Adjusting for Principal Components of Molecular Phenotypes Induces Replicating False Positives

Area of Science:

• Genomics
• Bioinformatics
• Statistical Genetics

Background:

• High-throughput molecular phenotyping generates complex, structured datasets essential for understanding cellular processes and disease.
• These datasets are often confounded, leading to false positives and reduced statistical power in association analyses.
• Principal component analysis (PCA) is widely used to estimate and correct for these confounders in genomic studies.

Purpose of the Study:

• To investigate the impact of PCA-based confounder correction methods on association tests in high-throughput molecular data.
• To theoretically and empirically assess the bias introduced by these correction approaches.

Main Methods:

• Theoretical derivation of an analytic bias approximation for PCA-based confounder correction.
• Realistic simulations to assess the performance of various confounder correction methods.
• Analysis of bias dependence on covariate and confounder sparsity, effect sizes, and their correlation.

Main Results:

• PCA-based confounder correction methods introduce a persistent bias, even in large sample sizes and out-of-sample validation.
• Perturbing basic parameters in simulations can lead to significant false positive rate (FPR) inflation.
• Standard two-step methods exhibit substantial FPR inflation (up to [Formula: see text]-fold) when the covariate and confounder have specific properties.

Conclusions:

• Confounder correction methods, particularly PCA-based approaches, can induce significant bias and inflate FPR in genomic association studies.
• The findings suggest that numerous false discoveries may have been made and replicated in differential expression analyses.
• This study informs best practices for confounder correction in high-throughput molecular data analysis to improve study reliability.