Metagraph: plasmid/chromosome classification enhancement using graph neural networks

Abstract

Chromosomes and plasmids are the main sites of genetic information in microorganisms. Separately identifying plasmids and chromosomes is essential for further metagenomic analysis. Computational tools have achieved significant results in classifying DNA into plasmids and chromosomes. However, there is often a trade-off between recall and precision in the currently available tools. Several graph-based tools have been proposed to improve the prediction precision and recall simultaneously by improving upon the results produced by existing tools. We propose MetaGraph, a Graph Neural Network (GNN) based tool for plasmid/chromosome classification enhancement. It uses the high confidence predictions of existing plasmid/chromosome prediction tools and improves the prediction accuracy of low confidence predictions using plasmid probabilities as features for the GNN. We evaluated MetaGraph for a set of simulated DNA sequences. The results significantly improved over stateof-the-art tools like PlasFlow and PlasClass. The results were increased up to 20% from the initial PlasClass predictions. The source code for MetaGraph is freely available at: https://github.com/MetaGSC/MetaGraph