Overcoming the inadaptability of sparse group lasso for data with various group structures by stacking

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Gene expression analysis is crucial for cancer classification and prediction.
• High-dimensional genomic data presents challenges for traditional prediction and feature selection models.
• Existing regularization methods often overlook biological group structures, leading to suboptimal predictive performance.

Purpose of the Study:

• To develop a robust classifier, Stacked SGL, for accurate cancer type prediction and stable gene selection.
• To address the limitations of standard sparse group lasso by integrating diverse ratio combinations through stacking.
• To enhance feature selection capabilities while maintaining high predictive performance.

Main Methods:

• Development of the Stacked SGL classifier utilizing sparse group lasso penalty and stacked generalization.
• Incorporation of a mixing parameter in sparse group lasso to balance feature group sparsity and within-group sparsity.
• Implementation of post hoc feature selection to refine gene identification after stacking.

Main Results:

• Stacked SGL demonstrates competitive and stable classification performance across various datasets.
• The method achieves a lower false discovery rate in feature selection compared to other regularization techniques.
• Experiments on public cancer datasets show improved accuracy and identification of potent discriminatory genes, including potential mutation drivers for thyroid carcinoma.

Conclusions:

• Stacked SGL offers a superior approach for cancer classification and gene selection in high-dimensional genomic data.
• The stacking strategy effectively overcomes the inadaptability of sparse group lasso for certain datasets.
• The identified genes hold potential for further research into cancer mechanisms and therapeutic targets.