How do PCR duplicates arise and why is it important to remove them for NGS analysis?

Question

I am trying to understand PCR duplicates in NGS analyses (actually whole-genome). I searched, and the best answer I found is in this blog.

However I don't understand if I understood how PCR duplicates arise correctly because I cannot see the problem of having them in the downstream analysis - aside from computational problems, i.e. unnecessary redundancy.

If I understood correctly, PCR duplicates arise during library preparation when PCR amplifies the fragments with adapters. In this case, if you have duplicate fragments, you'll end up amplifying some of them twice or more.

However, when you are doing the mapping, they should fall on the same region, most likely decreasing the quality of the mapping (since they increase the consensus on a specific sequence, which could have been subject to sequencing errors). But aside from that, you should have the same mapping with respect to the mapping where duplicates are removed, although with decreased quality.

Is the quality problem the actual reason for removing PCR duplicates or is there anything that I'm missing?

Answer 1

In any scenario where depth of coverage is an important factor, PCR duplicates erroneously inflate the coverage and, if not removed, can give the illusion of high confidence when it is not really there.

For example, consider the following hypothetical scenario.

                                         *
TTTCATACTAACTAGCCTGCGGTCTGTGTTTCCCGACTTCTGAGTCATGGGGTTTCAATGCCTATAGATTC
               ..........................C.
                ............................
                      .............T..............
                       ..................C.........
                           ............................
                               ............................
                                   ............................
                                    .....C......................
                                       ..C.........................

The marked position has a coverage of 9 (9 overlapping reads), 4 of which suggest the presence of an alternate allele. Since these 4 reads map to different positions, they are independent observations supporting the same putative single nucleotide variant (SNV). The T in the third read can safely be ignored as a sequencing error, since it only appears once.

Now consider the following hypothetical scenario.

                                         *
TTTCATACTAACTAGCCTGCGGTCTGTGTTTCCCGACTTCTGAGTCATGGGGTTTCAATGCCTATAGATTC
                ............................
                      .............T..............
                       ..................C.........
                       ..................C.........
                       ..................C.........
                       ..................C.........
                               ............................
                               ............................
                               ............................

In this case, the marked position also has a coverage of 9, but 7 of those reads appear to be PCR duplicates. If we remove the duplicates, the marked position has only a coverage of 4, and only one C at the marked position, which is not enough confidence to call the SNV.

Answer 2

PCR polymerases introduce errors. When an error arises in the first few cycles of amplifications, it will appear in a reasonably high fraction of DNA fragments in the library. After sequencing, you may see the same error occur to multiple reads. If you remove PCR duplicates when calling variants, all errors are reduced down to one read. For high-coverage data, you won't call the error as a false variant. However, if you keep PCR duplicates, the multiple occurrences of the error may make the error look like a real variant. This is particularly bad for calling SNVs from impure tumor samples.