Table 1 provides details on validation for our various network weighting algorithms and our network propagation algorithms. These methods have not changed since our previous NAR Web server article except as described in previous sections and validated here. Filtering unsupported co-expression interactions led to a significant increase (P < 10−10) in the cross-validated area under the precision-recall curve (AUPRC) versus the unfiltered co-expression networks on the task of classifying genes into Gene Ontology biological process terms with between 10 and 100 annotations using GeneMANIA for human, C.elegans and yeast (using GeneMANIA data release from 3 March 2011). These were the only species we tested. On the same benchmark but testing only in human and using an internal data build (between the 21 December 2011 and 19 July 2012 releases), adding up to 100 attributes and a small amount of L2-regularization to the network weight regression led to a significant increase (P < 10−10) in AUPRC compared with using just the default network selections in human (median non-zero change in AUPRC was an increase of 0.0078). In terms of software validation,
