Enhancer-Promoter Interaction Prediction

EPI prediction methods fall into three broad categories:

• Correlation-based: Leverage the correlation between enhancer-associated epigenomic marks (e.g., H3K27ac, ATAC-seq) and target gene expression across cell types or conditions. These methods assume that co-activated regulatory elements interact.
• Supervised learning: Train classifiers on experimentally validated EPIs using features such as genomic distance, CTCF motif orientation, conservation scores, and epigenomic signal at both loci. Random forests and gradient-boosted trees were early standards.
• Deep learning from 3D structure: Use models like Akita or DeepC that predict Hi-C contact maps from DNA sequence, then extract pairwise contact probabilities between annotated enhancers and promoters. Graph neural networks (GNNs) operating on chromatin interaction graphs represent the current state-of-the-art, learning latent representations of regulatory elements from their local sequence context and global 3D neighborhood.

Modern EPI prediction frames the genome as a heterogeneous graph where nodes represent genomic elements (enhancers, promoters, CTCF sites) and edges represent either linear adjacency or 3D contacts. Graph Neural Networks (GNNs) learn node embeddings by aggregating features from local neighborhoods through multiple message-passing layers. Key architectural components include:

• Graph Attention Networks (GATs): Learn to weight the importance of different neighboring elements, allowing the model to focus on functionally relevant contacts.
• Edge prediction decoders: Compute the probability of an enhancer-promoter link from the learned node embeddings using a dot-product or a small MLP.
• Negative sampling: Critical for training; non-interacting enhancer-promoter pairs from matched genomic distance distributions are used as negative examples to prevent the model from simply learning distance.

EPI prediction directly informs variant interpretation in clinical genomics. Non-coding variants identified by GWAS or whole-genome sequencing often fall within enhancers. Predicting which promoter a variant-linked enhancer contacts enables:

• Target gene identification: Linking risk variants to their effector genes, a prerequisite for drug target discovery.
• Structural variant annotation: Predicting how large deletions, inversions, or duplications rewire the 3D genome, creating enhancer hijacking events that drive oncogene activation.
• Cell-type specificity: EPIs are highly cell-type-specific. Models trained on diverse ENCODE and Roadmap Epigenomics data can predict tissue-specific regulatory consequences of non-coding mutations.