Geometric deep learning framework for de novo genome assembly.

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence

Background:

• De novo genome assembly relies on identifying correct paths in complex assembly graphs.
• Repetitive genomic regions create tangles in assembly graphs, leading to fragmented genome reconstructions.
• Existing algorithmic methods face significant challenges in accurately resolving these complex graph structures.

Purpose of the Study:

• To introduce GNNome, a novel framework for accurate path identification in genome assembly graphs.
• To develop an AI-driven approach that leverages geometric deep learning for genome reconstruction.
• To enable training on assembly graphs without dependence on pre-existing assembly strategies.

Main Methods:

• Developed GNNome, a framework utilizing geometric deep learning for path identification in assembly graphs.
• Trained AI models on assembly graph symmetries, bypassing traditional assembly methods.
• Utilized PacBio HiFi reads for genome reconstruction.

Main Results:

• GNNome achieves contiguity and quality in genome assemblies comparable to state-of-the-art tools across multiple species.
• The framework successfully reconstructs genomes from PacBio HiFi reads.
• Demonstrated the potential for GNNome to assemble complex genomes, including those with varying ploidy and aneuploidy.

Conclusions:

• GNNome offers a robust AI-based solution for accurate genome assembly, particularly in challenging genomic regions.
• The framework's ability to learn from assembly graphs and simulated data positions it as a cornerstone for future complex genome reconstruction.
• GNNome and its trained model are publicly available as a tool for assembling haploid genomes.