From months to minutes: AI agents build 3D tumor digital twins via natural language
Researchers at the Barcelona Supercomputing Center (BSC-CNS) developed a framework empowering AI agents to build 3D tumor digital twins via natural language, reducing setup times from months to under 10 minutes without requiring coding expertise.
image: Visualization of a 3D tumor model
Credit: BSC Data Analytics and Visualization Group
BARCELONA, Spain — September 23, 2026 — Until now, building a digital model of a tumor required scientists to review extensive literature, undergo training in specialized software, and master multiple programming languages. From start to finish, the process often took months, if not years, of specialized work.
The Barcelona Supercomputing Center – Centro Nacional de Supercomputación (BSC-CNS) took a pioneering step in applying artificial intelligence to biology by developing a technology that connects AI agents with specialized tools, enabling researchers to create digital tumor models in minutes without coding expertise.
Published in the Nature group journal npj Systems Biology and Applications, the study describes the development of specialized tools (MCP servers) designed to empower existing AI agents to act as intermediaries between researchers and advanced modeling tools such as NeKo, MaBoSS, and PhysiCell.
Instead of sifting through extensive documentation, writing code, and juggling multiple programs, researchers can now build models through a simple conversation with an AI agent equipped with these new capabilities, describing the biological problem they wish to explore in natural language.
“With this advance, we completely broke down the barriers separating experimental biologists from advanced modeling tools. Now, any researcher can create complex biological models simply by conversing with an AI agent that we have equipped with these new features. It is very similar to using ChatGPT; you do not need to write a single line of code,” explained Marco Ruscone, BSC researcher and first author of the paper.
According to the researchers, using these systems alongside domain expertise and software familiarity drastically reduces setup times. Thanks to the intermediaries developed in the study, any researcher with sufficient background knowledge can generate an initial draft of a tumor digital twin in under 10 minutes.
Open-Access AI at the Service of Science
The study also explores how scientists can interact with AI agents to tackle complex scientific problems.
“AI agents already automate routine tasks, but their true transformative potential lies in solving complex scientific challenges, such as creating digital twins of cellular systems. Automating these processes, which involve numerous critical decisions, is a significant hurdle, and our study is an important first step,” reflects Alfonso Valencia, lead researcher of the study and director of the Life Sciences Department at BSC.
The paper also addresses key challenges like reproducibility, a cornerstone of scientific publishing that can be difficult to guarantee with probabilistic language models. However, the authors demonstrate that targeted, iterative interaction with the AI yields consistent and reliable results that meet scientific standards.
“Different language models can give you different answers, just as different scientists might. The key is working with them iteratively by asking questions time and again. Used correctly, minor variations in their responses do not matter, because they ultimately converge on the same core facts,” explained Miguel Vázquez, BSC researcher and third author of the paper.
The MCP servers developed at BSC are openly available to the scientific community, democratizing access to advanced biological modeling tools. Ultimately, this study lays the groundwork for using intelligent agents that will accelerate treatment development and deepen our understanding of cancer and other complex diseases.