Fig. 2: EvoWeaver’s ensemble predictions outperform individual algorithms on the Modules benchmark.
From: EvoWeaver: large-scale prediction of gene functional associations from coevolutionary signals
Coevolutionary approaches were compared for their ability to discern adjacent proteins in KEGG modules (i.e., 899 positives) from proteins in distinct modules (i.e., 899 negatives). No single source of coevolutionary signal greatly outcompeted all other sources. However, EvoWeaver’s ensemble predictions that combine all component sources of coevolutionary signal substantially improved predictive accuracy, as seen by larger areas under the curves. Inset of the receiver operating characteristic highlights the region with low false positive rates. Scores from individual algorithms tended to have low correlation except within similar categories of coevolutionary signal (i.e., boxed groups in the heatmap), suggesting that the ensemble approach is superior because it combines semi-orthogonal coevolutionary signals. Spearman’s correlation from positive and negative sets is averaged to correct for artificial correlation among high-performing algorithms. Source data are provided as a Source Data file.
