PhenoEncoder: A Discriminative Embedding Approach to Genomic Data Compression

Abstract

AbstractExploring the heritability of complex genetic traits requires methods that can handle the genomes vast scale and the intricate re-lationships among genetic markers. Widely accepted association studies overlook non-linear effects (epistasis), prompting the adoption of deep neural networks (DNNs) for their scalability with large genetic datasets and ability to detect complex relationships. However, the curse of di-mensionality continues to limit the potential of DNNs, underscoring the critical need for dimensionality reduction for suitably sizing and shaping the genetic inputs, while preserving epistasis. Linkage disequilibrium (LD), a measure of correlation between genetic loci, offers a pathway to genome compression with minimal information loss. Using LD, the genome can be divided into smaller genomic regions, i.e., haplotype blocks, which can be locally compressed using deep au-toencoders. While autoencoders excel at preserving the main non-linear patterns, they still risk losing phenotype-relevant information when dom-inated by other sources of genetic variation. We propose a novel approach, PhenoEncoder, that incorporates pheno-typic variance directly into compression. This single nucleotide polymor-phism (SNP)-based pipeline employs multiple autoencoders, each dedi-cated to compressing a single haplotype block. The window-based spar-sity of the model eases the computational burden of simultaneously pro-cessing numerous SNPs. Concurrently, an auxiliary classifier predicts the phenotype from the compressed haplotype blocks. Epistasis is processed both within and between haplotype blocks by maintaining non-linearity in the autoencoders and the classifier. Through joint optimization of the compression and classification losses, PhenoEncoder ensures that disease-causing patterns are highlighted during compression. Applied to protein expression and simulated complex phenotype datasets, PhenoEncoder demonstrated enhanced generalizability in downstream classification tasks compared to standard autoencoder compression. By enabling phenotype-aware compression, PhenoEncoder emerges as a promis-ing approach for discriminative genomic feature extraction.