Comparative assessment of differential network analysis methods

Abstract

Differential network analysis (DiNA) denotes a recent class of network-based Bioinformatics algorithms which focus on the differences in network topologies between two states of a cell, such as healthy and disease, to identify key players in the discriminating biological processes. In contrast to conventional differential analysis, DiNA identifies changes in the interplay between molecules, rather than changes in single molecules. This ability is especially important in cases where effectors are changed, e.g. mutated, but their expression is not. A number of different DiNA approaches have been proposed, yet a comparative assessment of their performance in different settings is still lacking. In this paper, we evaluate 10 different DiNA algorithms regarding their ability to recover genetic key players from transcriptome data. We construct high-quality regulatory networks and enrich them with co-expression data from four different types of cancer. Next, we assess the results of applying DiNA algorithms on these data sets using a gold standard list (GSL). We find that local DiNA algorithms are generally superior to global algorithms, and that all DiNA algorithms outperform conventional differential expression analysis. We also assess the ability of DiNA methods to exploit additional knowledge in the underlying cellular networks. To this end, we enrich the cancer-type specific networks with known regulatory miRNAs and compare the algorithms performance in networks with and without miRNA. We find that including miRNAs consistently and considerably improves the performance of almost all tested algorithms. Our results underline the advantages of comprehensive cell models for the analysis of -omics data.

NetCoMi: Network Construction and Comparison for Microbiome Data in R

“…Differential network analysis allows not only to uncover the global role of a taxon in the overall network structure but also its changing influence under varying conditions. Differential association analysis [14], on the other hand, can directly assess which associations significantly change across conditions, providing concrete hypotheses for follow-up biological perturbation experiments. Similar to phyloseq's [94] plot_net function, NetCoMi also enables network representation and comparison of the amplicon data samples themselves, using popular sample dissimilarity or distance measures, such as the Bray-Curtis dissimilarity and the Aitchison distance.…”

“…Current approaches for comparing networks between two conditions can be divided into two types: (i) differential association analysis focusing on differences in the strength of single associations, and (ii) differential network analysis, analyzing differences between network metrics and network structure between two conditions [14]. Differential associations can further be used as the basis for constructing differential networks, where only differentially associated nodes are connected (see [14] for a comparative study of differential network analysis methods). Existing tools for network comparison either require pre-computed networks (adjacency matrix or edge list) as input (e.g.…”

Integrated transcriptomics reveals master regulators of lung adenocarcinoma and novel repositioning of drug candidates

“…Indeed, most reverse‐engineering methods model gene networks as static processes, in which interaction changes among elements in the network are not accounted across different conditions . To address this issue, recent studies have been working toward the development of such methods using a differential coregulation framework . However, the addition of dynamic approaches is no trivial endeavor and these methods can be hindered by network reconstruction stability problems, since network topology is of critical importance in these scenarios .…”

DCI: Learning Causal Differences between Gene Regulatory Networks

“…This is done to remove some edges to reduce the downstream computational burden. The difference undirected graph can be determined either using previous methods such as KLIEP [2][3][4][5][6], based on prior biological knowledge, or simply with the complete graph when the number of considered genes is small. In addition, to reduce the number of downstream hypothesis tests, the nodes to be considered as conditioning sets can be reduced to the nodes in the difference undirected graph as well as nodes whose conditional distribution changes between the two conditions, namely C = i | ∃j ∈ [p] such that Θ (1) i,j = Θ (2) i,j .…”

“…In addition, to reduce the number of downstream hypothesis tests, the nodes to be considered as conditioning sets can be reduced to the nodes in the difference undirected graph as well as nodes whose conditional distribution changes between the two conditions, namely C = i | ∃j ∈ [p] such that Θ (1) i,j = Θ (2) i,j . The reduced node set can be determined from the output of methods such as KLIEP [2][3][4][5][6], prior biological knowledge, or taken as the set of all nodes when the number of genes to be considered is small.…”