Robin: A multi-agent system for automating scientific discovery

TL;DR

The paper presents Robin, a first-in-class multi-agent system that automates hypothesis generation, experimental planning, and data analysis to accelerate therapeutic discovery. By integrating literature-search agents with an autonomous data-analytic agent in a lab-in-the-loop, Robin identified a novel dry AMD candidate (ripasudil) and elucidated a potential mechanism via ROCK inhibition–driven enhancement of RPE phagocytosis and ABCA1 upregulation. The study demonstrates a seamless end-to-end workflow—from literature synthesis and mechanistic hypothesis to automated experimental analysis and iterative hypothesis refinement—that can substantially speed drug repurposing and discovery efforts. This framework establishes a new paradigm for AI-guided scientific discovery with potential to generalize across diseases beyond ophthalmology.

Abstract

Scientific discovery is driven by the iterative process of background research, hypothesis generation, experimentation, and data analysis. Despite recent advancements in applying artificial intelligence to scientific discovery, no system has yet automated all of these stages in a single workflow. Here, we introduce Robin, the first multi-agent system capable of fully automating the key intellectual steps of the scientific process. By integrating literature search agents with data analysis agents, Robin can generate hypotheses, propose experiments, interpret experimental results, and generate updated hypotheses, achieving a semi-autonomous approach to scientific discovery. By applying this system, we were able to identify a novel treatment for dry age-related macular degeneration (dAMD), the major cause of blindness in the developed world. Robin proposed enhancing retinal pigment epithelium phagocytosis as a therapeutic strategy, and identified and validated a promising therapeutic candidate, ripasudil. Ripasudil is a clinically-used rho kinase (ROCK) inhibitor that has never previously been proposed for treating dAMD. To elucidate the mechanism of ripasudil-induced upregulation of phagocytosis, Robin then proposed and analyzed a follow-up RNA-seq experiment, which revealed upregulation of ABCA1, a critical lipid efflux pump and possible novel target. All hypotheses, experimental plans, data analyses, and data figures in the main text of this report were produced by Robin. As the first AI system to autonomously discover and validate a novel therapeutic candidate within an iterative lab-in-the-loop framework, Robin establishes a new paradigm for AI-driven scientific discovery.

Figures (10)

• Figure 1: Architecture and workflow of the Robin system. A) Given the name of a target disease, Robin generates hypotheses and selects top therapeutic candidates to test experimentally. Robin can autonomously analyze raw data from these experiments to synthesize scientific insights and generate updated therapeutic hypotheses. B) Robin interacts with language agents to generate hypotheses and analyze experimental data. C) Crow and Falcon are used to conduct concise and deep literature searches, respectively, to gather information to guide hypothesis generation. Finch performs analyses of experimental data, which Robin uses to derive insights to inform the next round of hypothesis generation.
• Figure 2: Robin generates therapeutic candidate hypotheses for dry AMD and analyzes experimental data from in vitro tests A) Robin proposes several experimental assays and ultimately decides to use an RPE phagocytosis enhancement assay. Robin synthesizes this strategy into an overall goal and then generates several novel therapeutic candidates to enhance RPE phagocytosis. B) Schematic representation of the phagocytosis assay. RPE cells are incubated with the drug for 1 hour before pHrodo beads are added. The cells are incubated with the beads for 3 hours and phagocytic activity is measured via flow cytometry. C-F) Example plots from a Finch flow cytometry analysis trajectory, formatted for readability in publication by a human. C) Finch performs gating to discard debris using a FSC-A vs SSC-A plot. D) Finch gates singlet cells from the FSC-H vs FSC-A plot. E) Finch identifies the DAPI signal and excludes dead cells. F) Finch performs statistical tests to compare candidate drugs to the DMSO control and plots the results.
• Figure 3: RNA-sequencing analysis of ARPE-19 cells treated with ROCK inhibitor Y-27632. A) Robin interprets results from the first experiment and proposes follow-up assays. B-D) Example plots from a Finch RNA-seq analysis, formatted for readability in publication by a human. B) Finch-made volcano plot showing differentially expressed genes between Y27632-treated and wildtype cells after phagocytosis. C) Finch-made consensus findings from eight RNA-seq analysis trajectories, showing the percentage of analyses that identified the same genes as consistently up- or down-regulated. D) Finch-made GO-term enrichment of differentially expressed genes.
• Figure 4: Ripasudil significantly enhances RPE phagocytosis. A) Excerpt of Robin proposal for ripasudil. Drawing from the insights from the first round of experimental analysis, Robin proposes ripasudil as a therapeutic candidate for treating dry AMD. B) Analyzed flow cytometry data from the second round of experiments shows that ripasudil significantly enhances phagocytosis in RPE cells, inducing an even greater effect than Y-27632.
• Figure :