Hive Panel Explorer: an interactive network visualization tool

Abstract

Motivation

Networks are used to relate topological structure to system dynamics and function, particularly in ecology systems biology. Network analysis is often guided or complemented by data-driven visualization. Hive one of many network visualizations, distinguish themselves as providing a general, consistent and coherent rule-based representation to motivate hypothesis development and testing.

Results

Here, we present HyPE, Hive Panel Explorer, a software application that creates a panel of interactive hive plots. HyPE enables network exploration based on user-driven layout rules and parameter combinations for simultaneous of multiple network views. We demonstrate HyPE’s features by exploring a microbial co-occurrence network constructed from forest soil microbiomes.

Availability and implementation

HyPE is available under the GNU license: https://github.com/hallamlab/HivePanelExplorer.

Supplementary information

Supplementary data are available at Bioinformatics online.

1 Introduction

Networks model relationships (edges) between components (nodes) of a system and aid in relating topological structure to dynamics and function (Chumpitazi et al., 2014; Giovannelli et al., 2017; Thompson et al., 2017). For example, nodes within a biological network may represent foodwebs, or pathway and taxonomic relationships in microbial data (Chumpitazi et al., 2014; Giovannelli et al., 2017; Thompson et al., 2017) while the edges may represent their physical or trophic interactions, respectively. Graph theory and measures such as degree distributions, modularity and connectance are commonly used to quantify data associations, inform predictive models or generate hypotheses associated with network topology (Chumpitazi et al., 2014; Giovannelli et al., 2017; Thompson et al., 2017).

HyPE is based on Hive plots (Krzywinski et al., 2012): a rule-based layout that positions nodes using a circular coordinate system (Fig. 1). This coordinate system is based on rules driven by relevant node and edge properties (i.e. degree, clustering coefficient, betweenness, etc.). Nodes are placed onto radially arranged axes and edges are drawn between nodes using curves. The rules determining a node’s assignment and position along an axis can be user defined.

Fig. 1.

HyPE panel of network calculated using SSU rRNA gene pyrotag operational taxonomic units (OTUs) from unmanaged (N) and harvested (H) soil profiles. Co-occurrence represents positive correlations. Mutual exclusion represents negative correlations. Node color represents module membership. Node placement represents graph properties

Hive plots are a generalizable, flexible and extensible network visualization modality; however, understanding how to harness their versatility for a particular research question can be challenging. HyPE addresses this issue by providing users with the ability to explore different hive plot constructions in a grid, or ‘hive panel’ (Krzywinski et al., 2012). This design circumvents the need to determine optimal a priori rule sets and enables users to explore multiple combinations of network properties and data dimensions simultaneously. More information related to HyPE’s navigation features are in Supplementary Material.

2 Materials and methods

HyPE was constructed using D3 (Mike Bostock’s hive plot plug-in) (Bostock et al., 2011), JavaScript (Powell et al., 2004) and Python v2.7 (Rossum, 1995). HyPE accepts two comma separated or tab separated (‘.tsv’ or ‘.txt’) files as input, a node file and an edge file, making it easy to include node and edge properties and creates a webpage with the interactive interface (Fig. 1).

Here, we use HyPE to reveal patterns in a microbial co-occurrence network. Generally, such networks are constructed by evaluating correlations among taxa represented by small subunit ribosomal RNA (SSU or 16S rRNA) gene sequences clustered in operational taxonomic units (OTUs) or amplicon sequence variants (ASVs). Details on data generation and network construction are in Supplementary Material. Correlations capture potential ecological interactions between taxa such as trophic exchange, predation or competition. In this light, a microbial co-occurrence network models community structure and network properties can be related back to environmental metrics (Chumpitazi et al., 2014; Giovannelli et al., 2017; Thompson et al., 2017).

Network properties associated with 26 soil microbiome samples sourced from a natural reference (N) and a harvested soil plot (H) near Williams Lake British Columbia, Canada were explored. Detailed sample descriptions, methods and network properties are in Supplementary Material. Two hypotheses guided this exploration, (i) that differences in soil microbial community co-occurrence patterns are due to changing resource availability with depth and harvesting, and (ii) network composition can reveal design principles shaping soil ecosystem functions (Fig. 1). Each hive panel node represents an OTU, and each edge a positive or negative correlation. The network contains 1880 nodes, connected by 13 605 edges. A total of 6967 edges correspond to positive correlations (co-presence) while 6638 edges correspond to negative correlations (mutual exclusion).

To address the first hypothesis, we identified modules (sub-graphs of highly clustered microbes within the network) in the positive edges. Three large (>99 nodes) modules and 11 smaller components (3–61 nodes) were identified. To guide interpretation, we calculated weighted mean depth of each OTU and highlighted ‘indicator OTUs’ (a method by which to identify OTUs driving differences among groups of samples or in this case soil depth and treatment). Based on both weighted mean depth and indicators, the data suggest components A and B are representative of the H and N surface horizons, respectively. For example, Module A was the shallowest (7 cm), and 55% of its nodes were indicators for the H surface horizons, while Module B was deeper (13 cm) and 16% of its nodes were indicators for the N surface horizons. Thus, modules partitioned with harvesting treatment and soil depth (Fig. 1).

To address the second hypothesis, we compared taxonomy of OTU network modules to taxonomic and metabolic data from cognate metagenomes. Detailed information on metagenomes and processing is in Supplementary Material. We observed that differences in taxonomy among the metagenomes were consistent with taxonomic structure of the network modules. For example, Proteobacteria were significantly more abundant within the N LFH microbial community compared to the H LFH microbial community. Similar patterns were observed for Bacteriodetes and Actinobacteria and the annotated open reading frames attributed to these groups. Thus, metabolic function from the metagenomes could be taxonomically linked to network components associated with distinct soil horizons. Together, the data suggest that within LFH horizons, forest harvesting alters the architecture of connectivity, resulting in novel interactions with potential implications for nutrient and energy flow within the ecosystem.

3 Conclusion

HyPE can be used for different network types, sizes and topologies to showcase system properties. The forest soil use case demonstrates HyPE’s versatility in resolving relevant patterns within a complex biological network to support hypothesis development and testing.