An Efficient Lightweight Hybrid Model with Attention Mechanism for Enhancer Sequence Recognition

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Enhancers are crucial noncoding DNA sequences that regulate gene expression but are challenging to identify due to their variable locations and scattered nature.
• Existing bioinformatics tools struggle with cell line-specific enhancer identification based solely on DNA sequences.
• Chromatin accessibility offers insights into regulatory element function, but integrating this with sequence-based methods remains complex.

Purpose of the Study:

• To develop an advanced deep learning model for accurate identification and strength assessment of gene enhancers.
• To overcome the limitations of current computational methods in pinpointing enhancers and their regulatory potential.
• To leverage chromatin accessibility information implicitly through sequence analysis for improved enhancer prediction.

Main Methods:

• A deep learning framework combining Convolutional Neural Networks (CNN) with attention-gated Recurrent Units (AttGRU) was developed.
• One-hot encoding was used to represent DNA sequences for model input.
• The model, termed enhancer-CNNAttGRU, was trained to identify enhancers and classify their strength.
• Performance was evaluated against state-of-the-art methods using cross-species analysis.

Main Results:

• The enhancer-CNNAttGRU model achieved high accuracy in predicting enhancer sequences (87.39%) and their strengths (84.46%) in cross-species analyses.
• The model demonstrated superior performance compared to existing methods for stage one and stage two enhancer prediction.
• Results indicate the model's effectiveness in identifying regulatory elements based on sequence and inferred accessibility features.

Conclusions:

• The proposed deep learning model (enhancer-CNNAttGRU) offers a powerful and accurate approach for identifying gene enhancers and assessing their regulatory strength.
• This method addresses key challenges in enhancer identification, particularly their scattered nature and positional variability.
• The findings highlight the potential of deep learning in advancing genomic regulatory element discovery and analysis.