Genetic Engineering – The main feature this week is a deep dive into genetic engineering, which sucked up pretty much all of my week.
Quick Links – The Apennine Colossus, Alex Jones is a beautiful woman, and how about we stop trying to make viruses worse?
Genetic Engineering
Dwarkesh Patel’s recent interview with Steve Hsu on genetic engineering, made me feel a bit like Neo getting a decade’s worth of updates about the state of the field smashed into my brain all at once. I had not realized just how much knowledge had been unlocked once we were able to start using machine learning to crunch the human genome. The main takeaway from all of it is that our genetic architecture is much simpler than a lot of people (including myself) had previously been assuming, leading to three big insights for me:
- Making straightforward changes to polygenic traits is much easier than I had realized.
- The potential magnitude of the changes we can engineer is much larger than I had realized.
- We are probably going to be able to start doing this much sooner than I had realized.
Hsu is an entrepreneur in the space, so obviously all of this should be taken with a grain of salt as he has plenty of incentives to encourage hype. But even if his claims are exaggerated by 50%, much of this news would still count as revolutionary for me. Below I have summarized the discussion as a condensed and reformulated Q&A to highlight the key points, with some supplementary material from Gwern and Ryan Beck worked in for additional context.
Q1: What is the relationship between polygenic traits and the genes that code for them?
Monogenic traits are simple. A condition like Huntington’s disease is a function of one, and only one gene. Nothing else in a person’s DNA except that one gene determine whether a person has that trait. By flipping that gene in one direction or the other, you have full control over the expression of that trait.
Polygenic traits are complicated. Things like height and intelligence are not the product of a single gene, but of the combined effect of thousands. But the relationship between these genes and the traits they code for has not always been clear. Is it merely additive? If you wanted to find out how tall a person will be using just their DNA, could you simply add up how many of the 10,000 associated genes were flipped in the “tall” direction? Or is the relationship more complex and nonlinear, requiring a deeper understanding of how different parts of the genetic code interact in order to discern how a trait will manifest?*
It would certainly be easier if the relationship were simple and additive, but for many years the assumption was that nonlinear effects were dominant. As we have begun using machine learning to analyze genomic data, however, we’ve discovered that polygenicity is mostly additive. It’s slightly more complicated than just looking at how many of the 10,000 correlated genes are flipped to 1 instead of 0, but in general this model is not that far off.
It is possible to understand why this may be the case by considering how the two options would fare in an evolutionary setting. If traits were dependent on complex, non-linear interactions between genes, it would be very hard for organisms to pass on those traits to their offspring. Sexual reproduction mixes up the genetic code, which would break most of those fragile nonlinear interactions. Simple additive relationships, on the other hand, would be much more robust to gene mixing. This would inevitably lead to additive mechanisms becoming the dominant means of controlling polygenic traits.
*Obviously environment also plays a role in most traits, but for the sake of this discussion I am only concerned with genetic contributions. Twin studies have shown that height is about 80% heritable, and intelligence 60-80% heritable, so for the purposes of discussion, these traits can be treated as being genetically determined.
Q2: How much potential room is there for the genetic improvement of any given trait?
A frightening amount. Take height. Height is controlled by about 10,000 genes. You only need to flip the square root of that number – 100 – to make a person one standard deviation taller. Flip 500 and you’d be 7 ft tall. And you’d still have thousands of genes left to flip.
We know this is true because we have analyzed enough people’s genetic code to find the correlations. But the length of our potential runway here can also be grasped intuitively by reflecting on our experience engineering crops and animals. Take chickens. The average chicken in 2005 was more than 4x the weight of the average chicken from 1957, and lay ~30x as many eggs as chickens in the wild. Now, these extremes are not necessarily good for the chickens, but we’ll get to that next. The main point here is that there is a lot of room for variance.
Q3: How big of a deal is pleiotropy (the condition where a single gene affects multiple unrelated traits)? How far can we modify genes to optimize for a particular trait before we start trading off against other traits?
It turns out that pleiotropy is not as big of a problem as biologists had previously assumed. With 3 billion base pairs, you can do the math to find that there are about 1000 independent traits that are fully disjoint in their genetic dependencies. There are many more degrees of freedom in our genome than we thought.
That also means that even in places where we do encounter pleiotropy, it’s possible to route around it. There will be hundreds or thousands of pieces of genetic code that affect most traits. If one of those pieces of code forces a negative tradeoff against some other beneficial trait, we can just flip a different gene instead. As discussed above, you don’t need to switch all that many to have a big effect. So it should take a while before we get to the point where we run out of switches that we can flip without any serious adverse consequences. Why is this not true for chickens, who definitely suffer as a result of artificial selection pressures? Frankly, because we don’t care enough to find the least harmful path for optimizing them. The same won’t be true for humans.
Q4: If it really is this simple, why haven’t humans already been optimized through natural selection?
For some traits, it is because they never faced selection pressures. Genes that code for Alzheimer’s or cancer or heart disease, for example, would never have been selected against in ancient humans who rarely lived past 40.
But the more important reason is that evolution hasn’t had enough time to optimize humans. If for every gene you flipped, the effect was to raise your chance of reproducing by 50%, you’d get a very rapid selection for those genes. But for most traits, changing a handful of genes just doesn’t give a big enough adaptive advantage for the gene pool to shift very quickly. It takes many, many generations for those effects to build up. In machine learning terms, evolution is working in a high-dimensional space with a very small slope. Gradient descent simply takes a long time to find the optimum in an environment like that.
Q5: So let’s get to the question we’ve all been waiting for: how easy will it be for us to boost intelligence?
Right now, we do not understand the polygenicity of intelligence as well as we do many physical traits. The best analysis of cognitive ability has managed to achieve a polygenic score of 9.7%. That means that we have currently only identified the genes responsible for 1/6 to 1/8 of the full genetic component of intelligence. In contrast, the best analysis of height achieved a polygenic score of 40%, meaning we have identified the genes responsible for fully 1/2 of the genetic component of height.
Part of the difficulty with intelligence seems to be that it is simply a very complicated trait, so coming to understand it will take time. But another obstacle is the difficulty of measuring it in the first place.
Right now, the main bottleneck for understanding the polygenicity of a given trait is data. You don’t just need lots of examples of genomes, you also need each of those genomes to be linked to a measure of the trait you are looking at. Take height. If you know the genomes of a group of people, but don’t know how tall each person is, there is no way to connect the genetic information you have to the trait you’re investigating. This lack of data linkages isn’t really a problem for widely measured and uncontroversial traits like height. But it is a big problem for intelligence.
The teams that are working to improve genetic screening technologies have access to piles of genomic profiles, but little of that data is linked to an intelligence score for the person it comes from. According to Hsu, this is often because the very concept of measuring intelligence is so controversial that no one even attempts it for fear of attracting criticism. While some data can be linked to cognitive assessment scores, for most datasets the best measure we have is educational attainment. This is imperfect, though, because educational attainment is as much a measure of conformity and conscientiousness as it is intelligence.
As we get better at measuring intelligence and build larger databases of genomic data linked to cognitive ability, our understanding of the genetic component of intelligence will inevitably improve. But for the reasons outlined above, progress on this front will be slower than for most physical traits.
For what it is worth, as of writing Metaculus gives a 57% probability that by 2050 we will have genetic engineering techniques capable of raising IQ by 10 points.
Q6: What does the future of genetic engineering look like?
The near future of genetic engineering will be closely tied to the IVF industry. Today, genetic screening is primarily used as a way to screen out IVF embryo candidates with a high risk of certain diseases. While it is also possible to screen embryos in utero, the real power of using this technology is in an IVF setting due to the optionality it unlocks. With a natural fertilization, you have the embryo you have. Screening can let you predict how that embryo might turn out, but the only thing you can do with this information is decide to either keep the embryo, or take another random spin of the wheel. With IVF, you can fertilize and screen a handful of embryos (right now, the median is 5), then select the most positive outlier from among them.
This optionality gives a much more direct lever for being able to optimize for the traits you are after. However, currently the overall effect of this optionality is still relatively modest. As of the state of the art in 2019, it is only possible to raise height by around 2.5 cm or intelligence by 2.5 IQ points. There are three lines of work that could potentially let us have even greater control.
The first is to increase the number of embryos you select from. Just as going from 1 embryo to 5 increases your optionality to select for desirable traits, going from 5 embryos to 500 will make it possible to exploit an even greater variance. The figure below, taken from Ryan Beck’s analysis of Karavani et al. shows how IQ gain varies as a function of polygenic score (how well we can screen for the genetic component of intelligence), and the number of embryos you have to select from.
With 5 embryos, you would need to understand almost half of the polygenicity of intelligence to achieve a 10pt IQ boost. With 500, you would only need to understand an eighth.
Right now, the bottleneck for how large you can make the embryo pool is the number of eggs you can harvest from a donor. One way to increase this is by using stem cells. By taking skin cells, inducing them to revert to their pluripotent state, and then turning those into eggs, it may become possible to create batches of hundreds or even a thousand embryos to select from. According to Hsu, this process has already been mastered for mice and rats, so getting it to work for humans should not be very far off.
The second line of work is in iterated embryo selection. Here you start with a set of embryos like normal, but instead of directly implanting the best one, you harvest stem cells from it to generate another round of embryos. By doing this several times in succession, you can compound the benefits of variation and achieve a hundred years of genetic progress in a single go. The science behind IES is still mostly theoretical, with Hsu noting that he is not familiar with any group pursuing this strategy full-time. But Gwern considers it to be probably the most promising of all approaches to genetic engineering when taking the perspective of the 2016-17 state of the art.
The third and most important line of work is in CRISPR gene editing. Eventually, CRISPR could allow us to surgically manipulate the genome of an embryo, allowing us to (eventually) flip an arbitrary number of the additive switches that contribute to polygenic traits. CRISPR would allow us to graduate from exploiting random variation to directly engineering the traits we are seeking. Hsu predicts that within a decade we will be able to start making massively multiplexed edits. Notably, Gwern is much less optimistic. But if CRISPR really can deliver on its promise, it will make other work focused on exploiting random variation redundant.
It should also be noted that cloning could be considered a form of genetic engineering as well. Due to the taboos involved, human cloning will likely never be realized, but I do think we should really give more thought to Alvaro de Menard’s argument that an army of a million John Von Neumanns and a few hundred Lee Kwan Yews may not only make the world a substantially better place, but may in fact be the only viable path towards averting a variety of extinction-level risks to the human species. (For the record, Erik Hoel doesn’t think this would do much good, though I find his claim almost impossible to believe)
Q7: How will genetic engineering impact equality?
Right now, 3-5% of children born in developed countries are born via IVF, with a max of around 10% in nations like Denmark and Israel where there is greater gender equality and healthcare that will pay for fertility treatments. Hsu has elsewhere argued that within a decade, he thinks that 20-25% of high SES kids will be born through IVF. But as Gwern notes, high cost, hostility towards the idea of fertility technology, the hassle of advanced family planning, and the pain and difficulty of the procedure itself will likely limit the use of IVF to a minority of wealthy families.
As long as genetic engineering is limited by the need to exploit random variation in IVF embryos, its benefits will overwhelmingly accrue to a small segment of the population. But if CRISPR unlocks the ability to make deliberate edits on any embryo, it will be much easier to imagine a future where genetic engineering is an option for all families. In fact, a new belief this has led me to toy with is that universal healthcare should be considered as a moral and political imperative because it may be the only way to ensure that CRISPR-based gene editing does not lead to unacceptable and destabilizing levels of inequality.
Notably, because we are further ahead in understanding how to select for physical traits than we are for intelligence, it is likely that to the extent genetic engineering does become possible, it will first impact equality in health and medical outcomes. Only later will we begin seeing differences manifest in cognitive abilities. This also means that cultural and political debates about the acceptability of genetic engineering will initially be focused on the health benefits of the technology, possibly softening the ground for later arguments about enhancing cognitive ability.
Q8: How is genetic engineering going to be regulated?
The costs of genetic sequencing are falling rapidly, and there are few other barriers to screening embryos. Once the DNA has been sequenced, it can be uploaded to services that can instantly produce a report for the parents and physicians. Any attempt to regulate this process will be challenging. Sequencing is highly unlikely to ever be outlawed completely. It’s use in screening embryos for health risks is too obviously beneficial. Some governments may consider measures prohibiting clinics from screening for certain traits, but the ease of uploading a DNA sequence to companies based in other jurisdictions would instantly threaten that regime.
Moreover, any country trying to restrict genetic engineering will also face arms race dynamics with those who rush to embrace it. Particularly if we get to the point where cognitive ability can be easily enhanced, genetic engineering will quickly become a matter of national security and economic competitiveness.
Q9: How will genetic engineering impact geopolitics?
The most facile prediction one could make is that China will enthusiastically embrace the technology at a population-wide level, Europe will fret and wring its hands as the opportunity passes the continent by, and America will be caught in a tug of war between progressive cultural elements and an eventually victorious elite class with the money, knowledge, and will to ensure that every possible advantage accrues to their offspring. But I think it’s worth trying to imagine the ways in which each of these predictions may be proven wrong.
For Europe, the most obvious alternate path is one in which CRISPR arrives much more slowly than we anticipate, and IVF-based gene editing remains the only viable method for genetic engineering for an extended period of time. Because of Europe’s more robust social safety net, its citizens will be able to more immediately begin taking advantage of any any IVF-based innovations. Already, the country with the greatest uptake of IVF is Denmark, where 10% of kids are born as a result of the practice. If Europe were to embrace IVF-based embryo selection, it could turn the continent into a leader in both the research and practice of genetic engineering.
In America, one could imagine a future in which a grand alliance were formed between the progressive and establishment wings to ensure the universal availability of genetic engineering. The left could be enticed by the promise of using the technology to achieve greater genetic equity and the establishment could easily be brought on board by simply whispering the word “China.” Right populist elements are a wildcard here, but will likely be overwhelmed regardless if they end up in the minority.
China itself is trickier. While neither of the two scenarios outlined above are likely, it is particularly difficult to imagine China’s leaders rejecting the potential of the technology. The only scenario that comes to my mind is one in which they do not end up having a choice. One way this may happen is if, like the European hypothetical above, the potential of CRISPR fails to materialize. In this case, China may end up being unable to bear the costs of promoting the use of the more expensive IVF-based embryo selection methods while also trying to manage the costs of reversing its immanent demographic collapse.
I would still bet on the facile prediction if forced, but I will be eager to watch how this space develops to see whether other futures may be possible.
Quick Links
1) A major thing I have changed my belief on over the course of the last 5 years: gain-of-function research (scientists conducting lab experiments to make pathogens more deadly or virulent in order to study those variants’ behavior) is absolutely insane and should be subject to nuclear arms treaty-level efforts to ban globally.
2) Alex Jones is a beautiful woman: the conclusion of Sam Kriss in his astonishingly well-written essay musing on the inevitable fate of the modern American entertainer: to be consumed by the characters they summon.
Alex Jones is a beautiful woman. He is magnetic. He fascinates. When he speaks, everything else disappears: all that matters is contained in the cadences of his beautiful voice, the low grunt rising, heaving, swelling up to that gorgeous thick naked yell. He bares his flesh for the cameras and he is not ashamed to be seen. When he goes outside people beg his name. They reach out their hands to touch him, just to brush for a moment against his rich red leathery skin. Every camera swivels around to face him, like a flower in love with the Sun. Sometimes people will yell obscenities at him on the street, which is also a kind of adoration: they need to see this beautiful woman hurt and humiliated. His beauty is unbearable. This punctum, this rare glory, this bird of paradise… Alex Jones is an American and a patriot. But birds of paradise were killed for their feathers, and America is where the beautiful women die.
3) When concrete cracks, moisture can seep inside and rust away the steel reinforcements, causing catastrophes like last year’s the Miami condo collapse. Scientists have been working on a way to create self-reinforcing concrete that could heal after being damaged. This is achieved by embedding bacteria inside the concrete that are activated by water and excrete calcite to re-bond the material. Tests have apparently demonstrated that up to 90% of concrete’s strength can be restored in this way. Materials science is the best kind of science. h/t The Prepared
4) Anton Chekhov’s gentle criticism of Tolstoy: “Reason and justice tell me there’s more love for humanity in electricity and steam than in chastity or vegetarianism.” I love Tolstoy too much to take Chekhov’s side here, but I deeply appreciate the force of his insight.
5)
6) Noah Smith continues to argue that Xi Jinping is just not that competent. His most interesting argument is that Xi has been very effective at taking power, but that the steps Xi is taking to consolidate that power are undermining the systems that have led to China’s recent prosperity. I think this is a very important question, with significant implications for the rising authoritarian elements of the Deep Right. Overall, though, I think Smith’s analysis is limited by his commitment to liberal democratic values, and his resultant inability to take seriously the idea that China’s attempt to fight against the encroachment of those values could give it an important cultural edge in the coming century.
Ruet Caelum 