Efficient DNA sequence compression with neural networks.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• The exponential growth of genomic data necessitates efficient lossless compression methods for storage and analysis.
• Existing neural network approaches for DNA compression lag behind specialized tools like GeCo2.
• A gap exists in models specifically designed for the unique characteristics of DNA sequences.

Purpose of the Study:

• To develop a novel genomic sequence compressor, GeCo3, that integrates neural networks with DNA-specific models.
• To improve upon the compression efficiency of existing state-of-the-art DNA compressors.
• To leverage neural networks for enhanced context modeling in DNA sequence compression.

Main Methods:

• Developed GeCo3, a genomic sequence compressor utilizing neural networks for mixing multiple context models.
• Incorporated substitution-tolerant context models tailored for DNA sequences.
• Benchmarked GeCo3 on diverse datasets including human genomes, viral genomes, and ancient DNA, both in reference-free and reference-based modes.

Main Results:

• GeCo3 demonstrated significant compression ratio improvements over GeCo2, ranging from 2.4% to 7.1% in reference-free compression.
• In reference-based compression, GeCo3 outperformed the state-of-the-art by 10.1% to 12.4%.
• While computationally 1.7-3 times slower than GeCo2, GeCo3 maintains constant RAM usage and scales efficiently with sequence size.

Conclusions:

• GeCo3 represents a significant advancement in genomic sequence compression, offering substantial gains through its neural network-based mixing approach.
• The portable mixing method allows for easy integration into other data compression or analysis tools.
• GeCo3 is freely available under GPLv3, promoting its adoption in the scientific community.