VSOLassoBag: a variable-selection oriented LASSO bagging algorithm for biomarker discovery in omic-based...

Area of Science:

• Biomedical research
• Translational research
• Bioinformatics

Background:

• High-dimensional omics data presents challenges for biomarker discovery.
• Least Absolute Shrinkage and Selection Operator (LASSO) is popular but can overfit with high dimensions and low sample sizes.
• Existing methods may yield an excessive number of predictive variables.

Purpose of the Study:

• To introduce VSOLassoBag, an ensemble learning approach for robust biomarker selection.
• To improve the efficiency and stability of variable selection from omics data.
• To provide a reliable alternative for identifying high-confidence biomarkers.

Main Methods:

• VSOLassoBag integrates an ensemble learning strategy with LASSO.
• A bagging strategy is employed with parametric or inflection point search methods.
• Variables from multiple LASSO models are integrated and voted to select optimal candidates.

Main Results:

• VSOLassoBag effectively identifies biomarkers for binary classification and prognosis prediction.
• The algorithm demonstrates robust performance on both simulated and real-world omics datasets.
• Compared to existing algorithms, VSOLassoBag achieves comparable performance with fewer selected features.

Conclusions:

• VSOLassoBag offers a powerful strategy for selecting reliable biomarkers from high-dimensional omics data.
• The R package implementation includes multithreading for efficient computation.
• This method aids in overcoming overfitting issues common with LASSO in omics research.