multiDEGGs: a multi-omic differential network analysis package for biomarker discovery and predictive modeling

Modern clinical trials increasingly leverage high-throughput omic data for patient stratification and biomarker discovery. While traditional differential gene expression analysis disregards the networked nature of molecular entities and produces extensive gene lists with limited interpretability, differential network analysis has emerged as a crucial complementary analysis for comparative studies. Here we present multiDEGGs, a CRAN R package that enables differential network analysis in multi-omic scenarios.

multiDEGGs uses a multi-layer graph framework to model omic data by leveraging an internal network of over 10 000 literature-validated biological interactions. For each data type, differential networks are generated, and the statistical significance of each link (p-values or adjusted p-values) is evaluated through robust linear regression with interaction terms. These networks are then integrated into a comprehensive visualisation that allows interactive exploration of cross-omic patterns.

Beyond network visualization and exploration, multiDEGGs extends its utility to predictive modelling applications. The package facilitates seamless integration into cross-validation machine learning pipelines, serving as feature selection and augmentation tool.

We validated multiDEGGs using two cohorts of rheumatoid arthritis patients who underwent tocilizumab and rituximab therapy, respectively. For each treatment group, multi-layer differential interactions were identified, and seven machine learning models were trained to predict treatment resistance using synovial RNA-seq data. We systematically compared multiDEGGs against five traditional feature selection methods. On average, AUC values obtained with multiDEGGs showed an improvement of 0.10 compared to conventional filters.

KEY POINTS

• Traditional gene expression analysis leaves researchers with hundreds of ‘significant’ genes but no clear biological story. The multiDEGGs CRAN package shifts the focus: instead of asking which genes change, it asks which gene relationships change.

• It can be used with single or multi-omic data: differential networks are calculated separately for each data type, with results integrated into a comprehensive, interactive view.

• multiDEGGs can be combined with the nestedcv CRAN package (nested cross-validation) to serve as feature selection and augmentation tool.

• In comparative evaluations, machine learning models trained with multiDEGGs-selected features showed AUC improvements of 0.10 compared to other feature selection methods.