Motivation
Experimentally determined gene regulatory networks
can be complemented by computational inference from high-
throughput expression profiles. However, in particular for eukaryotes,
indirect and spurious effects impair the reliablity of predicted
regulatory interactions. Recently published methods aim to address
this problem by exploiting the a priori known targets of a regulator (its
local topology) in addition to expression profiles.
Results
We discover that the selection of candidate regulations may be
influenced by a high degree preference (HDP), such that an excessive
number of new interactions is predicted for regulators with
many a priori known targets.
In a cross-validation setup this effect inflates performance
estimates substantially. In particular, global evaluation criteria
like ROC curves prefer HDP results over the correctnes of individual
regulators. We argue that this is critical and, suggest Confidence Recalibration (CoRe), a method that reduces the false-discovery rate
of predictions on the level of individual regulators.
Simultaneously, it enables an integrated view of the complete network. Quality
estimates are consistent for this network, regardless of a regulator-wise
or network-wide point of view.
Supplementary Material
