Hacker News Reader

Comment by exdsq | Hacker News Reader

What’s the choke point of scanning an entire genome? Is it purely computational? Imaging? Biological unknowns?

For IVF embryos, it's the small amount of genetic material which makes accurate sequencing difficult. For standard human whole genome sequencing (WGS), the fastest times are down to five hours[1] (although that is exceptional), for which the imaging was massively parallelized and it seems that the choke point was computational for that particular case (sequencing like this basically multiples the DNA, shreds all the copies, images each shred in a reasonably-parallel manner, and then uses longest-common-subsequence to put all the shreds back together digitally).

Average WGS times for something a person can buy for a couple hundred dollars take a few weeks, and will usually be bottlenecked by logistics (mailing things, getting it to the lab, etc.) and then the imaging (a flow cell runs the fragments over an imaging device; parallelization is accomplished by running multiple flow cells at once).

I might be slightly off on my terminology here but the idea is broadly correct, I think: Differing grades of sequencing are determined by the average depth of fragments for each base-pair -- e.g. 1x sequencing implies an average coverage of one fragment per base-pair, which could lead to mis-aligned or improperly joined samples. 30x-50x sequencing is pretty common for consumer grade WGS these days, which gives a higher number of fragments that allows you to weed out bad reads or incorrect joins of the sequence, though errors can still occur. This is usually sufficient for diploid (having two values for each base pair, one maternal and one paternal) humans, since it's important to determine if the basepair is heterozygous (different bases from each parent) or homozygous (same base from each parent). However, if you're doing cancer sequencing, a tumor may contain many many more different genes, so a massively deeper sequencing depth is required to ensure you have reasonable statistical certainty on what variations occur within the different genomes. Higher depth == more material == longer time to sequence since you have more material to image.

Note that services like 23andMe are NOT doing whole genome sequencing; they are using specialized flow cells that only examine certain specific sites on the genome called single nucleotide polymorphisms[2], which are traits or indicators that are turned on or off by a single nucleotide, hence it being much cheaper and faster to run.

[1]: https://med.stanford.edu/news/all-news/2022/01/dna-sequencin.... [2]: https://en.wikipedia.org/wiki/Single-nucleotide_polymorphism

Average WGS times for something a person can buy for a couple hundred dollars take a few weeks,

I hadn't realize that we'd actually gotten the $1,000 genome. It was a big point of discussion 20 years ago, but seems to have arrived without the fanfare I would have expected. According to Wikipedia, it's been around for a few years, and claims the X-prize for it was canceled before that since it no longer looked like much of a challenge.

So we sailed right through it and are now pushing an order of magnitude below that. Thanks; I had missed it.

As someone who was interested in projecting that back in like 2013 and watched the $1k arrive, I got the impression that it never became headlines in part because there was never a crystal-clear moment for the "$1k genome".

Someone spends $1k and gets back a DVD of data ~2017. Well, is that $1k just the cost of the reagents? What about storage/assembly/analysis? How many samples did you run in that batch? Are you including the amortized cost of the sequencer (what utilization, discount rate, and upfront cost)? Are these long reads or short reads? To what depth? (There being energetic advocates for low depth / few passes + imputation.) Is this list price or institutional bulk price? Did Illumina give you a special deal (everyone knows they have hefty margins)? Are you even allowed to talk about the cost by Illumina? And so on. If the genome were $100k, or $10k, these could be rounded off, but as the cost drops, these nuisance factors become ever more of the price.

There just was never a specific moment where you could write a headline "$1k Genome Arrives At Long Last!", and by the time you get Dante Labs doing consumer sales for $500 or whatever (as long as you don't mind waiting half a year for the results), it's a bit late to write any articles as it is no longer 'news', now is it? So, people just stopped talking about the $1k genome because it had arrived quietly in the night, and switched to the $0.1k genome. And then you get similar issues, like that '$100 genome' BGI announced the other year - yes, there's a '$100' number there, kinda sorta, but there are so many caveats and questions that it discourages anyone from saying 'the $100 whole genome is here at long last!'

Different grades of sequencing are basically down to the technology and depth.

The short read sequencing you're describing makes reads that are like motes of dust in comparison to the genome.

More modern long read methods can read off sequences with high accuracy and/or length. These tend to be much higher in cost and substantially higher in the completeness and quality of the resulting bioinformatically inferred genome.

Thanks for your response! So is it true to say every genome could be sequenced, assuming time and money isn’t an issue?

Starting material is a huge problem. You need at least 50ng of DNA for an Illumina sequencing run, usually more if you want to sequence ultra deep, which you will need for whole genome sequencing. For PCR you need at least 0.1ng (usually higher) of DNA. But a single cell has about 80 pg of DNA. So you need to amplify the DNA several orders of magnitude before you can apply standard PCR. Several approaches can solve this problem, DOP-PCR, MDA, MALBAC. But each results in biased amplification, with differing biases with each method.

Source: I worked on the computational modeling of this problem in grad school, https://academic.oup.com/bioinformatics/article/30/22/3159/2...

The most interesting idea from the article, that was new to me, was using larger samples available from each parent, to augment and guide the analysis of the small amount of material available from the embryo. Obvious idea after you hear it, but very clever nonetheless.

if you're asking whether we could sequence 7 billion people, yes, it would be absolutely possible. Obtaining samples would be harder than the sequencing. If you really took this seriously there would be only 3 sequencing sites in the world and they would be located close to the largest datacenters. Each sequence is about 100G data (the raw data file that's shipped around) so if you really wanted to keep all of that it's 700 exabytes, but since you're sequencing everybody, you can instead create a tree structure of deltas from references with additional nodes representing major haplotypes and other structure. This would be more like ~1M-1G/person, so 7 petabytes to 7 exabytes, well within the storage limits of a modern datacenter cluster. To go along with that storage, you'd also need ~1M fast CPUs and ~250K fast GPUs, plenty of RAM, and a really good privacy and security story.

All of that could be formatted for large-scale analytics and the problem treated much like machine learning/data science problems in ads and other similar areas.

If you're really serious, this is just a bulk buy for one of the data center providers and they would be happy to make a fixed profit margin contract to do this for you. I helped set up all the infrastructure to enable google to do this ~10 years ago (I got a patent on compressing DNA for storage) and now they're at the point of hosting this sort of service (terra bio). I could do everything I described above for $25B but there is absolutely no medical reason to do anything of this scale.

It's very doable, but we only know what a small amount of the genome is actually doing. Whole Exome testing hits the majority of the points people are interested in and is way less computationally expensive, since it's a fraction of the genome.

Okay interesting - so reading the genome is basically solved nowadays and it's understanding what it's actually doing where we're working atm?

Reply on news.ycombinator.com