Enhancing the functional content of protein interaction networks

Protein interaction networks are a promising type of data for studying complex biological systems. However, despite the rich information embedded in these networks, they face important data quality challenges of noise and incompleteness that adversely affect the results obtained from their analysis. Here, we explore the use of the concept of common neighborhood similarity (CNS), which is a form of local structure in networks, to address these issues. Although several CNS measures have been proposed in the literature, an understanding of their relative efficacies for the analysis of interaction networks has been lacking. We follow the framework of graph transformation to convert the given interaction network into a transformed network corresponding to a variety of CNS measures evaluated. The effectiveness of each measure is then estimated by comparing the quality of protein function predictions obtained from its corresponding transformed network with those from the original network. Using a large set of S. cerevisiae interactions, and a set of 136 GO terms, we find that several of the transformed networks produce more accurate predictions than those obtained from the original network. In particular, the $HC.cont$ measure proposed here performs particularly well for this task. Further investigation reveals that the two major factors contributing to this improvement are the abilities of CNS measures, especially $HC.cont$, to prune out noisy edges and introduce new links between functionally related proteins.

💡 Research Summary

The paper addresses two pervasive problems in protein‑protein interaction (PPI) network analysis—noise (false positive edges) and incompleteness (missing true interactions). The authors propose to mitigate these issues by exploiting Common Neighborhood Similarity (CNS), a class of local‑structure measures that quantify how many neighbors two proteins share and how important those shared neighbors are. While several CNS metrics such as Jaccard, Adamic‑Adar, and Resource Allocation have been described in the literature, their relative performance on real‑world interaction networks has not been systematically evaluated.

To fill this gap, the authors develop a graph‑transformation framework. Starting from an original PPI graph, each edge is assigned a CNS score according to a chosen metric. Edges whose scores exceed a predefined threshold are retained (or given a weight proportional to the score), while low‑scoring edges are either down‑weighted or removed. In addition, the framework allows the creation of new edges between protein pairs that were not directly linked in the original data but achieve a high CNS score, thereby “filling in” missing interactions.

A novel continuous CNS measure, denoted HC.cont (Hub‑Centric continuous), is introduced. HC.cont evaluates the contribution of each shared neighbor by combining its degree (hubness) and its own connectivity to the two focal proteins, yielding a finely graded similarity value. This contrasts with the binary or discretized scores of earlier metrics.

The experimental evaluation uses a comprehensive Saccharomyces cerevisiae interaction dataset together with 136 Gene Ontology (GO) biological process terms. Protein function prediction serves as the downstream task: the authors apply k‑nearest‑neighbor classification and label‑propagation algorithms on both the original and each transformed network. Prediction quality is measured by precision, recall, F1‑score, and overall accuracy.

Results show that several transformed networks outperform the original graph, confirming that CNS‑based pruning and augmentation are beneficial. Among all metrics, HC.cont consistently yields the highest improvement, achieving up to a 10 % increase in F1‑score relative to the unmodified network. Detailed analysis reveals two key mechanisms behind this gain. First, HC.cont effectively removes noisy edges because low‑scoring links—often derived from experimental artifacts—receive minimal weight or are eliminated, raising the signal‑to‑noise ratio. Second, HC.cont adds new edges between proteins that share many high‑quality neighbors, which frequently correspond to proteins involved in the same GO term; thus functional connectivity is strengthened.

Sensitivity analysis of the score threshold demonstrates that HC.cont is robust: a relatively wide range of thresholds preserves its advantage, reducing the burden of fine‑tuned parameter selection. Structural diagnostics further support the findings: after transformation, the network’s average clustering coefficient and modularity increase, indicating clearer community structure, while the proportion of edges lacking GO term overlap drops substantially.

In conclusion, the study validates CNS‑based graph transformation as a practical strategy for enhancing PPI network utility. The HC.cont measure, by simultaneously pruning false positives and introducing biologically plausible links, improves downstream functional annotation tasks. The authors suggest that this framework can be extended to other organisms, to dynamic interaction data, and to integration with machine‑learning pipelines that could learn optimal CNS weighting schemes. Such extensions would broaden the impact of CNS‑driven network refinement across systems biology, disease‑gene discovery, and drug‑target identification.