[Paper Review] Inferring disease causing genes and their pathways: A mathematical perspective

Jeethu V. Devasia, Priya Chandran|arXiv (Cornell University)|Oct 30, 2016

Bioinformatics and Genomic Networks23 references3 citations

TL;DR

This paper proposes a graph-theoretic approximation algorithm with graph pruning and randomized rounding to identify disease-causing genes and dysregulated pathways in molecular interaction networks. By reducing network size through edge and vertex pruning based on edge weights and vertex degree, and applying randomized rounding to select high-impact paths, the method achieves higher accuracy and faster execution than existing approaches like random walks and electric circuit models.

ABSTRACT

A system level view of cellular processes for human and several organisms can be cap- tured by analyzing molecular interaction networks. A molecular interaction network formed of differentially expressed genes and their interactions helps to understand key players behind disease development. So, if the functions of these genes are blocked by altering their interactions, it would have a great impact in controlling the disease. Due to this promising consequence, the problem of inferring disease causing genes and their pathways has attained a crucial position in computational biology research. However, considering the huge size of interaction networks, executing computations can be costly. Review of literatures shows that the methods proposed for finding the set of disease causing genes could be assessed in terms of their accuracy which a perfect algorithm would find. Along with accuracy, the time complexity of the method is also important, as high time complexities would limit the number of pathways that could be found within a pragmatic time interval.

Motivation & Objective

• Address the computational challenge of identifying disease-causing genes and dysregulated pathways in large-scale molecular interaction networks.
• Overcome the high time complexity of existing methods, such as random walks and electric circuit models, which limit scalability.
• Improve both accuracy and execution time by integrating graph pruning with approximation algorithms.
• Enable practical analysis of large biological networks by reducing network size while preserving biologically relevant connections.
• Provide a computationally efficient yet biologically relevant method for identifying causal genes and pathways for drug development and gene therapy applications.

Proposed method

• Transform the problem of identifying causal genes and pathways into a graph theoretical problem using molecular interaction networks.
• Apply graph pruning by iteratively removing low-weighted edges and vertices of minimum degree to reduce network size.
• Use a min-heap data structure to efficiently identify and remove vertices with the lowest degree during pruning.
• Set the pruning threshold to 10% of the total number of vertices (|V|), reducing the network to a manageable size.
• Apply randomized rounding to the pruned graph to identify the most relevant paths and select high-impact causal genes.
• Leverage gene expression correlation (Pearson correlation) to assign edge weights, representing interaction strength between genes.

Experimental results

Research questions

• RQ1 Can graph pruning techniques significantly reduce the computational complexity of identifying disease-causing genes in large molecular interaction networks?
• RQ2 How does integrating randomized rounding with graph pruning improve the accuracy and efficiency of causal gene and pathway inference compared to existing methods?
• RQ3 To what extent does the proposed method outperform random walk and electric circuit-based approaches in terms of both accuracy and execution time?
• RQ4 Can the method identify novel disease-causing genes beyond known benchmark genes, enhancing biological relevance?
• RQ5 Does the use of min-heap-based vertex selection during pruning improve time efficiency without compromising the quality of identified pathways?

Key findings

• The proposed approximation algorithm with graph reduction outperformed all baseline methods in identifying a higher number of causal genes with improved accuracy.
• On benchmark datasets, the method achieved 4114 minutes (over 2800 hours) of execution time in one dataset, but this was still more efficient than the electric circuit and matrix inversion approaches, which took over 1800 hours.
• The randomized rounding approach identified more causal genes than the random walk and eQED methods, which had lower accuracy despite faster execution.
• The method with min-heap optimization reduced execution time to 8.61 seconds on one dataset, significantly outperforming the standard approximation algorithm.
• Biological relevance analysis confirmed that the identified genes and pathways were consistent with known disease mechanisms, validating the method’s biological plausibility.
• The algorithm successfully identified novel disease-causing genes not present in benchmark datasets, as detailed in the supplementary material.