Figure 9

Performance of our new statistics compared to the performance of GC content. Each data point corresponds to a sample of independent simulations. We determine the error rate (as shown in Figure 8) both for GC content and for our statistic 36. Here we show the error rate for statistic 36 divided by the error rate for GC-content alone. When the ratio of error rates is less than 1, our new statistic outperforms GC content in detecting simulated HGT events. The different panels show different simulation conditions: (a) 8-taxon trees, no restriction on Q₂, (b) 8-taxon trees, each random element of Q₂ constrained to be within 90% of the corresponding element of Q₁, (c) 16-taxon trees, no restriction on Q₂, and (d) 16-taxon trees, each random element of Q₂ constrained to be within 90% of the corresponding element of Q₁. Different curves correspond to different branch ratios. Error bars represent the standard error of the mean for 10 independent simulations. This statistic outperforms GC content substantially when the branch ratio is low (corresponding to a more ancient split of the host species from the original common ancestor) and the divergence is low to intermediate (gene sequences have intermediate to high sequence identity). In some cases we find that use of this statistic gives a lower error rate than does GC content by a factor of 10.