An agentic AI environment to support biomedical research in collaborative academic environments
The paper introduces Murmurent, an open-source, shared software environment designed to support collaborative biomedical research by providing specialized agents, tiered memory, traceability, and enforced standard operating procedures that lower technical barriers for academic labs using agentic AI, demonstrated through its application in identifying inhibitors for the Pin1 protein.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
In the modern laboratory, the tools of discovery have expanded far beyond microscopes and test tubes. Scientists now rely heavily on computers to sift through mountains of data, design new molecules, and predict how biological systems behave. This shift has brought powerful artificial intelligence into the fold, capable of reading scientific papers, writing code, and proposing experiments. However, a significant gap remains between the raw power of these intelligent systems and the daily reality of academic research. Most researchers are experts in biology or chemistry, not in the complex software engineering required to safely manage these digital assistants. Without careful oversight, an artificial intelligence might accidentally leak sensitive patient data, lose track of a crucial decision made weeks ago, or repeat a mistake that wasted days of computing time. The challenge is not just having a smart assistant, but having a system that understands the rules of the lab, protects private information, and helps a team work together without a single person needing to be a computer expert.
To bridge this gap, a team of researchers at Western University in Canada has developed a new software framework called Murmurent. They describe it not merely as a tool, but as an operating system for artificial intelligence, designed specifically for the messy, collaborative, and high-stakes environment of academic science. Instead of asking every lab to build its own safety protocols and data management systems from scratch, Murmurent provides a shared foundation. It sits beneath the artificial intelligence models, acting as a layer of governance that manages identities, secures data, and remembers the history of a project. The system organizes work into "projects" where multiple researchers can collaborate, each bringing their own methods to a shared question. It employs a suite of specialized digital assistants, each with a specific role: one handles data analysis, another searches scientific literature, a third creates visualizations, and others act as auditors to check for errors or ethical concerns. Crucially, the system enforces strict rules automatically. It prevents the artificial intelligence from deleting important files, from sending private data to the public internet, or from making decisions that violate the lab's safety standards, regardless of how the human researcher phrases their request.
The researchers tested this system by tackling a difficult problem in drug discovery: finding a molecule that can inhibit a protein called Pin1. This protein is involved in cancer and neurodegenerative diseases, but it is notoriously hard to target because its active site is a shallow groove rather than a deep pocket, making it difficult for drugs to stick to it. The team did not rely on a single method. Instead, they used Murmurent to coordinate four different approaches simultaneously. One team member used a deep learning model to generate new molecules from scratch. Another started with a known weak binder and modified its structure. A third focused on a specific chemical strategy to attach a molecule permanently to the protein, while the fourth combined different chemical parts to search for a better fit. In a traditional setting, these four groups might have worked in isolation, using different standards and losing track of each other's progress. With Murmurent, they worked in parallel within the same secure environment. The system ensured they all used the same reference data, the same safety checks, and the same way of recording their decisions.
As the four approaches ran, the system's built-in auditors played a vital role. One of the digital assistants, acting as a critic, noticed that the initial way the team was measuring success was flawed. The method they were using to rank the new molecules was actually just recognizing the shape of the starting materials rather than predicting how well they would bind to the protein. Because the system had recorded every step and decision, the team could trace this error back to its source. They adjusted their methods, re-ran the calculations, and found that the initial promising results were actually no better than random chance. This correction saved the team from wasting resources on molecules that were unlikely to work. The system also managed the complex logistics of the project, such as distributing the heavy computing tasks across multiple graphics processors to finish in hours rather than days, and ensuring that sensitive candidate molecules never left the secure local network.
The outcome of this experiment was not a single new drug, but a demonstration of how such a system can guide scientific inquiry. None of the four approaches produced a molecule that the system could confidently certify as a winner, largely because the protein is so difficult to target and there were not enough known examples to compare against. However, the process revealed the strengths and weaknesses of each approach, allowing the researchers to refine their strategies. The system successfully prevented the team from chasing false leads and provided a clear, auditable record of why certain paths were abandoned. The researchers found that the system allowed a wet-lab trainee with limited coding experience to participate in complex data analysis, while ensuring that a principal investigator could oversee the entire process without needing to micromanage every line of code.
Murmurent represents a shift in how academic science might be conducted in the age of artificial intelligence. It moves away from the idea of a single, central controller managing every experiment and toward a model where groups of researchers coordinate their own work within a shared, safe framework. The system does not replace the scientist; instead, it handles the administrative and safety burdens that often slow down research. By managing memory, enforcing security, and coordinating multiple agents, it allows researchers to focus on the science itself. The team has made the software open source, allowing other laboratories to adopt and adapt it. While the system does not guarantee that every experiment will succeed, it provides a robust structure that helps ensure that when experiments fail, the reasons are clear, the data is safe, and the lessons learned are preserved for the future. This approach offers a practical path forward for integrating powerful artificial intelligence into the collaborative, diverse, and often unpredictable world of academic biomedical research.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.