netCRS: Network-based comorbidity risk score for prediction of myocardial infarction using biobank-scaled PheWAS data

Yonghyun Nam, Sang‐Hyuk Jung, Anurag Verma, Vivek Sriram, Hong‐Hee Won, Jae‐Seung Yun, Dokyoon Kim

TL;DR

This paper introduces netCRS, a novel network-based approach to predict individual disease comorbidity risk by integrating genetic information with disease relationships. By constructing a multi-layered network of diseases and SNPs, netCRS improves upon traditional polygenic risk scores (PRS) by accounting for how diseases co-occur, leading to better prediction of myocardial infarction (MI) risk.

What Problem Does This Paper Address?

Predicting an individual's risk for developing diseases is crucial for precision medicine. Traditional Polygenic Risk Scores (PRS) are effective for single diseases by aggregating the effects of many genetic variants (SNPs). However, most people experience multiple diseases over their lifetime (comorbidity), and these conventional PRSs fail to capture the complex interplay and shared genetic underpinnings between different conditions. While some methods attempt to combine PRSs for multiple traits, they often suffer from selection bias in choosing which diseases to include. This paper addresses the need for a more holistic approach that considers the network of relationships between diseases and their shared genetic components to provide a more accurate, individual-level risk assessment.

Methods

The core innovation is the development of a Disease-SNP Heterogeneous Multi-layered Network (DS-Net). This network has three main components:

  1. SNP-layer: Represents relationships between SNPs based on linkage disequilibrium (LD), built from individual genotype data.
  2. Disease-layer: Represents relationships between diseases based on shared genetic components identified through Phenome-Wide Association Studies (PheWAS).
  3. Coupling graphs: Connect the SNP-layer and Disease-layer, representing the association between specific SNPs and diseases.

Using this DS-Net, the authors employ graph-based semi-supervised learning (SSL). Individual genotype data is used to "label" the SNP-layer. This label information then propagates through the network, leveraging the disease-disease and SNP-disease relationships, to predict comorbidity scores for each disease in the Disease-layer. Finally, these individual disease scores are aggregated using logistic regression to produce the final netCRS score for a specific index disease.

Datasets

  • UK Biobank (UKBB): Used to construct the reference network (DS-Net) from PheWAS summary statistics.
  • Penn Medicine Biobank (PMBB): Provided individual-level genotype data for model development and validation.

Key Insights and Ideas

  • Diseases are interconnected: The paper strongly emphasizes that diseases don't occur in isolation. Understanding these co-occurrence patterns, driven by shared genetic factors, is key to better risk prediction.
  • Network approach for comorbidity: Moving beyond single-trait PRS, a network-based approach (DS-Net) allows for a more comprehensive view of genetic influences across multiple diseases.
  • Bridging population-level data and individual risk: The DS-Net effectively translates population-level genetic associations (from UKBB PheWAS) into individual-level risk predictions using personal genotype data (from PMBB).
  • SSL as a powerful tool: Graph-based SSL proves effective in propagating genetic information through the complex multi-layered network to derive meaningful comorbidity scores.
  • Complementary nature of PRS and netCRS: The study highlights that netCRS and traditional PRS can work together, offering a more robust prediction than either method alone, especially for identifying high-risk individuals.

Key Results

  • Improved prediction for Myocardial Infarction (MI): The netCRS model, particularly when combined with PRS and covariates (sex, age), achieved a significant improvement in prediction performance.
  • AUC Improvement: The combined model (netCRS + PRS-PT + covariates) achieved an AUC improvement of 28.29% compared to the PRS-PT alone model in predicting MI cases. Even the netCRS + covariates model showed a 6.26% improvement over the PRS-PT + covariates model.
  • Enhanced Risk Stratification: The combined model could identify individuals with an MI risk up to approximately eight-fold higher than the low-risk group.
  • netCRS with μ=0.1: The best predictive performance for netCRS was observed with a hyperparameter value of μ=0.1.

Limitations and Future Work

  • Node Selection Criteria: The authors suggest that applying more precise criteria for selecting diseases and SNPs in the DS-Net could lead to even better comorbidity scores.
  • Variant Scope: The current network construction primarily uses common variants. Expanding to include rare variants and other clinical information could further enhance prediction.
  • Disease-Layer Construction: The disease-layer was built using shared SNPs. Future work could explore alternative methods for building the disease-disease network, such as using LD regression scores.
  • Generalizability: The authors plan to test netCRS in various diseases and compare it with more recent PRS approaches to assess its generalized prediction performance.

Summary updated

Abstract

The polygenic risk score (PRS) can help to identify individuals’ genetic susceptibility for various diseases by combining patient genetic profiles and identified single-nucleotide polymorphisms (SNPs) from genome-wide association studies. Although multiple diseases will usually afflict patients at once or in succession, conventional PRSs fail to consider genetic relationships across multiple diseases. Even multi-trait PRSs, which take into account genetic effects for more than one disease at a time, fail to consider a sufficient number of phenotypes to accurately reflect the state of disease comorbidity in a patient, or are biased in terms of the traits that are selected. Thus, we developed novel network-based comorbidity risk scores to quantify associations among multiple phenotypes from phenome-wide association studies (PheWAS). We first constructed a disease-SNP heterogeneous multi-layered network (DS-Net), which consists of a disease network (disease-layer) and SNP network (SNP-layer). The disease-layer describes the population-level interactome from PheWAS data. The SNP-layer was constructed according to linkage disequilibrium. Both layers were attached to transform the information from a population-level interactome to individual-level inferences. Then, graph-based semi-supervised learning was applied to predict possible comorbidity scores on disease-layer for each subject. The SNP-layer serves as receiving individual genotyping data in the scoring process, and the disease-layer serves as the propagated output for an individual’s multiple disease comorbidity scores. The possible comorbidity scores were combined by logistic regression, and it is denoted as netCRS. The DS-Net was constructed from UK Biobank PheWAS data, and the individual genetic profiles were collected from the Penn Medicine Biobank. As a proof-of-concept study, myocardial infarction (MI) was selected to compare netCRS with the PRS with pruning and thresholding (PRS-PT). The combined model (netCRS + PRS-PT + covariates) achieved an AUC improvement of 6.26% compared to the (PRS-PT + covariates) model. In terms of risk stratification, the combined model was able to capture the risk of MI up to approximately eight-fold higher than that of the low-risk group. The netCRS and PRS-PT complement each other in predicting high-risk groups of patients with MI. We expect that using these risk prediction models will allow for the development of prevention strategies and reduction of MI morbidity and mortality.
BibTeX
@inproceedings{nam2022netcrs,
  author = {Yonghyun Nam and Sang‐Hyuk Jung and Anurag Verma and Vivek Sriram and Hong‐Hee Won and Jae‐Seung Yun and Dokyoon Kim},
  title = {{netCRS: Network-based comorbidity risk score for prediction of myocardial infarction using biobank-scaled PheWAS data}},
  booktitle = {PubMed},
  year = {2022},
}
Figure