Researchers led by Kexin Huang report the development of a general-purpose biomedical artificial intelligence agent that can autonomously perform a range of research tasks. The authors tested their tool in various case studies, where it showed strong performance, and they say their results point “toward a future in which AI agents work alongside human researchers to accelerate biomedical discovery from basic research to translation.” Biomedical research drives discoveries in disease mechanisms, diagnostics, and therapeutics, but increasingly complex workflows, expanding datasets, specialized tools, and a rapidly growing literature make it difficult to fully leverage available knowledge and resources. As a result, valuable data and analytical opportunities often go underused because the demand for expert researchers exceeds the available workforce. This underscores the need for approaches that can scale expertise and accelerate discovery. The development of AI agents offers opportunities to reshape aspects of biomedical research, particularly if such tools could handle a broad range of biomedical tasks.

 

Here, Huang and colleagues introduce Biomni, an AI agent that mines biomedical literature to define comprehensive “action spaces.” The large language model also acts as the planning engine, interpreting a user request and preparing a multi-step series of sub-tasks to satisfy that request. In doing so, Biomni can compose and execute workflows without pre-set templates. The authors tested Biomni against a variety of biomedical benchmarks. They also evaluated how it performed compared to humans across several tasks, finding Biomni consistently matched expert accuracy while requiring substantially less time. They also tested the tool in five case studies ranging from analyzing wearable sensor data to designing laboratory protocols to assist wet-lab researchers. The authors note that while Biomni’s environment spans a wide range of biomedical tools and databases, “the evaluated tasks represent only a subset of the field, and key domains remain unexplored.” Even so, they say, by automating complex workflows, “Biomni enables researchers to redirect their efforts toward creative hypothesis generation, experimental innovation, and cross-disciplinary collaboration.”

 

***Related embargoed news briefing is planned for 11:00 a.m. U.S. ET on Tuesday, July 7, as a Zoom Webinar. Reporters interested in joining the briefing should please pre-register here https://aaas.zoom.us/webinar/register/WN_yEqbHT68TvqMItVMwEGOKg as soon as possible. See Other Information for Journalists section for full briefing information.***

A new statistical tool could help researchers and practitioners determine when personalized treatments, educational programs, or recommendations are truly worth the extra cost and complexity, according to a new study. The findings offer a rigorous tool to identify when tailoring interventions is likely to outperform more conventional one-size-fits-all approaches. Personalized interventions are increasingly used in fields such as medicine, education, marketing, economics, and computer science because individuals often respond differently to the same treatment or support. However, personalized interventions are more expensive, often require more data, and can be more difficult to implement than universal interventions. There are very few tools to evaluate and identify scenarios when the added cost and complexity of personalization are justified by meaningful improvements in outcomes. To address this, Zhaoqi Li and Emma Brunskill developed a statistical hypothesis test – the K-fold Personalization Test (KPT) – that analyzes existing datasets to evaluate whether a personalized intervention policy is expected to outperform the best universal intervention. Unlike previous methods, the test accommodates multiple interventions, many individual characteristics, and flexible machine learning models while maintaining rigorous statistical guarantees, including control of false-positive (Type I) error. Across four datasets spanning job training, depression treatment, education, and marketing systems, the authors found that the KPT approach consistently demonstrated broad applicability and outperformed previous statistical methods for evaluating personalization. Li and Brunskill note, however, that the test evaluates personalization only within a user-specified class of decision policies and does not identify the optimal personalized policy itself. Yet even with these limitations, the KPT provides a rigorous and broadly applicable framework for helping researchers and practitioners determine when personalization is likely to deliver meaningful benefits.

In experiments in which physicians made decisions about treating hypothetical patients, the physicians tended to trust incorrect advice presented as being generated by artificial intelligence (AI), even after given the opportunity to notice that patient recovery data contradicted the recommendations. Aranzazu Vinas of the University of the Basque Country, Spain, and colleagues present these findings in the open-access journal PLOS Digital Health. 

In creating a comprehensive, AI-enabled research agent for the biomedical sciences, Stanford researchers hope to speed innovation by eliminating the tedium of scientific legwork.

A research team in Freiburg has investigated how random sounds played during sleep affect the consolidation of new memories. The results show that non-targeted acoustic stimuli disrupt deep sleep. As a result, the slow brain waves typical of this sleep phase occur less frequently and spread more irregularly across the brain. It is precisely these changes that impair the consolidation of new memories.