Feature selection in finite mixture of sparse normal linear models in high-dimensional feature space

Area of Science:

• Genomics
• Statistical Modeling
• Bioinformatics

Background:

• Modern technology generates large, complex datasets, especially in genomics.
• Selecting relevant features from numerous variables is challenging with small sample sizes and multiple subpopulations.
• Accurate feature selection is crucial for reliable predictive modeling in high-dimensional biological data.

Purpose of the Study:

• To develop an effective feature selection method for finite mixture of sparse normal linear (FMSL) models.
• To address computational challenges and high false discovery rates in large feature spaces.
• To improve the accuracy of predictive models in genomics and related fields.

Main Methods:

• A two-stage procedure combining likelihood-based boosting and penalized likelihood methods.
• Likelihood-based boosting to reduce the dimensionality of candidate features.
• A novel scheme for initializing expectation-maximization estimation and an extended Bayesian information criterion for model selection.

Main Results:

• The proposed method successfully selects significant features while minimizing the inclusion of insignificant ones.
• Simulation studies demonstrate the procedure's effectiveness in feature selection.
• The method was applied to a real-world gene transcription regulation dataset.

Conclusions:

• The developed two-stage feature selection approach is effective for FMSL models in high-dimensional genomic data.
• The method overcomes computational hurdles and reduces false discovery rates.
• This approach enhances the reliability of predictive models in complex biological analyses.