Nearest-neighbor Projected-Distance Regression (NPDR) for detecting network interactions with adjustments for...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• High-dimensional data analysis in genomics and neuroimaging requires robust feature selection methods.
• Existing machine learning techniques struggle with complex interaction detection, false discovery control, and covariate adjustment.

Purpose of the Study:

• To develop a novel feature selection technique, Nearest-neighbor Projected-Distance Regression (NPDR), for high-dimensional data.
• To address limitations in detecting interaction-network effects and controlling for covariates.

Main Methods:

• NPDR calculates predictor importance using generalized linear model regression on projected nearest-neighbor distances.
• The method handles diverse data types (dichotomous/continuous outcomes and predictors).
• NPDR incorporates covariate adjustment, statistical inference, and penalized regression.

Main Results:

• Simulations show NPDR outperforms standard methods (Relief, Random Forest) in precision-recall.
• NPDR effectively removes confounding effects in RNA-Seq data for major depressive disorder (MDD) studies.
• Application to eQTL data identifies interacting variants regulating MDD-associated transcripts.

Conclusions:

• NPDR is a powerful tool for feature selection in high-dimensional data, particularly in genetic and neuroimaging research.
• The method offers superior performance in identifying interactions and controlling for confounding factors.
• NPDR demonstrates utility for Genome-Wide Association Studies (GWAS) and analysis of continuous outcomes.