Automated calibration for stability selection in penalised regression and graphical models

Area of Science:

• High-dimensional statistics
• Bioinformatics
• Genomics and epigenomics

Background:

• Feature selection is crucial for understanding complex biological data.
• High-dimensional data presents challenges for traditional statistical methods.
• Stability selection offers a robust framework for feature identification.

Purpose of the Study:

• To develop an automated calibration procedure for stability selection.
• To accommodate a priori-known block structures in multi-omic data.
• To improve feature selection accuracy in penalized regression and graphical models.

Main Methods:

• Introduced an automated calibration procedure maximizing an in-house stability score.
• Incorporated a priori-known block structures for multi-omic data integration.
• Applied to Least Absolute Shrinkage Selection Operator (LASSO) penalized regression and graphical models.

Main Results:

• The proposed method outperformed existing non-stability-based and original stability selection approaches.
• Application to multi-block graphical LASSO on epigenetic and transcriptomic data.
• Identified a novel cross-omic role of LRRN3 in the biological response to smoking.

Conclusions:

• The automated calibration procedure enhances stability selection for high-dimensional and multi-omic data.
• The method provides a robust approach for identifying relevant features in complex biological systems.
• The findings highlight LRRN3's significant role in smoking-related biological responses.