Related Experiment Video
Updated: May 10, 2026
07:35
Selecting Multiple Biomarker Subsets with Similarly Effective Binary Classification Performances
Published on: October 11, 2018
When complexity does not pay: benchmarking deep learning and ensemble methods for biomarker discovery
Cyrille Mesue Njume1, Irene Petracci2, Sonia Bellini2
1Department of Molecular Biology and Genetics, Ayazaga Campus, Istanbul Technical University, Reşitpaşa, Sarıyer, 34467 Istanbul, Turkey.
Briefings in Bioinformatics
|May 8, 2026
Summary
Benchmarking multi-omics biomarker discovery reveals ensemble feature selection enhances accuracy. Simpler models often match deep learning, highlighting the need for transparency and interpretability in computational methods.
Area of Science:
- Bioinformatics
- Computational Biology
- Translational Medicine
Background:
- Multi-omics data integration offers potential for biomarker discovery but faces challenges due to computational complexity.
- Evaluating the practical utility and interpretability of various computational methods is crucial for clinical application.
Purpose of the Study:
- To benchmark 27 feature selection strategies and 11 predictive models across Alzheimer's disease, progressive supranuclear palsy, and breast cancer cohorts.
- To compare traditional machine learning, ensemble methods, and deep learning for predictive performance, stability, and interpretability.
- To assess the impact of omics data integration levels (single, dual, triple) on biomarker discovery.
Main Methods:
- Developed a comprehensive benchmarking framework for feature selection and predictive modeling.
- Evaluated 27 feature selection strategies and 11 predictive models, including logistic regression, support vector machines, multilayer perceptrons, and deep learning.
- Utilized three real-world disease cohorts and validated biomarkers against five independent databases.
Main Results:
- Ensemble feature selection consistently improved robustness and accuracy, especially for compact biomarker panels.
- Simpler classifiers (e.g., logistic regression) achieved comparable or superior results to deep learning models, with lower computational cost and higher interpretability.
- Triple-omics data integration demonstrated the highest validation performance (Triple > Dual > Single).
- Identified both established and novel biomarkers with confirmed clinical relevance.
Conclusions:
- A pragmatic approach prioritizing transparency, reproducibility, and biological insight is recommended for biomarker discovery over algorithmic complexity.
- Ensemble methods and simpler classifiers offer effective and interpretable solutions for multi-omics biomarker identification.
- The developed web-based tool supports reproducibility and community adoption of the benchmarking pipeline.
