Predicting disease-associated substitution of a single amino acid by analyzing residue interactions

Yizhou Li, Zhining Wen, Jiamin Xiao, Hui Yin, Lezheng Yu, Li Yang, Menglong LiBMC Bioinformatics 2011 · DOI

TL;DR

This paper introduces a novel computational approach, NetSAP, that leverages protein structure networks to predict whether a single amino acid substitution (a non-synonymous SNP or SAP) is likely to cause disease. By analyzing the "interaction network" of amino acids within a protein, NetSAP achieves high accuracy, outperforming existing methods and offering a new perspective on understanding the functional impact of genetic variations.

What Problem Does This Paper Address?

Genetic variations, particularly single nucleotide polymorphisms (SNPs), are a major source of human diversity and can lead to diseases. Non-synonymous SNPs (nsSNPs), also called single amino acid polymorphisms (SAPs), directly alter the protein sequence. While millions of SNPs are known, only a fraction have been linked to specific diseases. Identifying which SAPs are disease-associated is crucial for understanding disease mechanisms and developing personalized medicine. Existing methods often rely on evolutionary conservation or sequence-based features, but these can miss the subtle, context-dependent effects of mutations. This paper addresses the challenge of accurately predicting disease-associated SAPs by exploring how the local and global structural context of an amino acid substitution influences its functional impact.

Methods

The core innovation of this paper is the use of Protein Structure Networks (PSNs). A PSN represents a protein as a graph where amino acid residues are nodes and interactions (defined by proximity within a certain distance threshold) are edges. The authors then extract several network topological features from these PSNs to characterize each residue:

  • Degree: The number of direct neighbors a residue has.
  • Closeness: How close a residue is, on average, to all other residues.
  • Betweenness: How often a residue lies on the shortest path between other pairs of residues.
  • Clustering Coefficient: A measure of how interconnected a residue's neighbors are.

These topological features are hypothesized to reflect the residue's importance in maintaining protein structure and function. The authors also incorporate neighboring environmental features, which capture the topological and conservation properties of residues directly interacting with the mutated site.

A Random Forest algorithm was employed to build predictive models. This ensemble learning method combines multiple decision trees to make predictions, offering robustness and good performance. The models were trained and evaluated using various combinations of features: conservation-based, topological, and environmental.

Datasets

The study utilized two main datasets:

  1. Training Dataset: Compiled from the Swiss-Prot variant page, containing 6527 SAPs from 1094 proteins. This dataset included 3953 disease-associated SAPs and 2574 polymorphic SAPs. A subset of 876 proteins with 3257 disease-associated and 2118 polymorphic SAPs was used for 5-fold cross-validation.
  2. Independent Test Dataset: A randomly selected set of 218 proteins with 696 disease-associated and 456 polymorphic SAPs, used to evaluate the final model's performance on unseen data.

The authors note that only variants with corresponding 3D structures were considered, and problematic structures were removed.

Key Insights and Ideas

  • Network Topology as a Proxy for Functional Importance: The study demonstrates that network topological features (degree, closeness, betweenness, clustering coefficient) can effectively distinguish disease-associated SAPs from polymorphic ones. This suggests that a residue's position and role within the protein's interaction network are strong indicators of its functional significance.
  • Disease SAPs are "Hubs" or "Bridges": Disease-associated SAPs tend to have higher degrees (more neighbors) and closeness values, indicating they are often central or highly connected residues within the protein structure. This implies that disrupting these critical nodes has a greater functional consequence.
  • Local Environment Matters: Analyzing the properties of neighboring residues around a mutation site (environmental features) significantly improves prediction accuracy. This highlights that the impact of a substitution isn't isolated but depends heavily on its immediate structural context.
  • Complementary Information: Network-based features provide information complementary to traditional conservation scores. While conservation indicates evolutionary pressure, network features reveal the structural and functional role of a residue within the protein's architecture.

Key Results

The developed model, NetSAP, achieved impressive performance:

  • 5-fold Cross-Validation: Using all features (conservation, topological, and environmental), the model achieved a total accuracy (ACC) of 83.0% and a Matthew's Correlation Coefficient (MCC) of 0.64.
  • Independent Test Dataset: The model achieved an ACC of 80.8% and MCC of 0.59 on the independent dataset.
  • Feature Importance: The study identified that features related to the "HLA family" had the highest importance score, suggesting its relevance in disease prediction. Environmental features, particularly closeness and conservation of neighbors, were also highly ranked.
  • Comparison to Existing Methods: On the independent dataset, NetSAP (using all features) outperformed SIFT (ACC 76.3%), PolyPhen-2 (ACC 75.7%), and Bongo (ACC 46.6%).

Limitations and Future Work

  • Structural Data Dependency: The method relies on the availability of 3D protein structures for the variants. This limits its application to proteins for which structural data exists.
  • Refining Interaction Definitions: The authors suggest that a PSN with more refined residue interaction definitions (e.g., considering different types of bonds like hydrophobic, hydrogen bonds, van der Waals, electrostatic) could further improve the accuracy in reflecting structural/functional roles.
  • Expanding Feature Sets: Future work could explore incorporating more sophisticated network analysis techniques and a wider range of environmental features.
  • Investigating Linkages: The authors envision this method aiding in understanding the intricate linkages between SAPs and diseases through extensive analysis of residue interactions.

Summary updated

BibTeX
@article{li2011predicting,
  author = {Yizhou Li and Zhining Wen and Jiamin Xiao and Hui Yin and Lezheng Yu and Li Yang and Menglong Li},
  title = {{Predicting disease-associated substitution of a single amino acid by analyzing residue interactions}},
  journal = {BMC Bioinformatics},
  year = {2011},
  doi = {10.1186/1471-2105-12-14},
}
Figure