Survival models with preclustered gene groups as covariates.

Area of Science:

• Genomics
• Bioinformatics
• Biostatistics

Background:

• High-dimensional gene expression data presents challenges in risk prediction and model interpretability due to the large number of genes relative to observations.
• Feature selection improves model accuracy and complexity but often leaves selected genes with limited biological understanding.
• Gene Ontology (GO) groups genes hierarchically, offering a way to summarize genes, but GO group expression profiles can be heterogeneous.

Purpose of the Study:

• To develop a novel method for creating coherent gene subgroups within GO groups.
• To integrate these preclustered gene groups as covariates in survival models for enhanced interpretability.
• To assess the impact of preclustered GO groups on the prediction accuracy and biological interpretation of survival models.

Main Methods:

• Genes within GO groups were preclustered based on the correlation of their expression measurements to identify coherent subgroups.
• Cox regression models were employed to analyze disease-free survival times in breast cancer patients.
• Models incorporated classical clinical covariates, individual genes, GO groups, and preclustered GO groups as genomic covariates.

Main Results:

• Survival models utilizing preclustered gene groups as covariates demonstrated prediction accuracy comparable to models using single genes or GO groups.
• The inclusion of preclustered gene groups did not compromise predictive performance.

Conclusions:

• Preclustering gene groups within GO annotations provides a more detailed analysis of the biological significance of selected covariates in survival models.
• This approach offers additional functional insights beyond individual genes and more coherent gene expression profiles than standard GO groups.