Feature selection for kernel methods in systems biology.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• High-throughput biotechnologies generate large, heterogeneous multi-omics datasets.
• Kernel methods offer powerful tools for analyzing diverse datasets from the same individuals.
• Current kernel methods often lack interpretability due to information loss during kernel embedding.

Purpose of the Study:

• To propose novel feature selection methods within the kernel framework.
• To extend feature selection to unsupervised learning and kernel output learning tasks.
• To enhance the interpretability of kernel-based multi-omics data analysis.

Main Methods:

• Development of novel feature selection methods adapted to the kernel framework.
• Formulation as a non-convex optimization problem with an L1 penalty.
• Solution using a proximal gradient descent approach.

Main Results:

• Demonstrated good performance in selecting relevant and less redundant features on systems biology datasets.
• Successfully identified key governmental measures influencing COVID-19 reproduction number time series.
• The method offers improved interpretability compared to existing alternatives.

Conclusions:

• The proposed feature selection methods effectively address challenges in multi-omics data integration and interpretability.
• The approach is versatile, applicable to both biological systems and epidemiological data.
• The methods are implemented in the R package mixKernel, available on CRAN.