UniBioNet: A Unified Inductive Frequency-Domain Framework for Biological Network Link Prediction

Biological link prediction is routinely evaluated on fixed graphs, although drug-discovery applications require reliable predictions for previously unseen entities. We present UniBioNet, a unified inductive frequency-domain framework that makes independently computed local spectra comparable across diverse biological networks. UniBioNet summarizes each spectral subspace through coefficient-Gram spectra, models interactions among frequency components, and reconstructs node states with eigenspace projectors. This design removes arbitrary eigenvector signs and within-eigenspace basis rotations while supporting homogeneous and heterogeneous networks without task-specific meta-paths. Among the evaluated methods, UniBioNet achieves the highest mean AUC on eight of nine datasets under both established benchmark protocols. A separate node-disjoint audit reveals a substantial benchmark-to-cold-start gap, clarifying that coordinate invariance does not eliminate information loss for unseen entities. Ablation and efficiency analyses provide descriptive evidence for the frequency pathway and lower local eigendecomposition cost while identifying protocol-dependent limitations.