LNGCN: a distance-aware continuous-time graph framework for PPI prioritization in experimental screening

Abstract

High-throughput PPI screening requires not only interaction discrimination but also stable, calibrated scores for candidate prioritization and efficient experimental allocation. Conventional graph neural networks may over-smooth deep residue representations, reducing score separation among high-confidence candidates. We therefore developed LNGCN, a distance-aware continuous-time graph framework for experimental PPI prioritization. LNGCN integrates residue-level structural graphs with liquid neural dynamics, uses radial distance to drive continuous graph evolution, and applies hierarchical calibration to convert raw outputs into prioritization scores. It performed robustly on balanced human, 1:10 imbalanced, and cross-species yeast bench-marks. On the imbalanced dataset, the top 5% of candidates achieved 85.1% precision. LNGCN also preserved residue-level representation diversity better than conventional GCN variants. In FGF23-FGFR1c--Klotho, SHP2 signaling, Tdk1 oligomer-dependent binding, and a TPR study, it enriched known partners near the ranking top, and experimentally supported novel interactions. Overall, LNGCN enables sustained enrichment of experimentally relevant PPI candidates.