We observed that the direct method worked poorly, and was only slightly better than random inference.
Some methods work very poorly whatever the type of HGTs and some depend on the conditions or on the metrics used.
We show that those methods are overfitting as they work poorly on independent validation datasets from multiple genes.
Likewise, methods based on variations of the G+C composition work poorly due to the intrinsic variations of base composition in eukaryote genomes [ 69].
As evident in the figure, existing global methods work very well in predicting highly connected proteins, whereas they perform quite poorly for loosely connected proteins, especially for those with degree less than 6.
This problem arises because all methods have conditions for which they work well, and other conditions for which they work poorly.
Filter methods are usually based on statistical measures like correlation and mutual information [ 15] and they usually work poorly when considering small datasets.
"Did these methods work?" Ms. Golub said.
The CC-MEM and C-MEM work poorly.
