PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs

TL;DR

PRING tackles the gap between pairwise PPI prediction and real biological networks by introducing a graph‑level benchmark built from a high‑quality, multi‑species interactome. It defines topology‑ and function‑oriented tasks to evaluate how well models reconstruct PPI networks and preserve biological modules, GO signals, and essential proteins. Extensive experiments across sequence, PLM, and structure‑based methods reveal that current approaches struggle to maintain sparsity, functional coherence, and centrality signals in reconstructed networks, despite strong pairwise accuracy. By providing a fully reproducible dataset and evaluation pipeline, PRING enables the community to develop PPI models with greater practical utility for biology and medicine, while highlighting risks of overreliance on classification metrics alone.

Abstract

Deep learning-based computational methods have achieved promising results in predicting protein-protein interactions (PPIs). However, existing benchmarks predominantly focus on isolated pairwise evaluations, overlooking a model's capability to reconstruct biologically meaningful PPI networks, which is crucial for biology research. To address this gap, we introduce PRING, the first comprehensive benchmark that evaluates protein-protein interaction prediction from a graph-level perspective. PRING curates a high-quality, multi-species PPI network dataset comprising 21,484 proteins and 186,818 interactions, with well-designed strategies to address both data redundancy and leakage. Building on this golden-standard dataset, we establish two complementary evaluation paradigms: (1) topology-oriented tasks, which assess intra and cross-species PPI network construction, and (2) function-oriented tasks, including protein complex pathway prediction, GO module analysis, and essential protein justification. These evaluations not only reflect the model's capability to understand the network topology but also facilitate protein function annotation, biological module detection, and even disease mechanism analysis. Extensive experiments on four representative model categories, consisting of sequence similarity-based, naive sequence-based, protein language model-based, and structure-based approaches, demonstrate that current PPI models have potential limitations in recovering both structural and functional properties of PPI networks, highlighting the gap in supporting real-world biological applications. We believe PRING provides a reliable platform to guide the development of more effective PPI prediction models for the community. The dataset and source code of PRING are available at https://github.com/SophieSarceau/PRING.

Figures (12)

• Figure 1: Overview of PRING benchmark. (a) Diverse databases are used to construct the PRING. (b) PRING includes two topology-oriented tasks and three function-oriented tasks for extensive evaluation. (c) List of baseline models, consisting of sequence similarity-based, naive sequence-based, structure-based, and PLM-based methods. (d) Evaluation metrics used for each task in the PRING.
• Figure 2: Data collection pipeline for the PRING. PPIs are first curated from comprehensive databases. Proteins are then filtered and mapped using SwissProt and NCBI Taxonomy to target species. Redundant interactions are removed through sequence and functional similarity checks to ensure data quality. The resulting PPI networks include four species: Human, Arath, Yeast, and Ecoli.
• Figure 3: Cross-species generalization performance evaluated via graph similarity score.
• Figure 4: Essential protein analysis. (a) The network centrality score of essential and non-essential proteins in the ground-truth PPI network. (b) Network centrality score distribution of PLM-interact (650M).
• Figure 5: Relationship between classification performance and graph-level metrics. The recall rate is positively correlated with the degree distribution (MMD), while the precision is negative.