UVA Study: AI Struggles to Predict Cellular Responses in Biomedical Research

Researchers at the University of Virginia (UVA) School of Medicine have identified a critical performance gap in the ability of artificial intelligence to conduct predictive biomedical research. The study reveals that while general-purpose AI models can identify individual cellular components, they struggle to synthesize that data to predict complex biological outcomes.

Testing popular large language models (LLMs), including ChatGPT, Gemini, and Claude, the research team evaluated the tools' ability to generate logic-based models of biochemical signaling networks. According to the findings, the AI-generated networks reconstructed between 24% and 65% of the structure of literature-curated signaling networks. However, the models' utility plummeted when tasked with predictive analysis; when challenged to predict perturbation responses—how cells respond to changes or new drugs—the AI's accuracy validated at only 5% to 26%.

The Synthesis Gap

The research, conducted by the UVA School of Medicine's Department of Biomedical Engineering, comes as medical institutions increasingly integrate AI into clinical education and research to assist with diagnosis and treatment planning. The disparity in the results suggests that LLMs are currently better suited for data retrieval than for scientific reasoning.

Jeff Saucerman, PhD, of UVA’s Department of Biomedical Engineering, noted that the models possess a foundational knowledge of cellular parts but lack the ability to connect them. "It’s pretty good already at knowing the individual pieces of cells, but it’s not very good at piecing them together," Saucerman said, adding that this synthesis is essential to predict how new drugs would function.

Risks to Research and Patients

This lack of predictive reliability poses significant risks for the biomedical industry. In a laboratory setting, relying on incorrect AI predictions can lead to millions of dollars in wasted funding and research hours. More critically, inaccurate models can provide false hope to patients awaiting the development of new treatments.

The study emphasizes that the current state of AI requires strict human oversight. Because the models can produce confident but incorrect answers, Saucerman warned that "just because a model makes a prediction doesn’t mean we should trust it."

The Path Forward

As AI continues to evolve, the UVA findings suggest that human scientists must remain the final authority in predictive analysis. Future research will likely focus on whether specialized biomedical AI can overcome the synthesis gap that currently hinders general-purpose models. For now, the study serves as a cautionary benchmark for the integration of LLMs into high-stakes medical discovery.