Prioritizing disease genes with an improved dual label propagation framework

Yaogong Zhang, Jiahui Liu, Xiaohu Liu, Xin Fan, YuXiang Hong, Yuan Wang, Yalou Huang, Maoqiang XieBMC Bioinformatics 2018 · DOI

TL;DR

This paper introduces IDLP, a novel framework for identifying genes associated with diseases. By modeling noise in protein-protein interaction (PPI) and phenotype similarity networks, IDLP significantly improves disease gene prioritization, especially for understudied phenotypes.

What Problem Does This Paper Address?

Identifying genes responsible for diseases is crucial for understanding disease mechanisms and developing treatments. While network-based approaches have shown promise, they often struggle with the inherent noise in biological data, particularly false positive interactions in protein-protein interaction (PPI) networks. Existing methods typically treat these networks as fixed inputs, leading to biased results. This paper addresses the challenge of accurately prioritizing disease genes in the presence of noisy biological networks, especially when dealing with phenotypes that have few or no known associated genes.

Methods

The core of this paper is the Improved Dual Label Propagation (IDLP) framework. IDLP builds upon label propagation, a technique that spreads "labels" (e.g., disease association scores) across a network. Here's what makes IDLP novel:

  • Dual Label Propagation: IDLP operates on a heterogeneous network that combines gene-gene interactions (PPI network) and gene-phenotype associations. It performs label propagation on both the PPI network (for phenotypes) and the phenotype similarity network (for genes) iteratively.
  • Modeling Network Noise: Crucially, IDLP treats the PPI network and phenotype similarity network matrices as learnable variables, rather than fixed inputs. This allows the algorithm to actively model and mitigate the impact of noisy interactions (false positives) within these networks.
  • Regularization: To prevent overfitting and ensure the learned network matrices remain consistent with initial estimates, IDLP incorporates regularization terms.
  • Closed-Form Solution: The paper proposes an efficient closed-form solution for updating the model parameters, making the algorithm computationally feasible.

Datasets

The study utilizes human gene-phenotype associations from the OMIM database. Two versions of the data were used:

  • Aug-2015 version: Used for training and cross-validation.
  • Dec-2016 version: Contains newly added associations and was used as an independent test set to evaluate the model's performance on predicting novel gene-disease links.

The human protein-protein interaction (PPI) network was obtained from BioGRID. A disease phenotype network was constructed using text mining on phenotype descriptions.

Key Insights and Ideas

  • Noise is a major bottleneck: The paper strongly emphasizes that false positive interactions in PPI networks significantly degrade the performance of gene prioritization methods.
  • Networks as learnable entities: Treating biological networks (PPI and phenotype similarity) as variables that can be learned and adjusted, rather than fixed inputs, is a key innovation. This allows the model to adapt to and correct for noise.
  • Dual propagation is better: Modeling noise in both the PPI network and the phenotype similarity network simultaneously leads to better results than focusing on just one. This suggests a synergistic effect.
  • Singleton phenotypes are challenging but important: The method shows particular strength in identifying genes for "singleton phenotypes" (those with no known disease genes), a critical area for future discoveries.

Key Results

  • Superior performance: IDLP significantly outperformed eight state-of-the-art methods in terms of AUC (Area Under the Curve) scores on the test set, particularly for top-ranked genes (AUC20 and AUC50).
  • Robustness: IDLP demonstrated good robustness to simulated noise in the PPI network, maintaining stable performance even when data was randomly deleted or added.
  • Predicting new genes: For Parkinson's disease, IDLP successfully predicted 8 out of the top 10 novel genes with supporting literature evidence, achieving an 80% prediction precision for this specific disease.
  • Efficiency: The use of a closed-form solution improved computational efficiency.

Limitations and Future Work

  • Sensitivity to very low noise: The authors note that IDLP might not perform optimally when there is very little noise in the data, as its design is geared towards handling noise. They suggest acquiring basic information about data noise before applying the method.
  • Future work: The paper implicitly suggests further exploration of the learned network matrices (S1 and S2) and their biological interpretations. The development of IDLP-G and IDLP-P variants (focusing on noise in only one network) also hints at potential future research directions for understanding the differential impact of noise.

Summary updated

BibTeX
@article{zhang2018prioritizing,
  author = {Yaogong Zhang and Jiahui Liu and Xiaohu Liu and Xin Fan and YuXiang Hong and Yuan Wang and Yalou Huang and Maoqiang Xie},
  title = {{Prioritizing disease genes with an improved dual label propagation framework}},
  journal = {BMC Bioinformatics},
  year = {2018},
  doi = {10.1186/s12859-018-2040-6},
}
Figure