TL;DR
This paper introduces CRAIN, a novel probabilistic model and bipartite graph framework to identify "co-regulated modules" in human biological networks. These modules are shown to act as "assemblers," integrating genes, protein complexes, and metabolic pathways, with significant implications for understanding complex diseases like cancer.
What Problem Does This Paper Address?
Understanding the intricate workings of cellular systems and the molecular basis of diseases is a major challenge in the post-genomic era. Biological processes are orchestrated by complex molecular networks, including protein-protein interaction networks (PPINs) and transcriptional regulatory networks (TRNs). While individual networks provide valuable insights, integrating them offers a more comprehensive view. Researchers hypothesize that groups of co-regulated genes, forming "functional modules," are responsible for packaging multiple biological entities and coordinating specific cellular functions. However, a robust method to identify these modules and understand their "packaging" capabilities – how they integrate diverse biological components – was lacking, particularly in the context of complex diseases like cancer.
Methods
The authors developed a probabilistic model called CRAIN (Co-Regulatory Analysis using Integrated Networks). The core of CRAIN involves:
- Integrative Weighted Network Construction: Combining human PPIN and TRN data into a single, weighted network. Edge weights reflect confidence based on evidence.
- Probabilistic Statistical Model: A log-likelihood ratio is used to assess whether a cluster of co-regulated proteins is likely to form a module. This model compares the fit of a sub-network to a "module" model (high interaction probability) against a "random" background model.
- Module Searching Algorithm: A greedy approach involving seed initialization, seed expansion, and overlap filtering is employed to identify candidate modules.
- Bipartite Graph Framework: To understand the "packaging" features, CRAIN uses bipartite graphs to map identified modules to three types of biological entities: genes, protein complexes, and metabolic pathways. This allows for the analysis of how modules integrate these components.
Datasets
- Human Protein-Protein Interaction (PPI) Data: Extracted from HPRD databases (Release 7), containing 34,083 interactions between 9,014 proteins. Edge reliability was determined by supporting evidence.
- Transcriptional Regulatory Data: Acquired from the Transfac Database (Release 11.4), comprising 281 TFs and 624 genes with 1,603 interactions.
- Cancer Mutated Genes: Obtained from the Cancer Gene Census, listing 384 genes causally implicated in various tumor types.
- Human Protein Complexes: Data from the MIPS database (1347 complexes) was used for complex packaging analysis.
- KEGG Pathway Database: Used for pathway annotation analysis.
Key Insights and Ideas
- Modules as "Assemblers": The study strongly suggests that co-regulated modules act as functional "assemblers," integrating diverse biological components (genes, complexes, pathways) to perform specific tasks.
- Cancer-Specific Packaging: Co-regulated modules are particularly relevant in cancer, where they can package cancer-mutated genes and regulate functions critical to tumorigenesis.
- Interconnectedness of Biological Layers: The research highlights how transcriptional regulation (TRNs) and protein interactions (PPINs) work together to package and coordinate higher-level biological entities like protein complexes and metabolic pathways.
- Robustness of Identified Modules: The identified modules demonstrate high internal connectivity and are robust against random node replacements, suggesting they represent genuine biological structures.
- Functional Coherency: The modules are significantly enriched in specific biological functions, indicating their biological relevance.
Key Results
- Identification of 96 Co-regulated Modules: CRAIN successfully identified 96 human co-regulated modules.
- Module Robustness: Analysis showed that internal connectivity of identified modules significantly decreased with random node replacement, while the ratio of connections inside vs. outside the module remained higher than random controls, indicating robustness.
- Functional Enrichment: 80% of the identified modules were significantly enriched in biological functions based on Gene Ontology (GO) analysis.
- Cancer Association: 42% of the modules were significantly associated with cancer (p < 0.05).
- Complex and Pathway Integration: 94% of modules could organize numerous protein complexes, and 82% had significant annotated pathway information. A specific example module (module 27) was shown to package 8 functional classes of protein complexes and 8 divergent metabolic pathways.
- Performance Comparison: CRAIN outperformed four other module identification algorithms (Connected, Biconnected, CPM, MCL) in terms of sensitivity, specificity, and F-score, indicating its superior ability to find modules with richer biological meaning.
Limitations and Future Work
- Greedy Algorithm Limitation: The greedy algorithm used for module searching aims for locally optimal choices and may not always find globally optimal solutions. However, the authors believe it was effective in practice.
- Incomplete Annotation: A significant limitation is the incomplete annotation of human genes, complexes, and pathways. Two-thirds of human genes lack full annotation, and the incompleteness of complex and pathway information might lead to missed overlaps or packing relationships.
- Future Directions: The authors anticipate that with the accumulation of more human data, their framework can be extended to identify additional modules and further elucidate their packaging features. They also suggest that the identified modules could serve as potential biomarkers for cancer.
Summary updated