Sparse Regression in Cancer Genomics: Comparing Variable Selection and Predictions in Real World Data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Evaluating gene interaction models in cancer genomics is difficult due to uncertain true distributions.
• Existing methods using synthetic or incomplete experimental data have limitations.
• A real-world data-driven approach is needed for robust model comparison.

Purpose of the Study:

• To develop and apply a novel benchmarking approach for genomic model inference algorithms using real-world data.
• To compare the performance of LASSO, elastic net, best-subset selection, L0L1, and L0L2 penalisation methods.
• To assess the efficacy of algorithmic preselection versus internal cross-validation for model selection.

Main Methods:

• Extracted five large genomic datasets (n=4000) from Gene Expression Omnibus.
• Trained 'gold-standard' regression models on data subspaces (n=4000, p=500).
• Trained penalised regression models on smaller samples (n=25, 75, 150) and validated against gold-standard models, assessing variable selection and prediction accuracy.

Main Results:

• L1L2 penalisation showed the highest cosine similarity for coefficient recovery.
• L0L2 penalisation explained the most variance in test responses and achieved the highest variable selection F1 score.
• Algorithmic preselection significantly outperformed internal cross-validation across all evaluated metrics.

Conclusions:

• The study presents a novel, data-driven approach for comparing model selection methods in cancer genomics.
• Benchmarking datasets are publicly available for future research.
• L0L2 penalisation is recommended for structural selection, and L1L2 penalisation for coefficient recovery in genomic data analysis.