Perspectives on benchmarking foundation models for network biology

Transfer learning has revolutionized fields including natural language understanding and computer vision by leveraging large‐scale general datasets to pretrain models with foundational knowledge that can then be transferred to improve predictions in a vast range of downstream tasks. More recently, there has been a growth in the adoption of transfer learning approaches in biological fields, where models have been pretrained on massive amounts of biological data and employed to make predictions in a broad range of biological applications. However, unlike in natural language where humans are best suited to evaluate models given a clear understanding of the ground truth, biology presents the unique challenge of being in a setting where there are a plethora of unknowns while at the same time needing to abide by real‐world physical constraints. This perspective provides a discussion of some key points we should consider as a field in designing benchmarks for foundation models in network biology.