Automated multigroup outlier identification in molecular high-throughput data using bagplots and gemplots

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Molecular high-throughput data analyses often lack robustness, making results sensitive to individual data points.
• Standard univariate outlier detection methods are unsuitable for high-dimensional data (thousands of features in small samples).
• Current automated outlier detection methods are limited to single experimental groups, hindering comparative analyses.

Purpose of the Study:

• To develop and present automated methods for identifying outliers in high-dimensional molecular data.
• To extend existing visualization techniques for more effective outlier detection across multiple experimental groups.

Main Methods:

• Utilized bagplots, a 2D extension of boxplots, to detect outliers in the subspace of the first two principal components.
• Introduced gemplots, a 3D extension of boxplots and bagplots, for outlier detection in the subspace of the first three principal components.
• Employed convex hulls to define regular observations, with points outside hulls identified as outliers, calculated separately for each experimental group.

Main Results:

• Demonstrated the effectiveness of bagplots and gemplots on artificial and real gene expression datasets.
• Showcased the ability to display multiple experimental groups within the same principal component subspace for comparative outlier analysis.
• Provided an R package implementation ('gemPlot') for the gemplot visualization.

Conclusions:

• Bagplots and gemplots in principal component subspaces offer automated and objective outlier identification for high-dimensional molecular data.
• These methods facilitate the visualization and detection of outliers across multiple experimental groups simultaneously.
• The approach enhances quality control for molecular high-throughput experiments, leading to more reliable downstream analyses.