Improved quality metrics for association and reproducibility in chromatin accessibility data using mutual information

Area of Science:

• Genomics
• Epigenomics
• Computational Biology

Background:

• Correlation metrics are frequently used in genomics, but often disregard underlying data assumptions like normality and homoscedasticity.
• Assays for transposase-accessible chromatin via sequencing (ATAC-seq) data commonly exhibit non-normality and zero values, complicating standard correlation analyses.
• Few studies have benchmarked correlation statistics' performance on ATAC-seq data, especially concerning reproducibility and outlier effects.

Purpose of the Study:

• To computationally simulate ATAC-seq data to investigate the behavior of various correlation and association statistics.
• To evaluate the accuracy of these statistics under controlled conditions of reproducibility and data differences.
• To identify robust methods for quantifying relationships in chromatin accessibility assays.

Main Methods:

• Developed a computational simulation framework for ATAC-seq data generation.
• Assessed the performance of Pearson's R, Spearman's rho, Kendall's tau, Top-Down correlation, R-squared, Kendall's W, and normalized mutual information.
• Investigated the impact of removing co-zero regions on statistical estimates.
• Utilized a random forest model to compare predictive accuracy for ATAC-seq replicate relationships.

Main Results:

• Most tested statistics (Spearman's rho, Kendall's tau, Kendall's W) showed insensitivity to increasing differences between simulated ATAC-seq replicates.
• Removing co-zeros substantially improved correlation and association estimates.
• R-squared and normalized mutual information demonstrated superior performance, closely mirroring known shared loci after co-zero removal.
• Normalized mutual information best predicted ATAC-seq replicate relationships when analyzed with a random forest model.

Conclusions:

• Standard correlation and association measures may not accurately reflect relationships in epigenomics data, particularly ATAC-seq.
• Removing co-zeros is a crucial preprocessing step for improving the reliability of correlation and association analyses in ATAC-seq data.
• Normalized mutual information and R-squared offer more robust strategies for quantifying relationships in chromatin accessibility assays.