To reiterate, the S. aureus problem highlights that Sativa is not perfect as it fails to remove sequences that are so far from what they're assigned to. My hypothesis is that this is because the reference tree is built from a number of sequences (in my case up to 30 per species) and one at a time is pulled out and placed. If there are, say, 5 S. aureus sequences among the E. coli, 4 will remain when one of the 5 is pulled out, which might be sufficient for the algorithm to place it among the E. coli. This is why I've suggested starting by clustering all sequences to proceed with Sativa only with the cluster representatives. This needs to be tested of course. I have started a couple of tests on clustering the 80,000 E. coli sequences. Disappointingly many clusters are formed at all identity thresholds (as always...). Will be back when I know more. I have started work on converting Sativa to Nextflow. My next step there will be to find a test dataset to be able to confirm that the Nextflow implementation is replicating the current implementation.