Deviance residuals-based sparse PLS and sparse kernel PLS regression for censored data.

Area of Science:

• Bioinformatics and Computational Biology
• Statistical Genetics
• Biostatistics

Background:

• Relating gene profiles to subject survival or cancer recurrence is a key challenge.
• High-dimensional data (e.g., transcriptomic, SNP profiles) require robust biomarker discovery methods.
• Traditional Cox proportional hazards models face limitations with high-dimensional, collinear, or incomplete data.

Purpose of the Study:

• To develop novel algorithms for variable selection and collinearity handling in high-dimensional survival analysis.
• To introduce sparse Partial Least Squares regression (sPLSDR) and its non-linear kernel version (DKsPLSDR) for Cox's model.
• To evaluate the predictive performance of sPLSDR and DKsPLSDR against existing state-of-the-art methods.

Main Methods:

• Development of two new algorithms: sPLSDR and DKsPLSDR, based on sparse PLS regression using deviance residuals.
• Comparison of prediction performance using simulated and real benchmark datasets.
• Utilized the R-package plsRcox, available on CRAN, for implementation.

Main Results:

• sPLSDR and DKsPLSDR demonstrated favorable performance in terms of computational time, prediction accuracy, and selectivity.
• These methods compare favorably against other state-of-the-art algorithms on benchmark datasets.
• The proposed methods offer additional benefits like biplots and handling of missing data within the PLS regression framework.

Conclusions:

• sPLSDR and DKsPLSDR are valuable additions to the statistical toolbox for analyzing high-dimensional survival data.
• They effectively address challenges posed by high dimensionality and low sample size in Cox's model.
• These methods provide improved estimation and prediction capabilities for biomarker discovery and survival analysis.