Analysis of Controls in ChIP-seq

The chromatin immunoprecipitation followed by high throughput sequencing (ChIP-seq) method, initially introduced a decade ago, is widely used by the scientific community to detect protein/DNA binding and histone modifications across the genome in various cell lines. Every experiment is prone to noise and bias, and ChIP-seq experiments are no exception. To alleviate bias, incorporation of control datasets in ChIP-seq analysis is an essential step. The controls are used to detect background signal, whilst the ChIP-seq experiment captures the true binding or histone modification signal. However, a recurrent issue is the existence of noise and bias in the controls themselves, as well as different types of bias in ChIP-seq experiments. Thus, depending on which controls are used, peak calling can produce different results (i.e., binding site positions) for the same ChIP-seq experiment. Consequently, generating "smart" controls, which model the non-signal effect for a specific ChIP-seq experiment, could enhance contrast and thus increase the reliability and reproducibility of the results. Our analysis aims to improve our understanding of ChIP-seq controls and their biases. We use unsupervised clustering and dimensionality reduction techniques to compare 160 controls for the K562 cell line in the ENCODE project, finding distincting groupings of controls which correlate to experimental characteristics. To customize a control for each ChIP-seq experiment, we use LASSO regression to fit a sparse set of controls to each of 500 ChIP-seq experiments (again, from ENCODE data for the K562 cell line). We look at how many controls are selected, which controls are used per ChIP-seq experiment, and how they are related to the different ChIP-seq experiment characteristics. Perhaps most surprisingly, we find that the LASSO models are not particularly sparse, often including half of the possible controls to model any given ChIP-seq. Cross-validation as well as testing with smaller sets of candidate controls proves that such large numbers of controls are beneficial for modeling ChIP-seq background distributions. We also observe clusters of ChIP-seq experiments that tend to rely on clusters of controls, and we look at the experimental characteristics that tend to cause a given control to be useful in modeling the background of a given ChIP-seq experiment. Through these analyses, we attempt to answer largely-unstudied questions regarding how much control data and of what types are useful in ChIP-seq analysis, and how suitable controls can be matched to ChIP-seq datasets.