Network-constrained group lasso for high-dimensional multinomial classification with application to cancer subtype...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Classic multinomial logit models struggle with high-dimensional genomic data due to limited predictors and ignoring variable relationships.
• Genomic features like gene expression are interconnected via biological networks, crucial for accurate classification and interpretation.

Purpose of the Study:

• To propose a novel multinomial logit model that addresses high dimensionality and incorporates underlying network information for genomic data analysis.
• To enhance classification performance and biological interpretability in genomic studies.

Main Methods:

• Developed a network-constrained multinomial logit model using group lasso for sparsity and network constraints for coefficient smoothness.
• Implemented a proximal gradient algorithm for efficient optimization of the non-smooth objective function.
• Compared the proposed model against traditional models using simulations and real TCGA gene expression data for cancer subtype prediction.

Main Results:

• The proposed network-constrained model demonstrated superior performance compared to traditional models lacking prior structure information.
• The model effectively handled high-dimensional genomic data and improved cancer subtype prediction accuracy.
• Network information significantly enhanced both classification performance and biological interpretability.

Conclusions:

• The network-constrained multinomial logit model offers a powerful approach for analyzing high-dimensional genomic data.
• Integrating biological network information improves the performance and interpretability of genomic data analysis.
• This method shows significant promise for applications like cancer subtype prediction.