Teaching AI to Understand Health How Foundation Models Are Shaping Biomedical Research
AI-powered foundation models like GPT have evolved from everyday tools for simple tasks to powerful systems capable of revolutionizing industries. Researchers at Helmholtz Munich are harnessing the potential of these models to drive advancements in healthcare, aiming to improve diagnostics, treatment planning, and biomedical research.
AI-powered foundation models like GPT have evolved from everyday tools for simple tasks to powerful systems capable of revolutionizing industries. Researchers at Helmholtz Munich are harnessing the potential of these models to drive advancements in healthcare, aiming to improve diagnostics, treatment planning, and biomedical research.
It started with simple interactions: "Hey ChatGPT, what’s a good recipe for dinner tonight?" or "Can you summarize this article for me?". AI-powered assistants had seamlessly woven themselves into daily life, helping with tasks ranging from drafting emails to tutoring students. They became indispensable digital companions, making life just a little easier with every query. But what began as a tool for convenience soon transformed into something far more powerful.
Foundation models, such as GPT – the AI system behind ChatGPT – are large-scale artificial intelligence systems trained on diverse and extensive datasets, enabling them to perform a wide range of tasks with minimal task-specific tuning. Their ability to generalize across different domains makes them highly adaptable for applications spanning from language processing to problem-solving in complex fields. These models power various applications, including content generation, coding assistance, and advanced data analysis, transforming industries by enhancing efficiency and decision-making.
As foundation models grow more advanced, experts continue to explore their applications beyond everyday tasks. Their ability to process vast data, recognize patterns, and generate insights is driving breakthroughs in research, finance, or creative fields. Among these, healthcare stands out as a particularly promising area, with opportunities to improve diagnostics, treatment planning, and medical research. Researchers at Helmholtz Munich recognize this potential and are investigating how AI-driven models can analyze complex medical data to transform healthcare.
Creating the Virtual Cell With AI
“One of my dreams in the next 10 years is to produce a virtual cell. What I mean by virtual cell is you model the whole function of the cell with an AI system,” stated Demis Hassabis, co-founder of DeepMind, in a 2022 interview in The Guardian. This bold vision has inspired researchers around the globe – including Prof. Fabian Theis and his team at Helmholtz Munich’s Computational Health Center, which is actively working to turn this into reality.
Theis, a pioneer in the field of AI-driven biomedical research, frequently speaks of the “virtual cell” as a long-term goal: a comprehensive digital model that simulates how individual cells behave in both health and disease. His team is developing AI-powered learning cell atlases to map cellular states across various forms of health and disease, advancing our understanding of cellular processes.
Koplexs-Stock - stock.adobe.comAt the heart of this research is single-cell sequencing, a technology that allows scientists to study individual cells rather than averaging data across entire tissues. While this method generates vast and complex datasets, foundation models have proven essential to unlocking their full potential. By leveraging AI, Theis’s team can process millions of individual cell profiles, identifying patterns, relationships, and transitions that would be impossible to detect manually.
"Before, this kind of analysis would take years. Now, with foundation models, we can do it in weeks – or even days."
Prof. Fabian Theis, Head of the Computational Health Center
The ultimate goal is to create a computational framework capable of predicting cellular behavior – a stepping stone toward the "virtual cell" Hassabis envisions. With AI-driven models, researchers can simulate how immune cells react to infections, predict how cancer cells resist treatment, or model the progression of neurodegenerative diseases at a cellular level. This ability is invaluable not only for fundamental research but also for the development of personalized medicine, where treatments can be tailored to a patient’s unique cellular structure.
One of the most transformative aspects of Theis’s work is its potential application in drug discovery and regenerative medicine. AI-powered cell atlases can help identify novel drug targets by revealing how different cell types interact and which genetic pathways drive disease. “Our models can analyze data from thousands of patients to identify biomarkers for cancer, diabetes, and rare genetic disorders, leading to earlier diagnoses and better-targeted therapies,” states Theis.
Teaching AI to Think Like a Human
Dr. Eric Schulz at the Institute of Human-Centered AI at Helmholtz Munich is on a mission to make AI psychologically aware. His team is integrating cognitive science into foundation models, allowing them to predict human behavior and improve medical decision-making.
Schulz works with Large Language Models (LLMs) – AI systems trained on vast amounts of text data that can understand and generate human-like language. These models, like GPT, are designed to understand context, generate responses, and complete tasks across many domains, making them highly adaptable for various applications, including medical research. However, the challenge lies in understanding how these models behave in complex scenarios. Schulz’s team was the first to apply paradigms from psychology and cognitive science to LLMs on such a large scale, exploring how they learn, make decisions, and differ from human cognition.
Filip - stock.adobe.comInitially focused on understanding their decision-making processes, Schulz’s lab now leads Europe in working with unique LLMs, improving them through human feedback. This feedback helps the models recognize their own limitations – a crucial aspect in medicine. For instance, if an LLM suggests a diagnosis, it should also provide the likelihood of its accuracy. Schulz also discovered that open-source models tend to take fewer risks compared to proprietary models, a significant insight for AI’s role in healthcare.
Schulz envisions applying these insights to psychiatry and medical sciences. By fine-tuning models based on individual patient data, AI could guide personalized therapies, such as reducing anxiety with tailored interventions. Just as the virtual cell aims to simulate cellular processes, Schulz seeks to create a "virtual human" model that captures the complexities of human behavior, health decisions, and emotions.
At its core, Schulz’s work bridges AI and psychology, developing models that not only diagnose diseases but also understand patient preferences. "Healthcare isn’t just about biology; it’s about people making choices," so Schulz.
Jörg Abendroth | Max-Planck-Institut für biologische Kybernetik"If AI can understand why patients behave the way they do, we can create better, more personalized treatments."
Dr. Eric Schulz, Director of the Institute of Human-Centered AI at Helmholtz Munich
AI-Powered Cancer Diagnostics
Sarawut - stock.adobe.comProf. Julia Schnabel from the Institute of Machine Learning in Biomedical Imaging at Helmholtz Munich is advancing medical imaging with AI-driven image reconstruction, analysis, and diagnosis. Her team is part of a nationwide initiative in Germany, where she is leading the development of foundation models to help answer one of oncology’s most pressing questions: Where does a cancer originate?
Identifying the primary cancer site when metastases appear in the body is a significant challenge. To address this, Schnabel’s team is building a Cancer Foundation Model – an AI system that integrates diverse medical data sources, including pathology, radiology, text reports, and electronic health records. Combining Vision Transformers (AI models specialized in image analysis) with LLMs, they are creating a system capable of seamlessly analyzing both imaging and textual data to trace the source of metastatic tumors. “By leveraging AI, we aim to develop a system that can pinpoint the origin of metastatic tumors with unprecedented accuracy,” says Schnabel.
Helmholtz Munich | Matthias Tunger Photodesign"This could transform cancer diagnostics, leading to earlier detection, more accurate treatment strategies, and improved patient outcomes."
Prof. Julia Schnabel, Director of Institute of Machine Learning in Biomedical Imaging at Helmholtz Munich
This work is part of the DECIPHER-M project, a national AI-driven initiative launched in March 2025 to understand cancer metastasis. Funded by the German Federal Ministry of Education and Research (BMBF) under the "National Decade against Cancer," it brings together experts in medicine, computer science, and biotechnology.
Ethical, Trustworthy and Sustainable AI
Best - stock.adobe.comYet, while groundbreaking research in foundation models is transforming healthcare, key challenges must be addressed. Ensuring AI technologies are developed and applied responsibly remains a critical priority. Issues such as data privacy, algorithmic bias, and integration into clinical workflows require not only technical innovation but also ethical oversight and sustainable practices.
“We want to build AI tools that are not only effective but also transparent, fair, and robust enough to be trusted in real-world healthcare settings,” says Julia Schnabel. “That’s why we actively contribute to international initiatives like FUTURE-AI, which provide clear principles and guidance for the development of trustworthy medical AI.”
The FUTURE-AI framework, developed by over 100 experts worldwide, outlines best practices for fairness, transparency, usability, and explainability throughout the AI lifecycle - from design to deployment. At Helmholtz Munich, these principles are also being applied to improve the environmental footprint of AI by refining and repurposing large models to reduce energy use without sacrificing performance.
Related Articles
Latest update: May 2025.