Fig. 2

Precision/recall plots for type classification of 70,000 Ion Torrent reads from an HPV16 amplicon sequencing reaction (a) and 3660 ONT minION reads derived from two HPV16 isolates (b, c) at various read sketch pruning levels M indicated by the label attached to each point. Read sketch pruning removes rare kmers in the read sketch which might be random sequencing errors. (a, b) were classified using a kmer size of 16 and (c) was classified using a kmer size of 10. Ion Torrent reads have low substitution error rates, so pruning removes few kmers and the precision boost is small (<0.001%) (a). ONT minION reads have a much higher error rate approaching 10% per-base. For minION reads, pruning is able to improve precision to roughly 99.8% when using a kmer size of 16 (b). A smaller kmer size of 10 combined with high levels of pruning lead to an increase in both precision and recall, with precision and recall increasing from slightly more than 97.0% to over 99% (c)