While network science provides a powerful framework for deciphering complex biological systems, static models often fail to capture the dynamic nature of cellular processes. Temporal network analysis addresses this by modelling biological relationships as time-indexed graphs, offering a more realistic representation of evolving biological interactions. But, biological data are often sparse, noisy, and heterogeneous, making temporal network reconstruction highly sensitive to modelling and preprocessing choices. This review synthesises temporal network analysis for systems biology with an emphasis on practical interpretability and trustworthy inference. We highlight how different notions of “time” (e.g., longitudinal measurements, condition/stage progression, or inferred trajectories) and different meanings of “edges” (e.g., physical interactions, statistical associations, or model-based influences) support different biological claims and therefore demand different validation strategies. Using a multi-scale perspective, we survey approaches for characterising local dynamics, tracking mesoscale reorganisation such as module and community changes, and quantifying global shifts in network topology, alongside common tasks including rewiring detection, network comparison, and community evolution. A central message is that inference is often the bottleneck, while prediction is the temptation. We therefore foreground validation and benchmarking practices needed to distinguish genuine biological dynamics from artefacts of sampling, windowing, or model class. Finally, we discuss temporal graph learning, including temporal graph neural networks. We highlight when these methods can enable meaningful forecasting in biology, and when performance is inflated by sensitivity to network construction choices or information leakage in evaluation.
Khazaal et al. (Mon,) studied this question.