This is an excellent question that I’d considered discussing in my original post. In general, I don’t believe that sequencing technology advances can account for the differences in mutations found across the five cancer genomes. Although the first (AML1) utilized only shorter single-end reads (36 bp), the genome was extensively characterized before and after our Nature paper – using traditional sequencing, and later using paired-end Illumina libraries – and I’m very confident that we’re not missing many coding mutations. AML2, the breast cancer genome, and Sanger’s melanoma study were all done on the same platform, Illumina 2×75. Every study ends up at around 40x haploid coverage per tumor. I doubt the costs were the same, but the endpoint coverage seems comparable.

Unless I’m mistaken, validation was essentially identical in all five studies – PCR and 3730 resequencing in normal and tumor DNA.

Are there somatic coding mutations that we’re missing? A few, almost certainly. Regions refractory to accurate short read alignment (during discovery) and/or primer design (during validation) are likely to have hidden a handful of coding mutations. However, I think that these issues are likely to have a similar effect on all short-read technologies. As reads get longer, and as third-generation sequencing technologies (e.g. PacBio, with >1kb reads) come to market, this problem may go away.

Finally, all sequencing questions aside, the number of somatic coding mutations seems consistent with my limited knowledge of the mutational profile of each cancer. AML was a challenge because we knew there were fewer genetic changes compared to many other cancers. Breast cancer was probably somewhere in the middle. I’d guess that small-cell lung cancer and melanoma were expected to harbor more mutations, because they’re associated with exposure to known mutagens.