A Network-Based Analysis of Disease Complication Associations for Obstetric Disorders in the UK Biobank

Vivek Sriram, Yonghyun Nam, Manu Shivakumar, Anurag Verma, Sang‐Hyuk Jung, Seung Mi Lee, Dokyoon KimJournal of Personalized Medicine 2021 · DOI

TL;DR

This paper introduces a novel network-based approach to predict long-term health complications in women who have experienced obstetric disorders. By analyzing genetic associations between diseases, they demonstrate that focusing on individual obstetric disorders (egocentric networks) significantly improves the prediction of future comorbidities compared to analyzing the entire disease network.

What Problem Does This Paper Address?

Obstetric disorders, such as preeclampsia and gestational diabetes, are known to increase the risk of various long-term health issues for women. However, the underlying biological mechanisms and genetic links connecting these pregnancy complications to subsequent chronic diseases remain poorly understood. This knowledge gap hinders the development of personalized risk assessments and preventative strategies. The authors aim to bridge this gap by leveraging network medicine and genetic data to uncover these complex associations.

Methods

The core of this research lies in constructing and analyzing Disease-Disease Networks (DDNs). Here's a breakdown of their approach:

  • Variant-Based DDN Construction: They started by using UK Biobank (UKBB) data to perform Phenome-Wide Association Studies (PheWAS). This process identifies genetic variants (SNPs) associated with various diseases. A DDN was then built where nodes represent diseases, and edges represent shared genetic variants between them.
  • Egocentric DDNs: Instead of using the entire, complex DDN, they created egocentric DDNs. For each specific obstetric disorder (the "ego-disease"), they built a smaller network containing only that disorder and its directly connected "alter-diseases" (those sharing at least one associated SNP). This focuses the analysis on the specific obstetric condition.
  • Graph-Based Semi-Supervised Learning (GSSL): They applied GSSL to these egocentric DDNs. GSSL is a machine learning technique that propagates labels (in this case, a "positive" label for the ego-disease) through a network to predict scores for unlabeled nodes (alter-diseases). These scores represent the likelihood of comorbidity. The key assumption is that diseases sharing more genetic variants are more likely to co-occur.
  • Disease Stratification Analysis: To validate their findings, they used individual-level genotype data from UKBB. They examined pairs of diseases (an ego-disease and an alter-disease) that shared SNPs and compared the comorbidity risk between individuals who possessed these shared SNPs and those who did not.

Datasets

  • UK Biobank (UKBB): This large-scale, electronic health record (EHR)-linked biobank was the primary data source.
  • PheWAS Summary Data: Used for constructing the initial variant-based DDN. This data included genetic and broad phenotype codes for approximately 400,000 British individuals.
  • Individual-Level Genotype and EHR Data: Used for validation and disease stratification analysis. This data was obtained through a specific application process.

Key Insights and Ideas

  • Focusing on the "Ego" is Key: The most significant insight is that creating smaller, focused networks around individual obstetric disorders (egocentric DDNs) dramatically improves the ability to predict future complications compared to analyzing the entire, vast "diseasome." This suggests that the genetic pathways leading to complications are often specific to the initial obstetric condition.
  • Genetics as a Bridge: The study highlights how shared genetic variants between obstetric disorders and other diseases can serve as a crucial link, providing a mechanistic explanation for observed comorbidity patterns.
  • Network Medicine for Precision: This work demonstrates the power of network medicine and graph-based learning for personalized risk assessment, moving beyond simple statistical associations to uncover complex biological relationships.
  • Obstetric Disorders as a "Window": The findings reinforce the idea that pregnancy-related disorders can act as an early indicator or "window" into a woman's predisposition to future chronic diseases.

Key Results

  • Improved Predictive Performance: The egocentric DDN approach, combined with GSSL, achieved a significant improvement in predicting disease comorbidities. The average Area Under the Receiver Operating Characteristic Curve (AUC) increased by 1.35 times, from 55.0% to 74.4%, when comparing the egocentric networks to the full DDN.
  • Significant Correlations: The study reported significant Spearman rank correlations between the predicted scores from their GSSL algorithm and ground truth disease co-occurrences derived from EHR data, validating their model.
  • Identification of Potential Pathogenic Mechanisms: The network analysis identified several disorders that may play roles in the pathophysiology of obstetric complications, including inflammatory diseases of the uterus and pelvic adhesions, as well as disorders that might co-occur or present alongside obstetric issues, such as coma and abnormal serum enzyme levels.
  • Disease Stratification Potential: The proof-of-concept experiment showed that individuals possessing shared genetic variants between placenta previa and hypertensive heart disease had a higher risk of comorbidity, suggesting the potential for genetic stratification of patient risk.

Limitations and Future Work

  • Causation vs. Association: The authors acknowledge that their DDN solely represents possible genetic associations between phenotypes and does not imply causation. To establish causation, methods like Mendelian randomization would be required.
  • PheCode System Imperfections: The use of the PheCode system for disease classification has inherent limitations that could affect the accuracy of the conclusions.
  • Population Bias: The study relied on the UK Biobank, which primarily consists of individuals of British European ancestry. This may limit the generalizability of the findings to other populations with different genetic diversity.
  • Source Dataset Overlap: Using the UKBB for both DDN construction and gold-standard validation might lead to an overestimation of accuracy.
  • Future Directions:
    • Broader Application: The authors plan to apply their GSSL approach to other disease categories beyond obstetric disorders.
  • Additional Datasets: They suggest using other EHR-linked biobanks (e.g., Penn Medicine Biobank) for external validation.
  • Exploring Different DDN Types: Future work will involve constructing DDNs based on rare variants or copy number variations, and potentially polygenic risk scores (PRS)-based DDNs.
  • Integrating Lifestyle Factors: The authors emphasize the need to incorporate lifestyle factors alongside genetics to gain a more comprehensive understanding of disease connections.

Summary updated

Abstract

Egocentric DDNs hold promise as a clinical tool for the network-based identification of potential disease complications for a variety of phenotypes.
BibTeX
@article{sriram2021network,
  author = {Vivek Sriram and Yonghyun Nam and Manu Shivakumar and Anurag Verma and Sang‐Hyuk Jung and Seung Mi Lee and Dokyoon Kim},
  title = {{A Network-Based Analysis of Disease Complication Associations for Obstetric Disorders in the UK Biobank}},
  journal = {Journal of Personalized Medicine},
  year = {2021},
  doi = {10.3390/jpm11121382},
}
Figure