imMeta: An Incremental Sub-graph Merging for Feature Extraction in Metagenomic Binning

Abstract

Metagenomic binning is a crucial step in understanding microbial communities without culturing. Many unsupervised binning methods follow the two-phase paradigm. In the first phase, specific features of metagenomic sequences, also known as reads, are extracted without relying on reference databases. The second phase involves applying clustering methodologies to group the reads into likely similar species, which are further studied in subsequent metagenomic steps, such as assembly and annotation. Specific well-studied methods refrain from building features for individual reads to improve computation performance and reduce input sensitivity. Instead, these methods create overlapping graphs that illustrate the closeness of reads based on their k-mer frequency distribution. Read nodes with high connectivity are then merged into sub-graphs, generating a feature for each sub-graph. This study introduces a novel unsupervised algorithm that incrementally merges sub-graphs into larger sub-graphs, expecting to obtain variable-sized groups. This approach differs from the fixed-size-based methods proposed previously. Empirical results demonstrate that the proposed approach achieves higher accuracy than other well-known short-read methods.