Kraken confidence threshold?

[miaumc]

Many thanks for developing this insightful tool. We will definitely use it in our research work. I do see that you have used a Kraken confidence threshold of 0.3. Do you think that this is a bit too relaxed?

I made a test using a negative control sample (sterile water), and I still see that lots of nutrients (more than 400) have been detected. I am wondering whether this predefined threshold is causing many false positives to pop up. Also, do you have any suggestions or other parameters to tweak to limit false positives?

[cdiener]

MEDI uses a different strategy based on read filtering because the confidence threshold was not very associated with true and false positives in our tests. For instance, if a read has 80 kmers annotated as Segatella and 10 as Segatella copri I would call this a pretty good assignment as it is pretty consistent. But this would not pass a threshold 30%. On the other hand, a read with 80% Bacteria assignments but 20% Eukaryotes would still be called as bacterial with a threshold of 0.6, even though I would call this pretty problematic. Due to this MEDI uses consistency and decoy mapping as main strategies. The paper has some more details on that.

Though you are right that MEDI mostly corrects for false detection of bacterial DNA as food DNA but not so much for false positive detection in negative samples. We also see some signal that negative samples often have a residual food DNA signal which is likely contamination from the lab environment during sample handling. If the contamination is itself coming from food than it would be hard to correct for it. How many reads mapped to food do you see in the negative samples?