DeepEPI: CNN-transformer-based model for extracting TF interactions through predicting enhancer-promoter...

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Enhancer-promoter interactions (EPIs) are crucial for gene regulation.
• Understanding EPIs is key to deciphering gene expression and disease mechanisms.
• Existing computational methods for EPI prediction have limitations in accuracy and interpretability.

Purpose of the Study:

• To introduce DeepEPI, a deep learning framework for direct prediction of EPIs from genomic sequences.
• To evaluate DeepEPI's performance against existing models and assess different encoding methods.
• To enhance the interpretability of EPI prediction by analyzing transcription factor (TF) interactions.

Main Methods:

• DeepEPI integrates Convolutional Neural Networks (CNNs) with Transformer blocks.
• The framework employs embedding layers for OneHot encoding and multihead attention mechanisms.
• A DNA2Vec-based version of DeepEPI was also developed and evaluated.

Main Results:

• DeepEPI consistently outperformed existing models across six cell lines.
• OneHot encoding achieved a 4% increase in AUPR and 1.9% in AUROC compared to prior methods.
• DeepEPI successfully extracted biologically relevant TF-TF interactions, aiding experimental validation.

Conclusions:

• DeepEPI offers a powerful and interpretable deep learning approach for EPI prediction.
• The framework advances the study of gene regulation and disease mechanisms.
• DeepEPI provides valuable insights for experimental validation in epigenomic research.