Hello,
I have about 25 mangrove metagenomic samples that I characterized using kraken2 with different databases including plusPF, nt and cutomized db like bacteria, archaea, fungi, protozoa, virus etc but only ~15-20% of the reads get classified. I have also tried Kaiju, metaphlan, phyloflash etc and obtained almost similar results.
Anyone has thought on how to deal with this large %age of unclassified reads?
Many thanks in advance!
•
written
by
arshad1292 •
You already got both possibilities covered by current responses: either something is wrong with your reads, or your sample contains novel organisms that are not in current databases and can't be classified at the read level. For the second possibility, I suggest you assemble the reads, bin the contigs, and then classify the MAGs using GTDBTk. This procedure will always be more accurate as the classification is done at the global sequence level rather than based on short reads.
I have already done that. I obtained some good quality MAGs but again about 75-80% data stay unbinned. I also assume this might be because of novel taxa as these samples are coming from an unexplored mangrove environment. Assuming, this is due to novel taxa, what is the best approach to go about it?
Binning has nothing to do with novelty. It simply groups together sequences that have similar tetranucleotide signatures, so it will work regardless of the environmental type or sequence novelty. I have had at most 10-20% of contigs unbinned in a sample of hundreds of assembled metagenomes. An educated guess here is that you have many short contigs (a fragmented assembly). Another possibility is that something is wrong with your binning procedure. If you have a large number of long-ish contigs (those > 3 kb) that are unbinned, I would consider that something is wrong with binning. Either way, even unbinned contigs are easier to classify by BLASTing than short reads.
Correction to my earlier comment: About 80-85% reads were actually binned. Then I applied >50% completion and <10% contamination criteria to these bins and mapped them back to reads. Only about 20% reads mapped to these bins. This tells that large portion of bins is actually of low quality i.e. low completion and high contamination. I was wondering if it is worth digging into these low quality bins in terms of finding more about taxonomic/functional info?
Then I applied >50% completion and <10% contamination criteria to these bins and mapped them back to reads.
It is important to understand that MAGs / bins that are < 50% complete or > 10% contaminated are not "defective" in any way. We simply have incomplete or less reliable information about them. It is up to you to decide whether you are interested in them or not.
This tells that large portion of bins is actually of low quality i.e. low completion and high contamination.
I would not say that 11% contamination is high or that a bin with that contamination is of low quality. If you have two related species in the community, or two subspecies of the same species, you may get a bin that is 100% complete and 30% contaminated. The relatively high contamination here comes from two related organisms that are not distant enough to be separated by binning, so some of the marker genes show up in multiple copies. I would consider that to be a perfectly good bin for the analysis purposes, even though I wouldn't necessarily submit it as such into public databases.
Here is something for you to consider: when you do a Kraken or Kaiju analysis on individual reads, how do you know whether a read belongs to a high- or low-quality bin? The answer is that nobody knows that without assembling the bins first, so we just do the analysis of all the reads and present them in the end without any regard for how good a quality their MAG is. Why doesn't a similar logic apply to analyzing everything even after the assembly, at the bin level? What's the harm in taxonomically classifying even those bins that you assume to be of low quality?
How does the read QC look? The most likely bad outcome candidates could be either poor read quality, or contamination with something not in the databases you're using (is human data in there, for example?).
A quick and simple starting point could be to take a a few thousand reads and blast them. If you're not even getting reads on there, then it's likely something is wrong with your reads.
Being able to classify only 20% of your reads is par for the course with random environmental samples as far as my experience with soil metagenomics goes. Doesn't have to be a quality issue, organisms in your sample are simply not recorded in the database and are thus not classifiable with this method.
yeah, I suspect there are some novel taxa there. I want to know about them. Any clue on how to find that out?
I quality trimmed the reads before running kraken2 i.e. removed low quality reads, removed human reads etc. The fastqc quality plots look good. I have also tried to run blast on some of the reads but blast gives nothing. Nothing found. I assembled the unclassified reads then I get some bacterial hits.
Log in to answer this question.
