Abstract
Detecting protein complexes is critical for studying cellular organizations and functions. The accumulation of protein-protein interaction (PPI) data enables the identification of protein complexes computationally. Although various computational approaches have been proposed to detect protein complexes from PPI networks, most of them ignore the signs of PPIs that reflect the ways proteins interact (activation or inhibition). As not all PPIs imply cocomplex relationships, taking into account the signs of PPIs can benefit the detection of protein complexes. Moreover, PPI networks are not static, but vary with the change of cell states or environments. However, existing protein complex identification algorithms are primarily designed for single-network clustering, and rarely consider joint clustering of multiple PPI networks. In this study, we propose a novel partially shared signed network clustering model (PS-SNC) for detecting protein complexes from multiple state-specific signed PPI networks jointly. PS-SNC can not only consider the signs of PPIs, but also identify the common and unique protein complexes in different states. Experimental results on synthetic and real datasets show that PS-SNC outperforms other state-of-the-art protein complex detection methods. Extensive analysis on real datasets demonstrate the effectiveness of PS-SNC in revealing novel insights about the underlying patterns of different cell lines.
Introduction
Proteins rarely function in isolation; they typically assemble into complex molecular machines to execute essential cellular processes. Identifying these complexes is a cornerstone of systems biology and is vital for understanding disease mechanisms, such as those driving cancer. Despite the proliferation of protein-protein interaction (PPI) data, computational identification of complexes remains challenging due to two primary factors.
First, traditional models often treat PPIs as simple binary links, ignoring the "sign" of the interaction. In biological systems, interactions can be positive (activation or co-complex membership) or negative (inhibition). Negative interactions provide a strong signal that two proteins are unlikely to belong to the same complex. Second, PPI networks are dynamic, changing significantly across different cell states or environmental conditions. Most existing algorithms analyze networks in isolation, failing to distinguish between "housekeeping" complexes shared across states and specialized complexes unique to specific conditions.
To address these limitations, the authors introduce PS-SNC (Partially Shared Signed Network Clustering), a framework designed to jointly analyze multiple state-specific signed PPI networks.
Methods
The PS-SNC framework is built upon Non-negative Matrix Factorization (NMF), enhanced with several novel regularization terms to capture biological nuances.
The core components of the model include:
- Partial Sharing Architecture: The model decomposes the membership matrices of different networks into a shared component ($C$), representing common complexes, and state-specific components ($S^1, S^2$), representing unique complexes.
- Signed Graph Regularization: By incorporating a signed Laplacian term, the model encourages proteins with positive interactions to cluster together while penalizing the inclusion of proteins with negative interactions within the same cluster.
- Diversity Constraint (HSIC): Using the Hilbert-Schmidt Independence Criterion, the model ensures that the state-specific components are statistically independent, preventing the "leakage" of shared information into the specific matrices.
- Low-Rank Constraint: This allows the model to adaptively estimate the number of complexes, which is typically unknown a priori.
Experimental Evaluation
The authors validated PS-SNC using both synthetic and real-world datasets. Synthetic data allowed for testing robustness against varying densities of positive and negative edges. For real-world application, the model was applied to PPI networks from BioPlex 3.0, specifically comparing the HEK293T (embryonic kidney) and HCT116 (colorectal cancer) cell lines. The CORUM database served as the gold standard for verifying predicted complexes.
Key Results
PS-SNC consistently outperformed state-of-the-art methods across various performance metrics, including Accuracy (ACC), F-measure, and Fraction of matching complexes (FRAC).
On real-world datasets, PS-SNC achieved an ACC of 0.656 and a FRAC of 0.378, representing a significant improvement over competing methods. Beyond quantitative metrics, the model demonstrated biological relevance. For instance, a case study revealed a complex involved in Wnt signaling that was specific to the HCT116 colorectal cancer line. This finding aligns with known biological drivers of colorectal malignancy, where Wnt pathway dysregulation is a hallmark.
Furthermore, analysis of the internal interaction distributions showed that complexes identified by PS-SNC had a higher proportion of positive interactions compared to unsigned methods, mirroring the characteristics of manually curated gold-standard complexes.
Conclusion
PS-SNC represents a significant advancement in protein complex detection by integrating interaction signs and joint network analysis. By mathematically mirroring the biological reality of shared and specialized cellular machinery, the framework provides more accurate and biologically meaningful insights. Future work will focus on improving scalability to handle dozens of networks simultaneously and extending the model to incorporate heterogeneous data types, such as transcriptomic or metabolic profiles.
Summary updated