A 3-fold kernel approach for characterizing Late Onset Alzheimer’s Disease

Margherita Squillario, Federico Tomasi, Veronica Tozzo, Annalisa Barla, Daniela Uberti, for the Alzheimer’s Disease Neuroimaging InitiativebioRxiv 2018 · DOI

Abstract

The purpose of this study is to identify a global and robust signature characterizing Alzheimer’s Disease (AD). Two public GWAS datasets were analyzed considering a 3-fold kernel approach, based on SNPs, Genes and Pathways analysis, and two binary classifications tasks were addressed: cases@controls and APOE4 task. In the SNP signature of the ADNI-1 and ADNI-2 datasets, chromosome 19 and 20 reached high classification accuracy. In addition, the functional characterization of ADNI-1 and ADNI-2 SNP signatures found enriched the same pathway (i.e., Neuroactive ligand-receptor interaction), with GRM7 gene in common with both. TOMM40 was confirmed linked to AD pathology by SNP, gene and pathway-based analyses in ADNI-1. Using this 3-fold kernel approach, a peculiar signature of SNPs, genes and pathways has been highlighted in both datasets. Based on these significant results, we retain such approach a valuable tool to elucidate the heritable susceptibility to AD but also to other similar complex diseases.

Introduction

Late-Onset Alzheimer’s Disease (LOAD) remains one of the most significant challenges in modern neurogenetics. Unlike early-onset AD, which is linked to specific mutations in the APP, PS1, and PS2 genes, LOAD is a multifactorial condition. While it is estimated to be 60-80% heritable, traditional Genome-Wide Association Studies (GWAS) often fail to account for this percentage—a phenomenon known as "missing heritability."

Standard statistical methods typically look for single mutations with large effects. However, LOAD likely results from the cumulative impact of numerous genetic variants, each exerting a subtle influence. To bridge this gap, this research proposes a multivariate machine learning framework designed to identify genetic patterns across multiple biological scales simultaneously.

Methodology: The 3-Fold Kernel Approach

The researchers developed a "3-fold kernel approach," described as a "telescope approach." This method allows researchers to start with a wide-angle view of the genome and progressively zoom in on relevant biological clusters. The analysis is integrated across three distinct levels:

  1. SNP-based (Single Nucleotide Polymorphisms): Utilizing the PALLADIO library, the team applied $l_1l_2$ feature selection to identify the most discriminative individual mutations.
  2. Gene-based: Using the Sequence Kernel Association Test (SKAT), the study examined the aggregate effect of variants within specific genes. This is often more powerful than SNP-level analysis because it captures the combined burden of multiple rare or common variants within a single functional unit.
  3. Pathway-based: Using "Group Lasso with overlap," the researchers analyzed groups of genes that function together in specific biological processes, such as metabolic pathways or signaling cascades.

By integrating these levels, the framework can identify "signatures"—complex groups of genetic features that accurately distinguish AD patients from healthy controls.

Datasets and Platform Challenges

The study utilized two primary datasets from the Alzheimer’s Disease Neuroimaging Initiative (ADNI):

  • ADNI-1: 393 samples (Illumina Human 610-Quad platform).
  • ADNI-2: 281 samples (Illumina Human OmniExpress-24v platform).

A significant hurdle in this research was platform inconsistency. Because the two datasets used different genotyping chips, they shared only approximately 300,000 SNPs in common. This discrepancy served as a "stress test" for the robustness of the 3-fold kernel approach.

Key Insights and Results

The Reproducibility Paradox

A critical finding of the study was that while specific SNP-level signatures showed little overlap between ADNI-1 and ADNI-2, the functional pathways associated with those SNPs were highly consistent. This suggests that while different individuals may possess different specific mutations, those mutations often disrupt the same underlying biological machinery.

Chromosomal Hotspots and Accuracy

Chromosomes 19 and 20 emerged as consistent hotspots for AD classification. On chromosome 19, the model achieved a Balanced Accuracy of 0.94 and a Matthews Correlation Coefficient (MCC) of 0.90 for the APOE4 task (distinguishing carriers of the high-risk APOE4 allele).

Multilevel Biomarkers

  • TOMM40: In the ADNI-1 dataset, the TOMM40 gene was flagged across all three levels (SNP, gene, and pathway). This reinforces its critical role in AD pathology, likely linked to its function in mitochondrial protein transport.
  • GRM7: This gene, involved in glutamate signaling, was identified in the signatures of both datasets, suggesting it is a robust marker for the disease regardless of the genotyping platform.
  • Neuroactive Ligand-Receptor Interaction: This pathway was enriched in both datasets, pointing toward fundamental disruptions in neuronal communication as a core driver of LOAD.
  • Cellular Stress: Pathways related to the "detoxification of reactive oxygen species" and "cellular senescence" were identified as common signatures across both classification tasks and both datasets.

Conclusion

The 3-fold kernel approach provides a more holistic view of the genetic landscape of Alzheimer’s Disease than traditional GWAS. By moving beyond individual SNPs to aggregate gene and pathway effects, the researchers identified robust biological signatures that persist across different data collection platforms. While limited by relatively small sample sizes, this multilevel framework represents a valuable tool for elucidating the heritable susceptibility of AD and other complex, multifactorial diseases.

Summary updated

Abstract

Summary The purpose of this study is to identify a global and robust signature characterizing Alzheimer’s Disease (AD). Two public GWAS datasets were analyzed considering a 3-fold kernel approach, based on SNPs, Genes and Pathways analysis, and two binary classifications tasks were addressed: cases@controls and APOE4 task. In the SNP signature of the ADNI-1 and ADNI-2 datasets, chromosome 19 and 20 reached high classification accuracy. In addition, the functional characterization of ADNI-1 and ADNI-2 SNP signatures found enriched the same pathway (i.e., Neuroactive ligand-receptor interaction), with GRM7 gene in common with both. TOMM40 was confirmed linked to AD pathology by SNP, gene and pathway-based analyses in ADNI-1. Using this 3-fold kernel approach, a peculiar signature of SNPs, genes and pathways has been highlighted in both datasets. Based on these significant results, we retain such approach a valuable tool to elucidate the heritable susceptibility to AD but also to other similar complex diseases.
BibTeX
@misc{squillario2018fold,
  author = {Margherita Squillario and Federico Tomasi and Veronica Tozzo and Annalisa Barla and Daniela Uberti and for the Alzheimer’s Disease Neuroimaging Initiative},
  title = {{A 3-fold kernel approach for characterizing Late Onset Alzheimer’s Disease}},
  year = {2018},
  doi = {10.1101/397760},
  howpublished = {bioRxiv},
}
Figure