Reproduce “PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xinzhe Zheng, Hao Du, Fanding Xu, Jinzhe Li, Zhiyuan Liu, Wenkang Wang, Tao Chen, Wanli Ouyang and 4 more

PRING introduces a comprehensive benchmark for evaluating protein-protein interaction (PPI) prediction models from a graph-level perspective, moving beyond traditional pairwise metrics. It assesses models' ability to reconstruct biologically meaningful PPI networks across topological and functional properties, revealing that current models struggle to accurately capture network structure and functional modularity despite achieving high pairwise prediction accuracy.

Key insights

  • Current PPI models, despite demonstrating high accuracy on pairwise prediction tasks, exhibit significant limitations in faithfully recovering the global topological structure and functional modularity of real PPI networks.
  • Protein Language Model (PLM)-based models consistently outperform other categories in graph-level structural reconstruction and cross-species generalization, indicating more robust protein representations.
  • Standard binary classification metrics (e.g., F1-score) do not fully align with network-level structural properties, and higher recall often correlates with increased distortion of network topology.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.