Current composite-feature classification methods do not outperform simple single-genes classifiers in breast cancer...

Area of Science:

• Bioinformatics
• Computational Biology
• Cancer Research

Background:

• Integrating gene expression data with secondary data (pathways, protein-protein interactions) is a common approach for cancer patient outcome prediction.
• Previous studies faced limitations due to small datasets, varied data, and inconsistent evaluation procedures, hindering objective performance assessment.
• This study addresses the need for standardized evaluation of methods that aggregate gene expression into composite features.

Purpose of the Study:

• To introduce the Amsterdam Classification Evaluation Suite (ACES), a Python package for objective evaluation of classification and feature-selection methods.
• To facilitate the comparison of new approaches against best-in-class methods for cancer outcome prediction.
• To assess the performance and stability of composite-feature methods compared to single-gene classifiers and network-based approaches.

Main Methods:

• Development and implementation of the Amsterdam Classification Evaluation Suite (ACES) for pooling and normalizing Affymetrix microarrays.
• Inclusion of established prognostic gene signatures for breast cancer, composite feature selection, and network-based gene ranking methods within ACES.
• Application of the ACES evaluation pipeline to compare various classification and feature-selection strategies.

Main Results:

• Composite-feature classification methods did not outperform simple single-gene classifiers in predicting breast cancer patient outcomes.
• The stability of features across different datasets was not improved for composite features compared to single-gene features.
• Prediction performance remained unaffected even when features were extracted from randomized protein-protein interaction (PPI) networks.

Conclusions:

• Current composite-feature approaches do not offer significant advantages over simpler methods for breast cancer outcome prediction.
• The integration of secondary data, particularly PPI networks, does not inherently enhance predictive accuracy or feature stability.
• The ACES package provides a standardized framework for robust evaluation of predictive models in cancer research.