Using adopted individuals to partition indirect maternal genetic effects into prenatal and postnatal effects on offspring phenotypes

Article activity feed

1. 

   Version published to 10.7554/elife.73671 on eLife

   Jul 13, 2022
2. 

   eLife

   Jul 1, 2022

   Author Response

   Reviewer #2 (Public Review):

   Summary

   The research paper presents a modeling approach aimed at disentangling mother's genetic effects on their offspring in two components: prenatal environment and postnatal environment. Specifically, the authors use SEM on adopted and non-adopted individuals from the UK Biobank and leverage the variation in genetic similarities from different family structures. Because the UK Biobank is not created as an adoption study, they build seven different family structures to include all possible family combinations that can provide information regarding the two parameters of interest: those representing prenatal and postnatal environment respectively. The model is used on two phenotypes (birthweight and education attainment) to illustrate it.

   The results indicate an 'expected pattern of …

   Author Response

   Reviewer #2 (Public Review):

   Summary

   The research paper presents a modeling approach aimed at disentangling mother's genetic effects on their offspring in two components: prenatal environment and postnatal environment. Specifically, the authors use SEM on adopted and non-adopted individuals from the UK Biobank and leverage the variation in genetic similarities from different family structures. Because the UK Biobank is not created as an adoption study, they build seven different family structures to include all possible family combinations that can provide information regarding the two parameters of interest: those representing prenatal and postnatal environment respectively. The model is used on two phenotypes (birthweight and education attainment) to illustrate it.

   The results indicate an 'expected pattern of maternal genetic effect on offspring birthweight' and 'unexpectedly large prenatal (intrauterine) maternal genetic effects on offspring education attainment. The authors mention this result can likely be explained by adopted offspring being raised by biological relatives. They then show simulations supporting this hypothesis.

   We praise the authors for the complex analyses executed and the work done to create the model and make the scripts available to the research community. The models can be a valuable addition to the behavior genetics literature and to researcher's toolkit. We do however have a few concerns regarding 1. the meaning of the results, 2. model building decisions and the choice of sample and 3. the way some limitations are addressed. We go into more details for each of these points.

   1. Interest to study mothers' genetic effects as acting via the prenatal environment or the postnatal environment and the meaning of the parameters tested by the model .

   I think this is an interesting question and a useful distinction for a number of phenotypes and the authors use the adoption design in an innovative way to define and estimate parameters that correspond to this distinction. However, I would suggest that the expressions of prenatal environmental effect and postnatal environmental effect (as distinct pathways for mother's gene to be expressed) seem to be an overstatement.

   The definition of mother genetic effects (effects of mother genotype on their child phenotype, over and above any genetic transmission) is citing Wolf & Wade 2009 (line 56) which mention the more general notion of 'maternal effect' that are defined as effect of genotype, phenotype (or both) on their offspring. I would argue that postnatal maternal genetic effects (as currently defined in the paper) are likely environmental effect and not only 'genetic effects'. These environmental effects are indeed partly influenced by mother's genes, but also strongly affected by other variables such as culture, generation, SES, education. It is not possible to disentangle these effects in the design(s) used here.

   Although we have referred to the maternal effects estimated in our manuscript as “prenatal maternal genetic effects” and “postnatal maternal genetic effects”- all of these effects on the offspring are mediated through maternal phenotypes (which as the reviewer correctly notes, will be influenced by both genes and the environment). In other words, the maternal PRS used in our study proxies some maternal phenotype/s that then forms part of the offspring’s prenatal and/or postnatal environment which then affects the offspring’s phenotype. We have referred to these effects as maternal genetic effects rather than just maternal effects to emphasize the causal link with the maternal genotype and the fact that we are only proxying that part of the maternal phenotype that is explained by the relevant genetic variation (NB. This is consistent with the Wolf & Wade 2009 definition of maternal effects i.e. “…the causal influence of maternal genotypes on offspring phenotypes…”). We agree with the reviewer that our model is not attempting to disentangle proportions of variance due to genetic and environmental factors (which is not its purpose).

   This consideration can affect the authors definition of the covariance between an adopted individual's genotype and phenotype as a function of prenatal (but not postnatal) maternal genetic effects (line 93-94). The authors current assumption does not consider the potential for environmental modulation of the effect of adopted mothers' genes (which are not zero for several phenotypes). Postnatal maternal genetic effects are thus also likely to capture and represent environmental differences.

   Assuming that adopted offspring are not biologically related to their adoptive mothers, then adopted individuals’ PRS should not be correlated with adoptive mothers’ PRS. The corollary is that adoptive mothers’ PRS should not influence the covariance between adopted individuals’ PRS and phenotype (i.e. regardless of whether there is environmental modulation of the effect of adopted mothers’ genes on offspring phenotype). It is true, however, that we do not consider genotype by environment interaction effects in our model, and that this is a limitation of our model. We allude to this important point several times in the Discussion:

   “Those assumptions explicitly encoded in Figure 1 include that the total maternal genetic effect can be decomposed into the sum of prenatal and postnatal components, that genetic effects are homogenous across biological and adoptive families, the absence of genotype x environment interaction…”

   And

   “In contrast, in our design it is more important that genetic effect sizes are homogenous across adopted and non-adopted individuals (i.e. no genotype by environment interaction)…”.

   At the request of the reviewer, we now include additional discussion of GxE and other assumptions of our model in further detail in Supplementary File 17.

   1. Model building decisions specific to the UK biobank. One of the main issues is that the method is tested on a sample that is not built as an adoption design. This forced the authors to make decision to circumvent this problem and lead to important limitations that are not inherent to their method, but to the specific sample they applied it to.

   a) Having adoptive parents partly genetically related to the child is breaking the logic of the adopted design. Thus, it brings back the genetic confound (passive gene-environment correlation) problem of usual family-based design. In their case, it alters their ability to differentiate between prenatal and postnatal environment.

   We agree that the UK Biobank was never designed for this purpose, and that data from it regarding adoption is less than perfect. Nevertheless, we think that an important conclusion of our paper is that large-scale biobanks (which because of their size) contain many hundreds/thousands of adopted individuals can be used to partition maternal genetic effects into prenatal and postnatal components, provided good quality data on the adoption process has been gathered and/or genetic information on their adoptive parents.

   To help address the reviewer’s concerns we have created a Supplementary Table (Supplementary File 17) that summarizes some of the main limitations/assumptions of our model, whether they are specific to the UK Biobank dataset or intrinsic to our method, their consequences on model parameters, and possible options for addressing them.

   b) In section starting on line 426, the authors have included simulations to show how this issue could be addressed. However, it does not help the fact that in their model applied to the UK biobank, the information regarding the degree of genetic similarity between adopting parents and biological parents and the child is unknown.

   We agree- but we feel it is important to demonstrate (a) that cryptic biological relatedness between adopted individuals and their adoptive parents is a potential issue not only for our study, but for other studies attempting to utilize this information in the UK Biobank, and (b) that cryptic relatedness can be dealt with effectively through appropriate modelling in our SEM framework (i.e. even if it is not possible with the current data from UK Biobank). The corollary is that we recommend that the UK Biobank (and other large-scale biobanks) attempt to acquire information on adopted individuals and their parents through e.g. questionnaire.

   c) To address this problem in their analyses of UK biobank, authors used (Lines 302 & 417) information regarding whether children were breastfed or not (on the basis that this knowledge would be more common if the child was raised by a biological family relative) to identify adopted singletons raised by biological relatives. However, this is, at best, a mediocre index of genetic relatedness. I can see other reasons for participants to have knowledge of if they have been breastfed: because they were adopted at an older age, because they are still (or have been) in contact with their biological mother. It is also possible, albeit rare, that adoptive parents may breastfeed a child via the use of drugs to stimulate milk production. Line 420: the fact that the prenatal maternal estimate became non-significant after removing participants that were breastfed do provide results more in-line with what would be expected. But we can't use expected results as a basis to evaluate the validity of the approach. The absence of GxE and rGE are two other strong assumptions of the model that could also produce this kind unexpected results.

   We agree that (a) the inclusion of adopted individuals whose adoptive parents are biologically related to them is only one possible reason for unexpectedly strong prenatal maternal genetic effect estimates, (b) attempting to remove these individuals from the analysis using a proxy like breastfeeding information is less than perfect. As indicated above, we now discuss in detail alternative explanations for our results including violations of assumptions regarding the absence of GxE and rGE, and other explanations (assortative mating, stratification etc) (see new text in the Discussion and Supplementary File 17).

   d) I would suggest discussing the issue of genetic relatedness between adopting parents and offspring in terms of passive rGE which is a common problem for the estimation of parental effects in every familial design.

   We now include mention of passive rGE in the Discussion:

   “Rather we hypothesize it is possible that our model could have been misspecified in that substantial numbers of adopted individuals in the UK Biobank may have in fact been raised by their biological relatives. This can be thought of as (unintentional) reintroduction of passive gene-environment correlation into the study. In other words, adopted children are brought up by their genetic relatives, who in turn provide the environment in which they are raised. This induces a correlation between adopted individuals’ PRS and their environment.”

   e) Line 291: why use an unweighted PRS for EY3 (Lee, 2018), while the usual way of computing PRS (as a weighted sum of risk alleles) was used for birthweight?

   We thank the reviewer for pointing this inconsistency out. We have now rerun the analyses using weighted and unweighted PRS for both birth weight and educational attainment. The reason for running both sets of analyses is that the GWAS on which the SNPs are selected (i.e. the weights are based), contains UK Biobank individuals. This may inflate the overall strength of association between the PRS and outcome through winner’s curse (although not differentially between individuals from adoptive and biological families). In contrast, unweighted scores should be much more robust to this inflation, and so are a useful sanity check on the results.

   1. Limitations

   As our Discussion is already very long, we have created a Supplementary Table (Supplementary File 17) that summarizes some of the main limitations/assumptions of our model, their consequences on model parameters, and possible options for addressing them. We also discuss specific concerns raised by the referee below.

   Assess other limitations of their method.

   a) limitation of the availability of birth father information,

   Our model does not require information on adopted individual’s birth fathers (although it does require PRS on non-adopted individuals’ birth fathers- which is typically readily available). It does, however, make the assumption that fathers do not contribute prenatally to offspring traits- which we think is a reasonable assumption for the majority of offspring phenotypes. If PRS for adopted individuals’ biological fathers were available, then prenatal paternal genetic effects could be estimated as part of the model. To accommodate the reviewer’s request, we have included and discussed this limitation/assumption in more detail in Supplementary File 17.

   b) prenatal events uncorrelated with birthmother's genes (disease or accidents),

   We agree that our model assumes that maternal genotype is uncorrelated with prenatal environmental factors. We now discuss this assumption/limitation further in Supplementary File 17.

   c) Inferring prenatal environment effect from higher birth mother correlation compared to birthfather is subject to bias from measurement differences between the two (Loehlin, 2016).

   Whilst this is a limitation of adoption designs that estimate prenatal effects using the difference between maternal and paternal correlations with offspring phenotypes, this is not actually a limitation of our model. In our model we do not use (phenotypic) mother-child and father-child correlations (we use PRS-phenotype correlations). Also, in our model, information on the size of the prenatal (and postnatal) maternal genetic effects primarily comes from the difference between the PRS-phenotype covariance in adopted singletons compared to the PRS-phenotype covariance non-adopted individuals (i.e. not from the difference between maternal and paternal correlations with offspring phenotypes). We state this in the Introduction and Methods e.g.:

   “Thus, the difference between the genotype-phenotype covariance in adopted and non-adopted singleton individuals provides important information on the likely size of postnatal genetic effects.”

   It is also worth noting, that in our model, the size of the paternal PRS-offspring association does not factor into the estimation of maternal genetic effects (nor does the difference between the maternal PRS-offspring phenotype association and the paternal PRS-offspring phenotype association). Also, our model takes into account if there are differences in the amount of (random) measurement error in adoptive and non-adoptive families.

   d) age at which the child is adopted (if the child has been partly raised by birth parents before adoption, it would bias (raise) the estimates of prenatal effects).

   We agree and now discuss this limitation further in Supplementary File 17.

   e) evocative rGE not mentioned. It has been shown that parents partly react to children's behaviors. Thus, the estimate of maternal genetic postnatal effects could be biased (lowered) by evocative gene-environment correlation. In other words, the model also assumes no evocative gene-environment correlation.

   We agree and now discuss this limitation in Supplementary File 17 (although we note that the effect that evocative rGE will have on the SEM parameters will depend on the direction of the gene-environment correlation).

   Final thoughts

   1. I would like a better case made for why it is important to distinguish genetic effects into prenatal and postnatal effect.

   We have included the following text in the Introduction:

   “Given the increasing number of variants identified in GWAS that exhibit robust maternal genetic effects, a natural question to ask is whether these loci exert their effects on offspring phenotypes through intrauterine mechanisms, the postnatal environment, or both. Indeed, resolving maternal effects into prenatal and postnatal sources of variation could be a valuable first step in eventually elucidating the underlying mechanisms behind these associations (Armstrong-Carter et al. 2020), directing investigators to where they should focus their attention, and in the case of disease-related phenotypes, yielding potentially important information regarding the optimal timing of interventions. For example, the demonstration of maternal prenatal effects on offspring IQ/educational attainment, suggests that if the mediating factors that were responsible could be identified, then improvements in the prenatal care of mothers and their unborn babies which target these factors, could yield useful increases in offspring IQ/educational attainment.”

   1. I would suggest the author make a clear distinction between the limits inherent to their sample (UK biobank) from those inherent to their methodological approach. I see important usefulness is plague by limits inherent to the sample used. At the same time, I am not aware of the availability of a big enough sample of adopted children with genotypic information available to compute PRS.

   One of the main limitations inherent to our sample (UK Biobank) is the fact that currently we cannot be certain that adopted individuals are not biologically related to their adoptive parents. As we demonstrate, this limitation could be addressed if information were gathered regarding the relationships, which at least in principle could be done relatively easily in the UK Biobank (e.g. by questionnaire, or even better, by genotyping adoptive parents where possible). The SEMs could then be adjusted to take these relationships into account. We discuss this limitation, and many others, in Supplementary File 17, and divide the table according to whether the limitation is primarily a consequence of the dataset (UK Biobank) or the method more broadly.

   We agree with the reviewer that the size of adoption studies is currently limited (e.g. Texas Adoption Project; Colorado Adoption Study etc). Nevertheless, it is likely that the number of adopted individuals available in large-scale Biobanks will increase over time, in which case models like the one espoused in this manuscript will become increasingly useful. Importantly, our method does not require adoptive families in order to partition maternal effects, merely adopted singleton individuals, and reliable information on the biological relatedness (or lack thereof) of their adoptive parents. We feel therefore that it is important that this sort of information be gathered so that the adopted individuals within these large-scale resources can be leveraged to examine interesting questions like the ones discussed in our manuscript.

   We have added these points to the Discussion:

   “We argue that of greater consequence for the validity of our model is that any genetic relationship between adoptive and biological parents is accurately modelled and included in the SEM. Through simulation, we have shown that the consequences of model misspecification depend upon which biological and adoptive parents are related, the nature of this relationship, and the proportion of adopted individuals in the sample who have had their relationship misspecified. Our simulations also showed that correctly modelling this relationship returns asymptotically unbiased effect estimates and correct type I error rates. Clearly, knowing these cryptic relationships in the UK Biobank would allow us to properly model them and better estimate prenatal and postnatal maternal genetic effects using this resource. We emphasize that accurately modelling these relationships does not require that actual genotypes for adoptive and/or biological parents be obtained (although this would be advantageous in terms of statistical power) as our SEM allows us to model these relationships in terms of latent variables. Indeed, as large-scale resources like the UK Biobank become more common, we expect that the number of adopted individuals who have GWAS will also increase, and consequently models like the one espoused in this manuscript will become increasingly useful. High quality phenotypic information on these adopted individuals and their adoptive parents including whether they share any biological relationship will be critical to making the most of these resources.”

   Read the original source
3. 

   eLife

   Jul 1, 2022

   Evaluation Summary:

   This paper will be of interest to scientists interested in intergenerational transmission of phenotypes through genetic pathways. The authors propose an innovative and sound method to leverage the adoption of a design for disentangling prenatal and postnatal genetic effects. Additional analyses are needed to address the limitations of the model applied to the specific dataset that was used to illustrate the method.

   (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #2 agreed to share their name with the authors.)

   Read the original source
4. 

   eLife

   Jul 1, 2022

   Reviewer #1 (Public Review):

   This article presents important new findings, of particular interest to those concerned with a) estimating parental indirect genetic effects, b) distinguishing pre- and post-natal maternal effects, c) optimizing adoption designs, and d) developing cohort studies strategically. Notably, prior work has investigated pre-natal 'genetic nurture' (Armstrong-Carter et al., 2020) and used the adoption sub-sample of the UK Biobank to distinguish pre-natal and post-natal indirect genetic effects (Demange et al., 2021). Here, the authors present the first structural equation model for estimating pre- and post-natal parental indirect genetic effects using polygenic scores and adoption data. The authors found, as expected, pre- but not post-natal maternal genetic effects on birthweight. However, pre-natal maternal …

   Reviewer #1 (Public Review):

   This article presents important new findings, of particular interest to those concerned with a) estimating parental indirect genetic effects, b) distinguishing pre- and post-natal maternal effects, c) optimizing adoption designs, and d) developing cohort studies strategically. Notably, prior work has investigated pre-natal 'genetic nurture' (Armstrong-Carter et al., 2020) and used the adoption sub-sample of the UK Biobank to distinguish pre-natal and post-natal indirect genetic effects (Demange et al., 2021). Here, the authors present the first structural equation model for estimating pre- and post-natal parental indirect genetic effects using polygenic scores and adoption data. The authors found, as expected, pre- but not post-natal maternal genetic effects on birthweight. However, pre-natal maternal genetic effects on educational attainment were unfeasibly large. Their simulations convincingly suggest this is because estimates of maternal pre-natal indirect genetic effects are inflated when adoptive parents and biological parents are related. It is nice to see the authors' practical suggestions on how the UK Biobank resource should obtain more information on the adopted individuals. The key caveats are the low sample size of adoptees (especially when restricting to adoptees who have breastfeeding data), and the inclusion of only genome-wide significant SNPs in polygenic scores. Given the evidence that population stratification and assortative mating can bias estimates of parental indirect genetic effects, the authors should consider how these factors would affect their model.

   Read the original source
5. 

   eLife

   Jul 1, 2022

   Reviewer #2 (Public Review):

   The research paper presents a modeling approach aimed at disentangling mother's genetic effects on their offspring in two components: prenatal environment and postnatal environment. Specifically, the authors use SEM on adopted and non-adopted individuals from the UK Biobank and leverage the variation in genetic similarities from different family structures. Because the UK Biobank is not created as an adoption study, they build seven different family structures to include all possible family combinations that can provide information regarding the two parameters of interest: those representing prenatal and postnatal environment respectively. The model is used on two phenotypes (birthweight and education attainment) to illustrate it.

   The results indicate an 'expected pattern of maternal genetic effect on …

   Reviewer #2 (Public Review):

   The research paper presents a modeling approach aimed at disentangling mother's genetic effects on their offspring in two components: prenatal environment and postnatal environment. Specifically, the authors use SEM on adopted and non-adopted individuals from the UK Biobank and leverage the variation in genetic similarities from different family structures. Because the UK Biobank is not created as an adoption study, they build seven different family structures to include all possible family combinations that can provide information regarding the two parameters of interest: those representing prenatal and postnatal environment respectively. The model is used on two phenotypes (birthweight and education attainment) to illustrate it.

   The results indicate an 'expected pattern of maternal genetic effect on offspring birthweight' and 'unexpectedly large prenatal (intrauterine) maternal genetic effects on offspring education attainment. The authors mention this result can likely be explained by adopted offspring being raised by biological relatives. They then show simulations supporting this hypothesis.

   We praise the authors for the complex analyses executed and the work done to create the model and make the scripts available to the research community. The models can be a valuable addition to the behavior genetics literature and to researcher's toolkit. We do however have a few concerns regarding 1. the meaning of the results, 2. model building decisions and the choice of sample and 3. the way some limitations are addressed. We go into more details for each of these points.

   1. Interest to study mothers' genetic effects as acting via the prenatal environment or the postnatal environment and the meaning of the parameters tested by the model

   I think this is an interesting question and a useful distinction for a number of phenotypes and the authors use the adoption design in an innovative way to define and estimate parameters that correspond to this distinction. However, I would suggest that the expressions of prenatal environmental effect and postnatal environmental effect (as distinct pathways for mother's gene to be expressed) seem to be an overstatement.

   The definition of mother genetic effects (effects of mother genotype on their child phenotype, over and above any genetic transmission) is citing Wolf & Wade 2009 (line 56) which mention the more general notion of 'maternal effect' that are defined as effect of genotype, phenotype (or both) on their offspring. I would argue that postnatal maternal genetic effects (as currently defined in the paper) are likely environmental effect and not only 'genetic effects'.

   These environmental effects are indeed partly influenced by mother's genes, but also strongly affected by other variables such as culture, generation, SES, education. It is not possible to disentangle these effects in the design(s) used here.

   This consideration can affect the authors definition of the covariance between an adopted individual's genotype and phenotype as a function of prenatal (but not postnatal) maternal genetic effects (line 93-94). The authors current assumption does not consider the potential for environmental modulation of the effect of adopted mothers' genes (which are not zero for several phenotypes). Postnatal maternal genetic effects are thus also likely to capture and represent environmental differences.

   2. Model building decisions specific to the UK biobank

   One of the main issues is that the method is tested on a sample that is not built as an adoption design. This forced the authors to make decision to circumvent this problem and lead to important limitations that are not inherent to their method, but to the specific sample they applied it to.

   a. Having adoptive parents partly genetically related to the child is breaking the logic of the adopted design. Thus, it brings back the genetic confound (passive gene-environment correlation) problem of usual family-based design. In their case, it alters their ability to differentiate between prenatal and postnatal environment.

   b. In section starting on line 426, the authors have included simulations to show how this issue could be addressed. However, it does not help the fact that in their model applied to the UK biobank, the information regarding the degree of genetic similarity between adopting parents and biological parents and the child is unknown.

   c. To address this problem in their analyses of UK biobank, authors used (Lines 302 & 417) information regarding whether children were breastfed or not (on the basis that this knowledge would be more common if the child was raised by a biological family relative) to identify adopted singletons raised by biological relatives. However, this is, at best, a mediocre index of genetic relatedness. I can see other reasons for participants to have knowledge of if they have been breastfed: because they were adopted at an older age, because they are still (or have been) in contact with their biological mother. It is also possible, albeit rare, that adoptive parents may breastfeed a child via the use of drugs to stimulate milk production. Line 420: the fact that the prenatal maternal estimate became non-significant after removing participants that were breastfed do provide results more in-line with what would be expected. But we can't use expected results as a basis to evaluate the validity of the approach. The absence of GxE and rGE are two other strong assumptions of the model that could also produce this kind unexpected results.

   d. I would suggest discussing the issue of genetic relatedness between adopting parents and offspring in terms of passive rGE which is a common problem for the estimation of parental effects in every familial design.
   e. Line 291: why use an unweighted PRS for EY3 (Lee, 2018), while the usual way of computing PRS (as a weighted sum of risk alleles) was used for birthweight?

   3. Limitations
   Assess other limitations of their method.

   a. limitation of the availability of birth father information,

   b. prenatal events uncorrelated with birthmother's genes (disease or accidents),

   c. Inferring prenatal environment effect from higher birth mother correlation compared to birthfather is subject to bias from measurement differences between the two (Loehlin, 2016).

   d. age at which the child is adopted (if the child has been partly raised by birth parents before adoption, it would bias (raise) the estimates of prenatal effects).

   e. evocative rGE not mentioned. It has been shown that parents partly react to children's behaviors. Thus, the estimate of maternal genetic postnatal effects could be biased (lowered) by evocative gene-environment correlation. In other words, the model also assumes no evocative gene-environment correlation.

   Final thoughts:

   1. I would like a better case made for why it is important to distinguish genetic effects into prenatal and postnatal effect.

   2. I would suggest the author make a clear distinction between the limits inherent to their sample (UK biobank) from those inherent to their methodological approach. I see important usefulness is plague by limits inherent to the sample used. At the same time, I am not aware of the availability of a big enough sample of adopted children with genotypic information available to compute PRS.

   Read the original source
6. 

   Version published to 10.1101/2021.08.04.455178v1 on bioRxiv

   Aug 6, 2021