Relative abundance threshold filtering

Frequently within taxonomic classification, relative abundance thresholds are enforced for a detected taxon in order to be acknowledged as truly present^28,40,67. This mitigates the effects of FPs and their penalty on precision, because FPs often turn up with low relative abundances. Since the relative abundance was computed from classified reads, abundance filtering was likewise done exclusively based on classified reads. The effect of a relative abundance filtering on performance was investigated using precision-recall (PR) curves. These display the tradeoff between precision and recall when shifting the relative abundance threshold. By increasing this threshold, precision typically increases but recall decreases. The PR curve can be summarized with the area under the precision-recall curve (AUPRC), which is calculated using the trapezoid rule. A higher AUPRC represents a better model performance. A minimum value of 0 indicates the worst possible performance, while a maximum value of 1 denotes a perfect model where the TPs and FPs can be clearly separated by a specific relative abundance threshold.