Somatic error correction for low allele fractions

[davidbenjamin]

Read error correction tends to follow the basic strategy of 1) collect kmer counts 2) replace rare kmers with their closest non-rare match. For germline calling where there is a huge gap between error rates and diploid het allele fractions this is sufficient. Mutect, however, must contend with cases where counts alone do not discriminate perfectly between errors and real mutations.

Without committing to an approach, it seems like phasing might help. That is, we could construct haplotypes of rare kmers and error correct those. This should work because sequencing errors are unphased and real variants are. There are phased artifacts, of course, but we handle those in downstream filtering.