The FENRIR framework included two parts: first, a Bayesian integration model created tissue-specific enhancer functional networks; and second, a network-based machine learning model prioritized disease-enhancer association in a tissue-specific manner.
This work was predicated on the hypothesis that an enhancer will have a strong tissue-specific functional relationship with another enhancer if they share particular attributes: interacting in the 3D genome, co-regulated by TFs, and/or regulating genes functionally related in a tissue. Using a Bayesian integration model, FENRIR integrated TF co-regulation for pairs of enhancers, chromatin interactions between enhancers and genes with tissue-specific functional relatedness for pairs of genes. Each resulting FENRIR network included probabilistically weighted interactions between pairwise enhancers, representing the functional similarity of these enhancers in a given tissue; and probabilities for physically interactive enhancers and genes, representing their tissue-specificity. All pre-built tissue-sppecific enhancer networks are downloadable from https://fenrir.flatironinstitute.org/download/
For the second FENRIR model of disease-enhancer association prediction, a learned context-specific enhancer network was combined with disease knowledge (e.g. GWAS SNPs or disease genes). To do this we trained a network-based classifier using elastic net logistic regression and prioritized disease associations for all enhancers. A web interface for disease-associated enhancer perdiction is available at https://fenrir.flatironinstitute.org/predict/.