The impact of strain heterogeneity on genomic profiling accuracy in mixed-strain Escherichia coli metagenomes
Metagenomic sequencing is increasingly used for surveillance of bacterial populations, enabling simultaneous characterisation of antimicrobial resistance, virulence, and epidemiologically relevant strain markers. However, the ability of commonly used bioinformatic tools to recover these features from heterogeneous populations remains incompletely understood. Here, we developed a controlled benchmarking framework based on simulated single-species Escherichia coli metagenomes generated from genomes with PCR-determined annotations. Assembly-based and read-based tools were evaluated for antimicrobial resistance gene (ARG) detection, virulence gene (VG) detection, multilocus sequence typing (MLST), and serotype prediction. In addition, datasets with varying strain compositions were used to assess the effect of population complexity on strain-level characterisation. We found that recovery of individual genetic determinants of interest (7 ARGs, 3 VGs) was consistently more successful than recovery of strain-level markers. ARG detection achieved high recovery rates across all evaluated tools (84.3-90.0%), while short read sequencing typer SRST2 achieved the highest virulence gene recovery rate (100.0%). In contrast, MLST and serotype prediction were strongly affected by strain composition. None of the evaluated MLST tools correctly assigned the dominant sequence type in the multi-strain log-normal dataset, despite successful partial recovery of MLST alleles, while complete serotype recovery remained uncommon across all methods. Decreasing the complexity of the simulated metagenome by increasing the abundance of the dominant strain improved both MLST and serotype prediction performance, while long-read analyses indicated that increased read length was insufficient to overcome the challenges associated with mixed-strain communities. These findings demonstrate that performance estimates derived from gene detection do not necessarily translate to strain-level characterisation. While individual resistance and virulence determinants could often be recovered reliably, accurate reconstruction of sequence types and serotypes remained strongly constrained by strain heterogeneity. More broadly, the results highlight current limitations in resolving strain composition from metagenomic data and provide a framework for investigating the abundance thresholds and population structures required for reliable strain-resolved surveillance.