The eleven panels in Figure 3—figure supplements 2, 3 and 4 show the full dependence of mutation spectrum on allele frequency in the UK10K data. If we let denote the fraction of SNVs of frequency that are of type and let denote the fraction of all mutations that are of type , the enrichment of mutation type as a function of frequency is . This function is expected to fluctuate around unless the rate of has recently increased or decreased. All 96 mutation types are visualized in every panel, but most corresponding lines are greyed out to enhance readability. Some lines deviate from due to the effects of biased gene conversion (BGC)–this occurs when one of the ancestral or derived alleles is a weak base (A or T, abbreviated W) and the other allele is a strong base (G or C, abbreviated S). WS mutations are more abundant at high allele frequencies, while SW mutations are more abundant at low frequencies. These effects are visible but modest in panels D, G, H, and I, but much more pronounced in panels B, C, and F, which focus on mutations in the CpG context. Transitions of the type CpACpG, which create CpG motifs, are extremely enriched at high frequencies, and this pattern may be an artifact of ancestral misidentification (Hernandez et al., 2007). CpG motifs have such high mutation rates that CpGCpT transitions often happen at the same site in humans and chimps, and these low-frequency double mutations are misclassified as high-frequency CpTCpG mutations. Although it is not surprising to see a peak of CpTCpG transitions at high frequencies in panel F, it is somewhat surprising to see CpGGpG transversions peak in abundance at high frequencies in panel C. This might be a signature of recent declines in the rates of these mutations, since neither ancestral misidentification nor biased gene conversion is thought to produce such a pattern. In addition, neither of these processes can explain the strong enrichment of certain AT mutations at high frequencies that is observed in panel K.