Back

Explanation of Models

Most genes in the Human Evolution Browser have been analyzed with PAML using five evolutionary models:
  • Model 1a: two classes of sites on all lineages: dN/dS<=1, dN/dS=1
  • Model H: one dN/dS ratio for human branch, one for the rest of the tree
  • Model Hnull: dN/dS=1 on human branch, one rate for the rest of tree
  • Model A: four site classes -- two classes with identical rates in all lineages (dN/dS<=1, dN/dS=1), two classes with dN/dS<=1 or dN/dS=1 on the background lineages but dN/dS>1 on the human lineage
  • Model Anull: as Model A, but with dN/dS fixed at 1 on the human lineage in the latter two classes
Tests of selection can be thought of as posing the following questions:
  • Branch tests:
    • Model H vs Model 1a: Is the dN/dS rate on the human branch significantly different from the average rate across the rest of the tree?
    • Model H vs Model Hnull: Is the dN/dS rate on the human branch significantly different from 1?
  • Branch+Site tests:
    • Model A vs Model 1a: Is there a class of sites (codons) that is evolving with a different dN/dS ratio in human than the same sites across the rest of the tree?
    • Model A vs Model Anull: Is the dN/dS ratio for those sites >1?
    • Model A vs Model 1a corresponds to Test 1 in and Model A vs Model Anull corresponds to Test 2 in this paper. Test 2 is the more conservative test of positive selection.
About 180 genes have also been analyzed with additional models that test for selection at certain sites (codons) across the tree:
  • Model 0: one dN/dS ratio across all lineages
  • Model 2: three classes of sites on all lineages: dN/dS<=1, dN/dS=1, dN/dS>1
  • Model 7: 10 classes of sites on all lineages, all with dN/dS<=1
  • Model 8: 11 classes of sites on all lineages, 10 with dN/dS<=1, one with dN/dS>1
  • Model 8a: 11 classes of sites on all lineages, 10 with dN/dS<=1, one with dN/dS=1
Tests of selection at a subset of sites:
  • Model 1a vs Model 0: Does allowing a second class of sites with dN/dS <=1 fit the data better than requiring one dN/dS ratio across all sites?
  • Model 2 vs Model 1a: Does adding a third class of sites with dN/dS>1 fit the data better than a model with two classes, each having dN/dS <=1?
  • Model 8 vs Model 7: Like Model 2 vs Model 1a, except ten classes with dN/dS <=1 and one class with dN/dS>1
  • Model 8 vs Model 8a: Is the dN/dS in the eleventh class of sites significantly >1?