Virtual cells face hurdles in predicting drug responses
AI-based virtual cells are emerging as a tool for predicting how entire cells respond to gene deletion, genetic modification or drug treatment. Speaking at the 2026 Seoul S&T Forum, Seoul National University professor Baek Min-kyung highlighted their potential while warning that predictive performance remains below expectations.
In the first “Virtual Cell Challenge” held in 2025, simpler computational algorithms produced lower prediction errors than more complex AI models. Overfitting can make models perform well on training data but struggle with unfamiliar biological conditions. Researchers are developing common evaluation metrics and large-scale experimental datasets, with the U.S. government, Google Isomorphic Labs and NVIDIA among those working to secure data.
AI agents are also expanding research automation by connecting language models with specialized life-science tools. Baek cited Anthropic’s use of more than 1,000 AI agents alongside protein design tools to design a protein that binds to scorpion neurotoxin, illustrating progress toward autonomous laboratories.
Despite advances in protein structure prediction and design, drug development still requires addressing safety, solubility and manufacturability. Baek emphasized clearly defined problems, reliable evaluation metrics, sufficient training data, computing resources and experimental feedback as essential foundations. Virtual cells lack the mature data and benchmarking infrastructure that helped protein prediction advance.