Congratulations on your interesting finding!

I will admit to only having limited knowledge in this area. I have some concern with allowing reads that map to multiple regions. Lets say we had two regions A and B, and that both have identical sequence. If we do alignment and reads are mapped randomly, these areas will get reads in a 50-50 ratio. Suppose that in the original data set only A has a TF binding site and B does not at all. In that case, we might create a signal for a false binding site at B.

One solution is to attribute duplicate reads to A and B in a ratio consistent with the unique reads mapped to both. This means they won't be under or over-populated with reads relative to their positions in the genome, but it still has the problem that one might be imposing a pattern that might not be there.

Any thoughts?

To be honest with you this is always one of those things that never ceased to irk me - people simply ignoring repetitive regions in the genome when producing large scale reports on the potential functionality of genome. Glad to see papers that point this out.

Congrats on this manuscript.

Didn't you use any maximum number of multiple hits? The output bam file must have been huge.

By default we always map multi-reads. However, one should take the following considerations into account:

• Is is important to use and index that contains all sequences, even if they are not properly placed on the assembly. For human and mouse,
  those are the _random chromosomes. If all sequences are not included, the random placement of multi-reads will be biased towards those in
  the assembly, thus creating artificial enrichments. Because some reads will be considered unique by the assembler if these extra regions are not included in the index it is a general good policy to always include those regions.

• Removal of duplicates its complicated. We don't do it. If you remove duplicates you are biasing your results towards repetitive regions.
  Imagine a read that maps to two distinct positions an a read that is unique. Now, assume that these two reads are duplicated. By random
  placement of the multi-read, each of the repetitive regions will get one copy. For the unique read, the two duplicates will map to the
  same position. A duplicate removal algorithm will only detect the
  unique read as duplicate because they are based on position. On the
  other hand, if reads having identical sequences are removed before
  mapping, repetitive regions have a higher chance of producing
  identical reads. A duplicate removal of this sort will target repetitive regions preferentially, specially if they are highly repetitive. To reduce the duplication rates we use low PCR cycles for the sample preparation.

• Some tools, specially peaks callers, use the so called 'effective genome size', which is the size of the genome after you remove duplicates. If you are planing to do peak calling you need to upgrade these values and don't use the default ones. Furthermore, it is important to tell the peak caller not to remove duplicates.

• When looking at the mapping results we always keep an eye on the mappability track to detect spurious findings. For mouse and human they are easily available from USCS. For other species they are more difficult to obtain. In our case the authors of the

Piero Carninci pointed me to an interesting publication he co-authored. I think, it is of interested for this post, since it highlights the immense importance of repeat regions.

Carrieri, C., Cimatti, L., Biagioli, M., Beugnet, A., Zucchelli, S., Fedele, S., ... & Gustincich, S. (2012). Long non-coding antisense RNA controls Uchl1 translation through an embedded SINEB2 repeat. Nature, 491(7424), 454-457.

Their message: SINE elements have so many other important aspects. They identified a role of SINEB2 elements in the stimulation of protein translation.