Partial least squares Cox regression for genome-wide data.

Area of Science:

• Genomics
• Biostatistics
• Bioinformatics

Background:

• Survival prediction from high-dimensional genomic data often uses the Cox proportional hazards model with dimension reduction techniques like partial least squares regression (PLS).
• Existing methods for applying PLS to the Cox model, such as Park et al.'s approach, have limitations in estimation and flexibility.

Purpose of the Study:

• To propose a modified Cox PLS method that separates baseline hazard and gene effect estimation.
• To enhance survival prediction accuracy and computational efficiency for high-dimensional genomic data.
• To enable the incorporation of non-genomic variables and improve survival probability estimation for new patients.

Main Methods:

• A modified Cox PLS approach is proposed, separating baseline hazard and gene effect estimation into distinct steps.
• The method allows for the integration of lower-dimensional non-genomic variables (e.g., disease grade, tumor thickness).
• A supervised version is introduced, incorporating an initial gene selection step based on Cox scores.

Main Results:

• The proposed method offers advantages over existing Cox PLS approaches, including improved survival probability estimation for new patients and reduced memory requirements.
• Simulations demonstrate that both unsupervised and supervised versions of the proposed Cox PLS method outperform other existing Cox PLS methods.
• The supervised Cox PLS method, with initial gene selection, shows superior performance in simulations.

Conclusions:

• The modified Cox PLS method provides a more flexible and efficient approach for survival prediction using high-dimensional genomic data.
• The proposed method enhances the utility of Cox PLS by allowing for the inclusion of clinical variables and improving predictive capabilities.
• The supervised Cox PLS approach offers a promising strategy for robust survival prediction in genomic studies.