Link prediction for biomedical network reconstruction

Networks offer an important framework for elucidating the intricate interplay among diverse entities in nature. Link prediction (LP) emerges as a powerful tool for network analysis, enabling researchers to infer potential connections and their interaction strengths in networks. This paper reviews LP techniques (including link weight prediction, LWP) with a focus on biomedical applications, formalizing problem definitions and summarizing state-of-the-art methodologies. We detail their applications across four representative biomedical networks: gene regulatory networks, protein–protein interaction networks, brain functional networks, and epidemic contact networks, demonstrating their capacity to enhance data quality and uncover mechanistic insights. Critical challenges in biomedical network analysis are discussed, particularly fusing multiomics data, processing complex biomedical networks, and resolving cross-network inconsistencies. To address these challenges, we highlight promising research directions, including: multiomics-integrated network analysis, algorithms tailored to complex network types, cross-network prediction methods, and large language model (LLM)-driven biomedical network techniques. • A review of link prediction for reconstructing biomedical networks is presented. • The evolution of link prediction from heuristics to physics-informed GNNs is detailed. • Applications span gene, protein, brain, and epidemic networks for mechanistic insights. • Key challenges include multiomics fusion and modeling complex network structures. • Future directions feature cross-network prediction and LLM-GNN hybrid architectures.