Benchmarking strain-level profiling of Escherichia coli in short-read gut metagenomes

Abstract

Metagenomes facilitate Escherichia coli strain-level profiling in human gut microbiomes, informing our understanding of colonisation diversity and the genetic features distinguishing infection from carriage. Among numerous reference-based tools for short-read metagenomic strain-level profiling, the best approach remains unclear. Here, we benchmarked six publicly available tools--PanTax, PathoScope 2.0, StrainGE, Strainify, StrainR2 and StrainScan--for their ability to detect co-existing strains of E. coli and estimate their relative abundance across real and simulated metagenomes of increasing complexity with varying reference database composition. We show that the tools varied substantially in both the number and relative abundance of strains predicted, indicating that despite being designed for the same broad purpose, they produce markedly different representations of strain composition from the same metagenomic samples. StrainR2 achieved the highest F1 score across our experiments (0.995) and successfully distinguished all strains within a complex, 10-strain spike-in mixture containing strains sharing up to 99.9% average nucleotide identity and included in a 304-strain E. coli reference database. However, when ground-truth strains were removed from the reference database to mimic the profiling of samples of unknown strain content PathoScope 2.0 and StrainGE performed strongest: both estimated phylogroup abundances accurately and frequently correctly assigned strain identities to the closest available genomic neighbours to the ground truth. Computational requirements also varied substantially (median runtime: 0.11-68.45 h; peak memory: 2.08-351.53 GiB). Collectively, these results demonstrate that tool selection should reflect the intended application, and we provide practical guidance on the relative strengths and use cases of each approach. These findings will facilitate characterisation of E. coli strain-level diversity within short-read gut metagenomes with greater accuracy than previously possible.