Ensemble Feature Selection for Robust Multi-Omics Biomarker Discovery

A Systems-Level Approach to Biomarker Discovery

Abstract

Biomarker discovery often suffers from instability — different feature selection algorithms produce different gene lists, reducing reproducibility and clinical trust.

This project addressed that challenge by designing and implementing a comprehensive ensemble feature selection framework that integrates multiple ranking strategies and machine learning classifiers to identify robust biomarkers across heterogeneous omics datasets.

The framework was later integrated into the BioMark user interface platform:

Balikci, M.A., Njume, C.M. & Cakmak, A.
BioMark: biomarker analysis tool.
BMC Bioinformatics 27, 42 (2026).
https://doi.org/10.1186/s12859-025-06346-3

🧠 Problem Statement

Single feature selection methods are:

• Sensitive to noise
• Biased toward model assumptions
• Unstable across datasets
• Often poorly reproducible

Research questions addressed:

1. Do ensemble feature selection methods improve stability and predictive performance?
2. Which aggregation strategies perform best across multi-omics cohorts?
3. Do complex models outperform simpler interpretable classifiers?
4. How stable are biomarker panels across diseases?

🛠 Feature Selection Methods Implemented

🔹 Embedded & Regularized Models

• Lasso
• Elastic Net
• Logistic Regression (L1/L2)
• Random Forest importance
• XGBoost importance

🔹 Wrapper Methods

• SVM-RFE
• Boruta

🔹 Statistical Methods

• T-Test
• Fold Change ranking

🔹 Explainability-Based Methods

• SHAP (mean absolute importance)
• LIME (local interpretability aggregation)

🔗 Ensemble Aggregation Strategies

Two major ensemble paradigms were implemented:

📊 Rank-Based Aggregation

• Min Rank
• Mean Rank
• Median Rank
• Geometric Mean Rank
• Robust Rank Aggregation (RRA)
• Stuart Aggregation

⚖ Weight-Based Aggregation

• Mean Weight
• Max Weight
• Median Weight
• Geometric Mean Weight
• Threshold Algorithm

This allowed systematic comparison between rank-level and importance-level integration strategies.

🤖 Classification Models Evaluated

To evaluate biological relevance and predictive performance, selected biomarker panels were tested using:

• Random Forest
• XGBoost
• Support Vector Classifier (SVC)
• Logistic Regression
• Multi-Layer Perceptron (MLP)

Performance metrics:

• Accuracy
• F1-score
• ROC-AUC
• Stability scores
• Cross-validation consistency

🧬 Multi-Omics Benchmark Study

The framework was evaluated across three independent multi-omics cohorts, including:

• Alzheimer’s disease
• Progressive supranuclear palsy
• Breast cancer

Data types included:

• Transcriptomics
• microRNA
• Metabolomics
• Combined multi-omics

Key finding:

Ensemble feature selection consistently improved robustness and performance compared to single-method selection.

Surprisingly:

Simpler classifiers (Logistic Regression, SVC) often matched or outperformed deeper models, particularly when combined with ensemble feature selection.

This work formed the basis of a large-scale benchmarking manuscript currently submitted to Briefings in Bioinformatics.

💻 Engineering Contributions

I implemented:

• Modular ranking pipelines
• Automated ensemble aggregation modules
• Cross-validation benchmarking framework
• Stability analysis metrics
• End-to-end evaluation pipelines
• Integration-ready backend architecture

My codebase was later integrated into the BioMark platform, providing researchers with a user-friendly interface to perform advanced biomarker discovery without deep programming knowledge.

📊 Key Insights

• Ensemble selection reduces method-specific bias
• Rank aggregation improves stability more than weight aggregation
• Deep learning did not consistently outperform classical ML
• Triple-omics integration provided the highest validation rates
• Small, well-selected biomarker panels can outperform large noisy feature sets

📣 Impact

This work contributed to:

• A published software platform (BioMark)
• A large-scale benchmarking manuscript
• Reproducible, modular biomarker discovery pipelines
• Practical guidance for researchers working with multi-omics data

🚀 Interested in Robust Biomarker Discovery?

If you’re working with:

• Multi-omics datasets
• Disease classification tasks
• Translational biomarker research
• Feature instability problems

I can help design and benchmark reproducible, ensemble-based biomarker pipelines tailored to your dataset.