A New Unsupervised Machine Learning Method for Metabarcoding Analysis Based on Selected Mock Community Data

DNA metabarcoding analysis has emerged as a powerful tool for biological monitoring, enabling the identification and assessment of biodiversity across ecosystems. Conventional analytical methods, including operational taxonomic unit (OTU) clustering and amplicon sequence variant (ASV) analysis, have demonstrated strengths and limitations, including bias in estimating biodiversity and challenges in detecting rare sequences. To address these challenges, we developed unsupervised machine learning (UML)-based methods utilizing hierarchical density-based spatial clustering of applications with noise (HDBSCAN) and ordering points to identify the clustering structure. This study focused on 3 genetic markers: 12S rRNA, cytochrome c oxidase subunit I, and 18S rRNA. Mock community datasets comprising diverse taxa were analyzed using UML-based methods, OTU clustering, and ASV analysis. Nineteen biological feature descriptors were evaluated across all datasets. Nucleic acid composition and composition of k-spaced nucleic acid pairs consistently produced the best clustering performance. A comparative analysis of the clustering methods was performed. HDBSCAN produced relative abundance profiles broadly consistent with OTU clustering and ASV analysis. Additionally, comparisons of sensitivity and precision based on detected species demonstrated that UML-based methods, especially HDBSCAN, provided a favorable balance between false-positive and false-negative detections in several datasets, suggesting their promise as alternative clustering approaches for DNA metabarcoding. However, their applicability to real environmental systems remains to be validated.