Fig. 3

MD-plot for the CRC dataset. a Four clusters of features can be identified in the MD-plot, corresponding to features detected in zero, one, two and all of the three blank samples. For this dataset, all high quality features (in red) are in the cluster of features with the highest average abundances that are detected in all three blank samples. b Because all of the high quality features in the training set are detected in all three blank samples, we remove any features detected in less than three blank samples. We filter features detected in all blank samples (shown here) by using the distribution of the known noise below the zero difference line (in blue) to estimate the noise above the zero difference line. We use the absolute value (green lines) of the lower quartile of the negative differences (purple lines) within each partition (20th, 40th, 60th, and 80th percentiles) as filtering cutoffs. Any features above the green lines are retained