Addressing false discoveries in network inference

Motivation: Experimentally determined gene regulatory networks can be enriched by computational inference from high-throughput expression profiles. However, the prediction of regulatory interactions is severely impaired by indirect and spurious effects, particularly for eukaryotes. Recently, published methods report improved predictions by exploiting the a priori known targets of a regulator (its local topology) in addition to expression profiles. Results: We find that methods exploiting known targets show an unexpectedly high rate of false discoveries. This leads to inflated performance estimates and the prediction of an excessive number of new interactions for regulators with many known targets. These issues are hidden from common evaluation and cross-validation setups, which is due to Simpson's paradox. We suggest a confidence score recalibration method (CoRe) that reduces the false discovery rate and enables a reliable performance estimation. Conclusions: CoRe considerably improves the results of network inference methods that exploit known targets. Predictions then display the biological process specificity of regulators more correctly and enable the inference of accurate genome-wide regulatory networks in eukaryotes. For yeast, we propose a network with more than 22 000 confident interactions. We point out that machine learning approaches outside of the area of network inference may be affected as well.