Robin Integrates 550 Papers in 30 min, Closing the AI‑Driven Research Loop and Discovering dAMD Therapies

Recent advances in high‑throughput biology have outpaced the ability of researchers to synthesize knowledge, especially in fields like drug repurposing where fragmented findings delay clinical translation. To address this, a joint team from FutureHouse (San Francisco), Oxford University, and Fordham University introduced Robin, a multi‑agent biomedical AI platform that uniquely closes the loop of scientific hypothesis generation, experimental analysis, result feedback, and hypothesis iteration.

Data Foundations

Robin’s knowledge base comprises three layers: (1) 551 curated English and Chinese papers on dry‑age‑related macular degeneration (dAMD), including 151 on disease mechanisms and 400 on retinal pigment epithelium (RPE) phagocytosis; (2) the public BixBench benchmark covering transcriptomics, genomics, functional enrichment, sequence analysis, and statistical testing; (3) a proprietary experimental dataset containing flow‑cytometry, RNA‑seq, cytotoxicity, immunohistochemistry, and VEGF ELISA results from ARPE‑19 cells and primary human RPE cells sourced from the New York Vision Restoration Eye Bank (donors > 60 years, no ocular disease).

System Architecture

Built on the Aviary framework and executed within Jupyter Notebook, Robin employs three cooperating agents:

Crow and Falcon – two literature agents powered by OpenAI o4‑mini that retrieve disease‑relevant papers, distill mechanisms, select experimental models, and propose candidate drugs; Falcon additionally validates and refines Crow’s suggestions while correcting hallucinated citations.

Finch – a data‑analysis agent that generates and runs Python or R code on demand, handling flow‑cytometry, differential expression, and gene‑set enrichment without relying on fixed scripts.

To mitigate stochasticity, Finch runs eight parallel analysis trajectories and aggregates the outcomes via meta‑analysis (“multi‑trajectory + consensus” mechanism). Evaluation is performed by a two‑layer model review: Anthropic’s Claude 3.7 Sonnet as the primary reviewer, complemented by Google Gemini 2.5 Pro to align with expert preferences. Pairwise comparisons and tournament ranking, optionally sampled for large candidate sets, produce weighted scores using the Bradley‑Terry‑Luce model.

Experimental Validation on dAMD

Robin first identified ten key pathogenic mechanisms of dAMD and singled out “enhancing RPE phagocytosis” as the primary therapeutic target. From an initial pool of 30 candidates, the team experimentally tested drugs such as Ezetimibe, Fingolimod, and Y‑27632, using MFGE8 as a positive control.

Finch‑driven RNA‑seq analysis revealed that Y‑27632 reprograms the RPE transcriptome by modulating actin dynamics, autophagy pathways, and the lipid‑transport gene ABCA1 , uncovering a previously unrecognized mechanism of action.

In a second iteration, Robin added ten more candidates and discovered that the glaucoma drug Lipasudil increased phagocytic activity by ~1.89‑fold, outperforming Y‑27632. Dose‑response experiments on primary human RPE cells confirmed Lipasudil’s efficacy without observable cytotoxicity, highlighting its translational promise.

Additional screening identified the circadian regulator KL001 as another enhancer of phagocytosis, expanding the therapeutic landscape for dAMD.

Benchmark Comparisons

Against a generic AI research agent (OpenAI Deep Research Agent) that generated 17 candidates but failed to detect any phagocytosis‑enhancing activity or the ROCK‑inhibition mechanism, Robin demonstrated superior domain adaptation.

On the BixBench benchmark, Finch achieved an overall accuracy of 22.8 ± 1.7 % versus 1.6 ± 1.2 % for a plain large language model. Sub‑task accuracies were 47.9 ± 1.5 % for biostatistics, 100 % for flow‑cytometry, and 86 % for RNA‑seq analysis, confirming the advantage of a purpose‑built scientific agent while noting remaining challenges for complex multi‑step bio‑informatics tasks.

Efficiency and Cost

Cost analysis showed a per‑workflow expense of roughly $10.76 USD. Robin processed the 551‑paper corpus in 30 minutes, a task that would require > 800 hours of manual effort. The complete end‑to‑end workflow (literature integration, hypothesis generation, experimental design, data analysis, and iteration) finished in under 2 hours, representing an efficiency gain of about 200‑fold over traditional manual pipelines.

Conclusion

Robin illustrates a shift from AI as a supportive tool to a semi‑autonomous research system capable of generating hypotheses, designing experiments, and iterating on results. While expert oversight remains essential for experimental design, interpretability, and cross‑scale biological reasoning, the platform proves that AI can now participate directly in scientific discovery, dramatically accelerating drug‑repurposing efforts such as those for dAMD.