Utilizing Deep Neural Networks to Fill Gaps in Small Genomes

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Next-generation sequencing significantly advances genome sequencing but faces challenges in small genome assembly, including gaps due to repetitive elements and low coverage.
• Existing assembly software has limited success in gap filling, often failing to leverage artificial intelligence for improved accuracy and efficiency.
• Complete small genome assembly is crucial for understanding biological functions and evolutionary processes.

Purpose of the Study:

• To propose DLGapCloser, a novel deep learning-based method to improve gap filling in small genome assembly.
• To develop an efficient and accurate prediction algorithm, Wave-Beam Search, to overcome limitations of existing methods.
• To establish new standards and evaluation methods for assessing gap-filling performance in genome assembly.

Main Methods:

• Creation of four enriched datasets using genomes of *Saccharomyces cerevisiae*, *Schizosaccharomyces pombe*, *Neurospora crassa*, and *Micromonas pusilla*, including homologous genomes.
• Development of the DGCNet deep learning model for feature extraction and contextual learning from sequences flanking gaps.
• Implementation of the Wave-Beam Search algorithm, an optimized prediction strategy balancing efficiency and accuracy for gap filling.

Main Results:

• The Wave-Beam Search algorithm enhanced the gap-filling performance of traditional assembly tools by 7.35% to 42.85% across tested genomes.
• DLGapCloser, utilizing the DGCNet model and Wave-Beam Search, increased the number of filled gaps by 1.4% to 15.3% compared to conventional methods.
• A novel evaluation method was established and validated, demonstrating DLGapCloser's superior performance in complete genome assembly.

Conclusions:

• DLGapCloser represents a significant advancement in addressing the challenge of gap filling in small genome assembly.
• The proposed deep learning approach and Wave-Beam Search algorithm offer a more effective solution than traditional methods.
• This work provides a robust framework for improving the completeness and accuracy of small genome assemblies.