I gave a talk last Wednesday at U. Michigan in the DCMB program where I included a slide estimating how much DNA sequencing (in base pairs) was needed for good de novo assembly of sequences from various biological environments or problems. The slide was there to motivate the challenges of soil metagenomics.

As you can see, you need about the same amount of sequencing for a soil or a marine metagenome as you do for a human genome. Or E. coli.

Wait, what? The E. coli genome is about 3 orders of magnitude smaller than the human genome -- they shouldn't compare!

Oh, sorry, the Y axis on that figure is in log format. Sorry, rookie mistake. Here's the figure with a linear Y axis.

Hmm. How did I say I got these numbers, again?

The basic calculations are on this Google Docs spreadsheet. For de novo assembly from Illumina reads, I estimated a requirement of 100x coverage (you can lower that to 20x if you want; it doesn't change the results all that much ;). The effective genome sizes are calculated as follows:

• for E. coli and the human genome, the numbers are well known: 5 Mbp for E. coli, and 3 Gbp for a (haploid) human genome.

• for the transcriptome, I adlibbed some numbers from this excellent blog post by Anthony Fejes. Briefly, I calculated that we needed about 100m reads, 100bp in length, per RNAseq sample; and that we would want to sample about 50 tissues to get a reasonably complete transcriptome.

• for the metagenomes, I calculated the dilution factor of the rarest species we wanted to see -- 1 in 200 for human gut, 1 in 1e3 for marine, and 1 in 1e6 for agricultural soil. The first number is from the MetaHIT papers and is probably wrong. The second number comes from Greg Dick at U. Michigan, who works on marine environments. The third number comes from Tracy Teal, who works on soil. The soil number is actually quite a bit lower than the estimate I cite in my papers, which is from Gans, Wolinsky, and Dunbar (Science, 2005). That paper claims a million distinct bacterial genomes in each gram of soil.

  I then multiplied this dilution factor by the estimated genome size for most bacteria (5e6 -- again, you can change this number down to 2.5e6, a good lower bound on the average, without much effect). I then multiplied that by 100x coverage for assembly from Illumina.

  The basic logic is that in order to assemble each metagenome (or transcriptome) "completely", we need to get sufficient shotgun coverage of the rarest genome in the metagenome (or transcript in the transcriptome) in order to assemble it.

  (This is absent considerations of noise, etc. -- read this excellent paper from @joe_pickrell et al., Noisy Splicing Drives mRNA Isoform Diversity in Human Cells. But I would argue you want to see the noise, then discard it as noise, rather than ignore it because you think you know it's noise.)

Feel free to tweak the numbers in the spreadsheet if you disagree with the assumptions made above, of course.

A few things to mention.

These are per individual. If you want a single bacterial or human genome, or a transcriptome for a single human (or vertebrate), or a complete metagenome from a single gut, marine, or soil sample, these numbers apply. It is an interesting question as to whether you can do shallower sequencing across multiple samples to detect rare transcripts (in transcriptomes) and rare genomes (in metagenomes), but I do not think we know the answer yet.

The numbers will be ~10x lower for detection, assuming you trust your reference genome/transcriptome/metagenome (hint: I wouldn't).

Second, yes, soil does seem to be that ridiculously diverse. In fact, it's probably more diverse; these estimates are from 16s, and probably ignore archaea and eukaryotes. They certainly ignore phage. Given that phage are likely to be high abundance even if they're rare (which we think we are seeing in some of our soil data) the above calculations are almost certainly an underestimate. This may be balanced by my failure to account for rRNA copy number in the bacteria, though, which would make my estimates a bit of an over-estimate, by maybe a factor of 5-10.

I'll talk more about soil in the future, as we start to post some of our ag soil papers. Soon, I promise!

Third, these calculations are an inescapable fact of shotgun sequencing, which samples randomly from the population of molecules. Sample enrichment approaches will certainly help lower this number, if you can target low abundance molecules in some way -- think library normalization, or cell sorting. On the other hand, they may also increase bias in your sampling... I tend to argue that, as sequencing costs continue to drop, you might as well just shove it all into a sequencer and use bioinformatics to sort it out.

Fourth, would increased read lengths help? Well, for de novo assembly, you can probably get away with 5-10x coverage with PacBio, if you assume that their error rate is going to decrease. That's in the graph below.

And the answer is, you still need to generate 10s of billions of reads for many of these samples. I do not yet have a clear cost estimate on PacBio or Nanopore reads -- I'd welcome a followup blog post or correct numbers! -- but I suspect that Illumina is still the only game in town for complex non-genomic samples, and will remain so for some time. (Also see this excellent blog post from EdgeBio: MiSeq 2x250 -- Does Length Really Matter?)

---

So, for our soil samples, we really do think we need about 50 Tbp or more (that's 50e12) per sample. And that's why we are working on scaling de novo assembly :).

Comments welcome!

--titus