Computing linkage disequilibrium aware genome embeddings using autoencoders.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genome-wide association studies (GWAS) have advanced our understanding of heritability but struggle to detect complex non-linear genetic effects like epistasis.
• Deep neural networks (DNNs) show promise for epistasis detection but face computational challenges due to large genomic datasets and the curse of dimensionality.
• Dimensionality reduction is crucial for efficiently analyzing complex genetic data with DNNs.

Purpose of the Study:

• To propose a novel dimensionality reduction method for single nucleotide polymorphism (SNP) data that preserves epistasis.
• To leverage linkage disequilibrium (LD) structure by clustering SNPs into haplotype blocks.
• To develop an autoencoder-based approach for compressing genetic data within haplotype blocks.

Main Methods:

• Clustering of correlated SNPs into haplotype blocks.
• Training of per-block autoencoders to learn compressed genetic representations.
• Application of the method to Project MinE genotyping data.

Main Results:

• Achieved 99% average test reconstruction accuracy, indicating minimal information loss.
• Compressed genetic data to approximately 10% of its original size.
• Demonstrated superior performance over Principal Component Analysis (PCA), with a 3% increase in chromosome-wide accuracy for reconstructed variants.

Conclusions:

• The proposed method effectively compresses SNP data using haplotype blocks and autoencoders, preserving crucial epistasis information.
• This approach offers a computationally efficient solution for analyzing large genomic datasets with DNNs.
• The method outperforms linear dimensionality reduction techniques like PCA in reconstructing genetic variants.