A non-negative spike-and-slab lasso generalized linear stacking prediction modeling method for high-dimensional omics...

Area of Science:

• Genomics
• Biostatistics
• Computational Biology

Background:

• High-dimensional omics data are crucial for disease risk prediction.
• Existing sparse methods often rely on single models, leading to limitations.
• Prior biological knowledge can guide omics data analysis.

Purpose of the Study:

• To develop a novel stacking strategy for disease risk prediction using high-dimensional omics data.
• To improve prediction accuracy and generalization compared to single-model approaches.
• To leverage biological group structure information effectively.

Main Methods:

• Proposed a stacking strategy using a non-negative spike-and-slab Lasso (nsslasso) generalized linear model (GLM).
• Segmented omics data into sub-data based on biological knowledge.
• Trained sub-models on each segment and ensembled predictions using a super learner with nsslasso GLM.

Main Results:

• The nsslasso stacking method demonstrated superior prediction performance over single-model methods on simulated and real-world breast cancer data.
• Outperformed traditional stacking methods in handling redundant and identifying important sub-models.
• Achieved robustly superior prediction in high-noise environments.

Conclusions:

• The nsslasso method offers enhanced predictive accuracy, stability, and biological interpretability.
• It effectively identifies key biological group structures and potential new biomarkers.
• This approach advances the application of omics data in clinical and public health research.