We normalize individual networks to make them suitable for GeneMANIA’s label propagation algorithm. In this transformation, each interaction in a physical or genetic network is assigned a link weight of 1/[sqrt(di)sqrt(dj)] where di and dj are the degrees (i.e. number of neighbors) of the two linked genes in the network. In networks where links already have weights (e.g. co-expression networks), the normalized link weights are wij / [sqrt(wdi)sqrt(wdj)] where wij is the initial link weight and wdi and wdj are the ‘weighted degree’ (i.e., the sum of the initial link weights to all neighbors) for the two genes. The post-normalization weighted degree of nodes is approximately one in each network. In addition to being required for the label propagation algorithm, this normalization helps to reduce the impact that the pleiotropy of high degree nodes [see, e.g. (29)] has on functional predictions (30).
