Thoughts on ChIP-seq

In this page I will start sharing some answers I have given in emails I get from C. elegans community on ChIP-seq.

Should I do spike-in ChIP-seq to determine changes in protein binding?

My answer to this question is maybe. I know it is frustrating that a powerful genome-wide mapping approach like ChIP-seq is not fully quantitative. Here is my personal experience on this.

Based on comparing ChIP-seq biological replicates, we found significant variability in signal to noise ratios between replicates. Over the years, many bioinformatics labs came up with strategies to estimate background to correct for these differences. As these tools came out, we gave them a try. While they do help a bit, their effect on variation was marginal. That is, if you have a large variability between your reps, bioinformatics cannot save you.

A quick search of literature will tell you that much of the ChIP-seq variation comes from how you prepare chromatin and how well the antibodies work. This has been our experience as well. Crosslinking and sonication are particularly important, and there is batch effect and differences between data sets generated by different people. As a result, I am extremely careful in using ChIP-seq for comparing binding between wild type and mutant.

1- When I am comfortable: I am comfortable when comparing relative changes in binding. This is where working on X chromosomes becomes handy. We study dosage compensation complex (DCC) binding and DCC-associated changes on the X chromosomes. Autosomes serve as an internal control for each data set. By comparing relative X signal changes to autosomes, we can confidently talk about what are the X-specific effects of any manipulation we are making.

Similarly, in other studies relative binding changes at different classes of binding sites helps a lot. Let’s say you have a protein A that binds to promoters and enhancers, and its binding to these sites are mediated by different mechanisms. Let’s say protein B recruits protein A to promoters. In a protein B mutant, you would expect signal at the promoters to reduce and signal at enhancers not to change. So, a ChIP-seq experiment would be excellent to see the effect of protein B on protein A binding.

2- When I am not comfortable: Let’s think of another scenario. Let’s say that protein C controls protein A binding to DNA generally, thus protein C mutation reduces protein A binding everywhere. Here, ChIP-seq may be helpful, depending on other considerations, such as genomic distribution of protein A (wide/focused) and the effect size of the protein C mutation (small/large effect on binding). Let’s say, protein C mutation completely eliminates protein A binding. By doing enough replicates, you can convince yourself and your reviewers that protein A does not bind. But if the effect is small, let’s say ~20%, it would be hard to convince anyone, because this is certainly comfortably within the differences between biological ChIP replicates.

So, is there a solution? We used spike-in normalization of ChIP-seq signal to look at absolute changes in binding (Kramer et al. 2015). Others developed elaborate spike-in strategies, particularly for histone modifications to quantify ChIP-seq data. I am not yet convinced if these strategies are worth the laborious molecular biology they need. The reason is that there is no good way to “scale” ChIP-seq signal to spike in signal. That is, enrichment is certainly nonlinear and may not follow a consistent scale between different protein targets. So, when would I use spike in normalization again? I would use it if I am interested in gross-fold changes for a protein that is widely distributed.

For example, we used spike-in for a histone modification H4K20me1, which is widely distributed across the genome. Instead of trying to “correct ChIP-seq scores” (which I don’t think we could), we wanted to know how a particular mutant affected the overall abundance of the modification on X and autosomes (Kramer et al. 2015). We found that H4K20me1 went up ~5 fold on X and ~10 fold on autosomes. These numbers made sense based on what we knew about H4K20me1 enrichment on the X, and the estimated abundance of H4K20 methylation containing nucleosomes in the genome measured by mass spectrometry. For another project, we did not use spike in. Here, we wanted to know how histone modifications on the X was different from autosomes. We expected small changes (X chromosome repression for dosage compensation is ~2-fold) and could use autosomes as controls (Street et al. 2019).

In summary, if you are starting a project where you want to measure binding changes of a protein in two conditions, I encourage you to be careful about using ChIP-seq. Given considerations above, you can certainly design an experimental strategy that fits your goals. If your goals require careful quantification tough, I also encourage you to supplement your ChIP-seq experiments with quantitative ChIP-seq (ChIP-qPCR) at a few single loci, doing multiple replicates that allow you to check %Input DNA pulled down in your ChIPs. This is not a definitive solution that solves the biological variation problem, but it eliminates /reduces the sonication/PCR/total read normalization that occur in ChIP-seq, which are the major reasons that make ChIP-seq less quantitative.