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

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Enhancer-enhancer interactions (EEIs) are crucial for gene regulation but difficult to predict computationally.
• EEIs are less understood compared to enhancer-promoter interactions.

Purpose of the Study:

• To develop a novel deep learning model, ETNet, for predicting EEIs from DNA sequences.
• To evaluate ETNet's performance against existing methods and assess its generalizability.

Main Methods:

• Developed ETNet, a deep learning architecture combining CNNs and Transformer modules.
• Evaluated ETNet on three cell lines (GM12878, K562, MCF-7) and validated using cross-validation and enhancer-level partitioning.
• Performed feature attribution analysis to identify regulatory motifs and cooperative mechanisms.

Main Results:

• ETNet significantly outperformed existing methods like EnContact across six cell lines.
• Demonstrated robust generalization, effective cross-cell type transfer learning, and transferability to enhancer-promoter interactions.
• Identified cell-type-specific regulatory motifs and provided computational evidence for super-additive cooperative mechanisms.

Conclusions:

• ETNet advances computational prediction of EEIs, offering superior performance and interpretability.
• The model provides a framework for understanding gene regulatory networks and identifying disease-associated variants.
• Findings suggest shared principles in chromatin interactions and highlight potential for experimental validation of cooperative mechanisms.