Determination of biomarkers from microarray data using graph neural network and spectral clustering.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Microarray data presents challenges like small sample size and high dimensionality.
• Effective feature selection is crucial for identifying disease biomarkers and improving classification accuracy.
• Existing methods struggle with redundant and irrelevant features in high-dimensional biological data.

Purpose of the Study:

• To develop an innovative feature selection method for microarray data analysis.
• To leverage a priori knowledge and graph neural networks for enhanced feature selection.
• To improve disease classification accuracy by identifying relevant biomarkers.

Main Methods:

• Constructing a graph structure using gene interaction networks as a priori knowledge.
• Employing a link prediction method based on graph neural networks to refine the graph structure.
• Utilizing spectral clustering for feature selection and biomarker identification.

Main Results:

• The proposed method significantly improved classification accuracy on DLBCL (10.90%) and Prostate (16.22%) datasets compared to traditional methods.
• Link prediction further enhanced average classification accuracy by 1.96% and 1.31%, outperforming published methods.
• The method effectively utilizes a priori knowledge to select high-accuracy disease prediction biomarkers.

Conclusions:

• The developed feature selection method effectively identifies disease biomarkers from high-dimensional microarray data.
• Integrating gene interaction networks and graph neural networks enhances the selection of relevant features.
• This approach offers a promising strategy for improving disease classification and biomarker discovery in bioinformatics.