On the limits of active module identification

Abstract

In network and systems medicine, active module identification methods (AMIMs) are widely used for discovering candidate molecular disease mechanisms. To this end, AMIMs combine network analysis algorithms with molecular profiling data, most commonly, by projecting gene expression data onto generic protein–protein interaction (PPI) networks. Although active module identification has led to various novel insights into complex diseases, there is increasing awareness in the field that the combination of gene expression data and PPI network is problematic because up-to-date PPI networks have a very small diameter and are subject to both technical and literature bias. In this paper, we report the results of an extensive study where we analyzed for the first time whether widely used AMIMs really benefit from using PPI networks. Our results clearly show that, except for the recently proposed AMIM DOMINO, the tested AMIMs do not produce biologically more meaningful candidate disease modules on widely used PPI networks than on random networks with the same node degrees. AMIMs hence mainly learn from the node degrees and mostly fail to exploit the biological knowledge encoded in the edges of the PPI networks. This has far-reaching consequences for the field of active module identification. In particular, we suggest that novel algorithms are needed which overcome the degree bias of most existing AMIMs and/or work with customized, context-specific networks instead of generic PPI networks.

NetMix2: Unifying network propagation and altered subnetworks

“…Defining a network-only baseline is a more subtle issue, and was discussed in two recent papers [68, 67]. Levi et al [68] benchmarks altered subnetwork algorithms on randomly permuted vertex scores while keeping the network G fixed.…”

“…The authors find that many existing methods output similar altered subnetworks (in terms of GO enrichment) on their permuted data, which suggests that these methods are utilizing the network G more than the vertex scores X v . Lazareva et al [67] benchmarks altered subnetwork algorithms on randomly permuted networks with the same degree distribution as G while keeping the vertex scores X v fixed. The authors find that many existing algorithms output similar altered subnetworks on permuted networks, indicating a degree bias in these methods.…”

“…We propose a more direct network-only baseline: we rank vertices v by their network centrality score N ( v ) for a network centrality measure N that is derived from the topological constraints used by the altered subnetwork algorithm. For example, for an algorithm that relies on the connected subfamily, we propose that degree centrality N ( v ) = d v is an appropriate measure, as in [67]. However, for network propagation algorithms that use random walk with restart, we claim that the PageRank centrality N ( v ) = ( M PPR · 1 ) v , where 1 ∈ ℝ n is an all-ones vector, is the more appropriate network-only baseline.…”

“…This is most likely due to the close relationship between the propagation family M δ,p and the network-only baseline: the network-only baseline ranks vertices by their PageRank centrality while the propagation family M δ,p is derived from the personalized PageRank matrix. Interestingly, ranking vertices by their degrees -as suggested in [67] -performed much worse than PageRank centrality (F -measure ≈ 0.05, not shown in Figure 2) for altered subnetworks A from the propagation family M δ,p , demonstrating the importance of using a network-only baseline that models the same topological properties as the network method.…”

“…While these comparisons are useful, they may also hide biases shared between algorithms. For example, Lazareva et al [67] observed that some well-known network algorithms have a bias towards high-degree vertices in the interaction network, while Levi et al [68] similarly observed a bias in GO term enrichment among well-known network algorithms. In order to quantify the potential biases of altered subnetwork algorithms, these algorithms need to be compared against carefully selected baselines, including baselines that do not use the interaction network and baselines that do not use the vertex scores.…”