Nucleic Transformer: Classifying DNA Sequences with Self-Attention and Convolutions.

Area of Science:

• Genomics and Bioinformatics
• Machine Learning in Biology
• Computational Biology

Background:

• Machine learning and deep learning applications in genomics often demand significant domain expertise.
• Existing models frequently lack interpretability, hindering biological insights.
• There is a need for accessible and understandable deep learning tools in genomic analysis.

Purpose of the Study:

• To introduce the Nucleic Transformer, a novel and interpretable deep learning architecture for DNA sequence classification.
• To demonstrate the model's effectiveness across diverse genomic prediction tasks.
• To showcase a model that requires minimal domain knowledge for training.

Main Methods:

• The Nucleic Transformer integrates self-attention mechanisms and convolutional layers, inspired by computer vision and natural language processing.
• The model architecture is designed for direct application to nucleic acid sequences.
• Training and evaluation were performed on several benchmark genomic datasets.

Main Results:

• The Nucleic Transformer achieved high performance in classifying *Escherichia coli* promoters.
• The model successfully identified viral genomes and classified enhancers.
• Accurate predictions were also obtained for chromatin profiles.

Conclusions:

• The Nucleic Transformer offers an effective and interpretable solution for DNA sequence classification tasks.
• The model's design reduces the reliance on extensive domain-specific knowledge.
• This approach facilitates broader application of deep learning in genomics research.