ETNet: an interpretable transformer framework for enhancer-enhancer interaction prediction with cross-context transferability.

Enhancer-enhancer interactions (EEIs) are critical regulatory components in transcriptional networks but remain computationally challenging to predict. While enhancer-promoter interactions have been extensively studied, EEIs remain comparatively underexplored. We developed ETNet (Enhancer-enhancer Interaction Explainable Transformer Network), a deep learning architecture integrating convolutional neural networks with Transformer modules to predict EEIs from DNA sequences. Evaluation across three cell lines (GM12878, K562, MCF-7) demonstrated superior performance compared to existing methods including EnContact, with statistical significance confirmed through DeLong tests across six cell lines. Rigorous validation through cross-validation and enhancer-level data partitioning confirmed robust generalization. ETNet exhibited effective cross-cell type transfer learning and showed transferability to enhancer-promoter interaction tasks, providing exploratory evidence for shared chromatin interaction principles. Feature attribution analysis recovered cell-type-specific regulatory motifs consistent with known transcription factors and revealed computational evidence for super-additive cooperative mechanisms, with cooperativity negatively correlating with sequence similarity-patterns representing hypothesis-generating observations requiring experimental validation. Proof-of-concept analysis demonstrated how single-nucleotide polymorphisms in JAK-STAT pathway genes may influence predicted interactions through motif alterations. ETNet advances computational approaches for studying enhancer interactions and provides a framework combining predictive capability with exploratory interpretability.