Using the Kriging Correlation for unsupervised feature selection problems.

Area of Science:

• Computational Biology
• Data Science
• Bioinformatics

Background:

• High-dimensional data, common in fields like single-cell RNA sequencing (scRNA-seq), presents challenges for clustering analysis.
• Identifying the most informative features is crucial for accurate and efficient clustering.
• Existing feature selection methods may not be optimal for complex, high-dimensional biological datasets.

Purpose of the Study:

• To propose a novel metric, the KC Score, for quantifying feature importance in clustering.
• To develop and evaluate a feature selection strategy based on the KC Score for high-dimensional data.
• To assess the effectiveness of this strategy on real-world scRNA-seq datasets.

Main Methods:

• The KC Score is calculated by correlating original features with their low-dimensional reconstructions in a latent space.
• A feature selection strategy is implemented using the KC Score to identify the most relevant features for clustering.
• The proposed strategy is tested on four diverse scRNA-seq datasets, comparing its performance against using all features.

Main Results:

• The KC Score-based feature selection strategy effectively identifies important features for clustering.
• In three out of four datasets, the strategy selected less than 5% of the original features.
• The reduced feature sets achieved clustering performance comparable to or better than using the complete dataset.

Conclusions:

• The KC Score provides a robust method for evaluating feature importance in high-dimensional clustering.
• Feature selection using the KC Score significantly enhances clustering efficiency and performance, especially for scRNA-seq data.
• This approach offers a computationally efficient and effective solution for analyzing complex biological datasets.