Abstract
BACKGROUND: Protein-protein interaction (PPI) network-based protein function prediction algorithms, which use the guilt-by-association principle, are widely applied in medical and human domains, but they are comparatively underused in predicting plant-related phenotypes. Although many algorithms are available for function prediction using PPI networks, an unbiased evaluation of these has not yet been conducted on plant networks to date. Here we evaluate seven network-based protein function prediction algorithms, including majority voting (MV), hishigaki method (HM), random walk (RW), random walk with restart (RWR), markov random fields (MRF), personalized page rank (PPR), and deep Network Fusion (deepNF) using PPI networks from six plant species: Arabidopsis thaliana, Oryza sativa japonica, Zea mays, Solanum tuberosum, Glycine max, and Medicago truncatula on four evaluation metrics. RESULTS: Ten-fold cross-validation was performed to evaluate the algorithms using the following performance metrics: Area Under the Receiver Operating Characteristic (AUROC) curve, Area Under the Precision-Recall Curve (AUPRC), Average Precision (AP), and F(max). The median metric values were used for comparison of the performance of algorithms for each species, which revealed RWR as the most consistent performer across all species, with deepNF becoming a close second. Metrics sensitive to positive ranking and threshold optimization (AP and F(max)) tended to yield comparatively higher values in sparsely annotated species such as Z. mays, likely reflecting the influence of small positive sets rather than inherently superior predictive performance. A comparison across Gene Ontology categories revealed that, except for AUROC scores, the algorithms performed generally better in the cellular component category over the biological process and molecular function categories. CONCLUSION: Although optimal performance depends on the network structure, the annotation density, the evaluation metric, and the plant species, network propagation methods (RWR and PPR) and deepNF showed a strong performance overall. Furthermore, the network propagation methods were seemingly robust to network perturbation. Therefore, those algorithms can be generally recommended for plant network-based protein function studies. However, it is encouraged that plant networks be evaluated on multiple algorithms and metrics, where feasible, to identify the most suitable approach for a specific network. The code and datasets used in this evaluation are publicly available at 10.5281/zenodo.18628981. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12870-026-08608-5.