinib (originally for melanoma) as a potential frontline treatment and proposed synergistic dual and triple-drug regimens (e.g., JNJ-64619178 + Selinexor) that were successfully validated in laboratory cell lines. These discoveries were made by analyzing thousands of papers in a fraction of the time required by human experts.
Precision Public Health: Outperforming the CDC Ensemble
Beyond bench research, AI is enhancing clinical forecasting and epidemiology. The Empirical Research Assistance (ERA) system was designed to optimize software for “scorable tasks”. In a retrospective study of COVID-19 hospitalization forecasting, ERA generated 14 distinct models that outperformed the CDC’s official CovidHub Ensemble—historically the gold standard for U.S. epidemiological prediction. By tirelessly exploring thousands of algorithmic variations, ERA identified “needle-in-the-haystack” solutions that improved prediction accuracy across a majority of U.S. jurisdictions.
“Lab-in-the-Loop”: A New Paradigm for Mechanisms and Targets
The Robin system introduces a “lab-in-the-loop” framework, where the AI not only generates a hypothesis but also autonomously analyzes the resulting laboratory data to refine its next set of questions.
- Novel Target Identification: When tasked with identifying mechanisms for chronic degenerative diseases, Robin proposed an experiment that, upon RNA-seq analysis, revealed the 3-fold upregulation of ABCA1, a critical lipid efflux pump.
- Liver Fibrosis: Co-Scientist identified novel epigenetic targets for liver fibrosis, proposing drugs that exhibited significant anti-fibrotic activity in human hepatic organoids.
- Antimicrobial Resistance (AMR): In a remarkable demonstration of accuracy, the AI independently recapitulated a breakthrough mechanism of gene transfer in bacteria—matching unpublished findings from human research teams—in just two days.
The Future: The “Scientist-in-the-Loop”
These systems are not intended to replace the clinician’s intuition but to amplify it. They operate under an “expert-in-the-loop” design, where physicians specify research goals in natural language and steer the AI toward clinical constraints. By reducing the “cognitive labor” of literature synthesis and data analysis from hundreds of hours to less than two, these tools allow physicians to focus on clinical translation and patient selection.
As these “thinking engines” continue to evolve, they promise to bridge the gap between disparate scientific fields, identifying non-obvious connections that could lead to the next generation of life-saving therapeutics.
The Future of Retinal Discovery: AI-Driven Breakthroughs for Dry AMD and Beyond
The rapid expansion of scientific literature presents a “breadth and depth conundrum” for retina specialists: the complexity of retinal diseases requires deep subject matter expertise, yet breakthroughs often require bridging broad knowledge across multiple disciplines. Traditionally, drug development is a protracted, multi-year process limited by the rate at which human experts can synthesize information. However, the emergence of multi-agent AI systems is now accelerating this cycle, offering a new paradigm for identifying therapeutic targets and repurposing drugs for high-priority ophthalmic conditions.
Automating Discovery: The “Robin” System and Dry AMD
A significant milestone for ophthalmology is the development of Robin, a semi-autonomous AI system designed to automate both hypothesis generation and experimental data analysis. When tasked with addressing dry age-related macular degeneration (dAMD)—the leading cause of irreversible sight loss in developed nations—Robin bypassed hundreds of hours of manual research to propose a novel therapeutic strategy.
After analyzing over 500 papers in just 30 minutes, Robin hypothesized that enhancing retinal pigment epithelium (RPE) phagocytosis could mitigate dAMD pathology. The system identified and prioritized two specific candidates for this purpose:
- Ripasudil: A Rho kinase (ROCK) inhibitor currently approved for glaucoma treatment in Japan.
- KL001: A circadian clock modulator that had never previously been proposed as an enhancer of RPE phagocytosis.
In laboratory validations using primary human RPE stem cells (RPE-SC) from geriatric donors, both drugs demonstrated a dose-dependent increase in phagocytic activity, with Ripasudil showing significantly higher potency than the research compound Y-27632.
Mechanistic Insights: From Drug to Novel Target
Beyond simply identifying drug candidates, these AI systems can provide “lab-in-the-loop” mechanistic insights. In a follow-up experiment proposed and analyzed by Robin’s data analysis agent, RNA-seq analysis revealed that ROCK inhibition induced a 3-fold upregulation of ABCA1, a critical lipid efflux pump.
Because ABCA1 facilitates the transport of phospholipids and cholesterol—and its lipid acceptor, Apo-E, is a known genetic susceptibility allele for macular degeneration—this discovery highlights a potential novel molecular target for dAMD that had previously remained unexplored.
A New Tool for the Surgeon-Scientist: Co-Scientist and ERA
Complementing these ophthalmic-specific breakthroughs are general-purpose scientific engines like Co-Scientist and Empirical Research Assistance (ERA).
- Co-Scientist uses a “generate, debate, evolve” paradigm, employing specialized agents to simulate scientific peer review and debate to refine research proposals. It has already been used to identify potent synergistic drug combinations for complex malignancies, a strategy that could eventually be applied to treatment-refractory retinal diseases.
- ERA focuses on optimizing “scorable tasks,” such as scRNA-seq batch integration. For retina surgeons involved in translational research, ERA’s ability to remove batch effects while preserving biological signals across disparate single-cell datasets is essential for identifying rare cell populations or subtle disease markers in high-dimensional genomic data.
The Clinical Outlook
These AI systems are designed for a “scientist-in-the-loop” collaborative paradigm. They do not replace the clinician’s intuition but rather amplify it by exhaustively searching the solution space for “low-hanging fruit” and non-obvious connections between disparate fields.
While in vivo validation and placebo-controlled trials remain the gold standard for clinical transition, the ability of AI to independently recapitulate breakthroughs—such as identifying a clinical niche for an approved drug in a new indication—marks the beginning of an era of AI-empowered discovery in ophthalmology.
References
- **** Gottweis, J. et al. Accelerating scientific discovery with Co-Scientist. Nature (2026).
- **** Ghareeb, A. E. et al. A multi-agent system for automating scientific discovery. Nature (2026).
- **** Aygün, E. et al. An AI system to help scientists write expert-level empirical software. Nature (2026).
