NeuralBeds: Neural embeddings for efficient DNA data compression and optimized similarity search

Area of Science:

• Bioinformatics
• Genomics
• Computational Biology

Background:

• High-throughput sequencing and decreasing costs generate vast omics data in databases like NCBI and EMBL.
• Identifying similar DNA sequences is crucial for homology detection, phylogenetic studies, and pathogen identification.
• Growing sequence repositories necessitate improved DNA similarity search methods beyond raw sequence comparison.

Purpose of the Study:

• To evaluate numerical sequence representation methods for optimizing DNA similarity searches.
• To compare the efficacy of various embedding techniques against traditional bioinformatics approaches.

Main Methods:

• Analysis of numerical embedding approaches: Chaos Game Representation, hashing, and neural networks.
• Comparison with classical methods, including principal component analysis.
• Utilizing embeddings as a distance measure for sequence similarity calculation.

Main Results:

• Neural network embeddings effectively capture DNA sequence similarity as a distance metric.
• Neural networks significantly outperformed other numerical and classical approaches in DNA similarity search.
• The study demonstrates the superiority of neural networks for large-scale sequence comparison.

Conclusions:

• Numerical embeddings, particularly from neural networks, offer a powerful optimization for DNA similarity searches.
• This approach enhances the efficiency and accuracy of analyzing large omics datasets.
• Neural network-based embeddings represent a significant advancement in bioinformatics sequence analysis.