Integrating Transcriptomics, Genomics, and Imaging in Alzheimer’s Disease: A Federated Model

Jianfeng Wu, Yanxi Chen, Panwen Wang, Richard J. Caselli, Paul M. Thompson, Junwen Wang, Yalin Wang, for the Alzheimer’s Disease Neuroimaging InitiativebioRxiv 2021 · DOI

TL;DR

This paper introduces GEIDI, a novel federated learning model that integrates brain imaging, genetic, and gene expression data to uncover complex relationships in Alzheimer's Disease (AD). By analyzing these multi-omics data in a privacy-preserving, distributed manner, GEIDI offers a more holistic understanding of AD and can identify novel genetic targets for drug development, outperforming existing methods.

What Problem Does This Paper Address?

Alzheimer's Disease (AD) is a complex, multifactorial neurodegenerative disorder with a growing public health impact. While structural brain imaging (sMRI), genome-wide association studies (GWAS), and transcriptomics (gene expression) are crucial for understanding AD, most research has focused on pairwise relationships (e.g., sMRI and GWAS, or sMRI and transcriptomics). There's a significant gap in methods that can effectively integrate all three data modalities to reveal deeper, personalized insights into how genetic variations influence brain structure and gene expression, ultimately impacting AD progression. Furthermore, the sensitive nature of biomedical data makes centralized data pooling challenging, necessitating privacy-preserving approaches.

Methods

The core innovation is the Federated Genotype-Expression-Imaging Data Integration (GEIDI) model. GEIDI leverages a generalized Chow test to detect differences in relationships between imaging biomarkers (X) and gene expression (Y) across different genetic subgroups (stratified by genotype, e.g., APOE status or specific SNPs). This allows for personalized inferences, as relationships can be genotype-dependent.

Key methodological aspects include:

  • Federated Learning: GEIDI is designed as a federated model, enabling distributed analysis across multiple hypothetical institutions without centralizing raw data. This addresses privacy concerns and allows for scaling to larger datasets.
  • Generalized Chow Test: The standard Chow test, typically used in econometrics for structural breaks, is adapted to assess differences in regression coefficients across genetic subgroups.
  • Linear Regression: Within each subgroup and for the overall population, linear regression is used to model the relationships between imaging biomarkers and gene expression.
  • Privacy Preservation: The federated approach ensures data privacy by only sharing aggregated gradients, not raw data, between local institutions and the global center.

Datasets

The primary dataset used for experiments is the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. This publicly available dataset includes:

  • Genotype data: Including known AD risk genes (e.g., APOE) and SNPs from GWAS.
  • Gene expression data: From microarray-based transcriptomic profiling of blood samples (for 20,211 genes).
  • Imaging biomarkers: From structural magnetic resonance imaging (sMRI) data of subjects' brains.

The paper also references data from the TADPOLE challenge for brain imaging biomarkers.

Key Insights and Ideas

  • Genotype-Specific Relationships: The most significant insight is that the relationship between brain structure and gene expression is not uniform across individuals. Stratifying by genotype (e.g., APOE status or specific SNPs) reveals stronger, more specific correlations that are diluted in a mixed population.
  • Personalized Medicine Potential: This genotype-dependent insight is crucial for personalized medicine, suggesting that therapeutic targets or drug responses might differ based on an individual's genetic makeup.
  • Federated Learning for Genomics: GEIDI demonstrates the feasibility and effectiveness of applying federated learning to complex multi-omics data integration, overcoming data privacy and accessibility barriers.
  • Holistic AD Understanding: By integrating imaging, genetics, and transcriptomics, GEIDI provides a more comprehensive view of AD's underlying mechanisms than pairwise approaches.

Key Results

  • Improved AD-Related Gene Discovery: GEIDI demonstrated superior performance in detecting AD-related gene expressions compared to state-of-the-art Matrix eQTL methods. For example, using hippocampus volume, GEIDI identified 1,625 significant gene expressions with a hypergeometric enrichment p-value of 0.00036 for known AD genes.
  • Enhanced AD-Related SNP Detection: GEIDI also showed better performance in detecting AD-related SNPs, achieving a higher True Positive Rate (TPR) compared to Matrix eQTL's linear regression and ANOVA models. For instance, in one experiment, GEIDI achieved a TPR of 0.57 for the top 500 SNPs.
  • Discovery of Novel Genes: The model identified genes not previously reported to be associated with AD, such as XRCC1 and SEC14L2, highlighting its potential for discovering new therapeutic targets.
  • Stability and Privacy: The federated GEIDI model demonstrated stable performance across different multi-site conditions and effectively preserved data privacy.

Limitations and Future Work

  • Dataset Size: The initial evaluation was performed on a relatively small dataset (697 subjects from ADNI). The authors plan to incorporate larger and more diverse datasets (e.g., Arizona APOE cohort, UK Biobank) to enhance robustness and reliability.
  • Imaging Biomarker Choice: While hippocampus volume showed strong performance, the authors acknowledge that other imaging measurements, such as surface-based morphometry (e.g., univariate morphometry index - UMI), might capture more nuanced brain pathology and could potentially lead to the detection of additional AD-related genes.
  • Blood-Based Biomarkers: Future work will explore integrating blood-based biomarkers (e.g., plasma amyloid-beta, P-Tau181) into the GEIDI model to further understand AD-related genotypes and their effects.

Summary updated

Abstract

Abstract Alzheimer’s disease (AD) affects more than 1 in 9 people age 65 and older and becomes an urgent public health concern as the global population ages. In clinical practice, structural magnetic resonance imaging (sMRI) is the most accessible and widely used diagnostic imaging modality. Additionally, genome-wide association studies (GWAS) and transcriptomics – the study of gene expression – also play an important role in understanding AD etiology and progression. Sophisticated imaging genetics systems have been developed to discover genetic factors that consistently affect brain function and structure. However, most studies to date focused on the relationships between brain sMRI and GWAS or brain sMRI and transcriptomics. To our knowledge, few methods have been developed to discover and infer multimodal relationships among sMRI, GWAS, and transcriptomics. To address this, we propose a novel federated model, Genotype-Expression-Imaging Data Integration (GEIDI), to identify genetic and transcriptomic influences on brain sMRI measures. The relationships between brain imaging measures and gene expression are allowed to depend on a person’s genotype at the single-nucleotide polymorphism (SNP) level, making the inferences adaptive and personalized. We performed extensive experiments on publicly available Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset. Experimental results demonstrated our proposed method outperformed state-of-the-art expression quantitative trait loci (eQTL) methods for detecting genetic and transcriptomic factors related to AD and has stable performance when data are integrated from multiple sites. Our GEIDI approach may offer novel insights into the relationship among image biomarkers, genotypes, and gene expression and help discover novel genetic targets for potential AD drug treatments.
BibTeX
@misc{wu2021integrating,
  author = {Jianfeng Wu and Yanxi Chen and Panwen Wang and Richard J. Caselli and Paul M. Thompson and Junwen Wang and Yalin Wang and for the Alzheimer’s Disease Neuroimaging Initiative},
  title = {{Integrating Transcriptomics, Genomics, and Imaging in Alzheimer’s Disease: A Federated Model}},
  year = {2021},
  doi = {10.1101/2021.09.14.460367},
  howpublished = {bioRxiv},
}
Figure