Selecting ChIP-seq Normalization Methods from the Perspective of their Technical Conditions

TL;DR

This paper examines the technical conditions affecting ChIP-seq normalization methods, highlighting their impact on differential binding analysis and proposing strategies for more robust peak identification.

Contribution

It identifies key technical conditions influencing ChIP-seq normalization and evaluates their effects through simulations and experiments, guiding better normalization choices.

Findings

High-confidence peaks are less sensitive to normalization method choice.

Approximately half of the peaks are differentially called across methods.

Understanding technical conditions improves differential binding analysis.

Abstract

Chromatin immunoprecipitation with high-throughput sequencing (ChIP-seq) provides insights into both the genomic location occupied by the protein of interest and the difference in DNA occupancy between experimental states. Given that ChIP-seq data is collected experimentally, an important step for determining regions with differential DNA occupancy between states is between-sample normalization. While between-sample normalization is crucial for downstream differential binding analysis, the technical conditions underlying between-sample normalization methods have yet to be examined for ChIP-seq. We identify three important technical conditions underlying ChIP-seq between-sample normalization methods: balanced differential DNA occupancy, equal total DNA occupancy, and equal background binding across states. To illustrate satisfying the selected normalization method's technical conditions for downstream differential binding analysis, we simulate ChIP-seq read count data where different combinations of the technical conditions are violated. We then externally verify our simulation results using experimental data. Based on our findings, we suggest that researchers use their understanding of the ChIP-seq experiment at hand to guide their choice of between-sample normalization method. Alternatively, researchers can use a high-confidence peakset, which is the intersection of the differentially bound peaksets obtained from using different between-sample normalization methods. In our two experimental analyses, roughly half of the called peaks were called as differentially bound for every normalization method. High-confidence peaks are less sensitive to choice of between-sample normalization method and could be a more robust basis for identifying genomic regions with differential DNA occupancy between experimental states when there is uncertainty about which technical conditions are satisfied.