HEAP: a task adaptive-based explainable deep learning framework for enhancer activity prediction.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Enhancers are critical cis-regulatory elements controlling cell-type-specific gene expression.
• Predicting enhancer activity and understanding their grammar remain significant challenges in genomics.

Purpose of the Study:

• To introduce HEAP (high-resolution enhancer activity prediction), an explainable deep learning framework.
• To accurately predict enhancer activity and explore enhancer grammar across diverse cell types.

Main Methods:

• HEAP utilizes a grammar-based deep learning approach incorporating DNA sequences and epigenetic modifications.
• A novel two-step multi-task learning method, task adaptive parameter sharing (TAPS), is employed for efficient cell-type-specific prediction.
• Post-hoc interpretation methods provide insights into prediction mechanisms.

Main Results:

• HEAP demonstrates superior performance compared to existing methods in enhancer prediction.
• The task adaptive parameter sharing (TAPS) method proves effective, particularly with limited training data.
• The explainable nature of HEAP offers insights into model decisions and enhancer grammar.

Conclusions:

• HEAP is a valuable tool for predicting enhancer activity and understanding enhancer grammar.
• The framework provides deeper insights into the complex mechanisms governing gene regulation.
• HEAP advances the field of computational genomics and regulatory element analysis.