Spikes in error rate by cycle and tile

[anders]

We consistently get individual cycles with error rates 5-20% above the 'baseline' error rate followed by cycles with normal low error rates in the per tile error plots. Below is a representative, "normal" error plot on our GAI machine. Are we alone with this problem?

These error 'spikes' occur in all 8 runs we've done (phiX controls, human genomic samples, small RNAs), in all lanes and in a majority of the tiles. They are more common in later cycles and become more common from lane 1 towards lane 8 and seem to be more common towards the short ends of the flowcell. Tiles imaged in succession do not consistently show spikes in the same cycles. The type of error causing the spikes is that As are read as Cs. Since spikes are randomly distributed, the intensity plot for each lane looks normal (both level and decline).

I have contacted Illumina repeatedly about this issue but they cannot seem to do anything about it. Needless to say I would be very happy if anyone could come up with an idea about the cause of this problem it can be fixed.

[new300]

One of the things we learnt at Sanger was that it's important to look at the images to diagnose these kinds of problem. If you can post a image from a bad tile/cycle and a good tile/cycle it would help figure out what's going on. Spikes like these can be down to a bunch of things:

Focusing issues
Bubbles
Filter jams
etc.

[Chipper]

No idea what is causing it, but A>C errors are known to be the most common in Solexa sequencing (http://www.pubmedcentral.nih.gov/art...?artid=2532726). Perhaps the TileQC R package can be of help to identify the cause?

[anders]

Unfortunately the images are too large to post and if I resize or compress them nothing of value can be seen.

I have spent many hours looking at the images of the cycles with error spikes and comparing them to the images in the preceding and following good cycles on the same tile. There are no apparent artefacts (like bubbles, shades, etc), however, the A image usually seems to be slightly less intense and slightly out of focus compared to A images of surrounding normal cycles, but there is no big difference. This is reflected in the A intensities calculated by firecrest which are about 10-25 % lower on average than the A intensities in the preceding and following cycles. The c,t and g images are normal and so are their average intensities. Curiously, the Bustard prb values of these A>C miscalls would lead you to believe that they are good calls.

[anders]

I haven't tried out tileQC yet, but I will. I forgot to mention that our filter wheel broke a few weeks ago. The filter wheel has now been replaced and we have performed one run with the new wheel. I had hoped that this would solve our problem but the new run still showed lots of the same spikes in the error plots. I guess the most likely cause is focusing. But why would it only affect some apparently random cycles/tiles and not others?