Genome-wide prediction and integrative functional characterization of Alzheimer’s disease-associated genes

Cuixiang Lin, Hong‐Dong Li, Chao Deng, Weisheng Liu, Shannon Erhardt, Fang‐Xiang Wu, Xing‐Ming Zhao, Jun Wang, Daifeng Wang, Bin Hu, Jianxin WangbioRxiv 2021 · DOI

TL;DR

This paper presents a novel computational approach to identify and functionally characterize genes associated with Alzheimer's Disease (AD) by integrating multiple biological networks. The developed pipeline, ALZLINK, not only predicts novel AD-associated genes with high accuracy but also provides functional evidence and potential regulatory variants, serving as a valuable resource for AD research.

What Problem Does This Paper Address?

Alzheimer's Disease (AD) is a complex neurodegenerative disorder with a significant genetic component, yet many of its underlying mechanisms and associated risk genes remain elusive. Current methods for identifying AD-associated genes, such as Genome-Wide Association Studies (GWAS), have identified some genes but explain only a fraction of the disease's heritability. This leaves a substantial gap in our understanding, necessitating more comprehensive approaches to discover and functionally validate novel AD-associated genes. The challenge lies in integrating diverse biological data to capture the complex interplay of genes involved in AD pathogenesis.

Methods

The core of this work is a machine learning pipeline that leverages brain-specific Functional Gene Networks (FGNs). FGNs represent genes as nodes and their co-functional probabilities as edges, capturing functional relationships. The authors constructed a feature matrix for each gene based on its co-functional probability with known AD-associated genes. They then employed state-of-the-art machine learning models, specifically ExtraTrees, to predict AD-associated genes.

Key methodological contributions include:

  • Pathway-based negative gene selection: Instead of random sampling, they identified non-AD genes by removing those enriched in AD-related pathways, aiming for more informative negative controls.
  • Integration of multiple biological networks: The approach goes beyond FGNs by validating predictions using protein-protein interaction (PPI) networks, coexpression networks, miRNA-target interaction networks, and transcriptional regulatory networks.
  • Identification of potential causal variants: By integrating brain eQTL and ATAC-seq data, they pinpointed specific genetic variants (SNPs) that might regulate the expression of predicted AD-associated genes.
  • ALZLINK web interface: A user-friendly platform was developed to make the predicted genes, their scores, and supporting functional evidence accessible to the research community.

Datasets

The study utilized several publicly available datasets:

  • AD-associated gene resources: AlzGene, AlzBase, OMIM, DisGeNet, DistiLD, UniProt, Open Targets, GWAS Catalog, and ROSMAP (for differentially expressed genes).
  • Brain-specific FGNs: Ten networks were obtained from the GIANT and BaiHui databases.
  • Independent datasets for validation:
    • Mayo RNA-seq dataset (for coexpression analysis).
  • Independent MSBB RNA-seq dataset (for correlation with neuropathological and clinical traits).
  • ROSMAP dataset (for differential expression analysis).
  • PPI networks: HuRI and STRING databases.
  • miRNA-target interaction data: miRTarBase and Human microRNA Disease Database (HMDD).
  • Transcriptional regulatory network: TRRUST database.
  • eQTL and ATAC-seq data: GTEx, PsychEncode, CommonMind Consortium, and BOCA database.

No new benchmark datasets were introduced; rather, the study integrated and analyzed existing resources.

Key Insights and Ideas

  • Brain FGNs are powerful predictors: The study demonstrates that brain-specific FGNs, when combined with machine learning, can effectively predict AD-associated genes, outperforming random selection.
  • Functional enrichment is key: Top-ranked predicted genes are not just statistically associated but are also functionally enriched in AD-related pathways and phenotypes, providing strong biological plausibility.
  • Multi-modal evidence strengthens predictions: Integrating data from various biological networks (PPI, coexpression, miRNA, TRN) significantly enhances the confidence in predicted AD-associated genes.
  • Potential for therapeutic targets: The identification of genes with strong functional links to AD pathology and potential causal variants opens new avenues for therapeutic target discovery.
  • ALZLINK as a community resource: The development of a web interface democratizes access to these predictions and the underlying evidence, accelerating AD research.

Key Results

  • The developed machine learning model achieved high performance with an average AUROC of 0.91 and AUPRC of 0.76 based on cross-validation.
  • Literature search confirmed that 12 out of the top 20 predicted genes had existing evidence of association with AD.
  • The top-ranked genes were significantly enriched in AD-related gene sets and pathways, including learning/memory, cognition, and amyloid-beta pathways.
  • Gene modules identified from an integrated network were significantly correlated with AD-related traits (CERAD, Braak, and CDR scores).
  • A set of 36 AD-associated genes were identified, supported by multiple lines of evidence, with potential causal variants identified for 7 of them.
  • The ALZLINK web resource was created, providing access to predicted genes and their functional evidence.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • No causality implication: The predictions indicate statistical association, not direct causality.
  • Gene-level predictions: The model predicts genes but does not differentiate between splice isoforms, which can have different or even opposite functions. Future work could involve isoform-level prediction.
  • Cell-type specificity: The human brain is heterogeneous, and the association of predicted genes with AD in different cell types needs further investigation. Integrating single-cell genomic data is suggested.
  • Generalizability of the pipeline: While the pipeline is presented as generic, its application to other diseases is a future direction.
  • Improving negative gene selection: The authors note that different negative gene selection strategies could introduce bias.

The study's findings are presented as a valuable resource for experimental validation and further research into AD pathogenesis and potential therapeutic interventions.

Summary updated

Abstract

Abstract The mechanism of Alzheimer’s disease (AD) remains elusive, partly due to the incomplete identification of risk genes. We developed an approach to predict AD-associated genes by learning the functional pattern of curated AD-associated genes from brain gene networks. We created a pipeline to evaluate disease-gene association by interrogating heterogeneous biological networks at different molecular levels. Our analysis showed that top-ranked genes were functionally related to AD. We identified gene modules associated with AD pathways, and found that top-ranked genes were correlated with both neuropathological and clinical phenotypes of AD on independent datasets. We also identified potential causal variants for genes such as FYN and PRKAR1A by integrating brain eQTL and ATAC-seq data. Lastly, we created the ALZLINK web interface, enabling users to exploit the functional relevance of predicted genes to AD. The predictions and pipeline could become a valuable resource to advance the identification of therapeutic targets for AD.
BibTeX
@misc{lin2021genome,
  author = {Cuixiang Lin and Hong‐Dong Li and Chao Deng and Weisheng Liu and Shannon Erhardt and Fang‐Xiang Wu and Xing‐Ming Zhao and Jun Wang and Daifeng Wang and Bin Hu and Jianxin Wang and Jianxin Wang and Jianxin Wang},
  title = {{Genome-wide prediction and integrative functional characterization of Alzheimer’s disease-associated genes}},
  year = {2021},
  doi = {10.1101/2021.02.09.430536},
  howpublished = {bioRxiv},
}
Figure