Weighted mutual information analysis substantially improves domain-based functional network models

Abstract

Motivation: Functional protein–protein interaction (PPI) networks elucidate molecular pathways underlying complex phenotypes, including those of human diseases. Extrapolation of domain–domain interactions (DDIs) from known PPIs is a major domain-based method for inferring functional PPI networks. However, the protein domain is a functional unit of the protein. Therefore, we should be able to effectively infer functional interactions between proteins based on the co-occurrence of domains.Results: Here, we present a method for inferring accurate functional PPIs based on the similarity of domain composition between proteins by weighted mutual information (MI) that assigned different weights to the domains based on their genome-wide frequencies. Weighted MI outperforms other domain-based network inference methods and is highly predictive for pathways as well as phenotypes. A genome-scale human functional network determined by our method reveals numerous communities that are significantly associated with known pathways and diseases. Domain-based functional networks may, therefore, have potential applications in mapping domain-to-pathway or domain-to-phenotype associations.Availability and Implementation: Source code for calculating weighted mutual information based on the domain profile matrix is available from www.netbiolab.org/w/WMI.Contact: Insuklee@yonsei.ac.krSupplementary information: Supplementary data are available at Bioinformatics online.

BarleyNet: A Network-Based Functional Omics Analysis Server for Cultivated Barley, Hordeum vulgare L.

“…Then, mutual information scores were computed between domain profiles. We used a weighted mutual information (WMI) scheme, which assigns more weight on rarer domains during mutual information computation (Shim and Lee, 2016;Shim and Lee, 2020). We calculated LLSs for gene pairs using a regression function between WMI and LLS.…”

Estimating mutual information under measurement error

“…We assume the distribution of measurement error is the same for all samples. The density of the observation p xy is different from the density of the true signals p µxµy as shown by equation (2). Our goal is to derive a corrected estimator for mutual information of I(µ x , µ y ) with a reduced bias.…”

“…Applications of mutual information in computational biology include analyzing the co-evolution relationship between amino acids or nucleotides [1], inferring co-occurrence patterns of protein domains [2], constructing gene regulatory networks [3], studying neural connectivity circuits [4], and so on. An accurate estimation of mutual information is critical to these studies.…”

From sequencing data to gene functions: co-functional network approaches

“…However, these methods require reference PPIs or known DDIs to identify new functional associations between coding genes. Recently, we proposed domain profiling (DP) ( Figure 1 , Inference step C), a domain-based method to infer functional links that requires only domain annotations for each protein coding genes (Shim & Lee 2016 ). In this method, the domain composition of each protein coding gene is represented as a domain profile, which is a vector of presence or absence of each domain of a comprehensive domain database, Interpro (Mitchell et al 2015 ).…”

“…There are various metrics to measure profile similarity. We developed a new metric, a weighted version of MI, and found that this metric outperformed other popular metrics including traditional MI (Shim & Lee 2016 ).…”