Protein foundation models: a comprehensive survey

Protein foundation models (pFMs) have emerged as pivotal tools in advancing protein science. By leveraging advanced deep learning architectures trained on large-scale protein datasets, pFMs learn generalizable patterns in proteins, enabling accurate prediction of key protein characteristics and generation of novel proteins with tailored properties. In this review, we provide a comprehensive exploration of the developments, applications, challenges, and prospects of pFMs. We systematically examine the multimodal dataset resources that underpin the development of pFMs, ranging from protein sequences to experimentally resolved and predicted three-dimensional (3D) structures, functional annotations, and interaction networks. We explore the advances in pFMs—spanning autoencoding, autoregressive, diffusion, and flow matching models—and highlight their representative applications across fundamental biological research, protein discovery and engineering, and biomedical applications, thereby illustrating their versatility and impact. We also discuss major challenges, encompassing data bottlenecks, evaluation complexities, and model interpretability. Looking forward, we outline promising research directions, including modelling protein dynamism and interactions, as well as developing integrated virtual cell systems, paving the way for next-generation bioengineering and therapeutic development. This survey offers both a roadmap for computational biologists and a strategic framework for experimentalists who are applying pFMs in their work.