KISL: knowledge-injected semi-supervised learning for biological co-expression network modules.

Area of Science:

• Bioinformatics
• Computational Biology
• Cancer Research

Background:

• Identifying cancer biomarkers is crucial for diagnosis, treatment, and prognosis.
• Gene co-expression network analysis aids in discovering synergistic gene sets for biomarker mining.
• Current methods like WGCNA have limitations in capturing complex gene relationships and lack biological knowledge integration.

Purpose of the Study:

• To develop a novel knowledge-injected semi-supervised learning approach (KISL) for enhanced co-expression network analysis.
• To address limitations of unsupervised methods by incorporating prior biological knowledge.
• To improve the identification and delineation of biologically relevant gene modules.

Main Methods:

• Introduced a knowledge-injected semi-supervised learning approach (KISL).
• Utilized distance correlation to measure both linear and non-linear gene dependencies.
• Applied KISL to eight cancer RNA-seq datasets, comparing its performance against WGCNA.

Main Results:

• KISL demonstrated superior performance over WGCNA across multiple evaluation metrics (silhouette coefficient, Calinski-Harabasz index, Davies-Bouldin index).
• KISL achieved better gene module aggregation and cluster evaluation values.
• Enrichment analysis confirmed the effectiveness of KISL-identified modules in revealing biological co-expression network structures.

Conclusions:

• KISL offers a robust and effective method for identifying biologically meaningful gene modules in co-expression networks.
• The integration of prior biological knowledge and semi-supervised learning significantly enhances biomarker discovery.
• KISL is a versatile method applicable to various co-expression network analyses.