Community-Reviewed Biological Network Models for Toxicology and Drug Discovery Applications

Abstract

Biological network models offer a framework for understanding disease by describing the relationships between the mechanisms involved in the regulation of biological processes. Crowdsourcing can efficiently gather feedback from a wide audience with varying expertise. In the Network Verification Challenge, scientists verified and enhanced a set of 46 biological networks relevant to lung and chronic obstructive pulmonary disease. The networks were built using Biological Expression Language and contain detailed information for each node and edge, including supporting evidence from the literature. Network scoring of public transcriptomics data inferred perturbation of a subset of mechanisms and networks that matched the measured outcomes. These results, based on a computable network approach, can be used to identify novel mechanisms activated in disease, quantitatively compare different treatments and time points, and allow for assessment of data with low signal. These networks are periodically verified by the crowd to maintain an up-to-date suite of networks for toxicology and drug discovery applications.

Keywords: COPD; biological network; crowdsourcing; drug discovery; toxicology.

Figure 1

Comparison of the NVC (A), KEGG (B), and Reactome (C) calcium/calmodulin signaling pathways. Shared portions highlighted in yellow with corresponding numbers.

Figure 2

Changes in network statistics as a result of NVC activity. Differences between the latest version of the networks and the original networks have been posted to the Bionet website. Notes: *Discussed in one jamboree. **Discussed in two jamborees. Networks are organized in the following biological categories: cell fate, cell proliferation, cell stress, inflammation, and tissue repair and angiogenesis. The details of the analysis and the description of the different statistics are described in the “Materials and methods” section.

Figure 3

Senescence (A) and cell cycle (B) networks scored with GSE28464 HRASV12 data from the NCBI GEO database. A selection from the TopoNPA-scored version is shown. Arrow edge indicates a positive relationship while ball and stick edge indicates a negative relationship (includes causal and correlative statements). Nodes are colored by their NPA score; yellow/orange indicates inferred increase and blue indicates inferred decrease in activity or abundance. Darker colors denote higher magnitude scores. Leading nodes contribute to 80% of the network score and are denoted by their shapes outlined in gray. Nodes added within this section of the network during the NVC are labeled in red. (A) Senescence network. Nodes boxed in yellow reflect experimental HRASV12 mutation, resulting in oncogene-induced senescence. (B) Cell cycle network. Predicted upregulated nodes (yellow) contain cell cycle inhibitors RB1, E2F4, and CDKN1A predicted increased. Predicted decreased nodes (blue) contain cell cyclins and E2Fs predicted decreased.

Figure 4

Macrophage signaling network scores in the E-MTAB-3150 dataset and pigmented macrophage counts in the same study. (A) Macrophage signaling network score increased with time with smoke exposure and decreased with switch or cessation. pMRTP did not have significant macrophage signaling network scores at any time point. Green, blue, and red asterisks indicate significant O, K, and experimental P-values, respectively. (B) Pigmented macrophage in the alveolar lumen increased with smoke exposure over time and decreased with switch or cessation. pMRTP did not induce an increase in pigmented macrophages. Notes: *P < 0.05 compared with sham. ^#P < 0.05 compared with smoke exposure.

Figure 5

Leading node contribution for macrophage signaling network in the E-MTAB-3150 dataset. Word size indicates relative contribution to network score. Notes: ^*significant score; (+) inferred increase; (−) inferred decrease.

Figure 6

Macrophage signaling network scores for seven-month smoke vs seven-month fresh air using the E-MTAB-3150 dataset. A selection from the TopoNPA-scored version is shown. Arrow edge indicates a positive relationship, while ball and stick edge indicates a negative relationship (includes causal and correlative statements). Nodes are colored by NPA score; yellow indicates inferred increase and blue indicates inferred decrease. Darker colors denote higher magnitude scores. Leading nodes contribute to 80% of the network score and are denoted by their shapes outlined in gray. Nodes added within this section of the network during the NVC process are labeled in red. Nodes boxed in yellow reflect prediction of TLR pathway.

Figure 7

Heat map of network scores comparing the impact of CS exposure, pMRTP, and cessation in the E-MTAB-3150 and GSE52509 datasets. Each treatment is compared to fresh air at the same time point. Scores are normalized to the maximum scores for each network. A network is considered impacted if, in addition to the significance of the score with respect to the experimental variation, the two companion statistics (O and K) derived to inform the specificity of the score with respect to the biology described in the network, are significant. Note: *O and K statistic P-values below 0.05 and NPA significantly nonzero.

Figure 8

Th17 signaling network scored with GSE52509 mouse lung exposed to 6 month smoke. The whole TopoNPA-scored version is shown. Arrow edge indicates a positive relationship, while ball and stick edge indicates a negative relationship (includes causal and correlative statements). Nodes are colored by NPA score; yellow indicates inferred increase and blue indicates inferred decrease. Darker colors denote higher magnitude scores. Leading nodes contribute to 80% of the network score and are denoted by their shapes outlined in gray. Nodes added within this section of the network during the NVC process are labeled in red. Nodes boxed in yellow reflect prediction of Il17 cytokines.