Metagenomic assemblies tend to break around antibiotic resistance genes.

Background: Assembly of metagenomic samples can provide essential information about the mobility potential and taxonomic origin of antibiotic resistance genes (ARGs) and inform interventions to prevent further spread of resistant bacteria. However, similar to other conserved regions, such as ribosomal RNA genes and mobile genetic elements, almost identical ARGs typically occur in multiple genomic contexts across different species, representing a considerable challenge for the assembly process. Usually, this results in many fragmented contigs of unclear origin, complicating the risk assessment of ARG detections. To systematically investigate the impact of this issue on detection, quantification and contextualization of ARGs, we evaluated the performance of different assembly approaches, including genomic-, metagenomic- and transcriptomic-specialized assemblers. We quantified recovery and accuracy rates of each tool for ARGs both from in silico spiked metagenomic samples as well as real samples sequenced using both long- and short-read sequencing technologies.

Results: The results revealed that none of the investigated tools can accurately capture genomic contexts present in samples of high complexity. The transcriptomic assembler Trinity showed a better performance in terms of reconstructing longer and fewer contigs matching unique genomic contexts, which can be beneficial for deciphering the taxonomic origin of ARGs. The currently commonly used metagenomic assembly tools metaSPAdes and MEGAHIT were able to identify the ARG repertoire but failed to fully recover the diversity of genomic contexts present in a sample. On top of that, in a complex scenario MEGAHIT produced very short contigs, which can lead to considerable underestimation of the resistome in a given sample.

Conclusions: Our study shows that metaSPAdes and Trinity would be the preferable tools in terms of accuracy to recover correct genomic contexts around ARGs in metagenomic samples characterized by uneven coverages. Overall, the inability of assemblers to reconstruct long ARG-containing contigs has impacts on ARG quantification, suggesting that directly mapping reads to an ARG database should be performed as a complementary strategy to get accurate ARG abundance and diversity measures.