Differential sensitivity of k-mer based identification tools using closely related database sequences affects relative abundance estimation

Abstract

Deep metagenomic sequencing is increasingly used for lineage profiling of complex biological mixtures, including applications such as wastewater virus surveillance and microbiome profiling. For profiling, sequence read sets are compared against a database of reference sequences to identify component genetic lineages, and estimate their relative abundance. However, there is currently a lack of clarity regarding the impact of two ubiquitous features of biological mixtures on compositional profiling: (1) population diversity, where reads contain genetic diversity not represented in the reference set, and (2) the presence of high sequence similarity between references, which causes ambiguity in read assignment. Here we show, using both simulated and real data, that performance of two commonly used tools for lineage profiling, KMA and Kallisto, is differently affected by sample-reference distance and the presence of highly similar competing reference sequences. Increasing sample-reference distance reduced read assignment for both tools, but more strongly for Kallisto, which led to directional bias in abundance estimates when distance differed between mixture components. In contrast, when sample-reference distance was low, Kallisto more specifically discriminated between highly similar competing references, whereas KMA was less able to partition reads accurately between them. This specificity-sensitivity trade-off was reflected in real datasets: KMA produced more accurate abundance estimates for mosquito COX1 mixtures, whereas Kallisto performed better for mixtures of closely related SARS-CoV-2 lineages, although neither tool fully resolved the most similar SARS-CoV-2 variants. Tool choice is therefore dependent on the degree of expected sample-reference distance and inter-reference distance within datasets.