Prediction of tumour pathological subtype from genomic profile using sparse logistic regression with random effects

Area of Science:

• Genomics
• Biostatistics
• Computational Biology

Background:

• Lung cancer research increasingly utilizes genomic data for understanding tumor heterogeneity.
• High dimensionality and inter-region correlations in genomic data pose challenges for predictive modeling.
• Sparse logistic regression offers a framework for feature selection in high-dimensional settings.

Purpose of the Study:

• To apply sparse logistic regression models for predicting tumor pathological subtypes using lung cancer genomic information.
• To introduce and evaluate novel penalized hierarchical likelihood (HL) methods for genomic data analysis.
• To compare the performance of new methods against established techniques like lasso and elastic net.

Main Methods:

• Utilized sparse logistic regression models to handle high-dimensional and correlated genomic data.
• Developed a hierarchical likelihood (HL) method assuming normal random effects and gamma-distributed variance.
• Extended the HL penalty to create 'HLnet', analogous to elastic net, incorporating ridge penalties.

Main Results:

• The HL penalty demonstrated greater sparsity than the lasso penalty with similar prediction performance.
• Both HLnet and elastic net penalties achieved the best prediction performance on real-world lung cancer data.
• The proposed methods effectively handle the complexities of genomic data for subtype prediction.

Conclusions:

• Sparse logistic regression, particularly with HL and HLnet extensions, is a powerful tool for lung cancer subtype prediction from genomic data.
• HLnet and elastic net provide superior predictive accuracy in practical applications.
• These methods offer advancements in analyzing complex genomic datasets for precision oncology.