The ability to classify patients based on gene-expression data varies by algorithm and performance metric.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Medicine

Background:

• Patient classification into subgroups improves healthcare efficacy compared to uniform approaches.
• High-throughput transcriptomic data presents challenges for traditional statistical models due to high dimensionality and complex gene dependencies.
• Numerous machine-learning algorithms with diverse hyperparameters exist, making optimal selection for gene-expression biomarker studies difficult.

Purpose of the Study:

• To benchmark the performance of various machine-learning classification algorithms for gene-expression biomarker studies.
• To identify which types of algorithms and preprocessing steps (hyperparameter optimization, feature selection) are most effective.
• To guide researchers in selecting optimal algorithms for analyzing transcriptomic data.

Main Methods:

• Applied 52 diverse classification algorithms from open-source libraries to 50 gene-expression datasets.
• Combined gene-expression data with clinical predictors where available.
• Evaluated the impact of hyperparameter optimization and feature selection using nested cross-validation.

Main Results:

• Kernel- and ensemble-based algorithms demonstrated superior performance compared to other methods.
• Hyperparameter optimization and feature selection generally enhanced predictive accuracy.
• Univariate feature selection methods often outperformed more complex feature selection techniques.
• Significant variability in algorithm performance was observed across datasets.

Conclusions:

• Algorithm selection is a critical determinant of success in gene-expression biomarker studies.
• While kernel and ensemble methods show promise, no single algorithm guarantees optimal performance.
• Systematic benchmarking and careful consideration of preprocessing steps are essential for robust patient classification.