Hacking DNA: CRISPR, Ken Thompson, and the Gene Drive
blog.ycombinator.com
blog.ycombinator.com
Some protein designs that undergird entire companies:
Cure hemophilia: https://serotiny.bio/notes/proteins/hbb/
Vaccinate against HIV: https://serotiny.bio/notes/proteins/ecd4ig/
Cure the American Chestnut of its blight: https://serotiny.bio/notes/proteins/oxo/
Make spider silk clothing: https://serotiny.bio/notes/proteins/adf3/
"There are plenty of constructs I'd love to have but don't have the time or expertise to make them. However, could this tool make them? There's so much expertise that goes into "if i hang this tag here, will the protein still traffic to where it needs to go/will it interfere with other domains/will it hang on the right side of the membrane/if I change literally one other amino acid the new tag will work but not if I leave it normal." The experts can already design proteins, but if this tool democratizes that, who's going to do QC?
AFAIK the real way people make substantive changes to proteins is to make 10 variants, test them all, and iterate the designs that worked best. Feels like we're miles away from being able to drag and drop functions."
And "In principle proteins are modular but I can give you ten examples where that principle ruined someone's life. The QC is hard and is case-by-case."
What is your editor aiming to help with? Is QC still the rate-limiting factor to innovating new protein designs with your tool? Does your tool help with QC in any way?
Thanks for sharing, I'm always interested to see new tech in this area.
- We keep track of how domains have been used before (so as to mitigate if not entirely predict that QC). We are trying to encode that expertise in a way that is intuitive and frictionlessly sharable. The more people use it the more helpful it becomes. We help encode biological institutional knowledge.
- We collaborate with DNA synthesizers so that you just do the designing, and we help get you the designed DNA to your bench at a reasonable cost. You get exactly the sequence that you want, not what someone gave your neighboring lab's buddy sometime ago, or what some restriction enzyme/quick-change let you that's probably good enough. We alleviate the pain of cloning.
- We help you build entire sets by dragging and dropping: "I want these three fluorophores upstream, these 4 linkers on these proteins of interest" - Boom! Buy! Done! (in literally as much time). We make efficient high-throughput design.
- It is true, there are many proteins that are note entirely modular. But there are many more than 10 examples of proteins designed exactly that way (see Addgene's database). Biology is hard. We help you prevent failure buy being able to see how other successes have been achieved, prevent stupid mistakes, and get you your material to you as quickly and cheaply as possible. If your design does fail, it has failed faster and with less sunk cost than if you had to do all the construction yourself.
- Finally, because combinatorial design and purchasing is reasonably priced (and you don't have to do the construction work), you can order an entire logical set of proteins to actually figure out which would work and which would not. When I cloned I made some tiny fraction of what I expected to work purely because cloning was so painful - and if that small fraction failed I didn't even bother to figure out why or what would have worked.
I'd appreciate any feedback you might have.
If I'd like to make trees that have a firefly glow at night, is this something that can be done by an amateur?
How about creating plants with higher efficiency of photosynthesis than 4-5% in most plants.
However the search-space to get to those solutions is enormous, and getting to those very simple conditions is still a significant (if no longer herculean) undertaking. If you're given the correct tools and a good protocol it's really pretty easy to accomplish useful and interesting biological tasks (again you already do this when you make beer/bread/yoghurt/kimchi). But for the time being engineering an entirely novel biological function would require a fairly sophisticated understanding of the context you're trying to fiddle with.
-User designed sequences (maybe I just couldn't find it). Everyone wants to drop their primer of interest in, their unique tag, their CRISPR site, etc. Also, some well-balanced GC spacers would be helpful, even make them unique so people could PCR off them if needed.
-Barcoded sequences. This will ruin / complicate the auto IDT pricing, but a nice drop-in feature would be a degenerate barcode sequence. This is all the range with MPRAs or other high-throughput methods, and could be really helpful
-Architectures. Say I want to design a lentivirus construct. Give the user the scaffold to drop their payload in, with set LTRs, etc. Screen for no polyA signals in their design, etc. An easier way to walk novice users into design (maybe best not to start with lenti...)
Anyway, keep up the good work, it's a cool product!
Email is my first name at my company's domain.
-justin
Key point:
One source of this resistance is the CRISPR system itself, which uses an enzyme to cut a specific DNA sequence and insert whatever genetic code a researcher wants. Occasionally, however, cells sew the incision back together after adding or deleting random DNA letters. This can result in a sequence that the CRISPR gene-drive system no longer recognizes, halting the spread of the modified code.
CRISPR/Cas9 is a nuclease, it cuts DNA. It does not stitch it back together, the cell does that afterwards. Sometimes that goes smoothly, other times it doesn't, but one thing that happens frequently when you cut DNA is mutation. It's why radiation exposure is bad, DNA breaks cause mutations. If mutations yield immunity, then something that depends on mutations NOT happening to keep operating, but itself CAUSES mutation, is unlikely to function for very long.
[1] http://www.nature.com/news/gene-drives-thwarted-by-emergence...
Yes, this is something that is mostly not talked about. CRISPR is a virus engineered to attack and penetrate human cells and modify its DNA. It can be used for good. But at the same time the technology is becoming so easy accessible that it can be used to create a virus capable of whipping out the people. Or put a dormant gene in there that gradually kills people.
Up to 80% of people have Oral Herpes by some estimates. Imagine a virus like that could be engineered and spread quietly that in 10-20 years cause a worldwide cancer.
My point is, CRISPR can be used for nefarious activities. It's inevitable! So we need to create an antidote for it to prevent unwanted viruses that might one day be created by CRISPR.
It does seem likely that an arms race will likely occur.
In particular, the gene driver really is only of interest in organisms with very rapid generational turnover. It 'rapidly' speeds the propagation of a mutation over sexual transmission proportional to the generation time. This works for mosquitos. It makes no sense to use a gene-driver on a human population. A virus itself is, well, virulent - they need no help from CRISPR to get there.
Further, there are always exceptions/antidotes/alternatives in biology. In fact the recent discovery of some (useful) Cas9 inhibitors was predicated on the idea that there are organisms out there that have CRISPR systems in them that are themselves flanked by excision sites. By all rights they should not be able to have both the CRISPR system and have that system be marked for destruction - but they did - simultaneously. So they went searching, and they found the inhibitor in those organisms. And there are many more out there. [2]
I mean, smallpox has been synthesized, and it spreads fast, has pretty high mortality rate, and hardly anyone is vaccinated against it any more.
I can't think of any military (or even terrorist) reason why you'd want that kind of delayed mortality, but maybe I'm overlooking something.
Not much movement in getting that into the clinic yet, however, for reasons that are entirely cultural. The first company to get one of these working via medical tourism will make a lot of money. CRISPR makes that goal pretty easy in comparison to the past.
These changes include things like extra follistatin or knocking out myostatin to gain greater muscle growth and less fat tissue, or removing or disabling ANGPTL4 to reduce heart attack risk by 50%. There are others along the same lines and more are being discovered as sequencing becomes ever cheaper.
The comparative lack of effort to make enhancement gene therapies a reality is nothing short of crazy, given the observed benefits in the few humans lucky enough to already have these variants or loss of function mutations.
(I have no affiliation with the product, I just like it): http://www.the-odin.com/diy-crispr-kit/
https://www.scientificamerican.com/article/u-s-military-prep...
This whole video is great btw, where Arati Prabhakar spins through CRISPR, Gene Drive and a heck of a lot more in a presentation at UW.
This road seems a much bigger existential risk than any AI holocaust scenario I can imagine. It's very scary stuff and everyone in the field seems to be brimming with glee at how easy it all is.
Just consider that the very mundane HeLa cell line alone managed to jump continents and invade the most stringent safety precautions at dozens of labs. At least they are pretty benign.
Now imagine a nice mosquito delivery vehicle with payload strapped on that nobody has a real clue what it is capable of. I shudder to think.
A lot of people talking about how to re-establish the baseline and that it's not a big deal and nobody should worry.
I get that there are very many wonderful things that can come from this as well, but on the other hand the research is going very fast and this is one area we should be exercising extreme caution.
Als; CRISPR can specifically induce a strand break (which is the FIRST step), but it is not quantitative, nor does it do selection. Generally as a part of a CRISPR protocol you have to counterselect for cells which have the DNA integration in them.
There is quite literally no computer science analogy to this. This would be like having editing a database that is sharded and replicated over multiple servers, randomly integrating the change you want on some (probably small) fraction of the servers, and then going through, scanning all database replicates for the existence of the change, and physically destroying with a hammer the servers that had shards that were unchanged, as the mechanism to insure eventual consistency across your database.
CRISPR is pretty exciting but we also need to make sure there's a rich commons to build on. I encourage everyone to look into and support projects that make free/libre genomic data more available.
[1] http://www.personalgenomes.org/harvard/about
It wouldn't have surprised me to hear that Thompson had moved into gene hacking (at the age of 74) but I think he still works for Google.
In the Trusting Trust attack, not only you change the compiler to miscompile the "login" program, but you change the compiler to miscompile the compiler itself so the "login" miscompilation persists even if you later revert all changes to the compiler source (as long as you use the new binary once, of course). Does this part of the attack have a gene drive analogy?
Wild type mosquito DNA: =================
Desired change to DNA: =============XX==
DNA encoding Cas9 and XX-Payload: =C9(XX)=
Gene Drivered Mosquito (before activation) (change to compiler): ====C9(XX)=============
Gene Drivered Mosquito (after activation) (change to login code): ====C9(XX)========XX===
------------------------------
A normal mosquito mating with a XX mutant mosquito:
Mother Gamete: -----------------
Father Gamete: -------------XX--
Mendelian fraction of children (and grandchildren) are either --, --/XX,, or XX/XX (if two chromosomal copies)
A normal mosquito mating with a Gene Drivered mosquito:
Mother Gamete: -----------------
Father Gamete: ---C9(XX)----XX--
All children: ====C9(XX)========XX=== (for all 'N' chromosomal copies)
Well, I had to give advice to my son, and my advice to him—to the next generation—was to get into biology.
Definitely worth a read through the whole interview even if you don't agree.
[0] http://genius.cat-v.org/ken-thompson/interviews/unix-and-bey...
I'd also like to mention that this technology can be imprecise/precise and due to the combinatorics of the systems you have to design in order to create edits, the ability to reliably edit different parts of the genome can be difficult to predict. It's not perfect. Furthermore, DNA self-replicates, mutates, and has external forces which dictate whether those changes actually should persist in the environment. What if a gene-drive mutates thus rendering it non-functional? The number of mutations that could render such a system to be non-functional vastly outnumbers the number of gene drives you would need to deal with said mutations. Although these types of mutations are unlikely, I'd just like to emphasize that DNA is constantly evolving, unlike say, a series of UNIX commands.
Also, spider-silk producing goats existed before CRISPR was used to create spider-silk producing goats. https://phys.org/news/2010-05-scientists-goats-spider-silk.h.... And so did glow-in-the-dark cats: http://www.bbc.com/news/science-environment-14882008. The only difference is that they didn't have to pay millions of dollars in licensing fees to the Broad Institute.
I get that this sounds like I'm saying CRISPR sucks. It doesn't. It's a very valuable technology, which has much to offer. I just want to readily identify the ceiling of the hype.
The floor to error is that a wild-type Cas9 protein with a naively generated guide RNA, and a payload DNA along with it, just dropped into a petri-dish of cells correctly edits some single-digit percent of cells. That's the crudest of experiments a grad student might do on their first try. And that it is that robust (in a biological laboratory) is precisely why it is so powerful. Single-digit percent activity on an unoptimized first try is actually pretty amazing in the context of biology.
All the pop science articles seem to make it out to be a flawless system, and the engineer in me is highly skeptical of that narrative.
The cutting splits the DNA in half completely. It has to rejoin itself (by a not-well understood mechanism termed 'NHEJ'). This rejoining of the DNA is the kind of 'well it works (mostly, sometimes!)' part that is actively being better understood. Cas9 really has no relevance to this part of the process - unless it is used as a platform to attach DNA-repair machinery onto.
Like in all things biology, concentration and time matter. So if you leave Cas9 on 24/7 at very high concentrations with nothing to do, it will go ahead and cut all over the place. The goal is now to bring to bear everything else we know about biological expression to put the Cas9 homing system in its place, only when it should be in place. We already know pretty good mechanisms to make sure Cas9 only turns on under certain conditions, at certain times, is inhibited under other conditions, is otherwise made even more specific. Those are engineering 'details' at this point. And by 'details' I mean you have a huge academic and corporate effort underway currently figuring out how most efficiently and most accurately control Cas9's function.
Common results with a naive wild-type Cas9 (which we have moved on from in a lot of ways):
(no change): ====== --> ======
(unhealed breakage): ====== --> === & ===
(forwards insert): ====== --> ===(insert)===
(backwards insert): ====== --> ===(tresni)===
(tandem insert): ====== --> ===(insert)(insert)(...)===
(off-target insert): ====== --> =(insert)=====
(intermediate failure: ====== --> ===(insert) & === || ....
There are also very powerful uses for Cas9 that do not cleave DNA at all. Sticking a gigantic 'BREAK()' command onto Cas9 (that cannot cut DNA) can be very useful where the natural break command has been lost.
2. Off target errors- this is when Cas9 cuts the genome at the wrong spot- this is the main error people are worried about. Computational efforts in guide design have attempted to reduce this, and will get more advanced as we learn more about the binding kinetics of Cas9. These also occur non-randomly, at highly quantifiable frequencies, and are due to biophysical rules. Also, newly discovered and engineered Cas variants have substantially better specificity and will work to further reduce this problem. Also off target effects are usually not problematic and can be tested for (will be bred out of crispr-ed crops)- however, if you're going to do CRISPR in live humans it is likely that regulatory bodies will require a demonstrated very low probability of off target effects.
Basically, errors are an issue to spend a lot of time thinking about when designing a crispr system intended as a human therapeutic. For most scientists, however, errors are not terribly important or common and they don't need anything beyond a bit of rational guide design for their experiments.
Why some guide RNAs work well, setting up Cas9 to cut and disable a gene nearly 100 percent of the time, while others bind but seldom or never knock out the gene, has been a puzzle since the technique was invented by Jennifer Doudna of UC Berkeley and Emmanuelle Charpentier of Umea University in 2012. The cutting efficiency varies with the type of cell and the particular cell line...
1: http://news.berkeley.edu/2016/01/20/advance-improves-cutting...
2: http://ko.cwru.edu/services/crispr_cas.shtml#successrates
3: https://phys.org/news/2016-08-crispr-cas9-genes-disrupt-dna....
For 'edits somewhere else', which are new mutations which might be harmful, as far as I can tell as an interested layman, in terms of the last category of errors, 'off-target' edits, the state of the CRISPR art is a low, effectively near zero mutation rate:
- "High-fidelity CRISPR-Cas9 nucleases with no detectable genome-wide off-target effects" Kleinstiver et al 2016, https://www.gwern.net/docs/genetics/2016-kleinstiver.pdf
- "Rationally engineered Cas9 nucleases with improved specificity", Slaymaker et al 2016 https://www.gwern.net/docs/genetics/2016-slaymaker.pdf
- Church, April 2016: "Indeed, the latest versions of gene-editing enzymes have zero detectable off-target activities." http://www.wsj.com/articles/should-heritable-gene-editing-be...
- Church, June 2016 "Church: In practice, when we introduced our first CRISPR in 2013,19 it was about 5% off target. In other words, CRISPR would edit five treated cells out of 100 in the wrong place in the genome. Now, we can get down to about one error per 6 trillion cells...Fahy: Just how efficient is CRISPR at editing targeted genes? Church: Without any particular tricks, you can get anywhere up to, on the high end, into the range of 50% to 80% or more of targeted genes actually getting edited in the intended way. Fahy: Why not 100%? Church: We don't really know, but over time, we're getting closer and closer to 100%, and I suspect that someday we will get to 100%. Fahy: Can you get a higher percentage of successful gene edits by dosing with CRISPR more than once? Church: Yes, but there are limits." http://www.lifeextension.com/Lpages/2016/CRISPR/index
In human embryos specifically, the published work is below state-of-the-art (unsurprising, given the taboo) but shows increasingly good performance in terms of making the desired edit and also not making undesired edits:
- Liang et al 2015 "CRISPR/Cas9-mediated gene editing in human tripronuclear zygotes" http://link.springer.com/article/10.1007/s13238-015-0153-5%2... http://www.nature.com/news/chinese-scientists-genetically-mo...
- Kang et al 2016, "Introducing precise genetic modifications into human 3PN embryos by CRISPR/Cas-mediated genome editing" https://www.gwern.net/docs/genetics/2016-kang.pdf http://www.nature.com/news/second-chinese-team-reports-gene-...
- Komor et al 2016, "Programmable editing of a target base in genomic DNA without double-stranded DNA cleavage" https://ase.tufts.edu/chemistry/kumar/jc/pdf/Liu_2016.pdf
- http://www.nature.com/news/chinese-scientists-to-pioneer-fir...
- "CRISPR/Cas9-mediated gene editing in human zygotes using Cas9 protein" Tang et al 2017 https://www.gwern.net/docs/genetics/2017-tang.pdf : no observed off-target mutations (but very small sample size); efficiency of 20%, 50%, and 100%
1) cut DNA (this is the part that CRISPR does)
2) add replacement DNA
3) kill everything that doesn't have the replacement DNA in there
4) optional: repeat the process to 'clean up' your modification [0]
Basically in the pre-CRISPR era, if you, say, wanted to make a transgenic mouse, you had to take some embryonic mouse stem cells and just add DNA and hope that it found its way into the "right place". If you were adding a totally new gene - usually not a big deal because you kind of don't care where it winds up, shoot first, do your experiment, ask questions later.
If you were knocking out or replacing a gene, (aka specific location addressing) it is a big deal. Suddenly step 3), while necessary, is not sufficient to guarantee successful DNA modification. Furthermore "checking" is really hard, You need to implant the modified stem cell into a mouse embryo, create a chimeric mouse (a mouse that has cells from two genomes), hope to hell that the cell you implanted into the embryo randomly got chosen to turn into an ovary, and then take the children of the mice which should have the gene modification. And then you check the genome of the mice and it turns out that it's all wrong and you have to start from scratch (literally had a friend who unluckily spent the first 5 years of grad school repeating this process about 10 times IIRC). This gets really complicated when step 1) is inefficient, or cuts in the wrong place, or has a spontaneous integration preferentially in the wrong place - you off the bat take a huge hit (a hit on the order of 90%, if you were lucky).
With CRISPR, you basically derisk step 1), pushing the 'hard part' of molecular biology elsewhere. Suddenly your step 3) selections instead of being a low-yield, probabilistic crapshoot, are near-quantitatively correct. Since it's also the first step, suddenly the one or two 'actually hard' parts of molecular biology, like in the case of the mice the complicated process of generating chimerics, etc, can really be tackled head on by sheer numbers much more easily.
[0] in order to do 3) you usually need some extra stuff (resistance genes) that you might not want in your 'final work', so you might have to come up with a second stage to 'clean this up'.
Keep in mind that Church and Mukherjee have not really ever 'been in the trenches' actually doing these things (church was a structural biochemist and later a yeast geneticist - yeast are easy peasy). Their grad students and postdocs have.
I've not read the article yet, but an answer to this question is, "not at the very start".
Most of the major research organizations working with CRISPR leverage Benchling -- the Broad Institute, various labs at Berkeley, companies like Editas, etc.
Shoot me an email: josh at benchling dot com
I work in the R&D side of the company, but our front end team has built a really nice website to design CRISPR gene knock-out/knock-in experiments from start to finish. Its all free to use until the final checkout where you order your CRISPR guides.
It seems to be such a huge body of knowledge and the field seems to be moving so quickly.
"CRISPR... is to genomics what vi (Unix’s visual text editor) is to software. "
This terrible comparison between DNA and code has been happening for years. In 2012, Thiel did the same thing in CS 183. We get to look back and see that he was wrong, but we'll have to wait a while before the 'DNA is code' group will admit this essay is bad. At least its not a TED talk and at least its better than Aubrey De Grey.
"Each VC on the panel made 2 predictions about technology in the next 5 years. The audience voted on whether they agreed with each prediction. One of my predictions was that biology would become an information science." -PT
The case of CS people commenting on biology via misinformed comparisons to computing eventually became known as the Andy's Grove Fallacy, named after his comments on the state of cancer therapies: