Weighted Sparse Partial Least Squares With Joint Sample and Feature Selection for Integrating Multi-Omics Data

Area of Science:

• Computational Biology
• Machine Learning
• Statistical Analysis

Background:

• Sparse Partial Least Squares (sPLS) is a dimensionality reduction technique for data fusion.
• Standard sPLS cannot identify latent subsets of samples or remove outliers.

Purpose of the Study:

• To develop a novel method for joint sample and feature selection in sPLS.
• To extend sPLS for identifying specific sample subsets and outlier removal.
• To adapt the method for multi-view data fusion.

Main Methods:

• Proposed an $\ell _\infty /\ell _{0}$-norm constrained weighted sparse PLS ($\ell _\infty /\ell _{0}$-wsPLS) for sample and feature selection.
• Proved the Kurdyka-Łojasiewicz property of the $\ell _\infty /\ell _{0}$-norm constraints for global convergence.
• Developed two multi-view wsPLS models and efficient iterative algorithms for multi-view data fusion.

Main Results:

• The proposed $\ell _\infty /\ell _{0}$-wsPLS method enables joint sample and feature selection.
• Globally convergent algorithms were developed for the proposed models.
• Numerical and biomedical data experiments demonstrated the efficiency of the multi-view wsPLS methods.

Conclusions:

• The novel $\ell _\infty /\ell _{0}$-wsPLS method effectively identifies sample subsets and outliers.
• The extended multi-view wsPLS models are efficient for multi-view data fusion.
• The developed algorithms ensure convergence and demonstrate practical applicability.