Application of statistical machine learning in biomarker selection

Area of Science:

• Oncology
• Biostatistics
• Genomic Medicine

Background:

• Advanced urothelial cancer (aUC) treatment benefits from maintenance therapy, as shown by the JAVELIN Bladder 100 trial.
• Identifying biomarkers is crucial for precision medicine in aUC, but challenges exist due to high collinearity and low signal in genomic data.
• The JAVELIN Bladder 100 dataset's characteristics make it difficult to reliably identify biomarkers using standard variable selection methods.

Purpose of the Study:

• To evaluate the performance of various variable selection methods in discovering prognostic and predictive biomarkers for aUC patients.
• To compare penalized regression, random survival forests, and Bayesian variable selection approaches.
• To propose a modified Bayesian Information Criterion (BIC) thresholding rule for Bayesian methods.

Main Methods:

• A simulation study assessed popular variable selection techniques for high-dimensional data.
• Penalized regression models, random survival forests, and Bayesian variable selection were employed.
• The methods were applied to the JAVELIN Bladder 100 dataset to identify survival-associated biomarkers.

Main Results:

• Variable selection methods generally had low false discovery rates but struggled with high collinearity.
• Lasso-related methods identified potentially biologically relevant variables in the JAVELIN Bladder 100 data.
• Stochastic search variable selection and random survival forest showed limitations with the dataset's collinearity and low signal.

Conclusions:

• No single variable selection method is universally superior for biomarker discovery in high-dimensional, collinear aUC data.
• Further research is needed to develop novel variable selection methods for robust biomarker identification in this patient population.
• The study highlights the complexities of biomarker discovery in advanced urothelial cancer and informs future research directions.