Would you be able to track down a log file for one of these runs? I think we should see the number of unique sequences printed to stdout which may be helpful in finding the issue.
I see that you even used a higher max-ee value, will try to use a higher one on mine to see if that helps. Additionally, I even used 'indepedent' pooling method but still came up short.
Perhaps a question for everyone, can one still proceed with downstream analyses despite this significant loss of reads after filtering?
Hi Mukhari_R,
I tried multiple max error, like 2, 5 but it did not make much difference in the reads that passed the final filter.
I am wondering if that might be the sequence quality itself. I am not entirely sure how that could happen.
Low rates of successful denoising, which is where you are losing most of your 16S reads, can happen when there is little duplication in the data, either due to relatively high error rates (more of an issue with long reads) or very high levels of sample diversity (so even real biological sequences are only being observed a handful of times or less). But what is surprising is that you aren't seeing the same issue in your ITS sequences.
Were there any differences in the library preparation/sequencing of your 16S data versus your ITS data? For example, was Kinnex used for one and not the other? How long is the ITS region you are amplifying?
Were you able to fiugre this out? I got my pac bio 16S hifi reads and now having the same problem. Only about 3-5% of the reads retained after denosing. Tried all the error models mentioned in the github and other places but same results every time.
Unfortunately, I couldn't. The sequencing lab confirmed that they used same methods for both 16S and ITS. I was seeing more loss in 16S compared to ITS. I am considering doing denoising in R or Python to see if that helps for my next datasets, which I still need to explore how to do it.
I tried both R and python ending up with the same results. I think the problem lies in the dada2 denoising step and specifically in the error learning method. error plots are almost flat and it seems the model doesn't have enough replication of reads to work with. I am seeing >90% of my reads as unique and dada2 needs at least 20% replication (in my case its <10%) of the reads to model to work. At the end it removes that ~90% of reads as errors and very few reads survive the denoising. As far as I see, I think dada2 is not a suitable method (unless something is changed) for denoising PacBio hifi reads and I might have to use something else.
Hi all. I used R in an attempt to resolve the read loss when running DADA2 via qiime, and there was average of about 32% of reads retained after primer removal. Which is an improvement from my initial runs. The loss seems to occur occurs at the filtering step inside DADA2, before denoising.
The parameters/ modifications I used for running in R are: minQ= 0, maxEE=10, nbases=5e8.