Fig. 3.

Prediction results on four real metagenomic datasets. a Boxplots of random forest AUC scores obtained by reference-based method and MicroPro (with assembler Megahit). Each random forest classification model was repeatedly trained and tested 30 times. Student’s t test p values are given. b Boxplots of random forest AUC scores obtained by MicroPro and de novo assembly-based method. Results of MicroPro with two different assemblers are shown. Each random forest classification model was repeatedly trained and tested 30 times. Student’s t test p values between pairs of methods are given