I second Vincent's answer, and I'd like to add that it likely also depends on your choice of primers (i.e., sequenced 16S subregion). Many Illumina-based studies have relied on (very) short reads; for example, the widely used Earth Microbiome's protocol is targeting the ~100bp of the V4 region (attached link). Are you, by any chance, using a similar setup? With such very short reads, I believe that deeper sequencing is necessary. If you are targeting longer reads (e.g., V3-V5 or V1-V3 which should be attainable in a MiSeq paired-end setup), you can likely get highly sensitive and useful data without having to go as deep as you can with a single-end approach – although I can't supply a paper to support this from the top of my head. An additional advantage of longer reads (paired-end) would be that they can in many cases provide for more rigorous, consistent and sensitive analyses of your data (taxonomic profiles, OTU demarcation, phylogenetic profiles, etc.). This has been shown in several papers, including one in which I have been involved; I attach some examples.

In the end the depth that you "need" will depend on what you want to do with the data. You can often pick up stark differences between groups in PCoA plots even with relatively shallow data of short reads (say, ≤100k per sample or 100bp reads). If you want to delve deeper and get highly accurate profiles to pick up subtle effects, longer reads and / or deeper sequencing will likely come in handy.

Hope this helps :-)

PS: one note on rarefaction curves as they were mentioned by Vincent. If you run a few (deep) pilot samples to find a sweet spot of sequencing depth on rarefaction curves, be aware that the shape of these curves (of the taxa abundance distributions in general) will be *very* strongly influenced by the parameters you choose in your analysis.

http://www.earthmicrobiome.org/emp-standard-protocols/16s/

http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0039315

Article The Effects of Alignment Quality, Distance Calculation Metho...

Article Evaluation of different partial 16S rRNA gene sequence regio...

Article Limits to Robustness and Reproducibility in the Demarcation ...

Hi Philip, 

the sequencing depth needed to correctly cover a diversity is very sample dependant. The best way to know if your sequencing effort is sufficient is to measure alpha diversity metrics like rarefaction curves. I would start with the recommended coverage given by the illumina protocol (100k sequences), then calculate the rarefaction curves to know if this is sufficient. If not, you should  re-sequence your sample to get more data and an acceptable coverage. 

Another thing that could help --> is there any work already done on fish gut that you could use as a guideline? Maybe previous work gave some idea about fishes' gut microbiota diversity...

Anyway, good luck for your work!

Thank you Vincent. The most recent research I have to go on is that of Franchini et al (2014) - 'The Gut Microbial Community of Midas Cichlid Fish in Repeatedly Evolved Limnetic-Benthic Species Pairs'. In this study an average of 278,931 sequences were obtained from each sample (12 samples total). Although I am looking at a different species (trout) and work by Ingerslev et al (2014) - 'The development of gut microbiota in rainbow trout is affected by first feeding and diet type' obtained 539,867 reads per sample after filtering, but using HiSeq not MiSeq. Therefore I am thinking a depth of 300,000 - 600,000 reads per sample ball-park as appropriate?

I second Vincent's answer, and I'd like to add that it likely also depends on your choice of primers (i.e., sequenced 16S subregion). Many Illumina-based studies have relied on (very) short reads; for example, the widely used Earth Microbiome's protocol is targeting the ~100bp of the V4 region (attached link). Are you, by any chance, using a similar setup? With such very short reads, I believe that deeper sequencing is necessary. If you are targeting longer reads (e.g., V3-V5 or V1-V3 which should be attainable in a MiSeq paired-end setup), you can likely get highly sensitive and useful data without having to go as deep as you can with a single-end approach – although I can't supply a paper to support this from the top of my head. An additional advantage of longer reads (paired-end) would be that they can in many cases provide for more rigorous, consistent and sensitive analyses of your data (taxonomic profiles, OTU demarcation, phylogenetic profiles, etc.). This has been shown in several papers, including one in which I have been involved; I attach some examples.

In the end the depth that you "need" will depend on what you want to do with the data. You can often pick up stark differences between groups in PCoA plots even with relatively shallow data of short reads (say, ≤100k per sample or 100bp reads). If you want to delve deeper and get highly accurate profiles to pick up subtle effects, longer reads and / or deeper sequencing will likely come in handy.

Hope this helps :-)

PS: one note on rarefaction curves as they were mentioned by Vincent. If you run a few (deep) pilot samples to find a sweet spot of sequencing depth on rarefaction curves, be aware that the shape of these curves (of the taxa abundance distributions in general) will be *very* strongly influenced by the parameters you choose in your analysis.

http://www.earthmicrobiome.org/emp-standard-protocols/16s/

http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0039315

Article The Effects of Alignment Quality, Distance Calculation Metho...

Article Evaluation of different partial 16S rRNA gene sequence regio...

Article Limits to Robustness and Reproducibility in the Demarcation ...

I couldn't agree more with Thomas' answer!

Thank you very much Thomas for your detailed and helpful answer to my question. I will analyze the V3/V4 region of the 16S gene, and  the amplicon length is 292bp with the current primer set

Greg,

I think you certainly have a point. Noise in the non-overlapping amplified region can indeed be a problem. However, there are ways and approaches to mitigate this, most prominently by de-noising (i.e., pre-clustering) the sequence data. One possible approach, although I've not given it a shot myself yet, is DADA2 (see attached link) which looks quite promising.

In general, de-noising will greatly reduce your memory-overhead, but there are some algorithms which can provide very good clusterings with efficient memory usage (e.g., SWARM or the hierarchical clustering software hpc-clust which, I should add as disclaimer, a lab member of mine is developing). So far, I have worked with several very large paired-end MiSeq datasets and never encountered "un-clusterable" problems.

http://benjjneb.github.io/dada2/

Hi Greg,

The memory limitations you are describing are specific to the scaling issues with mothur's implementation of hierarchical clustering. Other approaches, such as greedy clustering or our DADA2 method, don't scale in the same way.

I agree with the Schloss lab's point that overlapping is a very poweful way to reduce sequencing errors, but it is not required by computational tractability, and extending the sequenced region (eg. V4 -> V3V4) can improve taxonomic resolution. This can be especially important when different clades of interest have different conservation patterns across the variable regions.

I guess I'm saying long-windedly that full overlap is a good place to start, but less-overlapping primer sets are not intractable and can make it easier to differentiate various species.

To answer your DADA2 questions: While we haven't published benchmarks on other regions, we've used the method on a number of other genetic loci (eg. V1V2, hupL, cytB) with great success. To get at what I think your point is (will it work w/o overlapping?) our paper also showed that the residual error rates and numbers of FPs when processing the forward reads alone (no overlapping) were very low (in fact better than the other methods using merged reads!). Because dada2 models the error process, it is not sensitive to lower quality sequence in the way that fixed cutoff methods are (i.e. OTUs).

As for computational scaling, we've used DADA2 on a couple datasets with >1B paired reads. Because samples can be processed independently (like reference based methods) the method scales to very big data.