Slashdot Mirror
The DNA Data Deluge
the_newsbeagle writes "Fast, cheap genetic sequencing machines have the potential to revolutionize science and medicine--but only if geneticists can figure out how to deal with the floods of data their machines are producing. That's where computer scientists can save the day. In this article from IEEE Spectrum, two computational biologists explain how they're borrowing big data solutions from companies like Google and Amazon to meet the challenge. An explanation of the scope of the problem, from the article: 'The roughly 2000 sequencing instruments in labs and hospitals around the world can collectively generate about 15 petabytes of compressed genetic data each year. To put this into perspective, if you were to write this data onto standard DVDs, the resulting stack would be more than 2 miles tall. And with sequencing capacity increasing at a rate of around three- to fivefold per year, next year the stack would be around 6 to 10 miles tall. At this rate, within the next five years the stack of DVDs could reach higher than the orbit of the International Space Station.'"
In high energy physics, we rolled our own big data solutions (mostly because there was no big data other than us when we did so). It turned out to be terrible.
SIGSEGV caught, terminating
wait... not that kind of sig.
Everybody knows we should measure the pile height in Libraries of Congress. Or VW Beetles.
To put this into perspective, if you were to write this data onto standard DVDs, the resulting stack would be more than 2 miles tall.
Once that happens, they'll be able to stop storing it on DVDs and move it into the cloud.
...what a shitty storage medium DVDs are these days.
A cheap 3TB disk drive burned to DVDs will produce a rather unwieldy tower of disks as well.
G.
Publish a scientific, paper stating that potential terrorists or other subversives can be identified via DNA sequencing. The NSA will then covertly collect DNA samples from the entire population, and store everyone's genetic profiles in massive databases. Government will spend the trillions of dollars necessary without question. After all, if you are against it, you want another 9/11 to happen.
Actually ASCII files are the easiest to process. And since we generally use a handful of ambiguity codes, it's more like ATGCNX. Due to repetitive segments GZIP actually works out better than your proposed 2-bit scheme. We do a lot of UNIX piping through GZIP which is still faster than a magnetic harddrive can retrieve data.
Bit rot is also a big problem with data. So, the data has to be reduplicated to keep entropy from destroying it, which means a self corrective meta data must be used. If only there were a highly compact self correcting self replicating data storage system with 1's and 0's the size of small molecules...
My greatest fear is that when we meet the aliens, they'll laugh, stick us in a holographic projector, and gather around to watch the vintage porn encoded in our DNA.
Storage is not the problem. Computational power is.
Each genetic sequence is ~3GB but since sequences between individuals are very similar it is possible to compress them by only recording the differences from a reference sequences making each genome ~20 MB. This means you could store a sequences for everybody in the world in ~132 PB or 0.05% or total worldwide data storage (295 exabytes)
Now the real challenge is more in having enough computational power to read and process the 3 billion letters genetic sequence and designing effective algorithms to process this data.
More info on compression of genomic sequences
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3074166/
They should use a NoSQL multi-shard vertically intgrated stack with a RESTfull rails driven in-memory virtual multi-parallel JPython enabled solution.
Bingo!
putting the 'B' in LGBTQ+
Please... entire DNA genomes are tiny... on the order of 1Gb, with no compression. Taking into account the huge similarities to published genomes, we can compress that by at least 1000X. What they are talking about is the huge amount of data spit out by the sequencing machines in order to determine your genome. Once determined, it's tiny.
That said, what I need is raw machine data. I'm having to do my own little exome research project. My family has a very rare form of X-linked color blindness that is most likely caused by a single gene defect on our X chromosome. It's no big deal, but now I'm losing central vision, with symptoms most similar to late-onset Starardt's Disease. My UNC ophthalmologist beat the experts at John Hopkins and Jacksonville's hospital, and made the correct call, directly refuting the other doctor's diagnosis of Stargartd's. She though I had something else and that my DNA would prove it. She gave me the opportunity to have my exome sequenced, and she was right.
So, I've got something pretty horrible, and my ophthalmologist thinks it's most likely related to my unusual form of color blindness. My daughter carries this gene, as does my cousin and one of her sons. Gene research to the rescue?!? Unfortunately... no. There are simply too few people like us. So... being a slashdot sort of geek who refuses to give up, I'm running my own study. Actually, the UNC researchers wanted to work with me... all I'd have to do is bring my extended family from California to Chapel Hill a couple of times over a couple of years and have them see doctors at UNC. There's simply no way I could make that happen.
Innovative companies to the rescue... This morning, Axeq, a company headquartered in MD, received my families DNA for exome sequencing at their Korean lab. They ran an exome sequencing special in April: $600 per exome, with an order size minimum of six. They have been great to work with, and accepted my order for only four. Bioserve, also in MD, did the DNA extraction from whole blood, and they have been even more helpful. The blood extraction labs were also incredibly helpful, once we found the right places (very emphatically not Labcorp or Quest Diagnostics). The Stanford clinic lab manager was unbelievably helpful, and in LA, the lab director at the San Antonio Hospital Lab went way overboard, So far, I have to give Axeq and Bioserve five stars out of five, and the blood draw labs deserve a six.
Assuming I get what I'm expecting, I'll get a library of matched genes, and also all the raw machine output data, for four relatives. The output data is what I really need, since our particular mutation is not currently in the gene database. Once I get all the data, I'll need to do a bit of coding to see if I can identify the mutation. Unfortunately, there are several ways that this could be impossible. For example, "copy number variations", or CNVs, if they go on for over a few hundred base pairs, are unable to be detected with current technology. Ah... the life of a geek. This is yet another field I have to get familiar with...
Celebrate failure, and then learn from it - Nolan Bushnell
And we can double storage efficiency by using two stacks! Clearly, they need to hire one of us.
Bio questions? Ask me to start a Q&A journal. Computer analogies available for most topics!
This is very much the case. I work as a bioinformatician at a sequencing center, and I would say we see around 50-100G of sequence data for the average run/experiment, which isn't really so bad, certainly not compared to the high energy physics crowd and given a decent network. The trick is what we want to do with the data: some of the processes are embarrassingly parallel, but many algorithms don't lend themselves to that sort of thing. We have a few 1TB ram machines, and even those are limiting in some cases. Many of the problems are NP-hard, and even the for the heuristics we'd ideally use superlinear algorithms, but we can't have that either, it's near linear time (and memory) or bust which sucks.
I'm actually really looking forward to a vast reduction in dataset size and cost in the life sciences, so we can make use of and design better algorithmic methods and get back to answering questions. That's up to the engineers designing the sequencing machines though..
The snow doesn't give a soft white damn whom it touches. -- ee cummings
why aren't they storing it in digital DNA format
Because they need to be able to read it back quickly, and error-free. Add to that, it's actually quite expensive to synthesize that much DNA; hard drives are relatively cheap by comparison.
When I had to get the first draft of the human genome onto CD, I used 2 bit substitution and run length encoding on repeats. gzip definitely did not cut it.
A single finished genome is not the problem. It is the raw data.
The problem is that any time you sequence a new individual's genome for a species that already has a genome assembly, you need minimum 5x coverage across the genome to reliably find variation. Because of variation in coverage, that means you may have to shoot for >20x coverage to find all the variation. The problem is more complex when you are trying to de novo assemble a genome for a species that does NOT have a genome assembly. In this case, you often have to aim for at least 40x coverage (and in the 100x range may be better).
To get the data, we use next-gen sequencing. To give you an idea of the data output, a single Illumina HiSeq 2000 run produces 3 billion reads. Each "read" is a pair of genomic fragments 100 bases long. That means 600,000,000,000 bases are produced in a single run. The run is stored as a .fastq file, meaning that each base is stored as an ASCII character, and has an associated quality score stored as another ASCII character. So that's 1.2 trillion ASCII characters for a single run, or about 1.09 terabytes uncompressed. This does not include the storage for the (uncompressable) images taken by the sequencing machine in order to call the bases. They can be an order of magnitude larger. A single experiment may involve dozens of such runs.
There is an expectation that these runs will be made available in a public repository when an analysis is published. That puts great stress on places like NIH, where 1.7 quadrillion raw bases have been uploaded in about the last four years:
http://www.ncbi.nlm.nih.gov/Traces/sra/
You are correct when you say that computational power is a bigger problem, but again, this is not related to the three billion bases of the genome, which is trivial in size. Once again, the problem is the raw data. When assembling a new species' genome from scratch, you somehow have to reassemble those 3 billion pairs of 100-base reads. The way that is done is by hashing every single read into pieces about 21 nucleotides long, then storing them all, creating a de Bruijin graph, and navigating through it. The amount of RAM required for this is absolutely insane.
Scientists who viewed this sequence also viewed these sequences...