Exploring Graph Neural Networks for Metagenomic Assembly Graphs

Metagenomics has transformed microbial research by analyzing environmental DNA without the need for culturing. Metagenomic binning focuses on clustering the DNA fragments obtained from these environments based on their shared characteristics, providing insights into microbial community composition. Deep learning approaches offer a promising approach towards this outcome by leveraging graph representations of metagenomic data to improve clustering accuracy. However, binning remains challenging with long-read sequencing technologies. While long reads enable more comprehensive genomic coverage, their high error rates create obstacles that traditional binning algorithms struggle to overcome, limiting accurate genome reconstruction. In this thesis, we develop a framework for metagenomic binning of long-read sequencing data, integrating Graph Neural Networks (GNNs) with graph-based error correction techniques. We leverage the MetaFlye assembler to construct assembly graphs from errorprone long-read data and incorporate a GNN model based on GraphSAGE for unsupervised clustering, aiming to accurately and efficiently group DNA fragments into taxonomic bins. Additionally, we explore and evaluate graph-based error correction methodologies designed to refine the graph structure by reducing erroneous connections and reinforcing biologically meaningful ones, thereby improving the quality of metagenomic binning. Through this research, we seek to develop computational methods that enhance the accuracy and reliability of metagenomic analysis.