A Robust and Efficient Representation-based DNA Storage Architecture by Deep Learning.

Area of Science:

• Computer Science
• Biotechnology
• Data Storage

Background:

• Images are crucial multimedia data, necessitating efficient storage solutions.
• DNA data storage offers high density but faces challenges with data integrity and error correction.

Purpose of the Study:

• To propose a representation-based deep learning architecture for image reconstruction in DNA storage.
• To evaluate the architecture's performance in handling insertion-deletion-substitution (IDS) errors and optimizing image quality versus compression.

Main Methods:

• Utilized a deep learning architecture comprising an autoencoder and a U-Net network for image representation, construction, and refinement.
• Incorporated feature quantization to enable trade-offs between compression ratio and image quality.
• Simulated experiments with varying IDS error rates and validated with wet lab practice using plasmids.

Main Results:

• Successfully reconstructed images with moderate quality at IDS error rates below 6%.
• Demonstrated a flexible trade-off between compression and image quality through adjustable representation channels.
• Showcased improved image quality by leveraging multiple reads, a common scenario in DNA storage.
• Validated the architecture's feasibility through a wet lab experiment reconstructing an image stored in 14 plasmids.

Conclusions:

• The proposed representation-based architecture offers a competitive and feasible solution for robust and efficient image storage in DNA.
• This deep learning approach addresses the challenges of data integrity in DNA storage, enabling large-scale image applications.