Abstract
Currently, the ability to analyze large-scale whole genome sequence (WGS) data is limited due to both the size of the data and the inability of many existing tools to scale. To address this challenge, we use data “tiling” to efficiently partition whole genome sequences into smaller segments resulting in a simple numeric matrix of small integers. This lossless representation is particularly suitable for machine learning (ML) models. As part of the Artificial Intelligence for Alzheimer’s Disease (AI4AD) consortium, we analyzed 15,000+ genomes from the Alzheimer’s Disease Sequencing Project (ADSP) and the Alzheimer’s Disease Neuroimaging Initiative (ADNI). We identify and prioritize tile variants that are possible genetic signatures for AD. The model shows added predictive value from variants of genes previously found to be associated with AD risk, including the APOE variant (rs429358).
Introduction
Whole Genome Sequencing (WGS) provides a near-complete picture of an individual's DNA, but its scale—roughly 3 billion base pairs per human genome—presents a significant computational challenge. Traditional Genome-Wide Association Studies (GWAS) often simplify this by focusing on specific Single Nucleotide Polymorphisms (SNPs), potentially ignoring critical structural variations. Furthermore, the "curse of dimensionality" arises when millions of genetic predictors ($p$) are analyzed against only thousands of patients ($n$). The AI4AD consortium sought a lossless, compact representation of the entire genome that is reference-independent and optimized for machine learning.
Methods: Whole Genome Tiling and Analysis Pipeline
The core innovation presented is Whole Genome Tiling. This method partitions the genome into overlapping segments called "tiles" using fixed 24-base sequences (24-mers) as "tags." These tags act as anchors, and the sequence between two tags defines a "tile variant."
Instead of billions of base pairs, each genome is represented as a numeric matrix of integers, where each integer corresponds to a specific version of a tile in a global "tile library." This reduces the human genome to approximately 10 million tile positions, a 300x reduction in feature space.
To identify significant genetic signatures, the researchers employed a multi-stage analysis pipeline:
- Weight-of-Evidence (WoE) Encoding: Categorical tile variants are transformed into continuous numbers based on their statistical relationship with the target phenotype (AD case vs. control).
- Feature Selection: Initial filtering by p-values (threshold $1 \times 10^{-8}$) is followed by a Bootstrapped Adaptive Lasso. This regularized regression technique identifies the most "stable" genetic signatures by shrinking less important variables to zero.
Datasets
The study utilized a massive repository of genomic data, primarily from the Alzheimer’s Disease Sequencing Project (ADSP) and the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The final cohort consisted of 11,762 genomes (4,983 AD cases and 6,779 controls), with a mean age of 73.8 years. For validation and population structure analysis, data from the 1000 Genomes Project and other public benchmarks were also tiled.
Key Results and Insights
The tiling approach demonstrated several advantages:
- Reference Independence: Because tiles are defined by flanking tags rather than absolute coordinates, data from different reference genomes (e.g., GRCh37 and GRCh38) can be harmonized without lossy conversions.
- Validation of Known Loci: The pipeline successfully re-discovered the APOE variant (rs429358), which showed the strongest statistical significance.
- Novel Signatures: Stable variants were identified in genes such as TOMM40, APOC1, and MIB1, many of which are associated with brain morphology and amyloid deposition.
- Predictive Power: A baseline model using APOE4, age, and sex achieved an AUC of 0.61. Incorporating the "stable" tile variants improved the AUC to 0.63 on the validation set.
- Scalability: Unlike SNP-based methods where feature counts explode with the discovery of rare variants, the number of tile positions remains constant at ~10 million as cohort sizes increase.
Limitations and Future Work
Current limitations include the use of "unphased" WGS data, where maternal and paternal chromosome mutations are not distinguished. Future efforts will incorporate globally phased data to improve accuracy. Additionally, while this proof-of-concept focused largely on White (Non-Hispanic/Latino) sequences, expanding to diverse populations is a priority. The AI4AD consortium aims to eventually integrate these genetic tiles with multi-modal data, including MRI/PET imaging and clinical cognitive metrics, to enhance AD risk prediction and progression modeling.
Summary updated