Celedon, J. | Search bioRxiv

bioRxiv2024

Finding Salient Multi-Omic Interactomes Relevant to Multiple Biomedical Outcomes using Graph Ensemble Neural Networks

Although multi-omics integration relevant to patient outcome is typically characterized by an analyte interactome, current multi-omic integration methods either (1) model outcome without directly including associations between analytes, (2) model the interactome without directly evaluating the saliency of the model in the context of outcome, or (3) model outcome in a high-dimensional parameter space not suitable for small sample sizes (which are common in multi-omics studies). We introduce Graph Ensemble Neural Network (GENN), a methodology that learns the interactome most predictive of outcome in a low-dimensional parameter space built on complementary attributes for all possible analyte associations (metafeatures). We show that GENN is robust to noise in measurements using a theoretical model, outperforms the predictive performance of existing methods when evaluated on Tegafur drug response in NCI-60 cancer cell line data, and uncovers potentially novel multi-omic mechanisms driving total serum IgE levels in pediatric asthma and patient survival in glioblastomas.

bioinformatics↗

bioRxiv2020

A region-based method for causal mediation analysis of DNA methylation data

Exposure to environmental factors can affect DNA methylation at a CpG site or a genomic region, which can then affect an outcome. In other words, environmental effects on an outcome could be mediated by DNA methylation. To date, single CpG site-based mediation analysis has been employed extensively. More recently, however, there has been considerable interest on studying differentially methylated regions (DMRs), both because DMRs are more likely to have functional effects than single CpG sites and because testing DMRs reduces multiple testing. In this report, we propose a novel causal mediation approach under the counterfactual framework to test the significance of total, direct and indirect effects of predictors on response variable with a methylated region (MR) as the mediator (denoted as MR-Mediation). Functional linear transformation is used to reduce the possible high dimension of the CpG sites in a predefined methylated region and to account for their location information. In our simulation studies, MR-Mediation retained the desired Type I error rates for total, direct and indirect effect tests, for both continuous and binary outcomes. Furthermore, MR-Mediation had better power performance than testing mean methylation level as the mediator in most considered scenarios, especially for indirect effect (i.e., mediated effect) test, which could be more interesting than the other two effect tests. We further illustrate our proposed method by analyzing the methylation mediated effect of exposure to gun violence on total immunoglobulin E (IgE) or atopic asthma among participants in the Epigenetic Variation and Childhood Asthma in Puerto Ricans (EVA-PR) study.

genomics↗