A simulation to analyze feature selection methods utilizing gene ontology for gene expression classification.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Gene expression profile classification is crucial for personalized medicine.
• A key challenge is the high dimensionality of gene expression data (many genes, few samples).
• Gene Ontology (GO) semantic similarity is explored for improved feature selection.

Purpose of the Study:

• To investigate the optimal conditions for using GO-based feature selection in gene expression data classification.
• To develop a simulation to evaluate the efficacy of GO-based feature selection under various factors.
• To determine when GO-based feature selection is beneficial compared to traditional statistical methods.

Main Methods:

• Developed a novel simulation to generate binary class datasets with underlying GO relationships for differentially expressed genes.
• Investigated the impact of gene connectedness in GO, mean separation (δ), and training sample size.
• Defined Biological Condition Annotation Level (BCAL(G)) to quantify gene connectedness.

Main Results:

• GO-based feature selection efficacy is primarily determined by the connectedness of differentially expressed genes in GO.
• Classification accuracy improves with increased gene connectedness (BCAL(G) ≥ 0.696).
• Accuracy decreases with low connectedness (BCAL(G) ≤ 0.389) or when gene count exceeds 50 with δ ≥ 0.7.

Conclusions:

• GO-based feature selection is effective when gene connectedness (BCAL(G)) is high (≥ 0.696).
• It is not recommended for conditions with fewer than ten genes.
• Guidelines are provided for using GO-based feature selection based on BCAL(G) and δ values.