K-mer based genome wide association study

All non-clonal isolates were selected for analysis, whereas only a single isolate was randomly selected from each infection sample. In total 62 non-clonal isolates (non-clonal group) and 23 infection isolates (infection group) were analysed.

Contigs were split into overlapping 30-mers using Jellyfish (version 2.2.10, parameters used: -m 30 -s 150M -C) [123]. The k-mer sequences were saved in FASTA format and their abundances were compiled in a table.

A 2 x 2 contingency table was generated for each k-mer. To test for significant differences between the two groups, Fisher’s exact test was applied[38]. Only k-mers occurring with significantly (p-value ≤ 0.001) different abundances between both groups were considered for further analysis.

For each isolate a binary vector, indicating the presence of each k-mer, was constructed. Based on these vectors, a random forest model was trained for classifying whether an isolate belongs to the infection group or non-clonal group. The Python package Scikit-learn [124] was used for this task. Finally, k-mers, were sorted by their feature importance. Only the k-mers with a feature importance greater than zero were kept.

A pan-genome consisting of all isolates was constructed using Roary 3.12.0 [106] with results previously generated by Prokka [125]. The python packages biopython and regex were then used associate k-mers with the genomic loci they originated from.