RESEARCH ARTICLE SUMMARY
RESEARCH ARTICLE SUMMARY
ARTIFICIAL INTELLIGENCE Autonomous biomedical research with an artificial intelligence agent
INTRODUCTION: Modern biology and medicine research generate data far faster than can be analyzed. A single study can require the use of dozens of specialized software programs, combing through many years of literature results, designing detailed experimental protocols, and evaluating statistics. As a result, valuable datasets sit unexamined, and connections across existing knowledge go unmade.
RATIONALE:
RATIONALE:
system meant to handle many kinds of research. An AI agent-software that plans and acts on its own-needs to be able to access specialized tools and datasets. The first step to designing Biomni was assembling these digital resources into one shared workspace of one hundred fifty analysis tools and dozens of software packages and databases spanning twenty-five areas of biology. Given this base of established methods, we designed the Biomni system to process and respond to questions in plain English using large language models. The agentic framework enables complex workflow planning, writing, executing code to analyze data, error checks, and adjustments.
RESULTS: Across four hundred forty-three questions spanning ten kinds of biomedical tasks, Biomni achieves an average accuracy of fifty-seven percent, considerably higher than that achieved by other agentic systems on the same benchmarks. On three expert-level tasks, it matched specialists in accuracy while only requiring a fraction of the time. We also assessed Biomni in real-world experiments across very different disciplines. We used Biomni to develop an analysis pipeline to process previously published
Biomni is an AI agent that enables biomedical research. A scientist asks a question in plain English; Biomni picks the right tools, plans the steps, and writes and runs its own code to deliver an answer. [Illustration: Jasmine Zhang]
smartwatch readings and recovered known early warning signs of COVID-19 infection. Biomni designed a gene-editing cloning protocol that bench scientists ran exactly as written, with sequencing confirming success. We redesigned a protein for greater heat stability, with Biomni creating a computational optimization procedure that proposed three rational mutations consistent with state-of-the-art protein thermostability engineering principles. We also prompted Biomni to write ready-to-run code to drive a laboratory robot through a multistep drug dose-response experiment. Each of these experiments drew on computational and wet-lab methods curated during optimization of Biomni but without requiring extensive specialized knowledge and coding ability of the human deciding the research agenda.
CONCLUSION: As a general-purpose agent, Biomni can design workflows for experiments in fields ranging from genetics, immunology, pharmacology, to clinical medicine, without being retooled or retrained on additional, domain-specific protocols. We propose that such agentic AI systems will be able to aid in accelerating cumbersome analysis tasks, thus enabling scientists to focus on framing questions, judging results, and pursuing the creative leaps that machines do not make. Biomni is openly available, and its developers emphasize responsible use as these tools become more powerful.