Optimum input DNA sequencing depth

[captainentropy]

I've been told recently that I need to be generating at least 1X coverage of the genome for the input DNA sequencing.

It makes sense to do so but I'm wondering what other labs do. On a GAII that's about 5 lanes of sequencing assuming 20M reads per lane and 36bp reads.

[krobison]

For what are you using the data? The amount you need to generate is completely dependent on the application.

[captainentropy]

ChIP-seq. The input is for the normalization step.

[GKM]

I haven't seen anyone sequence that many reads from their input, 30-40M should be enough. Also, remember that your 1X will not at all be 1X because the input is far from a homogenous representation of the genome.

[captainentropy]

Right, but that was why it was suggested I need to make sure it's a real 1X, other regions would certainly be many more X for sure. That's the only real way to control for all the endogenous "peaks" of the genome. I think it's unrealistic to do 1X or more for each input but I want it right, not simply what's easiest. 40M reads is about 0.5X. I guess that'll be good enough for now.

[simonandrews]

Our experience is that 30 million reads of an input sample (40bp) is enough to use as a filter, but not enough to normalise against (assuming a normal sized eukaryotic genome).

The problem is that over any enriched region you'll have a huge discrepancy between the coverage of your enriched and your input measurement, and as such the input measurement will limit the accuracy of your corrected measure, and small changes to the input can have a big effect on the corrected value.

What you can do with this much data is to ignore ChIP peaks where you see a corresponding peak in the input sample. These are the things you really need to get rid of and you don't need a huge amount of data to find them. There's even a group (whose name escapes me for the moment, but I can look it up if you're interested) who have created a catalogue of genome regions where they see unusual enrichment in supposedly unbiased input samples - you can just filter against this.

I should also say that we now don't run input controls for every sample, since the biases you see are generally consistent across samples (as long as you're using the same protocol, species and aligner), so you can merge together the inputs you have and just use those against any new samples.