I am seeing strange behavior when comparing alignments generated by bwa mem 0.7.12 to two different reference genomes: hg19 and hs37d5 (basically hg19 plus additional decoy sequences). We have DNA-seq data. I noticed that when using hg19, there are very few alignments to the MHC region on chromosome 6, and even fewer of these have nonzero mapping quality. When using hs37d5, there are dramatically more alignments to the region and these have mostly high mapping quality scores. I have not observed this phenomenon anywhere else I've looked in the genome. The behavior is robust to multiple different choices of BWA parameters. Can anyone explain why the inclusion of the additional 35Mb of decoy sequences in hs37d5 would drastically improve the number and quality of alignments to this region of chr6?