Thoughts on Network deconvolution as a general method to distinguish direct dependencies in networks
The opposite of clique completion: #Network deconvolution.. to distinguish direct dependencies http://go.nature.com/dVzNwC via @taziovanni
Network deconvolution as a general method to distinguish direct dependencies in networks
Soheil Feizi, Daniel Marbach, Muriel Médard & Manolis Kellis
http://www.nature.com/nbt/journal/vaop/ncurrent/full/nbt.2635.html
My thoughts:
Indirect relationships in a network can confound the inference of true direct relationships in a network. T, so this paper sought to develop a quantitative framework, termed network deconvolution (ND), to infer direct relationships and remove false positives in a network by quantifying and then removing indirect transitive relationship effects. The mathematical framework assumes that (1) an indirect relationship (edge) can be approximated as the product of its component direct edges and that (2) the observed edge weights are the sum of the direct and indirect edge weights – a linear dependency. The main application seems to be in mutual information (MI) and
correlation-based (COR) networks. They applied ND to various scenarios such as local network connectivity prediction (FFL
prediction), gene regulatory network prediction (in E. coli), prediction of interacting amino acids in protein structures (MI network) and coauthorship relationship network and found that (1) it can be used with various networks beyond just MI and COR (2) it can be used alone or more powerfully in combination with existing
methods/algorithms to improve predictions. In a sense it is the opposite of clique and module completion approaches (such as k-core).