Transfer Learning of a Methylation-Based COVID-19 Severity Model for DNA Viral Infection by Eczema Herpeticum

Authors Wenyu Zhou, Meher Preethi Boorgula, Monica Campbell, Sameer Chavan, Genelle F. Harrison, Brett R. Peterson, Nicholas Rafaels, Bret Barnes, Rishi Porecha, Rasika A. Mathias, Alem Taye, Kathleen C. Barnes

Methylation risk scores (MRS) are an algorithmic aggregation of methylation states that are increasingly used to infer clinical phenotypes. An advantage of MRS is that it captures multi-factorial predispositions. This possibility is due to methylation states being influenced by both genetic and environmental factors. Whether MRS can be universally applied across closely relevant phenotypes remains unexplored. In addition to this, factors that need to be taken into account for reaching a satisfactory model generalization are undetermined. DNA and RNA viruses have evolved to utilize epigenetic modifications including DNA methylation as a means to maximize viral gene expression and to escape host immunity during infection. Previously, we customized Illumina’s MethylationEPIC array to enhance immune response detection. We profiled peripheral blood samples from patients infected with SARS-CoV-2 (a RNA virus) thereby developing a COVID-19 MRS model to predict disease severity. In the current study, we aimed to assess the utility of COVID-19 MRS model in DNA viral infections. The clinical outcome of focus was atopic dermatitis (AD)-associated eczema herpeticum (ADEH), a severe complication in AD caused by disseminated herpes simplex virus (HSV, a double-stranded DNA virus). We profiled peripheral blood DNA samples with the customized Infinium MethylationEPIC array from 258 ADEH+ and 100 ADEH-
patients with disease severity measures. We further compared data to the COVID-19 positive and negative controls from our recently reported study (Konigsberg et al 2021, Comm. Med.).

Firstly, we demonstrated the COVID-19 MRS model accuracy on ADEH severity prediction. We then stratified model performance based on cohort characteristics to understand contributing factors that influence model transferability and generalizability. Next, we trained a separate MRS model for specificity to ADEH cases alone. Both models were compared to quantify specificity of methylation patterns induced by different viral types. In summary, our study reveals the occurrence of both common and differential methylation states in infectious diseases caused by different viral classes. In addition, our study provides evidence to delineate host DNA modifications utilized by unique viruses to cause infections in humans. Overall, this work serves as a case study to evaluate the validity and utility of MRS in its implementation towards infectious disease healthcare.