Computing linkage disequilibrium aware genome embeddings using autoencoders

Abstract

Motivation: The completion of the genome has paved the way for genome-wide association studies (GWAS), which explained certain proportions of heritability. GWAS are not optimally suited to detect non-linear effects in disease risk, possibly hidden in non-additive interactions (epistasis). Alternative methods for epistasis detection using, e.g. deep neural networks (DNNs) are currently under active development. However, DNNs are constrained by finite computational resources, which can be rapidly depleted due to increasing complexity with the sheer size of the genome. Besides, the curse of dimensionality complicates the task of capturing meaningful genetic patterns for DNNs; therefore necessitates dimensionality reduction.

Results: We propose a method to compress single nucleotide polymorphism (SNP) data, while leveraging the linkage disequilibrium (LD) structure and preserving potential epistasis. This method involves clustering correlated SNPs into haplotype blocks and training per-block autoencoders to learn a compressed representation of the block's genetic content. We provide an adjustable autoencoder design to accommodate diverse blocks and bypass extensive hyperparameter tuning. We applied this method to genotyping data from Project MinE, and achieved 99% average test reconstruction accuracy-i.e. minimal information loss-while compressing the input to nearly 10% of the original size. We demonstrate that haplotype-block based autoencoders outperform linear Principal Component Analysis (PCA) by approximately 3% chromosome-wide accuracy of reconstructed variants. To the extent of our knowledge, our approach is the first to simultaneously leverage haplotype structure and DNNs for dimensionality reduction of genetic data.

Availability and implementation: Data are available for academic use through Project MinE at https://www.projectmine.com/research/data-sharing/, contingent upon terms and requirements specified by the source studies. Code is available at https://github.com/gizem-tas/haploblock-autoencoders.

Figure 1.

An overview of the workflow.

Figure 2.

Visualizations of rectangular and elliptic autoencoders. Both networks have the same input size, number of hidden layers, and bottleneck size ($\text{inputs}=10$, $hl=4$, $bn=3$) and only differ by the slope p, see also Equation (1).

Figure 3.

Elbow plot of validation reconstruction accuracies of the best models across bn$\in \{1,\dots ,10\}$, averaged over 221 haploblocks, with 95% confidence intervals. The black line is the elbow point.

Figure 4.

The highest validation accuracies obtained from the grid search for each block, with bottleneck values ranging from 1 to 10. The accuracies are plotted against the number of SNPs in each block (left) and against the within-block average pairwise LD (right). The zoomed-in plot on the right focuses on bottleneck values 2 and 3, offering a clearer comparison and highlighting the LD threshold at 0.4.

Figure 5.

Box plot of SNP accuracies obtained from 2854 autoencoders used for compression of all haploblocks in Chromosome 22. Accuracies were also calculated for three categories of genotype dosage values (0 s, 1 s, and 2 s), and for two phenotypic groups (ALS cases and controls) as displayed on the y-axis.

Figure 6.

Scatter plots of SNP accuracies obtained by AE and PCA for dosage values 0, 1, and 2 (top to bottom). The x-axes represent block sizes, shared across plots. Solid lines illustrate the moving average of SNP accuracies on training samples using a window size of 50, dotted lines depict the same for test accuracies.