Background: Growing evidence has demonstrated the effect of biological and environmental factors on the pathogenesis of COVID-19. We hypothesized that the accuracy of prediction analysis on COVID-19 susceptibility can be improved by considering multi-omics layers, including genomic, transcriptomic, metabolomic, and exposomic factors, through integrative multi-omics analysis.
Methods: First, we examined the prediction accuracy of single-omics, pairwise-omics, and multi-omics models in predicting COVID-19 susceptibility using the best linear unbiased prediction (BLUP) method. Second, we assessed the discriminatory power of omics layers in distinguishing between individuals with positive and negative COVID-19 status, employing the area under the receiver operating characteristic curve (AUC) as a performance metric. This analysis used case-control COVID-19 data (N=107857). Lastly, we evaluated improvements in genomic prediction by incorporating trait-specific biological priors through gene-ontology (GO) term enrichment analysis and a machine learning algorithm.
Results: Our analysis revealed that the exposome exhibited the highest prediction accuracy (R2=4%), followed by the transcriptome and metabolome, each contributing an R2 of 1.4%. Conversely, genomic data had negligible effect on COVID-19 prediction, contributing only 0.1%. However, leveraging biological priors, such as considering GO-enriched markers, significantly improved the accuracy of genomic prediction to 0.5%. Higher-level omics models combining transcriptome and exposome emerged as the best predictive model for COVID-19 (5.5%). This model also showed modest performance (AUC=67%) in classifying case-control COVID-19 data.
Conclusion: Our findings suggest that traits with low SNP-based heritability, such as COVID-19 susceptibility, necessitate the inclusion of multiple omics layers and leveraging prior biological knowledge to achieve optimal prediction accuracy.