Graph-based self-supervised learning for repeat detection in metagenomic assembly.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Repetitive DNA sequences present significant challenges for accurate genome assembly and sequence alignment, especially in complex metagenomic datasets.
• Metagenomic data complexity arises from genome dynamics like horizontal gene transfer, gene duplication, and gene loss/gain, further complicating repeat detection.

Purpose of the Study:

• To develop a novel computational approach, GraSSRep, for accurate and efficient detection of repetitive DNA sequences in metagenomic data.
• To leverage graph neural networks (GNNs) and assembly graph structures for improved repeat identification.

Main Methods:

• GraSSRep frames repeat detection as a node classification task within a metagenomic assembly graph.
• A self-supervised learning framework uses a high-precision heuristic to generate pseudolabels for training a GNN embedding and a random forest classifier.
• The method combines sequencing features with learned graph features for robust repeat classification.

Main Results:

• GraSSRep demonstrates state-of-the-art performance in repetitive DNA detection.
• Evaluation on simulated and synthetic metagenomic datasets highlights GraSSRep's robustness to repeat attributes and its effectiveness in handling complex sequences.
• Comparative analyses show GraSSRep outperforms existing repeat detection tools in both precision and recall.

Conclusions:

• GraSSRep offers a powerful and accurate solution for detecting repetitive DNA in challenging metagenomic environments.
• The integration of graph structure and GNNs significantly enhances repeat detection performance.
• This approach is crucial for advancing accurate genome assembly and sequence alignment from complex biological communities.