Efficient k-mer based curation of raw sequence data: application in Drosophila suzukii

Several studies have highlighted the presence of contaminated entries in public sequence repositories, calling for special attention to the associated metadata. Here, we propose and evaluate a fast and efficient k–mer-based approach to assess the degree of mislabeling or contamination. We applied it to high-throughput whole-genome raw sequence data for 236 Ind-Seq and 22 Pool-Seq samples of the invasive species Drosophila suzukii. We first used Clark software to build a dictionary of species-discriminating k–mers from the curated assemblies of 29 target drosophilid species (including D. melanogaster, D. simulans, D. subpulchrella, or D. biarmipes) and 12 common drosophila pathogens and commensals (including Wolbachia). Counting the number of k–mers composing each query sample sequence that matched a discriminating k–mer from the dictionary provided a simple criterion for assignment to target species and evaluation of the entire sample. Analyses of a wide range of samples, representative of both target and other drosophilid species, demonstrated very good performance of the proposed approach, both in terms of run time and accuracy of sequence assignment. Of the 236 D. suzukii individuals, five were reassigned to D. simulans and eleven to D. subpulchrella. Another four showed moderate to substantial microbial contamination. Similarly, among the 22 Pool-Seq samples analyzed, two from the native range were found to be contaminated with 1 and 7 D. subpulchrella individuals, respectively (out of 50), and one from Europe was found to be contaminated with 5 to 6 D. immigrans individuals (out of 100). Overall, the present analysis allowed the definition of a large curated dataset consisting of > 60 population samples representative of the worldwide genetic diversity, which may be valuable for further population genetics studies on D. suzukii. More generally, while we advocate careful sample identification and verification prior to sequencing, the proposed framework is simple and computationally efficient enough to be included as a routine post-hoc quality check prior to any data analysis and prior to data submission to public repositories.