Detecting disease comorbidity based on SNP association on PheWAS scale

Disease comorbidity presents special challenges to disease diagnosis and treatments. Much effort has been made in understanding the genetic causes, tracing back to individual genes, gene-gene interactions, and even to specific single nucleotide polymorphism (SNPs). In this work, we develop a machine learning method to detect comorbidity based on disease association with SNPs reported in UK BioBank PheWAS data. Due to the high dimensionality of SNP data, a common technique - principal component analysis (PCA) -- is applied to reduce the dimension. Despite the information loss, a neural network trained on the reduced SNP vectors outperforms the state-of-the-art method on using the same dataset. Further, we exploit the disease-disease network derived from SNP association to compensate the information loss and significantly improve the prediction performance in cross-validation experiments. Moreover, using our SNP based approach together with a random forest classifier, we are able to trace back from the informative features to the top SNPs that contribute most to the accurate prediction of comorbidity. Our analysis shows that these top SNPs exhibit statistically significantly different antagonistic and synergistic patterns as compared to randomly selected SNPs, which may help with the efforts in uncovering the root cause of comorbidity for disease pairs.