Optimizing hybrid ensemble feature selection strategies for transcriptomic biomarker discovery in complex diseases.

Area of Science:

• Biomedical research
• Computational biology
• Genomics

Background:

• Omic data, particularly transcriptomics, is crucial for understanding complex diseases.
• Identifying transcriptomic biomarkers traditionally relies on feature selection methods.
• Hybrid ensemble feature selection (HEFS) offers robustness but requires careful design choices.

Purpose of the Study:

• To extensively analyze HEFS scenarios for identifying cancer biomarkers from transcriptomic data.
• To evaluate different feature reduction and resampling strategies within HEFS.
• To highlight critical design considerations for HEFS in biomarker discovery.

Main Methods:

• Investigated four HEFS scenarios using transcriptomic data from Stage IV colorectal, Stage I kidney and lung, and Stage III endometrial cancers.
• Employed two feature reduction methods: differentially expressed genes and variance.
• Utilized two resampling strategies: distribution-balanced stratified and random stratified repeated holdout.
• Performed feature selection using an aggregation of thousands of wrapped machine learning models.

Main Results:

• HEFS approaches demonstrate advantages in identifying complex disease biomarkers.
• Selected features exhibit generalizable and stable results against data and functional perturbations.
• The study provides insights into optimal HEFS design for biomarker discovery.

Conclusions:

• HEFS is a powerful strategy for robust transcriptomic biomarker identification in complex diseases.
• Careful consideration of feature reduction and resampling methods is critical for HEFS success.
• This research aids in the design of more effective HEFS strategies for biomedical applications.