Generating Ensembles of Gene Regulatory Networks to Assess Robustness of Disease Modules

Abstract

The use of biological networks such as protein–protein interaction and transcriptional regulatory networks is becoming an integral part of genomics research. However, these networks are not static, and during phenotypic transitions like disease onset, they can acquire new “communities” (or highly interacting groups) of genes that carry out cellular processes. Disease communities can be detected by maximizing a modularity-based score, but since biological systems and network inference algorithms are inherently noisy, it remains a challenge to determine whether these changes represent real cellular responses or whether they appeared by random chance. Here, we introduce Constrained Random Alteration of Network Edges (CRANE), a method for randomizing networks with fixed node strengths. CRANE can be used to generate a null distribution of gene regulatory networks that can in turn be used to rank the most significant changes in candidate disease communities. Compared to other approaches, such as consensus clustering or commonly used generative models, CRANE emulates biologically realistic networks and recovers simulated disease modules with higher accuracy. When applied to breast and ovarian cancer networks, CRANE improves the identification of cancer-relevant GO terms while reducing the signal from non-specific housekeeping processes.

The Network Zoo: a multilingual package for the inference and analysis of biological networks

“…CONDOR [21] identifies communities in bipartite graphs (including eQTL networks and GRNs), while ALPACA [22] finds differential community structures between two networks, such as in a case versus control setting, by going beyond the simple difference of edge weights and using the complete network structure to find differential communities. CRANE [23] extends ALPACA’s differential community estimation by assessing significance of differential modules and comparing to a baseline of network ensembles that were generated by preserving the specific structure and constraints of GRNs. An important use case of CRANE is modeling the transition between an initial and a final condition such as between healthy and disease states.…”

“…Our methods for reconstructing networks include PANDA [14] and OTTER [15] for modeling TF-gene regulatory processes, DRAGON [16] for estimation of multi-omic networks based on GGMs, and PUMA [17] which adds miRNA-gene post-transcriptional regulation to TF-gene interactions In addition, SPIDER [18] reconstructs networks by accounting for chromatin state, EGRET [19] include genotype information in network inference, and LIONESS [20] estimates network models for individual samples. Another set of methods has been developed for network analysis including CONDOR [21] for modeling and detection of communities in expression quantitative trait locus (eQTL) networks, ALPACA [22] and CRANE [23] for identifying communities within networks and how communities change between states, and MONSTER [24] to estimate TFs that drive the transition between network states. SAMBAR [25] allows us to group biological samples based on how genetic variants alter functional pathways, and, finally YARN [26] is a tissue-aware implementation of smooth quantile normalization for multi-tissue gene expression data.…”

Reconstructing Sample-Specific Networks using LIONESS

“…These modules may be enriched for specific biological properties (Platig et al, 2016;Fagny et al, 2017Fagny et al, , 2020. In addition, community structure comparison methods (Padi and Quackenbush, 2018;Lim et al, 2020) can be used to identify differences between pairs of single-sample networks. Although these methods are generally designed to compare two networks, they could be used to iteratively compare the community structure of each of the single-sample networks with that of the aggregate network.…”

The Network Zoo: a multilingual package for the inference and analysis of gene regulatory networks

“…CONDOR [24] identifies communities in bipartite graphs [25] (including eQTL and TFgene networks), while ALPACA [26] finds differential community structures between two networks, such as in a case versus control setting, by going beyond the simple difference of edge weights and using the complete network structure to find differential communities. CRANE [27] assesses the significance of differential modules discovered by ALPACA based on a baseline of network ensembles that are simulated while preserving the specific structure and constraints of GRNs. In this regard, CRANE provides an efficient tool for hypothesis testing inference on differential community structures in GRNs.…”