I worked with 454 non assembled data and with newer Illumina assembled data. Even though you can extract information directly from the reads, such as taxonomic information with tools like Metaphlan, or functional information doing assignments with blast or similar tools, I found much more interesting to work with the assembled data. As you mention, you can reconstruct genomes or have much more clear views on protein families. The only issue is doing the actual assembly, which is not a minor problem. I have no experience there, but it's feasibility has to do with the complexity of your system and the sequencing depth. If you don't have a substantial amount of reads assembled, in the end you can't say much from the assembled data.

http://bioinformatics.oxfordjournals.org/content/30/5/629.abstract

In addition to the previous answer, there are also possibilities to estimate which reads belong to the same genome. Using that information, the assembly can be improved. It depends on your reads' length. Unfortunately, neither me has experience there.

http://www.nature.com/articles/srep04516

Lately, I've been experimenting with the HUMAnN2 pipeline.  I prefer to use Kraken instead of MetaPhlAn 2, and then I use MaAsLin and LEfSe for comparative analyses.  HUMAnN2 is still in development I believe - at least, it isn't published yet - so be forewarned.

I've attached links to all of these tools to this post.

https://ccb.jhu.edu/software/kraken/

http://huttenhower.sph.harvard.edu/humann2/manual#markdown-header-faqs

https://huttenhower.sph.harvard.edu/maaslin

https://bitbucket.org/biobakery/biobakery/wiki/lefse

With my assembled by SPAdes metagenomic Illumina data I successfully used Metawat Binner and MG-RAST.

http://sourceforge.net/projects/metawatt/

http://metagenomics.anl.gov

I think the workflow of your analysis is really dependent on your null-hypothesis! There are many ways to analyze your data and most of time the real problem is the complexity of your data. Often you cannot find clear patterns so you have to dig deep.

So I will just post some advices from my mentors in metagenomics here. Hope these comments will help you throughout your analysis:

1) Complement

  e.g. read centric + assembly centric

2) Adapt

  settings to microdiversity: e.g. subsampling, increase assembly   stringency, k-mer size

3) Revise

  assemblies (filtering by coverage and length)

4) Evaluate

  by reference – independent metrics (e.g. ALE, Reaper …)

  + taxonomic markers (e.g. Amphora)

And finally in general: binning some genomes from the metagenome - it's not the easiest way, but often it is worth the effort!

metaAmos is a good pipeline for shotgun metagenomic analysis.  The pipeline consists of 13 steps and provides a visual summary of the results.

It can be a bit tricky to install but provides a complete pipeline using a modular approach so that you can try different tools for each step in the pipeline.  

http://metamos.readthedocs.org/en/v1.5rc3/

Article MetAMOS: A modular and open source metagenomic assembly and ...

Hi there, 

The consensus I've been exposed to, whatever that's worth, is for taxonomically and functionally profiling the entire community, it's becoming better to work with the reads rather than assembly - as they get longer and as newer algorithms are designed for longer reads. But either way, I say process both ways and use information from all angles of course. 

I want to echo some of what Alex Grier mentioned with humann2 for functional profiling (or potential of course in this case). It takes reads as input and outputs gene families, and it also uses metacyc by default for identifying pathways outputting abundance and coverage info on them. They also provide a handy script for easily merging all of your samples' outputs together for analysis. And in the temp files generated you can find taxonomy with relative abundance performed using metaphlan2, so it's a one shot deal that will give you both. Though I'm looking forward to trying Kraken as Alex mentioned for taxonomy, thanks for that. Humann2 is out of the Huttenhower lab at Harvard and they have many useful meta'omics analysis tools that can be found here: https://bitbucket.org/biobakery/biobakery/wiki/Home

When you want to try assembly, one great assembler that hasn't yet been mentioned here is IDBA-UD: http://i.cs.hku.hk/~alse/hkubrg/projects/idba_ud/

And to test out different assemblies you get with various programs you can try this assembly validation tool: https://www.cbcb.umd.edu/software/valet

And another tool for binning I've had some luck with and is very easy to run, compared with something like ESOM, but hasn't been mentioned yet is MaxBin: http://sourceforge.net/projects/maxbin/

Cheers!

https://bitbucket.org/biobakery/biobakery/wiki/Home

http://i.cs.hku.hk/~alse/hkubrg/projects/idba_ud/

https://www.cbcb.umd.edu/software/valet

http://sourceforge.net/projects/maxbin/

Thank you all.

These days I have been trying to deal with MetAMOS although I haven't been very lucky producing results...

I'm trying to give HUMAnN2 a go and let's see what happens.

By the way, how do you preprocess the raw reads? Remove low quality reads, trimming ends, and removing potential contaminants, i.e. human? Do you remove eukaryotic sequences or do something else?

What do you use for removing contaminant seqs, and if it's the case, eukaryotic reads?

Thank you again for your help

I forgot to ask... I have paired-end data. Do you discard the unpaired reads during the quality preprocessing?

I use SolexaQA++ which includes DynamicTrim, which removes low quality ends of the reads, and LengthSort which gets rid off the very short reads resulting from DynamicTrim but also can remove the unpaired reads if you specify that your data is paired end. This is quite important for genome assembly but I'm not sure if I should apply that for shotgun metagenomic sequencing

Hello,

After reading all the answers, what I understood is that there are different possible approaches to analyzing shotgun data metagenomics.

My question is: If I use Kraken + MG-RAST + MaAsLin for comparate the analyse, is a good option?

Patrick Pascoal Ferreira Kraken only works with taxonomic annotation while MG-RAST works on taxonomic and functional annotation. I think that all comparative analysis to confirm which software results in more data or different kinds of analysis is good. However, in my opinion, as metagenomics (shotgun) the functional content is very important to the analysis, I suggest that you: (i) remove the adapters (cutadapt), (ii) perform the merjing of paired-ends, (iii) if they merge in a lower proportion it is useful work only with the R1 sequences (foward) (It is hard to just say, because some researchers discard their non consensus senquences (assembled contigs -R1+R2) and others prefer work only with R1 reads if they don't match. For this reason, is hard answer this question) ; (iv) perform the quality control of the sequences (trimmomatic is the best in my view); take your sequences and retrieve the 16S rRNA sequence reads to analyse them separatly (in QIIME or Mothur) and annotation may be conducted on the FragGeneScan software to predict short reads function. After that, you may correlate the taxa with their function.

If you opt by MG-RAST, it perform all these steps (using another defaults) and you may correlate the functional information with the taxonomy using the R package ASAR, using MG-RAST output for instance. (https://academic.oup.com/bioinformatics/article/34/8/1404/4683460)

Just little suggestions. Hope you appreciate.

Kind regards.

Hello Otávio Guilherme Gonçalves de Almeida,

Thank you so much for your answer, I really appreciate and I will try your suggestions.

Kind regards.