High dimensional classification with combined adaptive sparse PLS and logistic regression.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• High-dimensional genomic data presents classification challenges, often leading to over-optimistic predictions.
• Existing methods combining compression and variable selection lack computational stability and convergence.
• Developing robust classification methodologies is crucial for accurate genomic data analysis.

Purpose of the Study:

• To propose a computationally stable and convergent approach for classification in high-dimensional genomic data.
• To introduce sparse Partial Least Squares (sparse PLS) as a robust framework for genomic classification.
• To enhance data visualization and interpretation through effective variable selection.

Main Methods:

• Developed a novel sparse PLS solution using proximal operators for univariate responses.
• Introduced logit-SPLS, an adaptive sparse PLS method combining logistic regression and iterative optimization for stability.
• Validated the approach on synthetic and experimental datasets, including gene expression data.

Main Results:

• Demonstrated the critical importance of convergence and stability, especially with cross-validation.
• Successfully applied the method to predict breast cancer relapse using gene expression data.
• Developed a multicategorical version for predicting cell-types from single-cell expression data.

Conclusions:

• The proposed sparse PLS-based approach offers a computationally stable and convergent solution for high-dimensional genomic data classification.
• The logit-SPLS method effectively addresses limitations of previous techniques, improving prediction and interpretation.
• The plsgenomics R-package provides accessible implementation for researchers.