Self-supervised Learning for DNA sequences with circular dilated convolutional networks.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• DNA sequence analysis is crucial for biological insights.
• Existing deep learning models struggle with long DNA sequences and require large labeled datasets.
• Accurate modeling of nucleobase interactions is essential for sequence-based inference.

Purpose of the Study:

• To develop a deep learning method capable of analyzing long DNA sequences.
• To overcome the limitation of requiring massive supervised labels in DNA sequence modeling.
• To improve the accuracy of DNA inference tasks.

Main Methods:

• Utilized circular dilated convolutions as core components of the neural network architecture.
• Implemented a self-supervised learning framework to extract information from DNA sequences without labeled data.
• Designed a neural network backbone capable of processing extended DNA fragments without condensation.

Main Results:

• The proposed method successfully processed long DNA sequences, capturing long-range information.
• Demonstrated superior performance compared to five other deep learning models on two distinct DNA inference tasks.
• Achieved accurate results in predicting human variant effects and plant open chromatin regions.

Conclusions:

• The developed method effectively addresses limitations in current deep learning approaches for DNA sequence analysis.
• Circular dilated convolutions and self-supervised learning offer a powerful combination for DNA sequence modeling.
• The approach provides a more efficient and accurate solution for various DNA inference applications.