5 Apr 2012 titus   » (Journeyer)

What is digital normalization, anyway?

I'm out at a Cloud Computing for the Human Microbiome Workshop and I've been trying to convince people of the importance of digital normalization. When I posted the paper the reaction was reasonably positive, but I haven't had much luck explaining why it's so awesome.

At the workshop, people were still confused. So I tried something new.

I first made a simulated metagenome by taking three genomes worth of data from the Chitsaz et al. (2011) paper (see http://bix.ucsd.edu/projects/singlecell/) and shuffling them together. I combined the sequences in a ratio of 10:25:50 for the E. coli sequences, the Staph sequences, and the SAR sequences, respectively; the latter two were single-cell MDA genomic DNA. I took the first 10m reads of this mix and then estimated the coverage.

You can see the coverage of these genomic data sets estimated by using the known reference sequences in the first figure. E. coli looks nice and Gaussian; Staph is smeared from here to heck; and much of the SAR sequence is low coverage. This reflects the realities of single cell sequencing: you get really weird copy number biases out of multiple displacement amplification.

Then I applied three-pass digital normalization (see the paper) and plotted the new abundances. As a reminder, this operates without knowing the reference in advance; we're just using the known reference here to check the effects.


Coverage of genome read mix, calculated by mapping the mixed reads onto the known reference genomes.


Coverage post-digital-normalization, again calculated by mapping the mixed reads onto the known reference genomes.

As you can see, digital normalization literally "normalizes" the data to the best of its ability. That is, it cannot create higher coverage where high coverage doesn't exist (for the SAR), but it can convert the existing high coverage into nice, Gaussian distributions centered around a much lower number. You also discard quite a bit of data (look at the X axes -- about 85% of the reads were discarded in downsampling the coverage like this).

When you assemble this, you get as good or better results than assembling the unnormalized data, despite having discarded so much data. This is because no low-coverage data is discarded, so you still retain as much overall covered bases -- just in fewer reads. To boot, it works pretty generically for single genomes, MDA genomes, transcriptomes, and metagenomes.

And, as a reminder? Digital normalization does this in fixed, low memory; in a single pass; and without any reference sequence needed.

Pretty neat.


Syndicated 2012-04-05 02:59:51 from Titus Brown

Latest blog entries     Older blog entries

New Advogato Features

New HTML Parser: The long-awaited libxml2 based HTML parser code is live. It needs further work but already handles most markup better than the original parser.

Keep up with the latest Advogato features by reading the Advogato status blog.

If you're a C programmer with some spare time, take a look at the mod_virgule project page and help us with one of the tasks on the ToDo list!