Another fantastic post Sasha. I am always interested in what Turkheimer has to say, and having you present his work makes it twice as good.
Does he talk about FGWAS at all? In my opinion, it is a major methodological breakthrough that can control for a lot of confounding that plagues standard GWAS for behavioral traits.
His concept of "essence heritability" is an interesting one, and as a mechanism-oriented molecular systems biologist, I am sympathetic to it. However, to me "predictive heritability" is what really counts. If a variant truly influences a trait, then its presence or absence can be used to predict phenotype in a population. Pragmatically speaking, PGIs can complement existing clinical scoring systems.
Overall I am a lot less pessimistic than Turkheimer, and one reason for optimism are the "young guns" in the molecular human genetics field like you and Alex. I was always a bit dubious about some of the older human geneticists, and I won't name any names but you know who I am referring to, and so it is good to see the field moving in a more positive direction. There are a lot of exciting discoveries to be made.
Reading research and accounts on Mathematical Circles in the USSR convinced me of near-zero essence heritability for mathematical ability while at the same time being, in my opinion, a strong cause for optimism. At least for traits within the "education/intellectual skill-building" domain, this form of socialization seems to be a key ingredient. Of course, like most things it's likely significantly confounded and more complicated than "socialize kids properly, and they'll turn into Kolmogorov" (and along with being hard to quantify/test, it's probably not amenable to causally-informative research designs) but it can at least be somewhere to start, and may make us slightly less gloomy.
I'm not sure one can get away with arguing that there aren't known genes that directly influence cognitive traits (and underneath this, neurological traits - not that this argument was made exactly but it is perhaps being elided). We have a slew of known monogenic neurodevelopmental conditions. The cellular and animal modeling work that has followed up on clinical genetic findings has consistently implicated shared biological pathways and shared molecular neuropathology. Doesn't this quite plainly make the prospect of this work anything but gloomy?
There is often a tendency to dismiss monogenic findings as rare events that are not representative of the normal trait architecture, but this seems out-of-step what we often see we complex traits. The rare and common effects have the a shared directionality. Doesn't one simply have to show the replicated GWAS hits affect the pathways implicated by monogenic work? Hasn't this been done? Again, this isn't everything that needs to be done, but it would seem to foreclose entertaining claims like near zero "essence heritability."
As an aside, it is notable that Turkheimer's premise here diverges substantially from premises used to build arguments to reject hereditarian-like claims in anthropology (i.e. man is not a special animal). Here, man is a special animal. I'm not sure on my thoughts on this, but I do agree that the social sophistication of man does create a web of interactions that are not easily resolved by the approaches available to us.
Thanks for bringing this up as I did think a bit about how to accurately describe the essence of heritability of IQ. I settled on "near zero" for the following reasons: Chen et al. 2023 Nat Genet (https://www.nature.com/articles/s41588-023-01398-8) looked at the rare variant burden effect on an IQ test ("VNR") in the UK Biobank and found a variance explained (R^2) of 0.0015; Kingdom et al. 2024 Nat Genet (https://www.nature.com/articles/s41588-024-01710-0#Sec16) conducted a similar analysis using known developmental delay genes and although they do not report the R2, you can back it out of the Z-scores in Table S5 and it is ~0.0002 for IQ ("Fluid Intelligence"). In terms of "essence heritability", these are probably still upper bounds as we do not fully understand how all of the genes involved actually influence cognitive function in the general population (for example Kingdom et al. show that common variants can modify the penetrance of these rare burdens). Thus given two studies showing an R2 of <<1% for an expansive set of genes, I thought "near zero" was an accurate description.
To your question about pathway analyses, in general these have actually been quite lackluster and appear to be confounded by background effects. For example, Kim et al. AJHG (https://pmc.ncbi.nlm.nih.gov/articles/PMC6506868) looked at ~760k pathway-trait pairs and identified just 156 that were significant at a liberal 5% FDR, with most findings from previous studies not replicating because of confounding. For Educational Attainment, these pathways were very general and likely capturing broad brain-related enrichment ("GABA synthesis, release, reuptake and degradation", "Cam-PDE 1 activation", "positive regulation of cell-matrix adhesion", "calcium ion-dependent exocytosis", "increased aggression towards mice", "absent corpus callosum"), with most explaining <1% of the trait heritability. So here again I would say that we are still very far from understanding.
I agree that our understanding is poor at the moment, but I think we have enough information to inform our priors about genetic effects on cognitive traits, specifically that it is non-zero in a meaningful way.
There's something irreconcilable about the incidence of NDDs (of even moderate severity) exceeding the estimated trait heritability for cognitive phenotypes. Even if we halve the incidence, attributing it to de novo variants (the typical number usually estimated to be 40% of NDD patients), this is still the case. I just think it makes a lot more sense to say that current heritability estimates are low for technical reasons (on both the G and P sides).
For instance, in Chen et al. 2023, the EDU R^2 estimate (not VNR, which is said to be similar but appears modestly larger based on the data in the figure) includes just rare PTVs (the study itself is limited to rare SNV coding variation). Additionally, in my reading, from the perspective of proof-of-concept, it seems supportive of my overall claim that deleterious variant burden will come at a cognitive penalty and fleshing out the mechanisms of this isn't beyond contemporary methods. There is an illustrative passage in the discussion that's convenient to my point about the shared direction of rare and common variation (e.g KDM5B PTVs are a 1.51y penalty while the lead SNP is a 1.4 week penalty).
Part of your response raises the issue of penetrance, which I think is important to this question and is under-explored. I would like to read more and think more about that. We've seen it demonstrated in many different traits where the polygenic background modifies rare variant penetrance (the example top of mind is BRCA variants). The Kingdom 2024 is unfortunately small (based on <5 pLOFs across 599 genes from 419k subjects). Interesting as to what this suggests about ascertainment bias given what we expect generally about the incidence of NDDs in population.
I would agree that population findings for pathway stuff is mostly weak/non-specific from what I've seen, but the clinical genetics work continues to fit known NDD genes into the same pathways and if we agree that the rare and common effects typically converge there are population variants that converge on these pathways to the same effect.
Yeah I largely agree, I think there are potential paths forward but we are not there yet and my claims about near zero essence heritability were just about where we are now. I also have a draft post titled "We need to talk about penetrance" :)
In brief I tried to point to the definitive evidence we have to show that cognitive traits are subject to direct genetic input. It's not a profound point. Everyone would acknowledge that there are rare variants with dramatic effects on the development and function of the brain.
Follow-up research on different examples of the lesions that disrupt cognitive traits has consistently implicated similar biological pathways, PTEN-AKT-MTOR for example. This suggests that any common variant that modifies these convergent pathways will have effects on cognitive traits too.
The critique of course is that the effect on cognition is all in one direction, this isn't meaningful to population, and the pathway finding is non-specific or not fleshed out enough to be considered a mechanisms. There are many rejoinders to these, but I think the most salient is that it's likely that the human norm of cognition is likely close to optimized by evolution (head circumference is obviously constrained and besides having more tissue there's probably a hard limit to how much cognitive horsepower can be gamed by more efficient metabolism, nerve conduction, and/or cellular organization). So it seems likely that population variation in general is mostly just slightly damaging variant moving neurodevelopment away from the platonic norm.
"GWAS has turned the tables on the heritability of intelligence, from the 80 percent presumed by Jensen to something closer to 20 percent now; within-family analyses have reduced the heritability of intelligence even further."
This is very different from the impression I've gotten from reading people like Steve Stewart-Williams or Cremieux. Is this just a semantic distinction, where intelligence is still "80% genetic" if we include correlated-but-non-causal parental traits as "genetic", or is there a deeper divide in the field on this topic? I've turned to this field for direction on important life decisions like the relative importance of mate selection vs child rearing styles, so some clarification on this inconsistency would be very much appreciated.
There are basically three schools of thought on what's going on with the heritability of intelligence: (a) molecular geneticists like myself who treat twin study estimates with skepticism due to their assumptions about environment and interactions and put much more stock in molecular estimates that are generally 2-4x smaller than those from twins, with extensive evidence of environmental confounding; (b) the more classical behavioral geneticists like Turkheimer who think twin studies provide reasonably accurate estimates, but the interpretation of this heritability is very challenging because of environmental interactions and lack of mechanism and should not be treated as genetic essentialism (basically what Turkheimer's book is about); (c) the hard hereditarians / essentialists like Cremieux and other edgy online anons, who argue that intelligence as essentially all genes, basic processing, immutable, and highly value-laden. I'm a strong proponent of (a) (see my discussion of missing heritability here: https://theinfinitesimal.substack.com/p/comments-on-no-intelligence-is-not) and (b) (see my discussion of how heritability does not map to policy here: https://theinfinitesimal.substack.com/p/no-heritability-will-not-tell-you). But there are of course disagreements and other views out there. If you are using the findings from behavioral genetics to inform your life choices then I would highly recommend Turkheimer's book, it's an easy read and will give you a lot of important context.
I follow and read skeptics like Sasha and Turkheimer and proponents like Williams and Cremieux and I think I would say the field probably doesn't (at the moment) have much to give you in terms of direct guidance on personal life decisions.
Apart from "choose a healthy, happy, competent partner..." and "don't abuse and neglect your kids..." I don't think there's much else of value right now. Or atleast the disagreement is still strong.
I pretty much agree though I would add that other studies (which do not involve estimating variance components) have also shown fairly convincingly that: (1) adoption into a home with educated parents leads to higher IQ; (2) more education leads to higher IQ; (3) better early reading leads to higher IQ and transfer to non-verbal skills (demonstrated in genetically informed study of MZ twins). These are sort of obvious points but they bear repeating given the online discourse tends to treat IQ like skill points in a video game that are fixed at birth.
I definitely appreciate your reiteration of these "obvious" points, since I've seen other behavioral genetics writers claim that even actions like these were non-causal, and that it was a matter of "person X having high-IQ genes is more likely to do smart-seeming-thing Y". And in this way they wave away any action of the parent or child that seemingly impacts IQ directly.
“I’m not one of those people who shouts at the authors of GWAS papers that they only discover false positives — they clearly replicate.”
Well, I don’t think it’s clear at all. For starters, for EA2 to EA4 (and beyond), there is no attempt to directly analyze the new data independently (ie do an independent GWAS). They simply add the new data to the old and do a combined GWAS, mislabeling it a meta-analysis. So we don’t know if the correlations in EA2 are replicated in EA4 if we are just looking at the new data as a separate GWAS (I am referring to the traditional 5 X 10 -8 standard, not the sign concordance reinvention of “replication.”). This kind of “meta-analysis” is being done for virtually all the behavioral genetics traits, so I am not aware of a single genetic variant for any trait that consistently independently replicates. Moreover, the added data, while ostensibly independent from the previous data, often is just the latest accumulation of the same database (usually the UK Biobank), which makes confounding far more likely due to the particular, consistent biases from those accumulating databases.
I don't think this is accurate, EA4 specifically looked at out-of-sample replication:
"We also examined the out-of-sample replicability of the lead SNPs identified in the most recent previous meta-analysis. In the independent 23andMe data, the replication record is broadly in line with theoretical predictions derived from an empirical Bayesian framework described in the Supplementary Note (Extended Data Fig. 3)."
Note the use of "independent 23andMe data" which is not only a different cohort but also predominantly from the US, whereas the discovery data was predominantly from UK/Europe.
Again, no independent GWAS. You can’t just invent a “replication.” You don’t need an out of sample replication when you have hundreds of thousands of new data points. You don’t just add them onto the old ones and redo the GWAS and call it a meta-analysis. You never did an independent GWAS of EA4. That would be what you then add to a meta-analysis (leaving aside issues of the validity of meta-analyses). This leaves it impossible to compare new data with old. How many significant hits (5 x 10-8) were there in just the new data? How many matched the old study.
I don't understand what you're saying, they ran a GWAS using UK/European data and then they replicated it in an independent GWAS from 23andme. The number of significant hits in just the old and new data are shown in Extended Figure 3.
Extended Figure 3 from Okbay et al (https://www.nature.com/articles/s41588-022-01016-z/figures/8) shows the effect sizes estimated in EA3 on the x-axis and effect sizes from "the subsample of our data that did not contribute to the EA3 GWAS" (mostly 23andme) on the y-axis.
Another fantastic post Sasha. I am always interested in what Turkheimer has to say, and having you present his work makes it twice as good.
Does he talk about FGWAS at all? In my opinion, it is a major methodological breakthrough that can control for a lot of confounding that plagues standard GWAS for behavioral traits.
His concept of "essence heritability" is an interesting one, and as a mechanism-oriented molecular systems biologist, I am sympathetic to it. However, to me "predictive heritability" is what really counts. If a variant truly influences a trait, then its presence or absence can be used to predict phenotype in a population. Pragmatically speaking, PGIs can complement existing clinical scoring systems.
Overall I am a lot less pessimistic than Turkheimer, and one reason for optimism are the "young guns" in the molecular human genetics field like you and Alex. I was always a bit dubious about some of the older human geneticists, and I won't name any names but you know who I am referring to, and so it is good to see the field moving in a more positive direction. There are a lot of exciting discoveries to be made.
Reading research and accounts on Mathematical Circles in the USSR convinced me of near-zero essence heritability for mathematical ability while at the same time being, in my opinion, a strong cause for optimism. At least for traits within the "education/intellectual skill-building" domain, this form of socialization seems to be a key ingredient. Of course, like most things it's likely significantly confounded and more complicated than "socialize kids properly, and they'll turn into Kolmogorov" (and along with being hard to quantify/test, it's probably not amenable to causally-informative research designs) but it can at least be somewhere to start, and may make us slightly less gloomy.
This is your greatest article so far Gusev. Makes me interested to see what the future holds for this debate.
I'm not sure one can get away with arguing that there aren't known genes that directly influence cognitive traits (and underneath this, neurological traits - not that this argument was made exactly but it is perhaps being elided). We have a slew of known monogenic neurodevelopmental conditions. The cellular and animal modeling work that has followed up on clinical genetic findings has consistently implicated shared biological pathways and shared molecular neuropathology. Doesn't this quite plainly make the prospect of this work anything but gloomy?
There is often a tendency to dismiss monogenic findings as rare events that are not representative of the normal trait architecture, but this seems out-of-step what we often see we complex traits. The rare and common effects have the a shared directionality. Doesn't one simply have to show the replicated GWAS hits affect the pathways implicated by monogenic work? Hasn't this been done? Again, this isn't everything that needs to be done, but it would seem to foreclose entertaining claims like near zero "essence heritability."
As an aside, it is notable that Turkheimer's premise here diverges substantially from premises used to build arguments to reject hereditarian-like claims in anthropology (i.e. man is not a special animal). Here, man is a special animal. I'm not sure on my thoughts on this, but I do agree that the social sophistication of man does create a web of interactions that are not easily resolved by the approaches available to us.
Thanks for bringing this up as I did think a bit about how to accurately describe the essence of heritability of IQ. I settled on "near zero" for the following reasons: Chen et al. 2023 Nat Genet (https://www.nature.com/articles/s41588-023-01398-8) looked at the rare variant burden effect on an IQ test ("VNR") in the UK Biobank and found a variance explained (R^2) of 0.0015; Kingdom et al. 2024 Nat Genet (https://www.nature.com/articles/s41588-024-01710-0#Sec16) conducted a similar analysis using known developmental delay genes and although they do not report the R2, you can back it out of the Z-scores in Table S5 and it is ~0.0002 for IQ ("Fluid Intelligence"). In terms of "essence heritability", these are probably still upper bounds as we do not fully understand how all of the genes involved actually influence cognitive function in the general population (for example Kingdom et al. show that common variants can modify the penetrance of these rare burdens). Thus given two studies showing an R2 of <<1% for an expansive set of genes, I thought "near zero" was an accurate description.
To your question about pathway analyses, in general these have actually been quite lackluster and appear to be confounded by background effects. For example, Kim et al. AJHG (https://pmc.ncbi.nlm.nih.gov/articles/PMC6506868) looked at ~760k pathway-trait pairs and identified just 156 that were significant at a liberal 5% FDR, with most findings from previous studies not replicating because of confounding. For Educational Attainment, these pathways were very general and likely capturing broad brain-related enrichment ("GABA synthesis, release, reuptake and degradation", "Cam-PDE 1 activation", "positive regulation of cell-matrix adhesion", "calcium ion-dependent exocytosis", "increased aggression towards mice", "absent corpus callosum"), with most explaining <1% of the trait heritability. So here again I would say that we are still very far from understanding.
I agree that our understanding is poor at the moment, but I think we have enough information to inform our priors about genetic effects on cognitive traits, specifically that it is non-zero in a meaningful way.
There's something irreconcilable about the incidence of NDDs (of even moderate severity) exceeding the estimated trait heritability for cognitive phenotypes. Even if we halve the incidence, attributing it to de novo variants (the typical number usually estimated to be 40% of NDD patients), this is still the case. I just think it makes a lot more sense to say that current heritability estimates are low for technical reasons (on both the G and P sides).
For instance, in Chen et al. 2023, the EDU R^2 estimate (not VNR, which is said to be similar but appears modestly larger based on the data in the figure) includes just rare PTVs (the study itself is limited to rare SNV coding variation). Additionally, in my reading, from the perspective of proof-of-concept, it seems supportive of my overall claim that deleterious variant burden will come at a cognitive penalty and fleshing out the mechanisms of this isn't beyond contemporary methods. There is an illustrative passage in the discussion that's convenient to my point about the shared direction of rare and common variation (e.g KDM5B PTVs are a 1.51y penalty while the lead SNP is a 1.4 week penalty).
Part of your response raises the issue of penetrance, which I think is important to this question and is under-explored. I would like to read more and think more about that. We've seen it demonstrated in many different traits where the polygenic background modifies rare variant penetrance (the example top of mind is BRCA variants). The Kingdom 2024 is unfortunately small (based on <5 pLOFs across 599 genes from 419k subjects). Interesting as to what this suggests about ascertainment bias given what we expect generally about the incidence of NDDs in population.
I would agree that population findings for pathway stuff is mostly weak/non-specific from what I've seen, but the clinical genetics work continues to fit known NDD genes into the same pathways and if we agree that the rare and common effects typically converge there are population variants that converge on these pathways to the same effect.
Yeah I largely agree, I think there are potential paths forward but we are not there yet and my claims about near zero essence heritability were just about where we are now. I also have a draft post titled "We need to talk about penetrance" :)
> but it would seem to foreclose entertaining claims like near zero "essence heritability."
Could you expound here?
In brief I tried to point to the definitive evidence we have to show that cognitive traits are subject to direct genetic input. It's not a profound point. Everyone would acknowledge that there are rare variants with dramatic effects on the development and function of the brain.
Follow-up research on different examples of the lesions that disrupt cognitive traits has consistently implicated similar biological pathways, PTEN-AKT-MTOR for example. This suggests that any common variant that modifies these convergent pathways will have effects on cognitive traits too.
The critique of course is that the effect on cognition is all in one direction, this isn't meaningful to population, and the pathway finding is non-specific or not fleshed out enough to be considered a mechanisms. There are many rejoinders to these, but I think the most salient is that it's likely that the human norm of cognition is likely close to optimized by evolution (head circumference is obviously constrained and besides having more tissue there's probably a hard limit to how much cognitive horsepower can be gamed by more efficient metabolism, nerve conduction, and/or cellular organization). So it seems likely that population variation in general is mostly just slightly damaging variant moving neurodevelopment away from the platonic norm.
"GWAS has turned the tables on the heritability of intelligence, from the 80 percent presumed by Jensen to something closer to 20 percent now; within-family analyses have reduced the heritability of intelligence even further."
This is very different from the impression I've gotten from reading people like Steve Stewart-Williams or Cremieux. Is this just a semantic distinction, where intelligence is still "80% genetic" if we include correlated-but-non-causal parental traits as "genetic", or is there a deeper divide in the field on this topic? I've turned to this field for direction on important life decisions like the relative importance of mate selection vs child rearing styles, so some clarification on this inconsistency would be very much appreciated.
There are basically three schools of thought on what's going on with the heritability of intelligence: (a) molecular geneticists like myself who treat twin study estimates with skepticism due to their assumptions about environment and interactions and put much more stock in molecular estimates that are generally 2-4x smaller than those from twins, with extensive evidence of environmental confounding; (b) the more classical behavioral geneticists like Turkheimer who think twin studies provide reasonably accurate estimates, but the interpretation of this heritability is very challenging because of environmental interactions and lack of mechanism and should not be treated as genetic essentialism (basically what Turkheimer's book is about); (c) the hard hereditarians / essentialists like Cremieux and other edgy online anons, who argue that intelligence as essentially all genes, basic processing, immutable, and highly value-laden. I'm a strong proponent of (a) (see my discussion of missing heritability here: https://theinfinitesimal.substack.com/p/comments-on-no-intelligence-is-not) and (b) (see my discussion of how heritability does not map to policy here: https://theinfinitesimal.substack.com/p/no-heritability-will-not-tell-you). But there are of course disagreements and other views out there. If you are using the findings from behavioral genetics to inform your life choices then I would highly recommend Turkheimer's book, it's an easy read and will give you a lot of important context.
I follow and read skeptics like Sasha and Turkheimer and proponents like Williams and Cremieux and I think I would say the field probably doesn't (at the moment) have much to give you in terms of direct guidance on personal life decisions.
Apart from "choose a healthy, happy, competent partner..." and "don't abuse and neglect your kids..." I don't think there's much else of value right now. Or atleast the disagreement is still strong.
I pretty much agree though I would add that other studies (which do not involve estimating variance components) have also shown fairly convincingly that: (1) adoption into a home with educated parents leads to higher IQ; (2) more education leads to higher IQ; (3) better early reading leads to higher IQ and transfer to non-verbal skills (demonstrated in genetically informed study of MZ twins). These are sort of obvious points but they bear repeating given the online discourse tends to treat IQ like skill points in a video game that are fixed at birth.
I definitely appreciate your reiteration of these "obvious" points, since I've seen other behavioral genetics writers claim that even actions like these were non-causal, and that it was a matter of "person X having high-IQ genes is more likely to do smart-seeming-thing Y". And in this way they wave away any action of the parent or child that seemingly impacts IQ directly.
“I’m not one of those people who shouts at the authors of GWAS papers that they only discover false positives — they clearly replicate.”
Well, I don’t think it’s clear at all. For starters, for EA2 to EA4 (and beyond), there is no attempt to directly analyze the new data independently (ie do an independent GWAS). They simply add the new data to the old and do a combined GWAS, mislabeling it a meta-analysis. So we don’t know if the correlations in EA2 are replicated in EA4 if we are just looking at the new data as a separate GWAS (I am referring to the traditional 5 X 10 -8 standard, not the sign concordance reinvention of “replication.”). This kind of “meta-analysis” is being done for virtually all the behavioral genetics traits, so I am not aware of a single genetic variant for any trait that consistently independently replicates. Moreover, the added data, while ostensibly independent from the previous data, often is just the latest accumulation of the same database (usually the UK Biobank), which makes confounding far more likely due to the particular, consistent biases from those accumulating databases.
I don't think this is accurate, EA4 specifically looked at out-of-sample replication:
"We also examined the out-of-sample replicability of the lead SNPs identified in the most recent previous meta-analysis. In the independent 23andMe data, the replication record is broadly in line with theoretical predictions derived from an empirical Bayesian framework described in the Supplementary Note (Extended Data Fig. 3)."
Note the use of "independent 23andMe data" which is not only a different cohort but also predominantly from the US, whereas the discovery data was predominantly from UK/Europe.
Again, no independent GWAS. You can’t just invent a “replication.” You don’t need an out of sample replication when you have hundreds of thousands of new data points. You don’t just add them onto the old ones and redo the GWAS and call it a meta-analysis. You never did an independent GWAS of EA4. That would be what you then add to a meta-analysis (leaving aside issues of the validity of meta-analyses). This leaves it impossible to compare new data with old. How many significant hits (5 x 10-8) were there in just the new data? How many matched the old study.
I don't understand what you're saying, they ran a GWAS using UK/European data and then they replicated it in an independent GWAS from 23andme. The number of significant hits in just the old and new data are shown in Extended Figure 3.
What were the results of the GWAs just for the 23 and me data used for the replication? Do you have that available?
Just to reiterate, since EA1, to my knowledge, there isn’t an independent GWAS for EA that doesn’t include EA1 (and EA2, for EA3, etc.).
Extended Figure 3 from Okbay et al (https://www.nature.com/articles/s41588-022-01016-z/figures/8) shows the effect sizes estimated in EA3 on the x-axis and effect sizes from "the subsample of our data that did not contribute to the EA3 GWAS" (mostly 23andme) on the y-axis.