Genes or Environment? A meaningless distinction.
As Gage at al. (2016) argued, genes are promising causal anchors for psychiatric epidemiology, but clinically useful psychiatric epidemiology based on genotypes will be very hard.
Genetic variants in a stretch of genome that contains the CHRNA5, CHRNA3, and CHRNB4 genes, relate to smoking intensity in smokers. They do so for clear biological reasons, these genes are subunits (parts of) the nicotinic acetylcholine receptor, which in response to acetylcholine, but importantly also in response to nicotine, opens a ion channel in (among other cells) neurons. It’s as clear of a biological story as you are gong to get out of GWAS of a behavioural trait: “I smoke, neuron firing changes, brain tingle, tingle nice, smoke more”. That was my subjective experience as a smoker for years anyway (some of you may have seen me relapse after a number of drinks, which is why I now drink as infrequently as possible).
For obvious reasons these genetic variants also increase the probability of developing lung cancer. However, some people would, perhaps reasonably, still consider these genetic effects part an environmental cause of lung cancer: smoking. Social policy, stigma, social medicine, parenting, eduction and taxes all influence smoking rates and in doing so modify the effect the genetic variants have on lung cancer. In 1450’s there was no tobacco in Europe (its native to South America) and so while people would have the risk variant related to smoking intensity, they couldn’t smoke, and the variants would not have caused lung cancer. Once we completely eliminate smoking, we‘ll return to a world where these variants do not cause lung cancers, and we’ll do so with social policy, educational policy and laws not CRISPR/Cas9.
It is highly unlikely that every genetic effect has a socially or psychologically mediated causal path, at some point the effect of a genetic variant must act trough biology. But, in this case you can reasonable ask: Is this one GWAS finding pulling back the curtain on lung cancer biology? If you decide it is, is it also the kind of biological knoledge that leads to a cure that can be applied after lung cancer is diagnosed? Those are valid questions which highlight genetic variants identified in GWAS aren’t sufficient evidence there is actionable disease biology to be learned. At times researchers or lay people confuse GWAS results for evidence for biology, or confuse heritability as evidence against the existence of modifiable environmental causes.
Genetic and environmental effects have a technical definition in twin and genetic studies that doesn’t align very well with their colloquial meaning. In statistical genetics smoking is a path for both genetic and environmental influences on lung cancer, but colloquially smoking is reasonably viewed as an environmental cause of lung cancer. The conflation between the technical and colloquial terms obscures one of the most valuable uses of genetics for psychiatry: Genetics can serve as a causal anchor to study environmental causes. The use of genetics to study modifiable environmental causes is usually called Mendelian Randomisation (MR). A very clear paper of its logic and value, specifically to psychiatry, was written by Gage et al. (2016)1, it should be obligatory reading for all that read/interact with or work on psychiatric genetics of behaviour genetics:
Gage, S. H., Davey Smith, G., Ware, J. J., Flint, J., & Munafo, M. R. (2016). G= E: What GWAS can tell us about the environment. PLoS genetics, 12(2), e1005765.
Perhaps the nicotinic receptor gene signal can’t teach us about lung cancer biology, but Gage et al. argue it can teach us whether smoking is a contributory cause of other outcomes, like for example schizophrenia. Among 108 genome wide significant effect on schizophrenia identified in a 2014 GWAS where smoking related variants near the CHRNA5, CHRNA3, and CHRNB4 genes2. The authors point out that unless this colocalisation of the genetic effect for these two traits is coincidence (not as unlikely as you might think) its at least suggestive evidence of a causal effect of smoking on schizophrenia.
This insight is profound, as the causal study of the long term rare (schizophrenia is sort of rare) consequences of habits like smoking is hard and knows ethical bounds. We can’t ethically make people smoke, and while more we can try and make them smoke less, that turns out to be very hard. Fortunately, the fact that people vary in their risk of smoking based on genomes they didn’t select (i.e. condtionally random) that are unknown to themselves (i.e. blind) offers a sort of imperfect natural experiment.
The natural experiment has well known limits. I mentioned the risk of coincidental collocalisation of genetic effects between exposure (smoking) and outcome (schizophrenia), but there are other potential confounds: if you have the smoking risk allele, your parents are also every so slightly more likely to smoke (you got that allele from them). There are solutions to many of these biases and confounds, you could consider siblings who differ in their risk allele for example, they share parents keeping that influence constant. Despite the assumptions we have to make, and the potential confounds, genetics, by way of mendelian randomisation, offers a rare natural experiment to study the causal effects of smoking and other environmental exposures.
Making most of mendelian randomisation for psychiatry means we must take measurement seriously.
In this newsletter I only write about science I love, and especially the science you love needs strong criticism to get better. Mendelian randomisation has had its fair share of methodological criticism, I want to add specific measurement related criticism that is relevant to psychiatry. Gage et al. as discussed previously label the fact that variants related to smoking relate to schizophrenia, as suggestive evidence of a causal effect of smoking on schizophrenia. I’d relabel it as evidence that smoking relates to being in the case group relative to the control group in the schizophrenia GWAS. This sounds like nitpicking but its actually quite important.
There are many causal paths that lead into selection into a GWAS as a case. You have to meet diagnostic and other inclusion criteria, often these amount to more then a single contact with a psychiatrist, diagnosis or referral. Sometimes your diagnostically interviewed after being flagged for inclusion to further screen you. If you are ill longer, or more frequently, you are more likely to be in treatment, and therefore often more likely to be available for inclusion in a study. If your medications aren’t working or not working well you are also more frequently in a clicic or care setting and more likely to be included in a study.
These kinds of inclussion processes open up all kinds of causal paths. For example smoking appears to influence the metabolism of the anti-psychotic clozepine (an association that was also studied in a recent mendelian randomisation study3) which could cause smokers to respond worse, and be over represented in the pool of schizophrenia patients that can be included in GWAS. The difference between “smoking causes schizophrenia”, "smoking causes GWAS case inclusion" and “smoking influences anti-pychotic drug metabolism” seems trivial, but are relevant as the paths might require slightly different interventions. For example, if smoking cause schizophrenia case status, then cessation is a solid candidate for interventions, but if medication metabolism is part of the causal mechanism then cessation needs to be coupled with dose monitoring and potentially adjustment.
If we are to rely on mendelian randomisation in psychiatric epidemiology, and I think we should because we frequently have no viable alternative, we must study the causal path systematically and with great care. Let’s consider some examples of paths forward. There are studies like the Danish IPsych, that sample cases from population registers, and so may avoid conditioning on severity or length of disease. Whinin IPsych you could condition on disease severity and length of treatment to simulate the situation you expect to see in other less selective GWAS. Under various causal models the MR may or may not be expected to show smoking causes case status in Denmark. Alternatively if you suspect smoking interferes with drug metabolism than MR might show that variants that cause smoking also causes more frequent anti-psychotic dose or medication changes in smokers, but not in non-smokers.
These and other mendelian randomisation analyses can leverage the genome to systematically discover environmental causal relations in psychiatric epidemiology. To make this work well and for this science to be clinically relevant, we must more systematically formulate hypothesis, engage with clinicians, epidemiologists and pharmacists to define optimal phenotypes, get a better grip on study inclusion mechanisms, and integrate genetic data better with register or medical record outcomes. mendelan randomisation cant be an afterthought in a GWAS paper, or performed in isolation based solely on grossly measured exposures and outcomes. I’ll happily admit that I should do better in this respect in my own work going forward. We must also, as a field, in our curricula, training and outreach more clearly highlight that genetic effects pass trough environments, and that there is frequently no clear distinction, as Gage et al. cleverly noted: “G(enetics) = E(nvironment)”.
Gage, S. H., Davey Smith, G., Ware, J. J., Flint, J., & Munafo, M. R. (2016). G= E: What GWAS can tell us about the environment. PLoS genetics, 12(2), e1005765.
Schizophrenia Working Group of the Psychiatric Genomics Consortium (2014). Biological insights from 108 schizophrenia-associated genetic loci. Nature 511, 421–427 https://doi.org/10.1038/nature13595
Zeng, L., Lv, H., Li, J., Xue, R., Liu, X., Zhou, C., & Yu, H. (2022). Cigarette smoking, coffee consumption, alcohol intake, and clozapine metabolism: A Mendelian randomization study. Frontiers in Psychiatry, 2261.
I largely agree, but I guess when you refer to "psychiatric genetics," you're excluding NDDs entirely? For ASD, I struggle to see risk variants acting any other way other than biology.