Opening up the blackbox: an interpretable deep neural network-based classifier for cell-type specific enhancer...

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Gene expression regulation relies on cis-regulatory modules (CRMs), notably enhancers.
• Enhancers, often distant from gene promoters, are critical for transcription.
• Epigenomic studies reveal enhancers are enriched for specific histone modification patterns.

Purpose of the Study:

• To develop a deep neural network (DNN) model, EP-DNN, for predicting enhancers in the human genome.
• To utilize histone modification expression levels as predictive features.
• To compare EP-DNN performance against existing enhancer prediction models.

Main Methods:

• A deep neural network (DNN) architecture, EP-DNN, was designed to predict enhancers.
• EP-DNN uses histone modification levels at functional sites and adjacent regions as input features.
• The model was trained and tested on four cell types (H1, IMR90, HepG2, HeLa S3), using p300 binding sites as enhancers.

Main Results:

• EP-DNN achieved superior prediction accuracy and reduced prediction time compared to DEEP-ENCODE and RFECS.
• Feature importance analysis revealed cell-type-specific histone modification patterns crucial for enhancer prediction.
• H3K4me1 was important across all cell types, while H3K27ac varied in importance.

Conclusions:

• EP-DNN demonstrates high accuracy (over 90%) in enhancer prediction across multiple cell lines.
• The developed interpretability method identifies key histone modifications and their proximity to enhancer sites.
• Feature importance analysis can optimize DNN training by reducing computational bottlenecks.