Identifying Disease Related Genes by Network Representation and Convolutional Neural Network

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Identifying disease-related genes is crucial in bioinformatics.
• Deep learning methods show high accuracy but struggle with gene multifunctionality and network scale-free properties.

Purpose of the Study:

• To propose a novel network representation method for improved disease-related gene identification.
• To overcome limitations of existing deep learning approaches in handling complex biological network characteristics.

Main Methods:

• Developed a network representation technique converting node-induced sub-networks into image-like datasets.
• Integrated surrounding topological structures and environmental characteristics for low-dimensional representation.
• Applied the image-like datasets to a Convolutional Neural Network (CNN) for cancer-related gene identification.

Main Results:

• Achieved an Area Under the Curve (AUC) of 0.9256 in a single network.
• Reached an AUC of 0.9452 when analyzing multiple networks.
• Demonstrated superior performance compared to existing gene identification methods.

Conclusions:

• The proposed network representation method effectively addresses gene multifunctionality and network scale-free properties.
• This approach significantly enhances the accuracy of identifying disease-related genes using deep learning.
• The method shows strong potential for advancing cancer-related gene discovery.