TL;DR
This paper introduces NETMAGE, a web-based tool that generates interactive network visualizations of human disease relationships based on phenome-wide association study (PheWAS) data. It allows researchers to explore potential shared genetic etiologies underlying disease comorbidities and complications, aiding in the identification of new therapeutic targets.
What Problem Does This Paper Address?
Understanding why certain diseases tend to co-occur or lead to secondary conditions is a major challenge in medicine. While genetic factors are known to play a role, pinpointing the specific genetic drivers behind these cross-phenotype associations has been difficult. Existing tools for analyzing PheWAS data often lack interactive visualization capabilities for exploring these complex relationships at a network level. This paper addresses the need for a user-friendly platform that can transform raw PheWAS summary statistics into intuitive, searchable disease-disease networks (DDNs), thereby facilitating the exploration of the "diseasome" – the landscape of all human diseases.
Methods
The core of NETMAGE is its ability to process PheWAS summary statistics and generate interactive disease-disease networks (DDNs). The process involves:
- Data Upload and Filtering: Users upload their PheWAS summary statistics, which can include p-values and minor allele frequencies (MAF). NETMAGE allows filtering based on these parameters to focus on significant and common genetic variants.
- Network Generation: The tool converts the filtered data into an intermediate file representing phenotype-to-variant mappings. This is then transformed into node (disease) and edge (shared genetic variants) maps. Edge weights can represent either the count of shared variants or a marginalized fraction.
- Visualization: NETMAGE utilizes Gephi and InteractiveVis (built on sigma.js) to create a 2D network visualization. Users can customize network layout, node size, and edge thickness.
- Interactive Features: The tool offers node selection to highlight direct neighbors and display associated variants, phenotypes, and network statistics in an "Information Pane." It also includes a powerful search function to find phenotypes based on various attributes and a "search by variant" option to identify shared genetic links between diseases. Users can also group nodes by disease category for focused visualization.
Datasets
The primary dataset used for demonstration and evaluation in this paper is the UK Biobank (UKBB) PheWAS data. This dataset comprised approximately 1,403 binary phenotypes (expressed as PheCodes) from 400,000 British individuals of European ancestry. The data was imputed with the Haplotype Reference Consortium panel, yielding 28 million imputed SNPs.
Key Insights and Ideas
- Visualizing the "Diseasome": NETMAGE provides a novel way to visualize the complex web of relationships between diseases, moving beyond simple pairwise comorbidity lists to a network perspective.
- Uncovering Shared Genetic Etiologies: By representing edges as shared genetic variants, the tool directly links disease associations to potential underlying genetic causes, offering hypotheses for why certain conditions co-occur.
- Identifying Hub Diseases: The network analysis revealed that endocrine/metabolic and circulatory diseases appear to be "hub" categories, suggesting they play a significant role in connecting other diseases through shared genetic pathways.
- Bridging PheWAS and Network Medicine: The paper effectively bridges the gap between large-scale PheWAS studies and the principles of network medicine, providing a practical tool for researchers in both fields.
Key Results
- The NETMAGE tool successfully generated a DDN from the UKBB PheWAS data, which accurately reflected previously identified disease comorbidities.
- The generated network showed concentrations of hub diseases in the endocrine/metabolic and circulatory disease categories.
- The network analysis revealed that a significant portion of genetic associations (76%) connected diseases across different categories, indicating that genetic influences on disease are not strictly confined within traditional disease classifications.
- The tool's runtime for generating networks scales approximately with O(n²) where 'n' is the number of phenotypes, with a full UKBB DDN generation taking around 48.60 seconds for one layout.
Limitations and Future Work
- PheCode Limitations: The authors acknowledge that using "PheCodes" for phenotype definition can be problematic due to the inherent inaccuracies of ICD codes and the hierarchical nature of PheCodes, which can lead to redundant nodes and edges if not handled carefully.
- Causality vs. Association: The output DDNs highlight associations, not necessarily causal relationships. Further analysis, such as Mendelian randomization, would be needed to infer causality.
- Directionality: The current DDNs do not indicate the direction of association between phenotypes. Future work aims to incorporate beta-values to provide this directional information.
- Single Node Selection: Currently, only one node can be selected at a time. The ability to select multiple nodes simultaneously would enhance clinical interpretation.
- Gene-based and Variant-variant Networks: Future directions include enabling the construction of gene-based DDNs from variant data and creating variant-variant networks to explore genetic connections directly.
Summary updated