iEnhancer-Flow: Integrating Transformer-Based Sequence Learning with DNA Shape Insights for Robust Enhancer...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Enhancers are vital non-coding regulatory elements in the genome.
• Predicting enhancers is challenging due to sequence variability and lack of clear motifs.
• Existing sequence-only models have limitations in capturing regulatory complexity.

Purpose of the Study:

• To develop a novel framework, iEnhancer-Flow, for enhanced enhancer classification.
• To integrate DNA sequence and shape features for improved prediction accuracy.
• To overcome limitations of sequence-only models in diverse genomic contexts.

Main Methods:

• Proposed iEnhancer-Flow, a dual-branch model combining DNABERT-2 for sequence and a CNN for DNA shape.
• Utilized central-difference techniques for capturing local DNA shape variations.
• Employed a flow attention mechanism for fusing sequence and shape features, followed by MLP classification.

Main Results:

• iEnhancer-Flow demonstrated superior performance over competing methods in enhancer prediction.
• Achieved significant improvements in balanced accuracy (Bacc) and Matthews correlation coefficient (MCC) across multiple cell lines.
• Showed stability and robustness in diverse biological contexts, outperforming other models in six of eight cell lines.

Conclusions:

• Integrating DNA sequence and shape features represents a significant advancement in enhancer prediction.
• iEnhancer-Flow captures complementary regulatory signals beyond sequence alone.
• A comprehensive view incorporating sequence and structural DNA contexts is essential for understanding genomic regulation.