Improvement of variables interpretability in kernel PCA.

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Kernel methods are powerful for integrating and analyzing high-throughput data.
• Kernel PCA offers a nonlinear approach to principal component analysis for biological data.
• Existing methods lack effective feature importance interpretation for kernel PCA.

Purpose of the Study:

• To propose a novel methodology for data-driven feature importance using kernel PCA.
• To address the interpretability challenge of kernel PCA in high-dimensional datasets.

Main Methods:

• Kernel PCA Interpretable Gradient (KPCA-IG) was developed.
• The method relies on linear algebra calculations for computational efficiency.
• KPCA-IG was evaluated on benchmark datasets and a Hepatocellular carcinoma dataset.

Main Results:

• KPCA-IG achieved equal or greater accuracy compared to existing methods.
• The method demonstrated high computational efficiency.
• Biological validation confirmed the appropriateness of the selected features, identifying potential biomarkers.

Conclusions:

• KPCA-IG provides a necessary tool for interpreting kernel PCA in high-throughput data.
• The methodology effectively selects influential variables, aiding in biomarker discovery.
• This approach can potentially uncover new biological and medical biomarkers.