PCVR: a pre-trained contextualized visual representation for DNA sequence classification.

Area of Science:

• Bioinformatics
• Genomics
• Machine Learning

Background:

• DNA sequence classification is crucial for genetic analysis but faces challenges with speed and information loss in traditional methods.
• Existing machine learning techniques struggle with long sequences or sacrifice structural information.
• Frequency Chaos Game Representation (FCGR) converts DNA to images but often overlooks long-range dependencies.

Purpose of the Study:

• To introduce a novel method, Pre-trained Contextualized Visual Representation (PCVR), for enhanced DNA sequence classification.
• To leverage visual representations and advanced deep learning for capturing global sequence context.

Main Methods:

• PCVR encodes FCGR images using a vision transformer to extract contextualized features.
• A masked autoencoder is employed for pre-training the vision transformer encoder, ensuring robust feature learning.
• The model is evaluated on multiple datasets for DNA sequence classification tasks.

Main Results:

• Pre-trained PCVR demonstrates strong performance in unsupervised learning across three datasets.
• Fine-tuned PCVR surpasses existing methods in classification accuracy at superkingdom and phylum levels.
• Ablation studies validate the effectiveness of the vision transformer and masked autoencoder pre-training.

Conclusions:

• PCVR significantly enhances DNA sequence classification accuracy by effectively capturing global information.
• The method shows promise for new species discovery due to its robustness and ability to analyze complex sequence data.
• The PCVR framework and code are publicly available for further research and application.