Foundation models—large AI systems pretrained on broad, heterogeneous data—are transforming scientific discovery. These models (e.g., GPT, GenCast, AlphaFold) excel at learning generalizable representations and adapting to new tasks with limited data. Yet, epidemic modeling has not experienced a comparable transformation. Traditional models remain pathogen-specific and often struggle to generate rapid insights during emerging outbreaks, as starkly illustrated by the SARS-CoV-2 pandemic. This Perspective asks whether the foundation model paradigm can extend to epidemic science: Can we build a single, pretrained model that captures the shared principles of infectious disease dynamics across pathogens, populations, and settings? Such a model could be fine-tuned to new contexts with minimal data, enabling faster forecasting, inference, and response, especially valuable in resource-limited settings. We argue that the growing convergence of epidemiological insight and modern AI makes this goal both urgent and increasingly plausible. We outline the main challenges in building foundation models for epidemics—nonstationarity, fragmented surveillance data, presence of diverse dynamical regimes, and the need for interpretability. We then propose a roadmap toward epidemic foundation models, emphasizing both algorithmic innovations to address these challenges and progress beyond algorithms, including investments in open datasets and cross-disciplinary training and collaboration. Developing epidemic foundation models offers a potentially transformative opportunity to strengthen global health security, particularly by improving preparedness in underresourced settings. If successful, they will serve as powerful, generalizable tools that complement existing efforts. The process of building these models will itself be valuable, exposing critical data gaps and guiding investments in global surveillance.
Lau et al. (2026) studied this question.