DEEP LEARNING FRAMEWORK FOR SUPER ENHANCER PREDICTION USING CNN AND MULTI-HEAD ATTENTION WITH CROSS-SPECIES TRANSFER LEARNING

Abstract: This paper presents a Super Enhancer Prediction system using advanced deep learning techniques for genomic analysis. The system utilizes a hybrid CNN–Transformer architecture to analyze DNA sequences and classify them as Super Enhancers (SE) or Typical Enhancers (TE). Convolutional layers capture local sequence motifs, while Multi-Head Self-Attention models long-range dependencies within genomic sequences. The proposed system incorporates cross- species transfer learning by pretraining on combined human and mouse datasets and fine-tuning for species-specific prediction. The model achieves high accuracy with improved AUC scores compared to existing methods. In addition to classification, the system provides biological insights such as GC content, motif density, and important regulatory regions. The integration of an interactive web interface allows users to upload DNA sequences and visualize results efficiently. This approach enhances genomic research by providing a fast, scalable, and interpretable solution for super enhancer identification.

Keywords: Super Enhancer Prediction, Deep Learning, CNN, Transformer, DNA Sequence Analysis, Transfer Learning, Bioinformatics, Genomics