Nature Methods has retracted the paper. Two of the authors formally agreed with the retraction, while the rest did not.
UPDATE, 3/26/18:
Schaefer et al. have now uploaded a manuscript to bioRxiv, stating that they have now sequenced additional CRISPR mice but did not find off-target mutations in these animals, in contrast to what they claimed in their Nature Methods paper. My perception is that they (the authors) are not going to retract the first paper.
Original post:
FSSSSSSSSSSS……(sound of me being thawed out of cryogenic storage)
This paper got a lot of attention, and even temporarily eroded stock prices of CRISPR biotech companies. Columbia University first pushed the story in a press release on May 30. It got picked up and the story was off and running.
Cosmos: May 30. CRISPR gene editing causes hundreds of unintended, off-target mutations
The Conversation: May 31 by Ian Haydon. CRISPR Controversy Fuels Debate in the Science Community
Wired, June 4: Crispr’s Next Big Debate: How Messy Is Too Messy?
Seeker: June 6. CRISPR DNA editing can cause hundreds of off-target mutations.
The Scientist, June 7: Was a Drop in CRISPR Firms’ Stock Warranted?
New Atlas: June 12. The CRISPR controversy: Scientists skeptical over recent critical study
USA Today, July 24: CRISPR gene editing tool: Are we ready to play God?
Forbes, July 1: CRISPR Gene Editing Controversy: Does It Really Cause Unexpected Mutations?
I will not present a very detailed breakdown of the paper here. The serious problems in the paper have been addressed previously, and well, by others. Gatean Burgio has published a very good analysis on Medium. A publicly available preprint also deals directly with the key data and concludes the authors were mistaken: Questioning unexpected CRISPR off-target mutations in vivo. Kim et al, bioRxiv, 6/30/17.
However, I can contribute a few pieces of key genetic-y information that have not been described in much detail. And these really involve details, and they really do matter. They have to do with what inbred mouse strains really are and how CRISPR mice are made.
Inbred does not mean identical. First, there is a widespread assumption that separate mice from the same inbred strain are genetically identical. For almost all practical purposes, this is a pretty useful assumption. But technically, of course, it is wrong. Each new mammalian embryo has on the order of 50-100 new mutations (mostly single base changes - “SNPs” - but other types too). So even siblings from (theoretically) genetically identical parents will not be perfectly identical. You’ll have to look hard to find the differences though; the new mutations amount to fewer than 1 new base per million bases of genomic DNA. That’s why hardly anyone needed to worry much about this problem before the advent of whole-genome sequencing, and its resulting power to actually find these variants.
Let’s pause briefly to think about inbred mouse strains. By definition, these strains were initially derived by repeated brother-sister matings, carried out for usually 20 or more generations. This is adequate to eliminate almost all of the initial heterozygosity present in the founding mating pair. However, some new SNP mutations still arise in each new generation. Each new mutation might be weeded out by chance thanks to the continued inbreeding. Alternatively, it has a chance to rise in allele frequency or even become fixed.
How many polymorphisms? It’s good to do a thought experiment about how many variants we might expect to see within a population of inbred mice. Population genetics theory predicts that the number of existing polymorphisms in a fully inbred (brother-sister mating) mammalian population will be approximately 4N, where N is the mutation rate per individual per generation. Let’s assume N is 50; 4N = 200. So we’d expect about 200 polymorphisms to exist, all told, among the genomes of the brother and sister. (e.g. one or both is heterozygous for the variant in question, or both are homozygous but for different alleles).
But 200 polymorphisms is just a lower bound. In theory this is what we’d expect for an inbred population of two animals, which is impractical in reality. In practice, inbred strains have larger effective population sizes – JAX probably has hundreds of FVB mice in their actual stock colony at any moment, if not more. Although they take pains to maintain as close inbreeding as possible with this colony, I believe it’s safe to assume that the number of SNP polymorphisms extant in a vendor’s FVB stock is in the thousands, when the whole population is considered. Mind you, many or most of these variants may be at low allele frequencies - perhaps only in a single animal. But many will certainly exist in multiple animals. (These individuals are still incredibly similar to one another as compared to, say, wild mice - any two of which almost certainly will have several million SNP differences between them.)
Sibs have a lot in common. Now we can do another thought experiment: if we take two true siblings from the inbred strain population, they will share more of these ~2000 SNPs in common with each other than either would share with another mouse that is not a full sibling but is from the same population. This is one of the simplest concepts in genetics, but it’s still not something we usually consider when dealing with mice of the same colony of inbred mice.
Let's make CRISPR mice. But how? Ok. Now back to CRISPR. How are CRISPR mice actually made? The process begins with 1-cell zygotes, which must be physically injected with the CRISPR reagents. The process is essentially the same as was perfected over thirty years ago for the production of transgenic mice; but instead of injecting a DNA transgene, we inject the zygotes with a mixture of CRISPR guide RNA, Cas9 protein or mRNA, and (optionally) a targeting DNA molecule.
To inject them we need to get the zygotes into a culture dish under a microscope. But we’ll need to collect about 100 or more of them to be sure of getting several live CRISPR’d mice in the end. Most, but not all of the zygotes will survive the initial injection process; but only some of these will implant and develop normally, and only some of these will be CRISPR-edited. That’s why we need to start with a lot of zygotes.
Fine - so we need to have 10 females of the right age, freshly mated the night before the injections. We’ll order these from a vendor like JAX. The night before injections, we’ll set up each female mouse in a separate mating cage with a stud male of the same strain. We keep these males on hand at all times; they are proven fertile males, usually somewhere between 3 and 12 months old.
A batch of embryos. On injection morning, we euthanize the female mice and quickly dissect the zygotes out of the oviducts into one culture dish. Now, they are all in one batch, and at this stage nobody ever keeps track of which female produced which zygotes. They are all supposed to be genetically identical, right? (sarcasm)
Some are sibs. So finally, we can do our final two thought experiments: If we take any two of these embryos at random, what is the chance that they will be more closely related to each other than to a third, randomly chosen embryo from the same batch? Obviously it is quite possible that the first two might be true siblings. The batch of embryos is made of ten or so "families" of sibling groups. But we can’t tell under the microscope which ones are true siblings or not. The point is, any two that you pick might be siblings. And it is certainly possible that the first two might simply be more closely related to each other than they are to the third one, even if none of them are siblings.
And last, let’s assume that when we pick two embryos, that they are in fact true siblings and that the third one we pick is not a true sibling of the other two. Now let’s sequence their genomes and see how many SNPs they all share. If there are several thousand SNPs in the vendor’s mouse colony, we expect that full sibs will likely share a thousand or more SNPs in common that would not be shared with another randomly chosen mouse from the same colony. This is based on the idea, as explained above, that any mouse is likely to have on the order of a few thousand polymorphisms. Siblings, of course, will share about 50% of those between them.
Final thoughts. This is pretty exactly much what the authors found. I see no reason to conclude the mutations were caused by CRISPR. It certainly shines a bright light on the problem of distinguishing pre-existing variations from new mutations in gene editing experiments – which is exactly the problem that tripped up the authors of the first human embryo CRISPR paper. The genome is big, and there are lots of variants to keep track of. It’s actually a much tougher problem than doing CRISPR in the first place.
