TL;DR
This paper introduces a novel statistical framework to analyze the evolutionary relationships of interacting proteins within biological networks. By developing sophisticated null models, the authors demonstrate that while protein interaction networks don't show significantly higher phylogenetic similarity than expected by chance, their approach is crucial for disentangling true biological signals from noise and methodological biases in interactome studies.
What Problem Does This Paper Address?
Understanding how biological systems function relies heavily on mapping protein-protein interactions (PPIs) into networks. While these networks reveal functional relationships, their evolutionary underpinnings are less clear. Previous studies have explored whether interacting proteins tend to evolve in similar ways (co-evolve), often by examining their phylogenetic trees. However, these studies have yielded contradictory results, making it difficult to ascertain if there's a genuine evolutionary signature of interaction. This ambiguity arises from several factors: varying PPI datasets, potential biases in experimental data, and the use of different statistical approaches and null models. This paper aims to resolve these discrepancies by developing a robust statistical framework that accounts for both network structure and evolutionary history, allowing for a more nuanced investigation of phylogenetic similarities among interacting proteins.
Methods
The core of this paper lies in its statistical framework for comparing phylogenetic trees of interacting proteins. They introduce a method to quantify the similarity between phylogenetic trees of protein pairs. This involves calculating an "edit distance" between tree topologies, which measures the minimum number of changes needed to transform one tree into another. This distance is then normalized to produce a similarity score between 0 and 1.
Crucially, the authors develop a suite of "null models" or "graph ensembles" to establish a baseline for comparison. Instead of simply comparing interacting proteins to random graphs, they create ensembles that preserve specific properties of the empirical protein interaction network (PIN). These include:
- Random Graph Ensemble: A basic Erdös-Rényi model where edges are placed randomly.
- Node Shuffle Ensemble: Preserves the graph structure (degree of each node) but randomly shuffles biological node traits (like GO annotations) among nodes. This helps assess if observed similarities are due to traits being associated with specific network positions.
- Network Shuffle Ensemble: Preserves both the degree of each node and biological node traits, but rewires edges randomly. This is a more sophisticated null model that accounts for both network structure and biological properties.
- Biological Ensembles (Bipartite Node Shuffle and Bipartite Network Shuffle): These further constrain the shuffling and rewiring processes based on specific biological characteristics (e.g., GO annotations, protein complex membership), allowing for a finer dissection of potential confounding factors.
By comparing the observed phylogenetic similarity of interacting proteins to these various null models, the authors can determine if the observed similarity is statistically significant or merely a consequence of general network properties or evolutionary trends.
Datasets
The study utilizes three distinct protein-protein interaction (PPI) datasets for Saccharomyces cerevisiae (yeast):
- CORE: A subset of DIP interactions, characterized by higher clustering.
- DIP: The Database of Interacting Proteins, a comprehensive interaction database.
- LC: A literature-curated graph, representing a combination of small-scale and high-throughput experimental results.
Phylogenetic trees were generated for proteins and their orthologs across 10 yeast species, including S. cerevisiae. The evolutionary divergence times of these species range from approximately 10 million years to over 300 million years.
Key Insights and Ideas
- The importance of appropriate null models: The paper strongly emphasizes that the choice of null model is critical. Simple random graph models can lead to misleading conclusions, while more sophisticated ensembles that preserve network structure and biological traits are essential for robust analysis.
- No strong evidence for co-evolutionary "signatures" in topology: Contrary to some earlier findings, the study finds no significant evidence that interacting proteins exhibit higher topological similarity than expected by chance, even when accounting for various confounding factors.
- Phylogenetic profiles are more informative than topology for interaction: While topological similarity might not be a strong indicator, the study highlights that phylogenetic profiles (patterns of ortholog presence/absence across species) show clearer linkages with protein interactions, reinforcing previous findings.
- Co-expression and evolutionary rate might be better predictors: The authors suggest that factors like co-expression and evolutionary rate, rather than just topological similarity, might be better predictors of PPIs, aligning with some existing literature.
- Biological constraints matter: The "biological ensembles" demonstrated that when random graph generation is constrained by biological characteristics (like GO annotations), the resulting phylogenetic patterns more closely resemble empirical data, underscoring the influence of biological context.
Key Results
- The study found negligible evidence for increased levels of phylogenetic similarity among interacting protein pairs compared to what would be expected by chance, particularly when using sophisticated null models like the network shuffle ensemble.
- While topological similarity was not significantly elevated, the phylogenetic profile differences were found to be lower for interacting proteins compared to random graph ensembles, indicating a stronger evolutionary signal in the presence/absence of orthologs.
- The network shuffle ensembles, especially those constrained by biological information, produced results that were more similar to the empirical data than simpler random graph or node shuffle ensembles.
- The choice of phylogenetic tree construction algorithm (PAML, PROML, PARS) could influence results, with some algorithms showing more pronounced trends in similarity depending on the number of shared orthologues.
Limitations and Future Work
- Data incompleteness and noise: The authors acknowledge that PPI data is inherently incomplete and noisy, which can affect the reliability of inferences.
- Limited scope of "co-evolution": The analysis primarily focuses on topological similarity and doesn't directly account for other aspects of co-evolution, such as evolutionary rate or functional similarity, which might still be linked to interactions.
- Overfitting with biological ensembles: While biological ensembles provide valuable insights, there's a risk of overfitting if the constraints are too specific.
- Future work: The authors suggest employing their statistical framework for further functional and evolutionary analyses of biological networks. They also highlight the potential for using their approach to better understand the interplay between co-expression, evolutionary rates, and protein interactions. The development of more refined null models that capture additional biological complexities is also a promising avenue.
Summary updated