GnnDebugger: GNN based error correction in De Bruijn Graphs.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• High-throughput sequencing enables complete mammalian genome reconstruction, but scalability to diverse species and populations is limited.
• Low-quality draft assemblies and high coverage requirements are key bottlenecks.
• Accurate reconstruction of diploid genomes, especially diverged regions, is challenging due to insufficient sequencing depth.

Purpose of the Study:

• To explore machine learning, specifically graph neural networks, for scalable error correction in De Bruijn Graphs.
• To address limitations of existing heuristic methods in genome assembly.

Main Methods:

• Developed a Graph Neural Network model for error correction in De Bruijn Graphs.
• Applied the model to classify edges as correct or erroneous.

Main Results:

• The Graph Neural Network model reliably classifies edges in De Bruijn Graphs, particularly for diploid genomes with coverage depth 35 or lower.
• Model predictions can guide downstream read error correction and genome assembly algorithms.
• Achieved more accurate genome assembly through guided error correction.

Conclusions:

• Machine learning methods show potential to replace traditional heuristic approaches in genome assembly.
• Learning-based methods can improve the performance of existing assemblers in difficult cases.
• These approaches facilitate adaptation to newly sequenced species and population-level studies.