Normalization and Genome-Wide Comparisons with Chipseq

Question

[zbeddingfield3]

Hi all, and apologies if this has been asked before.

I have chipseq data (technical and biological replicates for 2 groups) and want to test whether my experimental group has a genome-wide increase in methylation compared to controls. Under the hypothesis, this group is predicted to have higher read counts overall.

I know that I would typically normalize using bamcompare before trying to compare chipseq files to eliminate genome-wide increased read count due to differences in chip quality. Is it possible to use chip data to assess whether read counts are up genome-wide for one group? If so, how can I normalize samples to remove read bias due to variable chipseq data quality without eliminating biological differences in total read count?

Thanks in advance

[Ankit]

Use "Diffbind". You can perform differential binding analysis using either edgeR or DESeq2 or both. It will perform normalization and will tell you up- or down-regulated (high or less enriched) sites.

For methylation part I don't know what you want to do ? Do you have methylation-seq data as well which you want to integrate with ChIP-seq? Please elaborate.

[zbeddingfield3]

Sorry for the confusion, I didn't run the original ChIP-seq so I'm super familiar with the process. The ChIP-seq used antibodies for H3K4M1/2 methylated sites. My concern was that by normalizing the read counts of the ChIP-seq between groups, if H3K4M1/2 was up uniformly across the genome in one group, we might not be able to see the difference between groups. Will normalizing the data using algorithms like CPM or RPKM result in group 1, which should have uniformly more H3K4M1/2 than group 2, appearing to have equal activity to group 2 by equalizing their read counts?

[Ankit]

Normalisation will alter the differences between two groups if there is a global up- or downregulation, otherwise No. You can check if there is a global defects by Visualization on IGV or some plots (eg. Scatter). The normalization you mentioned accounts for differences in reads counts between samples (due to difference in sequencing depth). If the defects are not global the assumptions for differential binding analysis will remain same as differential gene expression analysis and you can use tools like DESeq2/edgeR implemented in Diffbind.

[zbeddingfield3]

Ok, thank you! One more follow-up: are you aware of any methods for demonstrating that global up-regulation is a true effect, rather than an artifact of sequencing depth, by showing that the difference is maintained across technical/biological replicates? Or is this not possible due to the possibility that one group would naturally have better sequencing depth?

[Ankit]

I don't know any specific methods. But statistics might not work well at multiplicity correction if all your regions are changes in one direction. You might need to confirm this from experts as I have limited understanding of Statistical analysis.

I can think of some ways to conclude that the actual global upregulation are biological effect and not because of sequencing depth:

1. Check between replicates. If coverage remains similarly high among replicates from a group compare to replicates in other group then it might be possible you the global changes are real.
2. If you know some regions which should not changes (prior experience/literature support) then check those regions coverage among two groups.
3. Use spike-ins during sequencing /library prep and evaluate coverage of spike-ins.