[Paper Review] DPPIN: A Biological Repository of Dynamic Protein-Protein Interaction Network Data

Dongqi Fu, Jingrui He|arXiv (Cornell University)|Jul 5, 2021

Bioinformatics and Genomic Networks64 references4 citations

TL;DR

DPPIN introduces a novel biological repository of 12 dynamic protein-protein interaction networks in yeast, generated from time-resolved gene expression data to enable dynamic graph representation learning. The repository supports diverse tasks like dynamic spectral clustering and few-shot graph classification, with tdGraphEmbed + ProtoNet achieving state-of-the-art performance, highlighting challenges and opportunities for next-generation dynamic graph algorithms.

ABSTRACT

In the big data era, the relationship between entries becomes more and more complex. Many graph (or network) algorithms have already paid attention to dynamic networks, which are more suitable than static ones for fitting the complex real-world scenarios with evolving structures and features. To contribute to the dynamic network representation learning and mining research, we provide a new bunch of label-adequate, dynamics-meaningful, and attribute-sufficient dynamic networks from the health domain. To be specific, in our proposed repository DPPIN, we totally have 12 individual dynamic network datasets at different scales, and each dataset is a dynamic protein-protein interaction network describing protein-level interactions of yeast cells. We hope these domain-specific node features, structure evolution patterns, and node and graph labels could inspire the regularization techniques to increase the performance of graph machine learning algorithms in a more complex setting. Also, we link potential applications with our DPPIN by designing various dynamic graph experiments, where DPPIN could indicate future research opportunities for some tasks by presenting challenges on state-of-the-art baseline algorithms. Finally, we identify future directions to improve the utility of this repository and welcome constructive inputs from the community. All resources (e.g., data and code) of this work are deployed and publicly available at https://github.com/DongqiFu/DPPIN.

Motivation & Objective

• To address the scarcity of high-quality, labeled, and attribute-rich dynamic network datasets in the biological domain.
• To provide a scalable, well-documented, and publicly accessible repository of dynamic protein-protein interaction networks for advancing dynamic graph machine learning.
• To inspire new methods in dynamic graph representation learning by offering realistic evolution patterns, node features, and labels.
• To link the repository to real-world applications through benchmark experiments in dynamic graph classification, clustering, and similarity retrieval.
• To identify future directions for improving label coverage and algorithmic utility through community-driven updates and self-supervised learning.

Proposed method

• Dynamic networks are generated by decomposing a static PPI network into 12 time-resolved snapshots using gene expression data from GSE3431.
• Node features are derived from time-aware protein expression values, and edge features represent co-expression or interaction likelihood.
• Each dynamic network is structured as a sequence of snapshots $ G^{(t)} = (V^{(t)}, E^{(t)}) $, with evolving node sets and time-dependent features.
• Subgraphs of 5 consecutive timestamps are extracted to form dynamic sequences for few-shot learning, preserving original class labels.
• Experiments employ meta-learning frameworks (e.g., ProtoNet, kNN) with graph embedding models like Graph2Vec, GL2Vec, and tdGraphEmbed to evaluate performance.
• A pooling strategy aggregates snapshot-level representations into graph-level embeddings, with attention mechanisms suggested as a future improvement for better class distinctiveness.

Experimental results

Research questions

• RQ1 Can dynamic graph representation learning models effectively capture biologically meaningful evolution patterns in time-evolving PPI networks?
• RQ2 How do state-of-the-art dynamic graph embedding methods perform on real-world biological dynamic networks with complex structural and feature evolution?
• RQ3 What are the limitations of current aggregation strategies in dynamic graph representation learning when class-ambiguous snapshots exist across different biological conditions?
• RQ4 How can dynamic network repositories like DPPIN guide the design of next-generation graph algorithms with improved generalization and interpretability?
• RQ5 What role can self-supervised or semi-supervised learning play in mitigating incomplete labeling in biological dynamic networks?

Key findings

• tdGraphEmbed + ProtoNet achieved the highest graph-level classification accuracy of 0.6791 ± 0.1141 in the 3-way 3-shot setting, outperforming other baselines.
• GL2Vec + ProtoNet achieved 0.6625 ± 0.0407 accuracy, indicating that temporal-aware graph embedding improves performance on dynamic biological networks.
• Graph2Vec + kNN failed to converge within 48 hours on a V100 GPU, highlighting computational challenges in applying static methods to dynamic biological data.
• The subgraph sampling strategy enabled a robust 11-subgraph per-class setup for few-shot meta-learning, supporting reliable evaluation of generalization.
• The presence of globally shared snapshots across different classes suggests that simple pooling may reduce class distinctiveness, indicating a need for more sophisticated aggregation mechanisms.
• The repository reveals that current label coverage is incomplete, with some proteins lacking functional annotations, which limits supervised learning and motivates future self-supervised learning integration.