Genomic Language Model for Predicting Enhancers and Their Allele-Specific Activity in the Human Genome.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Predicting gene regulatory elements like enhancers is complex due to intricate sequence features.
• Existing machine-learning methods often overlook sequence context, leading to suboptimal enhancer prediction.
• Genomic language models offer a promising approach for analyzing DNA sequences.

Purpose of the Study:

• To develop a novel and accurate enhancer prediction method using a pre-trained genomic language model.
• To improve the identification of cis-regulatory elements for better genome interpretation.
• To predict candidate genetic variants with allele-specific regulatory activity.

Main Methods:

• Applied the DNABERT pre-trained language model to human genome sequences.
• Trained and fine-tuned two models using enhancers from the ENCODE registry.
• Evaluated model performance on independent datasets using accuracy and Matthews correlation coefficient.

Main Results:

• The best DNABERT-Enhancer model achieved 88.05% accuracy and 76% MCC on independent data.
• Performed genome-wide enhancer annotation and comparison with existing databases.
• Identified candidate single nucleotide polymorphisms (SNPs) with allele-specific enhancer and transcription factor binding activity.

Conclusions:

• DNABERT-Enhancer provides a powerful and accurate method for predicting enhancers.
• The tool offers valuable genome-wide enhancer annotations for functional and clinical genomics.
• Predicted genetic variants can aid in interpreting the functional impact of genomic variations.