The study of multi-type Protein-Protein Interaction (PPI) is fundamental for
understanding biological processes from a systematic perspective and revealing
disease mechanisms. Existing methods suffer from significant performance
degradation when tested in unseen dataset. In this paper, we investigate the
problem and find that it is mainly attributed to the poor performance for
inter-novel-protein interaction prediction. However, current evaluations
overlook the inter-novel-protein interactions, and thus fail to give an
instructive assessment. As a result, we propose to address the problem from
both the evaluation and the methodology. Firstly, we design a new evaluation
framework that fully respects the inter-novel-protein interactions and gives
consistent assessment across datasets. Secondly, we argue that correlations
between proteins must provide useful information for analysis of novel
proteins, and based on this, we propose a graph neural network based method
(GNN-PPI) for better inter-novel-protein interaction prediction. Experimental
results on real-world datasets of different scales demonstrate that GNN-PPI
significantly outperforms state-of-the-art PPI prediction methods, especially
for the inter-novel-protein interaction prediction.

本研究提出一种新的评估框架和基于图神经网络的方法，旨在改善多类型蛋白质相互作用预测中的性能下降问题，尤其是在处理未知数据中的新蛋白质相互作用预测。结果表明，该方法在不同规模的真实世界数据集上都比其他最先进的预测方法表现更好。