Slide 1 - title

I'm going to be talking about our assembly pipeline for soil metagenomes.

Slide 2 - acks

Much of this work was done by a talented former postdoc of mine, Adina Howe, who has now moved on to Argonne National Lab. Other students and postdocs contributed significantly, including most especially Jason Pell who has been involved with many of the underlying computational work. I have some excellent collaborators, including Jim Tiedje and Janet Jansson.

Slide 3 -- kids

Folker should feel free to call my sequencing data ugly, but even he has to admit that my daughteres are very attractive. They take after their mother, who bears the burden of me being gone from home to talk here.

Slide 4 -- open science

Everything I'm discussing here is available online!

Code: http://github.com/ged-lab/ (BSD license)

Twitter: @ctitusbrown

Grants on Lab Web site: http://ged.msu.edu/interests.html

Preprint available: on arXiv, q-bio: diginorm arxiv

Slide 5 -- a polemic position

I don't see any alternatives to short-read sequencing and assembly in soil metagenomics. We need deep sampling and total bases, not fewer but longer reads. Isolation and enrichment approaches are excellent complements to whole metagenome sequencing, but with the vast diversity present in agricultural soil, they won't give you everything anytime soon. Whole metagenome sequencing lets you study the nucleic acid content of communities in situ, which is one of the things we really want to do!

I would also like to point out that assemblies are computationally convenient: they represent a summary of the short read data, and often a much smaller summary.

Slide 6 -- holy heck that's a lot of diversity!

Even in this crowd, I rarely see an appreciation for just how diverse soil really is.

To sequence the things that we can reproducibly see in OTUs, we will need a minimum of about 50 Tbp of sequencing -- roughly 100 full Illumina HiSeq runs -- to sample it thoroughly. (That's whether or not you're doing assembly!)

Alternatively, if you guess that there are about 100,000 species in a soil sample, that's approximately 100x the genomic complexity of a haploid human genome.

We can do this sequencing, and it's becoming progressively easier. But current assembly and analysis approaches simply cannot handle this data!

Slide 7 -- coverage is essential.

To robustly assemble things, you really need high coverage. This is a graph of fraction of things recovered in 300 bp or 1kb contigs from the Human Microbiome Project mock community -- real sequencing of a fake community -- showing that you need at least 10x coverage, on average, to retrieve about 95% of genome content in 1kb contigs. That's where the 50 Tbp number comes from in the last slide.

The main reason you're not getting assemblies is because usually most of the microbes in your soil metagenomes have not reached this threshold.

Slide 8 -- scaling challenges

We've spent a lot of time figuring out how to analyze this essentially infinite amount of data. Our goal is to make metagenome assembly straightforward, and to develop evaluation techniques so that you can tell when your assembly is good.

Slide 9 -- DOE grand challenge data set

The data set that really motivated this work came from the JGI via Janet Jansson and Jim Tiedje. This is a two terabasepair Illumina data set from 9 different midwestern soil samples, including an Iowa cornfield and Iowa prairie. They took small, ~1g samples of soil and subjected them to both 16s and Illumina sequencing. It's one of the largest soil metagenome data sets around.

Today I'll be talking about our assemblies of the Iowa corn and Iowa prairie data sets, which are each on the order of 300 Gbp, or 3 billion reads each.

Slide 10 -- digital normalization

The first technique we developed for dealing with this data is a computational version of cDNA library normalization, in which we preferentially downsample high abundance components of the data for the purpose of assembly. This is predicated on the observation that in mixed samples, much of the data is essentially redundant -- by the time you robustly sample the low-abundance genomes that are in your data, you've seen the high abundance stuff quite a bit! This data consumes disk space and memory.

You can read more about digital normalization here and here.

Slide 11 -- data partitioning

The second technique we developed is a computational version of cell sorting, in which we use read-to-read connectivity to split reads into bins. These bins turn out to correspond really well to source genomes -- that is, reads that bin together tend to come from the same species. I'll show you an example of that later.

We can do this partitioning very efficiently due to some nifty computational techniques we've developed.

You can read more about our data partitioning approach here.

Slide 12 -- our computational strategy

So, our overall computational strategy is this: develop computational approaches as needed, critically evaluate them on test data sets, especially including those where we already know the answer, and see what we see! Our general experience and those of labs using our software is that our stuff works pretty well, especially for scaling. The digital normalization stuff Just Works, but relatively few people are using the partitioning.

Slide 13 -- partitioning on real data

When we look at the mock community data with partitioning, what we can see is that the vast majority of partitions, or bins of reads, contain reads from only one genome (in blue). A few reads, those from highly conserved genes in different species, tend to group together (green), but it's less than 2-3%.

When we do a computational spike-in of a single E. coli, we find that we can group all those reads together into a separate set of partitions.

When we do this with 5 different E. coli strains, we do get partitions that contain reads from all of those strains, and those reads do assemble together. This is essentially unavoidable, but we can detect it very easily.

Slide 14 -- assembling soil

So when we assemble soil, what do we actually get? A lot.

We recover approximately 3 Gbp of sequence in contigs > 300 bp, containing millions of contigs and genes. Overall we assemble only about 20% of the reads from the highest abundance critters, indicative of the ridiculous diversity lurking just below the coverage threshold.

Slide 15 -- contigs are low coverage

After assembling contigs we can go back and count them, by mapping raw reads back to them. Basically we see that, as expected, most of our contigs end up being low coverage; the corn is slightly small so we have, overall, less coverage.

Slide 16 -- even abundance distributions

When we line contigs up on a rank/abundance distribution, we see that after a bit of high abundance stuff, there is a long slow plateau that represents the underlying evenness of the microbial population. Essentially we see that it is very hard to pick off the highest abundance critters to any degree because there's so many that are all at similar abundance.

Slides 17 and 18 -- preliminary taxonomy

These represent the taxonomy of individual contigs, without accounting for abundance. So, for example, phage contigs are counted only once even though they may be very high abundance. This is something that the MG-RAST folk will be adding for assemblies.

Slide 19 -- strain variation

Another thing that we can do is assess the degree to which strain variation or polymorphism shows up, by mapping reads back to the assembled contigs. This graph represents one particular contig, with the top two alleles plotted for each position; at the ends, we see higher variation because of repeat content that tends to end contigs, while in the middle you can see more reliable variants. Here you can see that there's only one position with a minor allele present in > 5% abundance.

Overall, when we look at the 5000 most abundant contigs, only 1 of them has an average polymorphism rate of above 5%. This is less than some animal genomes, and tells us that we should expect to get decent non-chimeric assemblies of these regions.

So, for our samples -- which represent very small, localized soil samples -- we haven't yet seen much in the way of strain variation. We think that we can assemble what we do see, with the caveat that you may get chimeric assemblies that we can pick out later.

Slide 20 -- technology conclusion

Basically, we can assemble stuff.

We want you to be able to assemble stuff.

We have good ways to assemble lots more stuff, and are working on them.

Slide 21 -- assembly conclusion

The main message I want to convey is that, one way or another, we will be able to extract decent quality metagenomes from shotgun metagenomics.

Depending on sample, you may not need to worry a lot about strain variation. Either way, we can generate assemblies and figure out if they're chimeric.

Your job? To make sense of all of this :)