Source-linked AI summary
Robin: A multi-agent system for automating scientific discovery
Ali Essam Ghareeb, Benjamin Chang, Ludovico Mitchener, Angela Yiu, Caralyn J. Szostkiewicz, Jon M. Laurent, Muhammed T. Razzak, Andrew D. White, Michaela M. Hinks, Samuel G. Rodriques
TL;DR
Scientific discovery requires linking literature synthesis, hypothesis generation, experimentation, and analysis, but existing AI systems had not automated these stages in one workflow. Robin integrates specialized literature-search and data-analysis agents in an iterative lab-in-the-loop system, and identified ripasudil as a promising dAMD candidate while finding ABCA1 upregulation after ROCK inhibitor treatment.
Problem
Scientific discovery and therapeutic development require synthesizing rapidly growing, cross-disciplinary knowledge, while prior AI systems had not integrated hypothesis generation with experimental data analysis in one workflow.
Method
Robin coordinates literature-search agents and a scientific data-analysis agent to propose therapeutic candidates, analyze experiments, and generate refined hypotheses.
Results
Robin identified ripasudil as the most potent tested enhancer of RPE phagocytosis, increasing it 7.5-fold versus DMSO controls, while RNA-seq found 3-fold ABCA1 upregulation after ROCK inhibitor treatment.
Takeaways & Limitations
Robin demonstrates an iterative approach to AI-assisted therapeutic discovery, including a repurposing opportunity for ripasudil and a possible mechanistic target in ABCA1.
Takeaways & Limitations
Robin does not yet produce precise executable protocols, and its data analysis remains heavily reliant on domain-expert prompt engineering.
Abstract
from arXiv · showhide
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.
1 Introduction
Scientific knowledge is expanding faster than researchers can synthesize it into therapeutic hypotheses. Robin addresses this gap by combining literature-search and data-analysis agents in an iterative workflow that connects hypotheses, experiments, and refinement.
- Drug development is constrained by the rate at which experts synthesize biological, clinical, and pharmaceutical literature.
- Existing scientific insights can take years to crystallize into new treatments, as illustrated by dabrafenib’s delayed repurposing for hearing loss.Dabrafenib’s otoprotective effects were discovered roughly 10 years after its molecular action was characterized.
- LLMs may help synthesize knowledge across scientific fields, building on systems that retrieve, summarize, and generate hypotheses from literature.
- Robin integrates literature search with experimental data analysis to generate, test, interpret, and refine therapeutic hypotheses in one continuous workflow.
- The system is demonstrated on dAMD, a major cause of irreversible sight loss with limited treatment options.In the United States, 1.5 million people have vision-threatening dAMD and 600,000 are legally blind due to AMD.
2 Results
Robin coordinates literature search, hypothesis generation, experimental testing, and automated data analysis in an iterative workflow for therapeutic discovery. Applied to dAMD, it refined phagocytosis-based candidates, identified ripasudil, and used RNA-seq to reveal associated molecular changes.
- Robin: A multi-agent system for scientific discovery: Robin combines literature-search agents, therapeutic hypothesis generation, experimental data analysis, and follow-up hypothesis refinement in a structured workflow.Crow and Falcon search the literature, while Finch analyzes experimental data to inform subsequent hypothesis generation.
- Therapeutic hypothesis generation: Finch analyzed flow-cytometry data from candidate-drug experiments, including gating, statistical comparisons with DMSO, and quantification of effects on RPE phagocytosis.The assay used pHrodo beads, and Finch developed an analysis notebook to quantify compound effects.
- Therapeutic hypothesis generation: For dAMD, Robin reviewed 151 papers, selected enhanced RPE phagocytosis as a therapeutic strategy, and proposed existing drugs for testing.The system identified ten disease mechanisms before choosing an RPE phagocytosis assay and then proposed 30 drug candidates.
- Automated RNA-seq analysis: Robin recommended RNA-seq of Y-27632-treated RPE cells, and Finch’s analysis identified a 3-fold upregulation of ABCA1 with adjusted p=2.13x10^-83.The follow-up experiment profiled transcriptional effects of ROCK inhibition after phagocytosis.
- Automated RNA-seq analysis: ABCA1 was connected to dAMD because it is essential for healthy RPE function and belongs to the same transporter family as the macular-degeneration target ABCA4.The passage also links ABCA1 biology to Apo-E, a potential dAMD therapeutic target.
- Ripasudil as a drug repurposing candidate: Ripasudil outperformed Y-27632 and increased RPE phagocytosis 7.5-fold versus DMSO controls, although dose and incubation-time testing remained necessary for definitive comparison.This iterative result illustrates Robin’s refinement of therapeutic hypotheses through experimental feedback.
3 Discussion
Robin integrates hypothesis generation, experimental planning, and data analysis in a lab-in-the-loop workflow that identified ripasudil as a promising dAMD candidate. The system also exposes current limits in protocol generation, analytical autonomy, and hypothesis evaluation.
- Therapeutic discovery: Robin identified ripasudil as the most potent tested enhancer of RPE phagocytosis and proposed ROCK inhibition as a therapeutic strategy for dAMD.Ripasudil’s established safety profile and clinical approval for ocular use support its potential as a repurposing opportunity.
- Novelty: Robin was the first to propose ROCK inhibitors for dry AMD based on their effect on phagocytosis, extending prior retinal applications in wet AMD and neovascular disease.The hypothesis synthesized evidence that RPE phagocytic dysfunction is pronounced in AMD and that Y-27632 can enhance phagocytic efficiency.
- Broader significance: Robin addresses a broader need for scalable therapeutic discovery as FDA approvals have stagnated at approximately 50 novel drugs annually over the past decade.The system automates literature-grounded hypothesis generation and experimental data analysis in one workflow.
- Limitations: Robin currently generates experimental outlines rather than precise executable protocols, leaving human translation necessary for laboratory execution.The Finch agent also relies heavily on domain-expert prompt engineering, while LLM-based hypothesis evaluation remains imperfectly aligned with human scientific judgment.
- Broader significance: By integrating hypothesis generation, experimental planning, and data analysis, Robin represents a proposed paradigm for AI-driven scientific discovery.The reported scope includes therapeutic development and broader scientific workflows.
4.1 Robin Implementation
Robin is implemented as a structured multi-agent workflow that uses literature-search agents, an LLM judge, and data-analysis capabilities. Its hypothesis-ranking procedure combines pairwise comparisons with expert-informed judging and shows substantial agreement with expert preferences.
- Workflow: Robin uses Crow and Falcon for literature search, while an LLM judge ranks therapeutic hypotheses through pairwise comparisons.The implementation uses OpenAI o4-mini for literature synthesis and hypothesis generation and Claude 3.7 Sonnet as the pairwise-comparison judge.
- Hypothesis ranking: For 25 or fewer hypotheses, Robin compares every pair; for larger sets, it samples 300 pairwise comparisons before estimating strengths with the BTL model.The sampling strategy limits computational and time requirements for larger hypothesis sets.
- Expert alignment: The LLM judge’s prompt was generated from domain-expert pairwise evaluations to elicit decisions consistent with expert preferences and criteria.The expert evaluations were provided to Gemini 2.5 Pro Preview to generate the judge prompt.
- Evaluation: 7.25 of the judge’s top 10 hypotheses matched the experts’ top 10 on average, while identical comparisons produced the same choice in 88% of cases versus 61% for experts.The top-10 concordance was reported as more than double random-selection expectation, and the judge showed higher intra-rater consistency than human experts.
4.2 Finch Implementation
Finch is an autonomous, Jupyter-native data-analysis agent operating in Aviary’s controlled environment. It uses ReAct-style prompting and notebook tools to execute and submit bioinformatics analyses.
- Capabilities: Finch systematically processes bioinformatics workflows such as RNA-seq differential expression analysis and flow cytometry from a dataset and research question.Finch was initially introduced in BixBench and is designed for autonomous scientific data analysis.
- Execution environment: Aviary provides Finch with standardized software environments, consistent tool access, and structured multi-step reasoning for reproducible evaluation.The environment is intended to isolate analytical capability from software installation and dependency management.
- Execution environment: Finch uses ReAct-style agentic prompting within a pre-built Docker container containing bioinformatics-oriented Python, R, and Bash libraries.The container is identified as BixBench-env:v1.0.
- Tool interface: Finch interacts with Aviary through two tools: edit_cell for modifying and executing notebook cells, and submit_answer for finalizing its analytical conclusion.These tools define the agent’s notebook-based execution and submission interface.
- Development: Finch’s initial instructions were optimized through prompt engineering, and its analytical trajectories are available through the FutureHouse platform.An example prompt is provided in Supplementary Figure S10.
4.3 Cell Culture
ARPE-19 cells are cultured in supplemented DMEM/F12 at 37°C and 5% CO2, then seeded in 96-well plates for phagocytosis assays. The protocol specifies media composition and seeding density.
- Cell maintenance: ARPE-19 cells are maintained in DMEM/F12 supplemented with 10% fetal bovine serum, 2mM L-glutamine, 1mM sodium pyruvate, and 1% penicillin-streptomycin.Cultures are maintained at 37°C with 5% CO2.
- Assay setup: For phagocytosis assays, cells are seeded in 96-well tissue-culture plates at 1x10^4 cells per well in 100 µL of complete medium.Cells are grown to confluence before assay procedures.
4.4 Phagocytosis Assay
The phagocytosis assay tested drug effects on RPE-mediated particle uptake using fluorescent pHrodo beads and flow cytometry. Drugs were prepared at literature-derived concentrations, with AICAR and TUDCA tested in combination.
- Drug preparation: Drug working concentrations were selected from the literature, using the highest proposed concentration when multiple concentrations were available.Working stocks were prepared at 2X concentration in complete media on the experiment day.
- Drug library: AICAR and TUDCA were used in combination at the concentrations specified in the drug-working-concentration table.The supplied table is identified as Table 1, listing drugs and working concentrations.
- Assay readout: pHrodo beads enabled specific detection of phagocytosed particles because they fluoresce in the acidic lysosomal environment.Deep red E. coli pHrodo beads were resuspended in PBS and sonicated before use.
- Assay procedure: Cells received test compounds or vehicle control for 60 minutes before bead addition, followed by 3 hours of phagocytosis.The assay used 0.5% DMSO as vehicle control and incubated cells at 37°C with 5% CO2.
- Flow cytometry: Fluorescence from pHrodo beads and DAPI was measured by flow cytometry across the full aspirated well volume.The assay used 637 and 405 nm excitation lasers with 670/14 and 450/40 filters, respectively.
4.5 Data Analysis
Robin’s Finch agent analyzed flow-cytometry and RNA-seq data using automated gating, while human scientists performed complementary flow-cytometry analysis and read alignment. RNA-seq differential expression used a defined Y-27632-versus-untreated comparison.
- Flow-cytometry analysis: Finch performed end-to-end flow-cytometry analysis, excluding debris and aggregates before clustering the main cell population.It used forward- and side-scatter data with k-means clustering through the flowMeans R package.
- Human comparison: Human flow-cytometry analysis followed a similar strategy but additionally used the no-bead control to remove background Alexa 647 signal.The human workflow identified the main population, singlets, and dead cells.
- RNA-seq workflow: A human performed RNA-seq read demultiplexing and alignment, whereas Finch performed subsequent differential gene-expression analysis.
- Read processing: RNA-seq reads from twelve samples were aligned to the GRCh38 human reference genome using HISAT2 with splice-junction information retained.The reads were paired-end 2 × 150 bp and used a GENCODE v44-based HISAT2 index.
- Differential expression: Gene-level counts were generated with featureCounts and analyzed in DESeq2 after filtering, normalization, dispersion estimation, and Wald testing.The comparison was defined as Y-27632 versus untreated, using six samples across two conditions.
5 Supplementary Material
The supplementary material documents Robin’s disease-assay and therapeutic-candidate workflow, including literature queries, assay selection, candidate evaluation, data-analysis instructions, and expert-comparison results. It also includes disease-specific candidate-generation figures and workflow prompts.
- Assay concepts: The candidate-generation material includes complement, autophagy, inflammatory signaling, mitochondrial, barrier-integrity, and drusen-like-deposit assay concepts.The listed assay concepts span multiple cellular mechanisms relevant to dAMD.
- dAMD assay design: The dAMD workflow selected a phagocytic-clearance assay and aimed to identify compounds that enhance RPE-mediated clearance of photoreceptor material.The supplementary material lists assay generation, evaluation, and candidate-generation stages for this strategy.
- Candidate evaluation: Candidate evaluation considered safety, tolerability, delivery to target tissue, pharmacokinetics, and relevance to the disease mechanism.The supplementary prompts prioritize established clinical safety and a plausible route to therapeutic exposure.
- Evaluation: Robin’s LLM judge selected an average of 7.25 of human experts’ top 10 hypotheses and showed greater pairwise consistency than human experts.