BrainNetScore: Enhancing Alzheimer’s disease risk prediction using genetic‐guided brain volumetric phenotype network

Yonghyun Nam, Jakob Woerner, Sang‐Hyuk Jung, Erica H Suh, Haochang Shou, Li Shen, Dokyoon KimAlzheimer 2024 · DOI

TL;DR

This paper introduces BrainNetScore, a novel network-based approach that integrates genetic information with brain volume data to significantly improve Alzheimer's disease (AD) risk prediction. By moving beyond traditional polygenic risk scores (PRSs), BrainNetScore offers a more comprehensive and accurate way to identify individuals at higher risk for AD, paving the way for personalized medicine strategies.

What Problem Does This Paper Address?

Alzheimer's disease (AD) is characterized by significant brain volume reduction, and genetic factors play a crucial role in these volumetric changes. While genome-wide association studies (GWASs) have identified links between brain imaging-derived phenotypes (IDPs) and AD, existing methods like polygenic risk scores (PRSs) have limitations in fully capturing these complex genetic relationships. Specifically, PRSs often treat genetic associations in isolation, failing to account for the interconnectedness of various brain regions and their volumetric phenotypes. This paper addresses the need for a more sophisticated approach that can leverage the network structure of brain volumes and their genetic underpinnings to enhance AD risk prediction accuracy.

Methods

The authors developed BrainNetScore, a network-based model that integrates genetic associations with multiple brain IDPs and AD incidence. The core of their method involves constructing a brain connectivity network using 96 regional brain volume IDPs from UK Biobank GWAS summary statistics. This network is then expanded into a heterogeneous "BrainNet graph" that includes these 96 IDPs and 12,043 common variants (SNPs) associated with each IDP.

For individual risk prediction, they utilized genotype data from independent cohorts, specifically the Alzheimer's Disease Neuroimaging Initiative (ADNI). A label propagation algorithm was employed to generate individualized predicted scores for each IDP. These scores were then aggregated into a single BrainNetScore using logistic regression. The performance of BrainNetScore was compared against conventional PRS models and combined models.

Datasets

  • UK Biobank: Used for GWAS summary statistics to construct the initial brain connectivity network and the heterogeneous BrainNet graph.
  • Alzheimer's Disease Neuroimaging Initiative (ADNI): Used for individual genotype data to evaluate the predictive performance of BrainNetScore. This cohort comprised 914 samples, including 550 AD cases and 364 cognitively normal controls.

Key Insights and Ideas

  • Network structure matters: The interconnectedness of brain regions and their volumetric phenotypes is crucial for accurate AD risk prediction, a factor often overlooked by traditional PRS methods.
  • Integrating genetic associations with IDPs: BrainNetScore's innovation lies in its ability to directly link genetic variants to brain volumetric phenotypes within a network framework, providing a more nuanced understanding of genetic influence.
  • Beyond individual SNPs: By aggregating information across multiple IDPs and their associated genetic variants, BrainNetScore captures a broader genetic signal related to AD pathology.
  • Synergistic effect of combined models: The combination of genetic risk scores (PRS) and network-based predictions (BrainNetScore) yields superior predictive power, suggesting that these approaches are complementary.

Key Results

  • BrainNetScore significantly improved AD risk prediction compared to PRS alone.
  • The BrainNetScore model achieved an average Area Under the ROC Curve (AUC) of 0.666 ± 0.029.
  • PRS alone achieved an average AUC of 0.595 ± 0.075.
  • The combined PRS + BrainNetScore model showed a superior average AUC of 0.684 ± 0.034.
  • Further enhancement was observed when including sex as a covariate and APOE genotypes, leading to an AUC of 0.778 ± 0.043 for the PRS + BrainNetScore + APOE model.

Limitations and Future Work

The authors acknowledge that future work should focus on extending this model to diverse populations, implying that the current study's findings might be specific to the populations represented in the UK Biobank and ADNI datasets. The current model also excluded the APOE region from PRS calculations, which was then added back in later analyses, suggesting potential for further refinement by integrating APOE more deeply into the network construction. The study also highlights the potential of network-based approaches and comprehensive genetic analysis for developing personalized medicine strategies for AD.

Summary updated

Abstract

BrainNetScore significantly improves AD risk prediction when combined with PRSs. This highlights the potential of network-based approaches and comprehensive genetic analysis in developing personalized medicine strategies for AD. Future work should extend this model to diverse populations.
BibTeX
@inproceedings{nam2024brainnetscore,
  author = {Yonghyun Nam and Jakob Woerner and Sang‐Hyuk Jung and Erica H Suh and Haochang Shou and Li Shen and Dokyoon Kim},
  title = {{BrainNetScore: Enhancing Alzheimer’s disease risk prediction using genetic‐guided brain volumetric phenotype network}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.084351},
}
Figure