ChIP-seq duplicate reads/ Poisson distribution

[kathrin]

Hi everybody

I started analyzing my first ChIP-seq data set, it contains one ChIP-sample and one input sample. After mapping the reads to a reference genome by using Bowtie and additionally MAQs, around 70% of all reads were uniquely mapped to the reference, which should be a quiet good rate (I guess).
For the input sample ~20 mio reads were left, and for the ChIP-sample ~17 mio reads.
However, I found ~2 mio duplicated reads (matching the same chromosomal location) in the input sample and ~16 mio duplicated reads in the ChIP-sample, which might be due to amplification errors or library preparation.
In literature I read that the expected number of reads matching the same position and strand can be modeled by the Poisson distribution. Is this assumption also true for ChIP-samples, where we enrich specific chromosomal locations and get rid of those, where the TF doesn't bind? Wouldn't we expect to find more more duplicated reads in ChIP-samples than in the Input samples

To identify Peaks I used MACS that removes duplicated reads before calling the peaks. Does anyone know a good peak caller that parameters concerning duplicated reads can be adjusted by the user? I want to try to set a customized threshold for the number of duplicated reads depending on my duplicated read distribution and check the sequences of my peak regions. Fortunately the TFBS motif of my TF is already known, so I can verify my results.

It would be great to get some comments or ideas, as I am an absolute beginner in NGS analyses...

Besides that, thank you for the great forum, it's a great help

Thanks a lot in advance
Kathrin

[dawe]

Hello Kathrin,

After mapping the reads to a reference genome by using Bowtie and additionally MAQs, around 70% of all reads were uniquely mapped to the reference, which should be a quiet good rate (I guess).

You may also try other mappers (but this is another thread). I usually get ~80-90% of "good quality alignments" in ChIP-seq experiments.

In literature I read that the expected number of reads matching the same position and strand can be modeled by the Poisson distribution. Is this assumption also true for ChIP-samples, where we enrich specific chromosomal locations and get rid of those, where the TF doesn't bind? Wouldn't we expect to find more more duplicated reads in ChIP-samples than in the Input samples

Indeed ChIP-seq samples (but also Input samples) show reads distributed with a power law, which essentially means "rich gets richer". In other words, there are sequences that have higher chances to be enriched, and the more reads you have the more will match on those regions.
I see your point but consider also that genomic DNA (the Input) may contain some genomic features (open chromatin) that may be constitutively "enriched", plus a number of repeated sequences you may not find in the IP. Also, TF binding sites + background noise cover a slice of your genome which is probably "wide" enough to sparse your duplicates.
You may try do zap duplicates from your alignment using picard (assuming you have BAM files)

Does anyone know a good peak caller that parameters concerning duplicated reads can be adjusted by the user?

You may try FindPeaks4. It has a -duplicatefilter flag that can be omitted. Also, it offers a huge set of options to model your experiment.

HTH

d

[mattanswers]

I am used to getting only 30% unique reads when doing ChIP-Seq on Arabidopsis genome. The number goes up to around 80% when all alignments are counted.

I would be very interested if dawe could elaborate on the read distribution, because our Input DNA does not span the genome evenly and particularly shows enrichment in exons. We have been wondering why this would be so ?

[dawe]

Hi, check this paper: Zhang et al. Modeling ChIP sequencing in silico with applications. PLoS Comput Biol (2008) vol. 4 (8) pp. e1000158
Indeed the Input is not random nor is flat. There are regions (open chromatin, fragile sites) that may be preferentially enriched. One should try to use naked DNA to isolate sequence bias (I believe there's a paper on this... I can't get the reference right now).

d

[mattanswers]

dawe, thank you for the reference and alerting us to this.