Table 5.

TT-Hybrid improves upon both of its constituent components on in-species prediction

Sparsity	GLIDE	Topsy-Turvy	TT-Hybrid	Random
p = 0.8	0.380	0.038	0.387	0.004
p = 0.6	0.437	0.079	0.451	0.009
p = 0.4	0.412	0.105	0.423	0.014
p = 0.2	0.318	0.133	0.354	0.019

Note: We generated partitions of the fly network of varying sparsity, using the sparsified networks as training for GLIDE. Sparsity p corresponds to the proportion of edges retained in the training network (p = 0.8 is the least sparse). Topsy-Turvy was trained on human PPIs. TT-Hybrid combines the predictions from both GLIDE and Topsy-Turvy. Here, we report the AUPR of each method on the held out edges removed from each network subset. We also show the AUPR of the random control; due to varying class imbalances, AUPR scores increase slightly with increasing sparsity. Bold entries represent best performance.