Age and diet shape the genetic architecture of body weight in diversity outbred mice

  1. Kevin M Wright
  2. Andrew G Deighan
  3. Andrea Di Francesco
  4. Adam Freund
  5. Vladimir Jojic
  6. Gary A Churchill  Is a corresponding author
  7. Anil Raj  Is a corresponding author
  1. Calico Life Sciences LLC, United States
  2. The Jackson Laboratory, United States

Abstract

Understanding how genetic variation shapes a complex trait relies on accurately quantifying both the additive genetic and genotype–environment interaction effects in an age-dependent manner. We used a linear mixed model to quantify diet-dependent genetic contributions to body weight measured through adulthood in diversity outbred female mice under five diets. We observed that heritability of body weight declined with age under all diets, except the 40% calorie restriction diet. We identified 14 loci with age-dependent associations and 19 loci with age- and diet-dependent associations, with many diet-dependent loci previously linked to neurological function and behavior in mice or humans. We found their allelic effects to be dynamic with respect to genomic background, age, and diet, identifying several loci where distinct alleles affect body weight at different ages. These results enable us to more fully understand and predict the effectiveness of dietary intervention on overall health throughout age in distinct genetic backgrounds.

Editor's evaluation

This is an outstanding dissection of the genetic architecture of body weight at the genome-wide level across time and across environments. The use of a multiparental mouse population permits high-resolution mapping. The statistical analyses are advanced, leveraging new models, as well as tools developed specifically for this mouse population. The corresponding results are presented in nice and informative figures.

https://doi.org/10.7554/eLife.64329.sa0

eLife digest

Body weight is one trait influenced by genes, age and environmental factors. Both internal and external environmental pressures are known to affect genetic variation over time. However, it is largely unknown how all factors – including age – interact to shape metabolism and bodyweight.

Wright et al. set out to quantify the interactions between genes and diet in ageing mice and found that the effect of genetics on mouse body weight changes with age. In the experiments, Wright et al. weighed 960 female mice with diverse genetic backgrounds, starting at two months of age into adulthood. The animals were randomized to different diets at six months of age. Some mice had unlimited food access, others received 20% or 40% less calories than a typical mouse diet, and some fasted one or two days per week.

Variations in their genetic background explained about 80% of differences in mice’s weight, but the influence of genetics relative to non-genetic factors decreased as they aged. Mice on the 40% calorie restriction diet were an exception to this rule and genetics accounted for 80% of their weight throughout adulthood, likely due to reduced influence from diet and reduced interactions between diet and genes. Several genes involved in metabolism, neurological function, or behavior, were associated with mouse weight.

The experiments highlight the importance of considering interactions between genetics, environment, and age in determining complex traits like body weight. The results and the approaches used by Wright et al. may help other scientists learn more about how the genetic predisposition to disease changes with environmental stimuli and age.

Introduction

Quantifying the contributions of genetic and environmental factors to population variation in an age-dependent phenotype is critical to understanding how phenotypes change over time and in response to external perturbations. The identification of genetic loci that are associated with a complex trait in an age- and environment-dependent manner allows us to elucidate the dynamics and context dependence of the genetic architecture of the trait and facilitates trait prediction. For health-related traits, these genetic loci may also facilitate greater understanding and prediction of age-related disease etiology, which is an important step to genetically or pharmacologically manipulating these traits to improve health.

Standard approaches used to identify genetic loci associated with quantitative traits can be confounded by nonadditive genetic effects such as genotype–environment (G×E) and genotype–age (G×A) interactions (Robinson et al., 2017), contributing to the ‘missing heritability’ of quantitative traits (Sul et al., 2016). The linear statistical models routinely used in genetic mapping analyses do not account for variation in population structure between environments and polygenicity in G×E interactions. Population structure can substantially increase the false-positive rate when testing for G×E associations (Moore et al., 2019). Furthermore, not accounting for polygenic G×E interactions has the potential to incorrectly estimate the heritability of quantitative traits in the context of specific environments (Sul et al., 2016). To address these limitations, recent efforts have generalized standard linear mixed models (LMMs) with multiple variance components that allow for polygenic G×E interactions and environment-dependent residual variation (Sul et al., 2016; Runcie and Crawford, 2019; Moore et al., 2019; Dahl et al., 2020). Moreover, these LMMs substantially increase the power to discover genomic loci that are associated with phenotype in both an environment-independent and environment-dependent manner.

In this study, we used an LMM to investigate the classic quantitative trait, body weight, in a large population of diversity outbred (DO) mice. Body weight was measured longitudinally from early development to late adulthood, before and after the imposition of dietary intervention at 6 months of age. We expect diet and age to be important factors affecting body weight and growth rate; however, it remains to be determined how these factors will interact with genetic variation to shape growth. Two early studies found significant genetic correlations for body weight and growth rate during the first 10 weeks of mouse development, which supported the hypothesis that growth rates during early and late development were affected by pleiotropic loci (Cheverud et al., 1996; Riska et al., 1984). Subsequent experiments found that the heritability of body weight increased monotonically with age throughout development: from 29.3% to 76.1% between 1 and 10 weeks of age (Cheverud et al., 1996), from 6% to 24% between 1 and 16 weeks of age (Gray et al., 2015), and from 9% to 32% between 5 and 13 weeks of age (Parker et al., 2016). The heritability of growth rate also varied with age, but exhibited a peak of 24% at 3 weeks of age and then declined to nearly 4% at 16 weeks of age (Gray et al., 2015). The strength of association and effect size of QTLs for body weight and growth rate were specific to early or late ages and were inconsistent with the hypothesis that pleiotropic alleles affect animal size at early and late developmental stages (Cheverud et al., 1996; Gray et al., 2015). While these results are well supported, their interpretation is somewhat limited because body weight measurements ceased at young ages and significant QTLs encompassed fairly large chromosomal regions. Given these limitations, we were motivated to ask two questions: Will fine-mapping to greater resolution reveal single genes that function at either early or late developmental stages, or reveal multiple genes in tight linkage with variable age-specific effects? How will the effect of these loci change at later ages and under different diets?

We expect the interaction of dietary interventions, such as calorie restriction (CR) or intermittent fasting (IF), with genetics to greatly impact the body weight trajectories of mice. Researchers have observed genotype-dependent reductions in body weight in the 7–50 weeks after imposing a 40% CR diet, and variation in heritability of this trait with age from 42% to 54% (Rikke et al., 2006). A second study subjected a large genetic-mapping population to dietary intervention and identified multiple loci with significant genetic and genotype–diet interaction effects on body weight at 2–6 months of age (Vorobyev et al., 2019). These studies identified substantial diet- and age-dependent genetic variance for body weight in mice, similar to what has been found in humans (Robinson et al., 2017; Couto Alves et al., 2019; Wang et al., 2019). It, however, remains to be determined how the contribution of specific genetic loci to body weight changes in response to different dietary interventions and whether the effects observed in younger mice are indicative of the maintenance of body weight in adult mice.

In order to address these questions, we measured the body weight of multiple cohorts of genetically diverse mice from the DO population (Svenson et al., 2012) from 60 to 660 days of age (Figure 1A). At 180 days of age, we randomized mice by body weight and assigned each mouse to one of five dietary regimes – ad libitum (AL), 20% and 40% daily CR, and 1 or 2 days per week IF. Longitudinal measurements of body weight in these DO mice allowed us to discover the age-dependent genetic determinants of body weight and growth rate in the context of different dietary interventions.

Figure 1 with 1 supplement see all
Study design and body weight data summaries.

(A) Outline of study design. (B) Median (interquartile range) body weight in grams and (C) median (interquartile range) growth rate in grams per day for five dietary treatments from 60 to 660 days of age. Vertical gray dotted line denotes the onset of dietary intervention at 180 days of age.

In the following sections, we first describe the study design and collection of the genetic and body weight measurements. Second, we motivate the use of the gene–environment mixed model (G×EMM) (Dahl et al., 2020) to model genetic associations in these data and give an overview of our analyses. We next use this model to identify genetic loci having additive or genotype–diet interaction effects on body weight. We fine-map candidate loci and determine the scope of pleiotropy for age- and diet-specific effects. We find many, but not all, loci are associated with body weight in a narrow age range and localize to small genomic regions, in some cases to single genes. We utilize the full genome sequence of the DO founders and external chromatin accessibility data to further narrow the genomic regions to a small number of candidate variants at each locus. Interestingly, many diet-specific body weight loci localize genes implicated in neurological function and behavior in mice or humans.

Results and discussion

Study design and measurements

The DO house mouse (Mus musculus) population was derived from eight inbred founder strains and is maintained at Jackson Labs as an outbred heterozygous population (Svenson et al., 2012). This study contains 960 female DO mice, sampled at generations: 22–24 and 26–28. There were two cohorts per generation for a total of 12 cohorts and 80 animals per cohort. Enrollment occurred in successive quarterly waves starting in March 2016 and continuing through November 2017.

A single female mouse per litter was enrolled into the study after wean age (3 weeks old), so that no mice in the study were siblings and maximum genetic diversity was achieved. Mice were housed in pressurized, individually ventilated cages at a density of eight animals per cage (cage assignments were random). Mice were subject to a 12 hr:12 hr light:dark cycle beginning at 06:00 hr. Animals exit the study upon death. All animal procedures were approved by the Animal Care and Use Committee at The Jackson Laboratory.

From enrollment until 6 months of age, all mice were on an AL diet of standard rodent chow 5KOG from LabDiet. At 6 months of age, each cage of eight animals was randomly assigned to one of five dietary treatments, with each cohort equally split between the five groups (N = 192/group): AL, 20% CR (20), 40% CR (40), 1 day per week fast (1D), and 2 days per week fast (2D) (see Figure 1A). In a previous internal study at the Jackson Laboratory, the average food consumption of female DO mice was estimated to be 3.43 g/day. Based on this observation, mice on 20% CR diet were given 2.75 g/mouse/day and those on 40% CR diet were given 2.06 g/mouse/day. Food was weighed out for an entire cage of eight. Observation of the animals indicated that the distribution of food consumed was roughly equal among all mice in a cage across diet groups.

Mice on AL diet had unlimited food access; they were fed when the cage was changed once a week. In rare instances when the AL mice consumed all food before the end of the week, the grain was topped off midweek. Mice on 20% and 40% CR diets were fed daily. We gave them a triple feeding on Friday afternoon to last till Monday afternoon. As the number of these mice in each cage decreased over time, the amount of food given to each cage was adjusted to reflect the number of mice in that cage. Fasting was imposed weekly from Wednesday noon to Thursday noon for mice on 1D diet and Wednesday noon to Friday noon for mice on 2D diet. Mice on 1D and 2D diets have unlimited food access (similar to AL mice) on their nonfasting days.

From enrollment until 660 days of age, each animal underwent a number of phenotypic assays to assess a wide range of physiological health parameters. These included two 7-day stints in Sable Systems Promethion metabolic cages at 120 and 390 days of age, two blood draws at 160 and 480 days of age, and one of each of the following challenge-based assays at 300–330 days of age: 3-day stint in a cage with wheel, rotarod, grip strength, dual-energy X-ray absorptiometry, echocardiogram, Y-maze spontaneous alternation, and acoustic startle.

Body weight measurements

Body weight was measured once every week for each mouse throughout its life. The body weight measurements for this analysis were collated on February 1, 2020, at which point 941 mice (98%) had measurements at 180 days, 890 (93%) at 365 days, 813 (85%) at 550 days, and 719 (75%) at 660 days. For these analyses, we included all body weight measures for each mouse up to 660 days of age. We smoothed out measurement noise, either due to errors in measurement or swaps in assigning measurements to mice, using an 1 trend filtering algorithm (Kim et al., 2009), which calculates a piecewise linear trend line for body weight for each mouse over its measurement span. The degree of smoothing was learned by minimizing the error between the predicted fit and measurements at randomly held-out ages across all mice. In the rest of this article, body weight and growth rate refer to the predicted fits from 1 trend filtering.

We present the average trends in body weight and growth rate, stratified by dietary intervention, in Figure 1B and C, respectively. The body weight and growth rate trends without 1 trend filtering are presented in Figure 1—figure supplement 1. The most prominent observation from these trends is that dietary intervention contributes the most to variation in body weight in this mouse population. After accounting for this source of variation, there remain substantial and different quantities of variation in body weight trends within the different dietary interventions, suggesting the existence of heteroscedastic noise or plausible G×D interaction effects on body weight trajectories.

Genotype measurements

We collected tail clippings and extracted DNA from 954 animals (http://agingmice.jax.org/protocols). Samples were genotyped using the 143,259-probe GigaMUGA array from the Illumina Infinium II platform (Morgan et al., 2015) by NeoGen Corp. (genomics.neogen.com/). We evaluated genotype quality using the R package: qtl2 (Broman et al., 2019). We processed all raw genotype data with a corrected physical map of the GigaMUGA array probes (https://kbroman.org/MUGAarrays/muga_annotations.html). After filtering genetic markers for uniquely mapped probes, genotype quality, and a 20% genotype missingness threshold, our dataset contained 110,807 markers.

We next examined the genotype quality of individual animals. We found seven pairs of animals with identical genotypes, which suggested that one of each pair was mislabeled. We identified and removed a single mislabeled animal per pair by referencing the genetic data against coat color. Next, we removed a single sample with missingness in excess of 20%. The final quality assurance analysis found that all samples exhibited high consistency between tightly linked markers: log odds ratio error scores were less than 2.0 for all samples (Lincoln and Lander, 1992). The final set of genetic data consisted of 946 mice.

For each mouse, starting with its genotypes at the 110,807 markers and the genotypes of the 8 founder strains at the same markers, we inferred the founders-of-origin for each of the alleles at each marker using the R package: qtl2 (Broman et al., 2019). This allowed us to test directly for association between founder-of-origin and phenotype (rather than allele dosage and phenotype, as is commonly done in QTL mapping) at all genotyped markers. Using the founder-of-origin of consecutive typed markers and the genotypes of untyped variants in the founder strains, we then imputed the genotypes of all untyped variants (34.5 million) in all 946 mice. Targeted association testing at imputed variants allowed us to fine-map QTLs to a resolution of 1–10 genes.

Motivating models for environment-dependent genetic architecture

Genome-wide QTL analyses in model organisms over the last decade have predominantly employed linear mixed models (e.g., EMMA Kang et al., 2008), FastLMM [Lippert et al., 2011], GREML [Yang et al., 2011], GEMMA [Zhou and Stephens, 2012], and LDSC [Bulik-Sullivan et al., 2015], expanding on the heuristic that samples sharing more of their genome have more correlated phenotypes than genetically independent samples. We found that the distributions of covariances in body weight, measured at 500 days of age, between animal pairs within the AL treatment were nearly indistinguishable when we partition pairs into high-kinship (>0.2) and low-kinship groups (Figure 2, no significant separation between solid and dashed red lines). However, animal pairs in the 40% CR treatment exhibited significantly lower covariance in body weight in the low-kinship group compared to the high-kinship group (Figure 2, significant separation between solid and dashed orange lines). This observation suggests that the genetic contribution to body weight is different in distinct dietary environments. This observation also motivates the use of recently developed generalized linear mixed models to conduct genome-wide QTL analysis because they more fully account for environment-dependent genetic variances, reduce false-positive rates, and increase statistical power (Yang et al., 2011; Sul et al., 2016; Runcie and Crawford, 2019; Moore et al., 2019; Dahl et al., 2020).

Illustrating diet-dependent association between genetic and phenotypic similarities.

Phenotypic divergence between animal pairs is quantified by the covariance in body weight at 500 days of age. We plot the cumulative density of body weight covariance for all pairs of animals in the ad libitum (AL) and 40% calorie restriction (CR) dietary treatments, partitioned into high-kinship (>0.2) or low-kinship groups. We test for separability between the high- and low-kinship distributions using the Mann–Whitney U test, and report p-values from this test. The reported area under the curve (AUC) is a standard transformation of the Mann-Whitney U test statistic.

Simulations

To evaluate the accuracy of G×EMM and compare it against standard linear mixed models like EMMA (Kang et al., 2008), we simulated phenotypic variation under a broad range of values for proportion of phenotypic variance explained by genetics (PVE) in each of two environments. For all simulations, we fixed the total sample size to N=946 and used the observed kinship matrix for the 946 DO mice in this study. For these simulations, the environment-specific genetic contribution PVEe, environment-specific noise σe2, and the relative distribution of samples between environments are the tunable parameters. In order to explore how the two models behaved under a wide range of conditions, we varied PVEe{0.05,0.10,,0.95}, fixed σe12=σe22=1.0, and assigned samples to environments either at a 1:1 ratio or a 4:1 ratio (i.e., a total of 114 parameter values). For each setting of the tunable parameter values, we ran 50 replicate simulations and in each simulation, we randomly assigned samples to one of the two environments. Given the environment assignments, kinship matrix, and the set of tunable parameter values for each simulation, we computed the environment-specific genetic variance using Equation 13, Equation 14, and Equation 15. Next, we generated the vector of phenotypes from a multivariate normal distribution with zero mean and covariance structure dependent upon the observed kinship matrix, environment assignments, tunable parameter settings, and the environment-specific genetic variance for each environment. Using the simulated phenotypes, we computed estimates of PVEtot under both the EMMA and G×EMM models and estimates of PVEe under the G×EMM model.

We first evaluated the accuracy of PVEtot estimated from the two models and found that the G×EMM model estimated total PVE with little bias and lower variance compared to EMMA in nearly all simulations across the suite of parameter combinations that we investigated (Figure 3A and B). In particular, we found that EMMA substantially underestimated the total PVE in comparison to G×EMM when samples were equally distributed between environments and there was substantial difference in PVE or noise between environments. This is because PVEe will have a greater impact on PVEtot when both environments are more equally represented in the study sample.

Evaluating gene–environment mixed mode (G×EMM) and EMMA using simulated datasets.

(A). Comparison of true PVEtot to that estimated from EMMA (left panel) and G×EMM (right panel). Simulations were run with an equal number of samples in each environment (Ne1=Ne2) and with the same value for the environment-specific error terms (σe12=σe22=1). (B). Same as (A) but with a 4:1 ratio of samples between environments (Ne1=4×Ne2). (C). Plot of the true PVEe2 vs. the G×EMM estimate of PVEe2. Gray points are the results of individual simulations, orange lines denote the median and 95% interquartile range (IQR). The three panels differ in the true proportion of phenotypic variance explained by genetics (PVE) specific to environment 1, PVEe1{0.2,0.5,0.8}. Simulations were run with an equal number of samples in each environment (Ne1=Ne2) and with the same value for the environment-specific error terms (σe12=σe22=1). (D). Same as (C) but with a 4:1 ratio of samples between environments (Ne1=4×Ne2).

Next, we examined the sensitivity and specificity of the G×EMM model to quantify environment-specific genetic contribution to phenotypic variance. We found that when samples were equally distributed between environments, G×EMM accurately estimates PVEe under a wide range of parameter values (Figure 3C). When there was a skew in the number of samples between environments, the PVEe estimate for the environment with the larger sample size was as accurate as when there was no skew in sample size (similar to Figure 3C, result not shown). In contrast, the PVEe estimate for the environment with the smaller sample size was more variable and the median PVEe estimates were consistently lower than the true PVEe (Figure 3D). Finally, we observed that increasing the environment-specific noise terms, σe12 and σe22, increased the variance of our estimates of PVEe but introduced little bias (results not shown).

Overall, the G×EMM model outperforms the EMMA model in estimating total PVE and shows little bias in estimating environment-specific PVE across a broad range of scenarios relevant to our study.

Overview of analyses

Starting with body weight measurements in 959 mice from 30 to 660 days of age, and founder-of-origin alleles inferred at 110,807 markers in 946 mice, we first quantified how the heritability of body weight and growth rate changes with age and between dietary contexts. We used the G×EMM model (described in detail in ‘Materials and methods’) to account for both additive environment-dependent fixed effects and polygenic gene–environment interactions. We considered two different types of environments: diet and generation. The five diet groups were applied from 180 to 660 days and the 12 generations were applied from 30 to 660 days. Next, we performed genome-wide QTL mapping for body weight at each age independently, testing for association between body weight and the inferred founder-of-origin at each genotyped marker. For ages 180–660 days, we additionally tested for association between body weight and the interaction of diet and founder-of-origin at each marker. We computed p-values using a sequential permutation procedure (Besag and Clifford, 1991; Shim and Stephens, 2015) at each variant for each of the additive and interaction tests and used these to assign significance (Abney, 2015). Finally, for each significant locus, we performed fine-mapping to identify the putative causal variants and founder alleles driving body weight and underlying functional elements (genes and regulatory elements) to ascertain the possible mechanisms by which these variants act.

PVE of body weight across age and diet

First, we quantified the overall contribution of genetics to variation in body weight and, importantly, how this contribution changed with age. We applied both EMMA and G×EMM to body weight estimated every 10 days. Since the mice from each generation cohort were measured at the same time every week, we used generation as a proxy for the shared environment that mice are exposed to as part of the study design. We accounted for generation-specific fixed effects (α in Equation 1) in both models and genotype-generation random effects (γ) in G×EMM. For ages after dietary intervention (180 days), we additionally accounted for diet-specific fixed effects in both models and genotype-diet random effects in G×EMM. We estimated the variance components in the model at each age independently and computed the total and diet-dependent PVE using Equation 7 and Equation 13.

In Figure 4A, we observed that the PVE of body weight steadily increased during early adulthood and up to 180 days of age, when dietary intervention was imposed. The G×EMM model estimates a higher PVE than the EMMA model during this age interval because the former model specifically accounts for polygenic genotype-generation effects. Following dietary intervention, PVE decreased in four of the five dietary groups; the one exception was the 40% CR group, which maintained a high PVE (PVE400.8) (Figure 4A) and low total phenotypic variance (Figure 4—figure supplement 3A, top-right panel) from 180 to 660 days of age. In contrast, the AL group had the lowest PVE and greatest total phenotypic variance in the same age range. Notably, 1 trend filtering of the raw measurements proved useful in quantifying smoothly varying trends in the PVE of body weight and growth rate, and improving our estimates of the PVE of growth rate by reducing the effect of measurement noise (Figure 4—figure supplement 1). Moreover, these results were robust to variation in the genetic data used to calculate the kinship matrix and to survival bias at 660 days of age.

Figure 4 with 6 supplements see all
Proportion of phenotypic variance explained by genetics.

(A) Body weight proportion of phenotypic variance explained by genetics (PVE) (± SE) for 30–660 days of age. Total PVE estimates are derived from the EMMA (light gray) and gene–environment mixed mode (G×EMM) (dark gray) models. Diet-dependent PVE values were derived from the G×EMM model. Dotted vertical line at 180 days depicts the time at which all animals were switched to their assigned diets. (B) Growth rate PVE; details the same as (A).

The kinship matrix used for estimating these PVE values was computed based on the founder-of-origin of marker variants (Aylor et al., 2011). When using kinship estimated using biallelic marker genotypes (as is commonly done in genome-wide association studies [GWAS]), we observed largely similar trends in PVE; however, differences in PVE between diets after 400 days were harder to discern due to much larger standard errors for the estimates (Figure 4—figure supplement 3A, left panels). To test for bias or calibration errors in our PVE estimates, we randomly permuted the body weight trends between mice in the same diet group and recalculated the total and diet-dependent PVE values. Consistent with our expectations, PVE dropped to nearly zero for the permuted dataset (Figure 4—figure supplement 3B, left panel), indicating that the PVE estimates are well-calibrated. Finally, to evaluate the contribution of survival bias to our estimates, we recomputed PVE at all ages after restricting the dataset to mice that were alive at 660 days. We observed PVE estimates largely similar to those computed from the full dataset, suggesting very little contribution of survival bias to our estimates (Figure 4—figure supplement 3C).

When using kinship computed from genotypes, we note that the PVE in the permuted dataset does not drop to zero, suggesting that genotype-based kinship includes some background relatedness that can explain some of the phenotypic variation even after permuting the labels (Figure 4—figure supplement 3B, right panel). The genotype-based kinship includes two components: the genetic sharing between the founder strains and the genetic sharing arising from the breeding strategy used to develop the DO panel. Randomly permuting the phenotype labels breaks the link between phenotypic similarity and the latter component of genotype-based kinship, but does not effectively break the link with the former component. Kinship computed using the founder allele probabilities explicitly includes only the latter component; this component explains little phenotypic variation once the phenotype labels are randomly permuted.

Next, we quantified the age-dependent contribution of genetics to variation in growth rate, enabled by the dense temporal measurement of body weight. As before, we applied EMMA and G×EMM to growth rate estimated every 10 days. In Figure 4B, we observed that PVE of growth rate increases rapidly during early adulthood, and then decreases to negligible values around 240 days of age. In contrast to body weight, PVE of growth rate is substantially lower at all ages, and there is little divergence in PVE across diet groups for most ages. Notably, the decrease and subsequent increase in growth rate PVE coincide with specific metabolic, hematological, and physiological phenotyping procedures that these mice underwent at specific ages as part of the study (Figure 1—figure supplement 1). Due to lower values and greater variance in PVE of growth rate with age, we focus on body weight throughout the rest of the article.

In summary, the 40% CR intervention produced the greatest reduction in average body weight and maintained a high PVE after dietary intervention. This is because the genetic variance in body weight remained relatively high and the environmental variance remained relatively low throughout this interval. In contrast, body weight PVE steadily decreased with age in each of the four less restrictive diets, which was due to a steady increase in noise (nongenetic variance) and not a decrease in the genetic variance of body weight (Figure 4—figure supplement 3A). Even though the genetic variance in body weight is nearly constant across diets from 180 to 660 days of age, the effect of specific variants may change with age.

Genome-wide QTL analysis of body weight across age and diet

We sought to identify loci significantly associated with body weight in a diet-dependent and age-dependent manner. To this end, we tested for association between the inferred founder-of-origin of each typed variant and body weight at each age independently. We note that body weight measurements taken at different ages are not independent; that is, we may detect a locus at a specific age if it has small effects on body weight acting over a long period of time or a large effect on body weight, resulting in rapid bursts in growth. Thus, a locus identified as significant at a specific age indicates its cumulative contribution to body weight at that age.

We identified 29 loci significantly associated with body weight at any age in the additive or interaction models using a genome-wide Bonferroni-corrected p104 threshold for both the additive and the genotype–diet interaction tests (Figure 5—figure supplement 1). Additionally, we report suggestive genotype–diet interaction QTLs using a weaker threshold of p103. The Bonferroni correction was computed based on the expected number of linkage-disequilibrium (LD) blocks in our dataset; as a proxy for the number of LD blocks, we used the number of top eigenvalues of the LD matrix between genotyped variants that explained at least 90% of the variance across markers. This threshold also corresponds to a 10% variant-level expected false discovery rate, using the Benjamini–Hochberg method (Benjamini and Hochberg, 1995). In order to focus on highly significant associations, we also used the number of eigenvalues explaining 99.5% of the variance across markers (Gao et al., 2010), resulting in a more stringent threshold of p1.1×105 (see Table 1). Notably, these thresholds hold for the interaction tests as well because we have only one test per marker. Using all biallelic variants imputed from the complete genome sequences of the eight DO founder strains (Keane et al., 2011), we retested the genetic association for the 29 candidate loci with the additive and interaction models. Specifically, we tested for association between all imputed genotype variants and body weight, accounting for kinship estimated using founder-of-origin inferred at genotyped variants as before (see file “emma_vs_gxemm.xlsx” in Data Release: Results for a comparison of the number of significant loci with those obtained using the EMMA model).

Table 1
Candidate diet-independent and diet-dependent body weight loci.

For each locus, we identified the variant with the lowest p-value at any age, the founder allele pattern (FAP) of this variant, the number of significant variants that comprise this lead FAP, the genomic positions spanning these significant FAP variants, and all ages for which at least one FAP variant is associated with body weight. For loci in which the lead FAP is comprised of fewer than 10 significant variants, we also present results for the second FAP. We list candidate genes for loci where the FAP spans three or fewer genes. Highly significant associations (p1.1×105) are denoted with ***, significant associations (p104) with **, and suggestive associations (p103) with *.

FAP position (Mb)Significant variantsSignificant age range (days)Age-dependent nonlinearityLead candidate genes (if ≤3)
ChrmFounder Allele Pattern (FAP)FAP rankp-ValueStartEndTotalOpen chromatin
Diet-independent
1AJ/NOD13.24E-05**152.046626152.04662611601.44E-05Trmt1l, Edem3
AJ/NOD/12923.05E-05**151.473677152.280212581060
2129/CAST/PWK12.17E-05**77.15496277.357295808120–3602.83E-09Ccdc141, Sestd1
3AJ/NOD/NZO/CAST11.89E-05**50.53359950.59507355200–2605.03E-12Slc7a11
4AJ16.81E-06***58.95036460.26712852260–3603.82E-19Ugcg
PWK/WSB22.80E-05**59.46124859.981846393100–300
6AJ/NOD12.75E-07***53.6176155.55597789260–2001.55E-26Creb5
7B6/CAST14.20E-06***71.37500772.78684947180–1602.48E-11Mctp2
7AJ/129/NZO/PWK15.70E-06***134.08785134.70446520280–2009.17E-10Adam12
10129/NZO/PWK/WSB14.99E-06***9.0780549.0780541540–6609.52E-06Samd5
129/NZO/WSB22.25E-05**8.9033879.092694102540Sash1, Samd5
10CAST/PWK15.98E-06***91.16319191.9052872,77981120–6604.09E-10Anks1b, Apaf1
11AJ/NZO/PWK/CAST15.89E-05**58.15542458.1554241802.35E-05
B6/CAST24.91E-05**56.98564559.0353096037080–100
12NZO/CAST16.47E-05**99.52055999.907182433160–2603.83E-12Foxn3
15B6/129/NZO14.77E-07***99.39060399.652959220260–6001.50E-13Aqp2, Aqp5, Aqp6
17AJ/NOD/WSB18.56E-06***6.7532778.853111491160–4201.23E-10Pde10a
19AJ/129/NZO/PWK17.90E-05**23.02504323.17612581480–1204.46E-09Trpm3, Klf9
Diet-dependent
1NOD/CAST14.89E-04*151.032114153.716837283360–6603.32E-04
2AJ/CAST/PWK/WSB17.74E-04*22.19222222.786648684801.32E-04
2PWK13.34E-04*73.59014275.3715641,67244280–3004.62E-03
3NOD/CAST/PWK/WSB13.87E-05**50.18474650.4198215420–6607.90E-04Slc7a11
3B6/PWK18.29E-05**156.74554156.7455416601.60E-05Negr1
129/CAST/WSB27.33E-04*156.133466156.3878251708660
4129/NOD/PWK19.61E-04*57.69630157.846521012005.10E-07Palm2, Pakap, Akap2
NOD29.76E-04*57.66982757.6698271200
5NOD13.88E-05**19.21390421.5700715316420–6602.26E-03Magi2, Ptpn12
5PWK15.84E-04*68.98665570.490341277360–4802.28E-03Kctd8
5AJ/129/NOD/WSB14.00E-05**117.543498118.0504537480–5401.55E-03
B6/CAST/PWK21.04E-04*116.797769118.2707456910360–600
6B6/CAST/PWK11.04E-04*54.14613255.452827507101540–6607.51E-04Ghrhr
6B6/CAST11.68E-05**139.190506140.1416941115240–4209.63E-03Pik3c2g
7NZO/PWK19.60E-04*133.838117133.92402321834203.19E-07Adam12
12AJ/B616.29E-04*78.0610279.734912102420–5404.23E-04Gphn
WSB29.84E-04*78.10798878.9328421228360–420Gphn
12AJ/CAST/PWK/WSB13.53E-04*102.447268102.56532434420–4802.59E-05
13NZO/WSB11.00E-04**117.100864118.777463991200–3001.14E-03Fgf10
15AJ/129/NOD13.52E-04*10.73803412.040137534240–3001.27E-04
17B6/129/NZO17.52E-04*6.7558657.598818263300–3603.23E-03
18129/CAST13.86E-04*71.3608771.58844161480–5406.92E-03Dcc
19AJ/B6/WSB11.24E-04*21.80666522.051277321220–3003.54E-03

Our fine-mapping analysis confirmed 24 candidate loci: 5 loci unique to the additive model, 10 loci unique to the interaction model, and 9 loci significant in both models (Table 1). We identified body weight associations with age-dependent effects from early adulthood to adulthood: four diet-independent loci were associated with body weight exclusively during early adulthood (ages 60–160 days) and three were associated exclusively during adulthood (ages 200–660 days); the remaining seven loci were associated prior to the imposition of dietary restriction at day 180 and continued to be associated into adulthood (Table 1). The majority of diet-dependent loci (12 of 19) had a detectable effect on body weight at least 240 days after dietary intervention (ages 420–660 days; Table 1). For each candidate body weight locus, we sought to determine the likely set of causal variants and estimate the effect of the eight founder alleles in a diet- and age-dependent manner.

Improved mapping resolution with founder allele patterns

In order to facilitate characterization and interpretation of the genetic associations at each locus, we sought to represent these associations in terms of the effects of founder haplotypes (Zhang et al., 2022). First, we determined the founder-of-origin for each allele at every variant that was significant in at least one age. This allowed us to assign a founder allele pattern (FAP) to each significant variant in the locus. For example, if a variant with alleles A/G had allele A in founders AJ, NZO, and PWK and the allele G in the other five founders, then we assign A to be the minor allele of this variant and define the FAP of the variant to be AJ/NZO/PWK. Next, we grouped variants based on FAP and define the logarithm of the odds (LOD) score of the FAP group to be the largest LOD score among its constituent variants. (Note that, by definition, no variant can be a member of more than one FAP group.) By focusing on the FAP groups with the largest LOD scores, we significantly reduced the number of putative causal variants (Table 1, see file all_significant_snps.csv in Data Release: Results), while representing the age- and diet-dependent effects of these loci in terms of the effects of its top FAP groups. We further narrowed the number of candidates by intersecting the variants in top FAP groups with functional annotations (e.g., gene annotations, regulatory elements, tissue-specific regulatory activity, etc.). For many loci, this procedure identified candidate regions containing 1–3 genes (Table 1), and we provide the full list of all genes within candidate regions in the file all_significant_genes.csv in Data Release: Results. In order to demonstrate the utility of this approach, we first examined a single locus on chromosome 6 strongly associated under the additive model.

We used this approach to examine a diet-independent locus on chromosome 6 significantly associated with body weight during early adulthood (Figure 5A). Interestingly, this same locus was nominally associated with body weight in certain dietary treatments at later ages. We hypothesized that the functional variant(s) responsible for the diet-dependent body weight association at this locus are among the variants in the respective lead FAP groups because they exhibit the strongest statistical association and it is unlikely any additional variants are segregating in this genomic interval beyond those identified in the full genome sequences of the eight founder strains. For the diet-independent locus, at age 120, we identified 87 significantly associated variants; of these, 79 could be assigned to the lead FAP group and shared a similarly high LOD score (Figure 5B). All of these variants are single-nucleotide polymorphisms (SNPs located in the gene Creb5, 78 are intronic and 1 a synonymous exon variant). For the diet-dependent locus, at age 600 days, we identified 617 variants as significant; of these, 507 could be assigned to the lead FAP and shared a similarly high LOD score (Figure 5D). Also, 2 of the 507 variants were intergenic structural variants; of the remaining SNPs, 5 were noncoding exon variants, 167 were intronic, and the remainder were intergenic. Given that all candidate variants were noncoding, we next sought to determine whether they were located in regulatory elements across a number of tissues, identified as regions of open chromatin measured using ATAC-seq (Cusanovich et al., 2018) or DNase-seq (Gorkin et al., 2020) (see file "chromatin_accessibility_celltypes.xlsx" in Data Release: Results for a description of the cell types for each of the assays, and their grouping into tissues). For the diet-independent and diet-dependent loci, we found 2 and 101 variants, respectively, that were located in regions of open chromatin (Figure 5C and E). Notably, both variants with diet-independent effects lay within the same muscle-specific regulatory element located within Creb5, suggesting that these variants likely affected body weight by regulating gene expression in muscle cells.

Figure 5 with 4 supplements see all
Distinct founder allele patterns (FAPs) contribute to diet-independent and diet-dependent effects on body weight within a region on chromosome 6.

(A) Manhattan plots of additive genetic associations and genotype–diet associations on chromosome 6 at multiple ages. (B, D) Fine-mapping loci associated with body weight: diet-independent association at 120 days of age and diet-dependent association at 600 days of age. Solid circles indicate significant variants. Colors denote variants with shared FAPs; ranks 1, 2, and 3 by logarithm of the odds (LOD) score are colored red, orange, and yellow, respectively. (C, E) Significant variants, colored by their FAP group, along with the gene models (shown in blue) and the tissue-specific activity of regulatory elements near these variants (shown in gray). Significant variants that lie within regulatory elements are highlighted as diamonds, and regulatory elements that contain a significant variant are highlighted in black.

Next, we identified a locus on chromosome 12 with strong diet-dependent effects on body weight from 300 to 420 days of age (Table 1; Figure 6). Upon fine-mapping the locus with the strongest association, at 420 days of age, we identified 77 significant variants partitioning into two distinct lead FAP groups, with similarly high LOD scores and centered at the same gene (Figure 6A and B). The rank 1 FAP group contained variants with the minor allele specific to AJ and B6, whereas the rank 2 FAP group contained variants with the minor allele specific to WSB (Figure 6A). We found that the minor allele of the lead imputed variant in the AJ/B6 FAP group was associated with increased body weight in the 2D fasting diet, but had little effect on the other four diets (Figure 6C). In contrast, the minor allele of the lead imputed variant from the WSB-specific FAP group had the largest positive effect on body weight in the AL and 1D fast diet and largest negative effect on body weight in the 40% CR diet (Figure 6C). The striking difference in diet-specific effects for WSB and AJ/B6 alleles suggests that there are multiple functional variants affecting body weight at this locus (e.g., allelic heterogeneity; Singh, 2013). Of the 77 variants, the AJ/B6 FAP group contained six intergenic SNPs and four intronic SNPs spanning three genes: Gphn (gephyrin), Plekhh1, and Rad51b (Figure 6A). One of these 10 variants is located in a regulatory element active specifically in adipose tissue. The remaining 67 significant variants all belonged to the WSB-specific FAP group; of these, 23 SNPs were intergenic and 44 were intronic and centered at the gephyrin gene. Also, 4 of these 67 variants were located in regulatory elements active in adipose tissue as well as other tissues relevant to metabolism (Figure 6D).

Multiple founder allele patterns (FAPs) at a body weight QTL on chromosome 12 show distinct diet-dependent effects.

(A) Fine-mapping loci, under the interaction model, at 420 days of age. Significant variants are marked as solid circles. Colors denote variants with shared FAPs; ranks 1, 2, and 3 by logarithm of the odds (LOD) score are colored red, orange, and yellow, respectively. (B) Log odds ratio as a function of age, for a single variant that exhibits the strongest association from each of the top three FAPs. (C) Age and diet-dependent effect size of variants with the strongest associations specific to either the WSB minor allele or AJ/B6 minor allele. (D) Significant variants, colored by their FAP, along with the gene models (shown in blue) and the tissue-specific activity of regulatory elements near these variants (shown in gray). Significant variants that lie within regulatory elements are highlighted as diamonds and regulatory elements that contain a significant variant are highlighted in black.

No evidence for pleiotropic alleles affecting diet-independent and dependent effects

We identified a diet-independent locus on chromosome 6 significantly associated with body weight during early adulthood and nominally associated with body weight in certain dietary treatments at later ages (Figure 5A). One explanation for this result is a single pleiotropic allele affecting body weight at two distinct stages of life: early adulthood and mid-late adulthood. Alternatively, this result could be explained by multiple alleles at a single locus (perhaps similar to the allelic heterogeneity observed in Figure 6) or multiple loci in tight LD. Fine-mapping this locus using the additive model, we identified the variant with the highest LOD score to be at 53.6 Mb. The minor allele at this lead variant was common to the AJ and NOD founders, while the remaining six founders possessed the alternate allele; this defined the lead diet-independent FAP at this locus to be AJ/NOD (Figure 5B). Separately, we fine-mapped this locus using the interaction model, identified the lead variant at 55.1 Mb, and determined the lead FAP to be B6/CAST/PWK (Figure 5D). These results do not support the pleiotropy hypothesis and were consistent with the hypothesis that distinct body weight alleles derived from different DO founders were responsible for the diet-dependent and diet-independent body weight associations.

We next fine-mapped other QTLs that had diet-dependent and diet-independent associations located in adjacent genomic regions to assess whether the associations were due to a single pleiotropic allele or, as we found on chromosome 6 (Figure 5B and D), the additive and interaction effects were due to distinct alleles. Of the eight additional diet-dependent and diet-independent associations that colocalized to the same genomic region (Table 1), all of them were consistent with the hypothesis that distinct FAPs rather than a single pleiotropic allele were responsible for diet-dependent and diet-independent associations. The first six loci were located in adjacent genomic regions (Figure 5—figure supplement 3). The remaining two loci were composed of similar (although, not identical) FAP groups located in overlapping genomic regions (Figure 5—figure supplement 4). At one of these loci, on chromosome 3, the FAP driving the diet-dependent association was NOD/CAST/PWK/WSB, whereas the diet-independent association was due to AJ/NOD/NZO/CAST (Figure 5—figure supplement 4A). The diet-dependent and diet-independent chromosome 7 loci are due to differential effects between NZO/PWK and AJ/129/NZO/PWK, respectively (Figure 5—figure supplement 2D). In summary, we identified multiple instances of diet-dependent and diet-independent associations that mapped to the same genomic region and for all of these loci we determined that distinct FAPs were responsible for diet-dependent or diet-independent effects on body weight.

Nonlinear changes in genetic effects with age

We found that the relationship between model LOD score and age was similar for each of the three lead FAPs at the diet-independent and diet-dependent loci on chromosome 6 (Figure 7A and B). The minor allele of the lead variant at the diet-independent locus was associated with increased body weight at young ages (Figure 7C), whereas the minor allele for the lead diet-dependent variant had a positive effect on body weight under the 20% CR diet, a nearly neutral effect under the 40% CR, and a negative effect under the AL diet (Figure 7D). We next measured the effect of each founder allele at the lead diet-independent genotyped marker and, consistent with the lead FAP group for the diet-independent association, the AJ and NOD alleles had large positive effects at young ages (Figure 7E). For the diet-dependent association, B6, CAST, and PWK alleles were associated with decreased body weight in the AL diet and increase in body weight in the 20% CR diet (Figure 7F), consistent with their role in defining the lead diet-dependent FAP group. This example clearly demonstrates the insight gained by focusing on lead FAP variants to link specific founder alleles to variation in body weight and narrow the number of potential functional variants underlying body weight.

Figure 7 with 2 supplements see all
Body weight loci on chromosome 6 exhibit age- and diet-specific effects.

(A) Log odds ratio of additive body weight association as a function of age for the lead variant from each founder allele pattern (FAP) group. Red, orange, and yellow colors denote lead variant for rank 1, 2, and 3 FAPs. (B) Same as (A) for interaction body weight association on chromosome 6. (C, D) Estimated mean (SE) effect on body weight (grams) of the minor allele for the lead imputed variant. For the diet-dependent locus (D), effects are shown for each diet treatment. (E) Estimated mean effect (±SE) on body weight of each founder allele for the genotyped marker with the highest logarithm of the odds (LOD) score from the additive model. (F) Estimated effect (±SE) on body weight of each founder allele under the ad libitum (AL) and 20% calorie restriction (CR) diets.

Given high temporal resolution in the measurements of body weight, we tested for nonlinearity in the trends of effect size with age for all significant loci in Table 1. Specifically, for each diet-independent locus, we tested for nonlinearity in the effect size trend with age in at least one founder allele of the lead variant, and for each diet-dependent locus, we tested for nonlinearity in at least one (founder, diet) pair at the lead variant (referred to as ‘age-dependent nonlinearity’). In addition to diet-dependent nonlinearity, we also observed age-dependent nonlinearity at the diet-independent locus on chromosome 6 (Figure 7C; p=1.35×1013). Surprisingly, the nonlinearity at this locus appears to be largely driven by the AJ founder background (Figure 7E; AL, yellow trace, p=1.55×1026 vs. NOD, navy trace, p=6.5×106), indicating that the effects of different founder alleles on body weight can have substantially different trends with age. Overall, we observed age-dependent nonlinearity to be predominant, with all 14 diet-independent loci exhibiting nonlinear trends with at least one founder allele (Table 1). In contrast, we observed nonlinear trends with age for diet-dependent effects to be less common, with only 4 of 19 diet-dependent loci exhibiting nonlinear trends with age in the effects of at least one founder allele under one diet. Additionally, nonlinearity was often specific to a subset of founder strains that are driving the associations at each locus, suggesting that the genetic background plays an important role when interpreting the dynamics of the genetic architecture of body weight in DO mice. Notably, such age-dependent nonlinearity often cannot be discerned even with large cross-sectional data that are typical of modern GWAS in humans.

Nonlinear changes in diet-specific genetic effects

The diet-dependent locus on chromosome 6 illustrated a rather counterintuitive result; while we observed an approximately linear reduction in median body weight between the AL, 20% CR, and 40% CR diets in response to a linear reduction in calories (Figure 1B), the effects of this locus on body weight were nonlinear in the context of each diet (Figure 7D). We defined diet-dependent effects to have a nonlinear trend with respect to diet (referred to as ‘diet-dependent nonlinearity’) if the genetic effects in the context of either the 20% CR or 1D fast treatments deviated substantially from the average of the genetic effects in the context of AL and 40% CR diets or AL and 2D fast diets, respectively. We observed a second instance of diet-dependent nonlinearity at a locus we mapped to chromosome 5. This locus had the strongest diet-dependent association observed in the genome. The lead FAP group, containing variants with an allele private to the NOD strain, was associated with large positive effect on body weight in the 20% CR diet, a small positive effect in the 1D fast diet, and nearly neutral effects in the 40% CR, 2D fast, and AL diets (Figure 7—figure supplement 1). In total, we found this pattern to be quite common, with 9 of the 19 significant loci identified under the interaction model exhibiting diet-dependent nonlinearity (Figure 7—figure supplement 2).

Candidate body weight genes linked to neurological and metabolic processes

Of the 24 significant loci from the additive and interaction models, we identified 14 loci with lead FAP groups spanning 1–3 genes (Table 1). Many of these genes implicated in modulating body weight are also known to affect neurological and metabolic processes.

One example is the neuronal growth regulator 1 (Negr1) gene, a candidate linked to body weight in mid-late adulthood via the lead 129/CAST/WSB FAP group within a diet-dependent locus on chromosome 3. This gene is highly expressed in the cerebral cortex and hippocampus in the rat brain (Miyata et al., 2003) and is known to regulate synapse formation of hippocampal neurons and promote neurite outgrowth of cortical neurons (Hashimoto et al., 2008; Sanz et al., 2015). Negr1 has also been implicated in obesity (Lee et al., 2012), autistic behavior, memory deficits, and increased susceptibility to seizures (Singh et al., 2018) in mice, and body mass index (Speliotes et al., 2010), and major depressive disorder (Hyde et al., 2016) in humans. Another example is the gephyrin gene (Gphn) implicated by two distinct FAP groups in the diet-dependent locus on chromosome 12 (Figure 6A). Gephyrin is a key structural protein at neuronal synapses that ensures proper localization of postsynaptic inhibitory receptors. Gephyrin is also known to physically interact with mTOR and is required for mTOR signaling (Sabatini et al., 1999), suggesting two plausible pathways for influencing body weight. On chromosome 1, murine Trmt1l (Trmt1-like), a gene with sequence similarity to orthologous tRNA methyltransferases in other species, was linked to body weight during early adulthood (Figure 5—figure supplement 3A). Mice deficient in this gene, while viable, have been found to exhibit altered motor coordination and abnormal exploratory behavior (Vauti et al., 2007), suggesting that the association at this locus is possibly mediated through modulating exploratory behavior. The association of these three candidates with mouse body weight is consistent with human body mass index and obesity GWAS that found an enrichment of genes active in the central nervous system (Locke et al., 2015; Yengo et al., 2018; Schlauch et al., 2020).

Among candidates that affect metabolic processes, Creb5, a gene linked to diet-independent effects on body weight (Figure 5B), has previously been reported to be linked to metabolic phenotypes in humans, with differential DNA methylation detected between individuals with large differences in waist circumference, hypercholesterolemia, and metabolic syndrome (Salas-Pérez et al., 2019). Another gene important for metabolic control in the liver, Pik3c2g was linked to diet-dependent effects on body weight in mid-adulthood. Pik3c2g-deficient mice are known to exhibit reduced liver accumulation of glycogen and develop hyperlipidemia, adiposity, and insulin resistance with age (Braccini et al., 2015). Edem3, another candidate gene at the locus on chromosome 1 linked to body weight in early adulthood (Figure 5—figure supplement 3A), has previously been linked to short stature in humans based on family-based exome sequencing and differential expression in chondrocytes (Hauer et al., 2019). Possibly sharing a similar mechanism, Adam12, a candidate gene in both a diet-independent and diet-dependent locus on chromosome 7, is known to play an important role in the differentiation, proliferation, and maturation of chondrocytes (Horita et al., 2019).

Future considerations

To summarize, we found that the effects of age and diet on body weight differ substantially with respect to the genetic background and type of dietary intervention imposed. These results highlight that with knowledge of these environment-dependent effects, we can generate more accurate predictions of body weight trajectories than would be possible from knowledge of genotype, age, or diet alone. Moreover, with the elucidation of specific candidate genes and variants underlying these effects, we are not limited to predicting how this complex quantitative trait changes with age, but can also identify specific targets for genetic or pharmacological manipulation in an effort to improve organismal health.

One important limitation to our study is the lack of direct measurements of food consumption and feeding behaviors for each mouse in the population; this makes it difficult to ascertain how much variation in body weight can be ascribed to variation in such behaviors. Additionally, CR was imposed based on the average food consumed by a typical DO mouse rather than a per-mouse baseline of food consumption. Furthermore, CR interventions were imposed on a per-cage basis, not a per-mouse basis, because all animals were housed in groups of eight. Therefore, the social hierarchy within each cage likely contributed to additional variation in body weight (Baud et al., 2017). Additionally, we note that over the course of this study each animal underwent a number of phenotypic assays (Figure 1—figure supplement 1) that would further contribute to study-specific variation in body weight. Accounting for these sources of variation will be a promising avenue for future research, helping interpret many of the associations identified in our study.

A second important consideration for this study is the potential for survival bias to lead to inflated PVE values and false-positive associations at later ages. To evaluate the presence of survival bias, we computed PVE at all ages restricting to animals that survived to 660 days of age (75% of animals in our study). We observed similar PVE values to those from the full dataset, across the full age range, suggesting that survival bias has very little effect on the results presented in this paper (Figure 4—figure supplement 3C, top-left panel). However, as these animals age and the survival bias of the population increases, genetic analyses of body weight at ages past 660 days will need to explicitly account for this effect.

In this study, we have elucidated the dynamics and context dependence of the genetic architecture of body weight from 60 to 660 days of age. By 660 days of age, nearly all surviving animals have realized their maximum adult body weight and the majority of animals have yet to experience appreciable loss of body weight indicative of late-age physiological decline. Under AL diet, reduced body weight has been associated with greater longevity, whereas under 40% CR conditions, greater longevity is associated with the maintenance of high body weight (Liao et al., 2011). Our future research will assess whether we observe a similar result in this DO population and determine whether alleles at body weight loci are predictive of life span in a diet-dependent manner.

Materials and methods

Polygenic models for gene × environment interactions

Request a detailed protocol

G×EMM (Dahl et al., 2020) is a generalization of the standard linear mixed model that allows for polygenic G×E effects. Under this model, the phenotype is written in terms of genetic effects as follows:

(1) Y=α0+e=1EZeαe+v=1VGvβv+e=1Ev=1VGvZeγve+ε,

where YRN is the vector of phenotypes over N samples, Gv,v=1,,V are genotypes of V biallelic SNPs, Ze,e=1,,E are binary vectors over E environments, and ε denotes the residual vector.

In our application, Y is the vector of 1 trend filtered body weights at a specific age; we do not standardize the body weights so that the estimated effects are interpretable and comparable across ages. When testing for association with the founder-of-origin of markers, Gnv[0,1]8;Gnv1=1 is a vector denoting the probability that the two alleles of the marker came from each of the eight founder lines from which the DO population is derived. 1 denotes the L1-norm of a vector. Alternatively, when testing for association with the allele dosage of a variant, Gnv[0,2] is the expected allele count at variant v. Finally, Zne=1 denotes that sample n is subject to environment e. Prior to dietary intervention, the environments in our model are the 12 generations over which the DO samples span (i.e., E=12). After dietary intervention, the environments further include the five diet groups (i.e., E=17).

The effects of covariates, α, are modeled as fixed while the genetic effects, β, and genotype–environment effects, γ, are modeled as random. Assuming heteroscedastic noise, εN(0,Θ), a normal prior on the random genetic effects, βvN(0,ϱ2/V), and a normal prior on the random G×E effects, γvN(0,1VΩ), we get YN(μ,Λ), where μ=α0+eZeαe and Λ=Θ+ϱ2K+e,eΩee(K(ZeZeT)) is a diagonal matrix with entries specified as Θnn=eZne2σe2, K is the kinship matrix with entries defined as Kmn=1VvGmvTGnv, ϱ2 is the variance of environment-independent genetic effects, ΩE×E is the variance–covariance matrix representing the co-variation in environment-dependent genetic effects between pairs of environments, and AB denotes the Hadamard product of matrices A and B. We do not include an environment-independent noise term since it is nonidentifiable.

Off-diagonal elements of Ω quantify the covariance in the interaction effect sizes between pairs of environments; thus, a diagonal Ω assumes that the interaction effect sizes between pairs of environments are uncorrelated. This assumption of diagonal Ω could lead to some bias in the estimates of the variance components, although bias in PVE estimates are expected to be minimal. We computed PVE estimates using a full Ω and a diagonal Ω and observed little difference (Figure 4—figure supplement 5). For the sake of computational simplicity, we assumed all off-diagonal elements of Ω to be zero in this study. Note that each of the above parameters and data vectors in the model may be distinct at different ages.

Recent work has shown that non-negativity constraints on the variance components are both important in order to obtain unbiased estimates and unnecessary to achieve valid estimates of heritability (Steinsaltz et al., 2020). Nevertheless, we constrained the variance components to be non-negative since they are more interpretable in the context of our study. While it is likely that this constraint contributes to the small bias observed in our simulations (Figure 3), we observed almost no bias in heritability estimates for body weight due to the constraint, likely because of strong signals in the data (Figure 4—figure supplement 4). We have also included a flag in our published code to relax this constraint when estimating the variance components.

Proportion of phenotypic variance explained by genetics

Request a detailed protocol

Decomposing the phenotype into genetic and nongenetic effects, Y=YG+Yε, the expected proportion of phenotypic variance explained by genetic effects is approximately given as

(2) PVE=E[V[YG]V[Y]]E[V[YG]]E[V[Y]]VarGVarY,

where V[] denotes the sample variance. The expected sample phenotypic variance conditional on environment e can be written as

(3) E[V[Y|e]]=E[nYn2Zne]Ne-E[(nYnZne)2]Ne2,

where Zne is an indicator variable denoting whether sample n belongs to environment e, and Ne=nZne is the number of samples in environment e. Under the G×EMM model, starting from Equation 1 and integrating out the random effects, we can write the numerator of the first term in the expectation as

(4) E[nYn2Zne]=nμn2Zne+ϱ2Vn,vGnv2Zne+e,e′′Ωee′′Vn,vGnv2ZneZneZne′′+nΘnnZne,

and the numerator of the second term in the expectation as

(5) E[(nYnZne)2]=(nμnZne)2+ϱ2Vvn,nGnvGnvZneZne+e,e′′Ωee′′Vvn,nGnvGnvZn,eZne′′ZneZne+nΘnnZne,

where e and e′′ iterate over all environments and n and n′′ iterate over all samples. Therefore, the expected sample phenotypic variance can be decomposed as follows:

(6) E[V[Y|e]]=(nμn2ZneNe(nμnZne)2Ne2)+ϱ2(tr(KWe)Nesum(KWe)Ne2)+eΩee(tr(KWeWe)Nesum(KWeWe)Ne2)+tr(ΘWe)(Ne1Ne2),

where We=ZeZeT, tr() denotes that trace of a matrix, sum() denotes the sum of all entries of the matrix, and AB denotes the Hadamard product of matrices A and B. The first term quantifies the phenotypic variance explained by fixed effects, the second and third terms together quantify the phenotypic variance explained by genetic effects (E[V[YG|e]]), and the fourth term quantifies the residual (unexplained) phenotypic variance. The total proportion of variance explained by genetics in the entire sample can now be computed using Equation 2.

(7) PVEtot=VarGVarY,
(8) VarG=ϱ2(tr(K)Nsum(K)N2)+eΩee(tr(KWe)Nsum(KWe)N2),
(9) VarY=VarG+((nμn2)N-(nμn)2N2)+tr(Θ)(N-1N2).

The two terms in VarG are genetic contributions to phenotypic variation that are shared across environments and specific to environments, respectively. The second and third terms in VarY are phenotypic variation explained by fixed effects and unexplained residual phenotypic variation, respectively.

The expected total sample phenotypic variance (across all environments) again has two terms, as in Equation 3; the numerator of the first term is written as

(10) E[nYn2]=nμn2+ϱ2VnvGnv2+e,eΩeeVnvGnv2ZneZne+nΘnn,

and the numerator of the second term is written as

(11) E[(nYn)2]=(nμn)2+ϱ2Vvn,nGnvGnv+e,eΩeeVvn,nGnvGnvZneZne+nΘnn.

The expected total sample phenotypic variance can be decomposed as follows:

(12) E[V[Y]]=(nμn2N(nμn)2N2)+ϱ2(tr(K)Nsum(K)N2)+eΩee(tr(KWe)Nsum(KWe)N2)+tr(Θ)(N1N2).

The proportion of variance explained by genetics conditional on environment can be computed by substituting the above in Equation 2.

(13) PVEe=VarG|eVarY|e,
(14) VarG|e=ϱ2(tr(KWe)Nesum(KWe)Ne2)+eΩee(tr(KWeWe)Nesum(KWeWe)Ne2),
(15) VarY|e=VarG|e+((nμn2Zne)Ne-(nμnZne)2Ne2)+tr(ΘWe)(Ne-1Ne2).

The proportion of phenotypic variance explained by genetic effects is equivalent to narrow-sense heritability, once variation due to additive effects of environment, batch, and other study design artifacts have been removed. In this work, we use the more general term, proportion of variance explained, to accommodate variation due to effects of diet and environment.

Under the EMMA model, the expected total sample phenotypic variance simplifies to

(16) E[V[Y]]=(nμn2N(nμn)2N2)+ϱ2(tr(K)Nsum(K)N2)+θ2(N1N2),

where θ denotes the homoscedastic noise. The first component quantifies the phenotypic variance explained by fixed effects, the second component quantifies the phenotypic variance explained by genetic effects (E[V[YG]]), and the third component quantifies the residual (unexplained) phenotypic variance. Substituting these into Equation 2 gives us the proportion of variance explained by genetics under the EMMA model.

Genome-wide association mapping

Additive genetic effects

Request a detailed protocol

To test for additive effect of a genetic variant on the phenotype, we include the focal variant among the fixed effects in the model while treating all other variants to have random effects.

(17) Y=cXcαc+ϕsGs+vsGvβv+v,eGvZeγve+ε

Applying the priors described above for βv, γv, and ε, we can derive the corresponding mixed-effects model is as follows:

(18) YN(cXcαc+ϕsGs,Λs),

where Λs=Θ+ϱ2Ks+eΩee(KsWe) and Ks is the kinship matrix after excluding the entire chromosome containing the variant s (leave-one-chromosome-out [LOCO] kinship). Leaving out the focal chromosome when computing kinship increases our power to detect associations at the focal variant (Lippert et al., 2011). The test statistic is the log likelihood ratio Φa(Y,Gs) comparing the alternate model :ϕs0 to the null model 0:ϕs=0.

(19) Φa(Y,Gs)=logmax(ϕs,α,σ,ϱ,Ω)max(ϕs=0,α,σ,ϱ,Ω)

Genotype–environment effects

Request a detailed protocol

To test for effects of interaction between genotype and environment on the phenotype, we include a fixed effect for the focal variant and its interactions with the set of all environments of interest () while treating all other variants to have random effects for their additive and interaction contributions:

(20) Y=cXcαc+ϕsGs+eχseGsZe+vsGvβv+vs,eGvZeγve+ε.

The corresponding mixed-effects model is

(21) YN(cαcXc+ϕsGs+eEχseGsZe,Λs),

and the test statistic is:

(22) Φi(Y,Gs)=logmaxL(χs,ϕs,α,σ,ϱ,Ω)maxL(χs=0,ϕs,α,σ,ϱ,Ω),

where Xs is a vector notation for all χse. Note that this statistic tests for the presence of interaction effects between the focal variant and any of the environments of interest.

Likelihood and gradients for G×EMM model

Request a detailed protocol

Under the full G×EMM model, the phenotype Y depends on fixed and random effects as follows:

(23) YN(μ,Λ),

where μ=α0+eZeαe captures all fixed effects and Λ=Θ+ϱ2K+eΩeeK(ZeZeT) captures the covariance after integrating out the random effects. The parameters to be estimated in this model are α, σ, ϱ, and Ω. The complete log likelihood can be written as

(24) =-N2log(2π)-12logdetΛ-12(Y-Xα)TΛ-1(Y-Xα).

Maximizing the log likelihood over α gives us

(25) α^=argmaxα=(XTΛ-1X)-1(XTΛ-1Y).

Substituting this into the log likelihood, we get

(26) Lα^=N2log(2π)12logdetΛ12YTPY,

where P=Λ1Λ1X(XTΛ1X)1XTΛ1.

Computing the gradient of α^ involves evaluating the following gradients:

(27) Λσe2=IZe
(28) Λϱ2=K
(29) ΛΩee=K(ZeZeT)
(30) logdetΛ=tr(Λ1Λ)
(31) Λ1=Λ1ΛΛ1
(32) P=PΛP,

where Ze is a diagonal matrix with elements of the vector Ze on the diagonal. Thus,

(33) Lα^σe2=12tr(Λ1IZe)+12YTPIZePY
(34) Lα^ϱ2=12tr(Λ1K)+12YTPKPY
(35) Lα^Ωee=12tr(Λ1Ke)+12YTPKePY,

where Ke=K(ZeZeT).

Permutations

Request a detailed protocol

Since the assumptions underlying the G×EMM model are unlikely to hold exactly in practice, we used permutations to assess the significance of the observed values of Φa and Φi. Specifically, for each variant s, we generated M independent random permutations μ1,,μM of (1,,N) and computed test statistics Φma/iΦa/i(μm(Y),Gs) for each permutation μm. The p-value associated with Φa and Φi is then given as

(36) pa/i=#{m:Φma/iΦa/i}+1M+1

We used a sequential procedure (Besag and Clifford, 1991; Shim and Stephens, 2015) to compute the p-value that helps avoid large numbers of permutations for nonsignificant results, substantially speeding up computation. Specifically, instead of a fixed large number of permutations M, we performed Ms permutations for each variant s until Csa/i#{m:Λma/iΛa/i}=10 or Ms=108. The p-value for the sth variant is then a random sample from the interval [CsMs,Cs+1Ms+1]. This scheme allows us to estimate p-value quickly without carrying out large numbers of permutations at every marker. Although the precision of this estimate is lower for markers that are not significant (i.e., most of the genome), precision increases as the significance of association increases.

Tests for nonlinearity in trends of effect size with age

Request a detailed protocol

We test for nonlinear dependence between effect size and age by comparing two models: a null model that only allows for linear dependence between age and effect size and an alternate model that allows for dependence between effect size and higher-order powers of age. While this test can be applied to evaluate nonlinear effect-size trends for any covariate (e.g., dietary intervention, genotype, etc.), we focus our attention to tests for nonlinear trends in the effect sizes of the founder alleles and the effect sizes of the interaction between founder alleles and diet.

Specifically, the estimated effects ϕ^st and their standard errors ψ^st can be modeled as

(37) ϕ^stN(ϕst,ψ^st);t{40,50,,660}

and the two models of the effect size trends are

(38) Hn:ϕst=φ0+φ1t+φ2t2+φ3t3H:ϕst=φ0+φ1t,

where t denotes the ages (in days) at which we test for association between a genetic variant and body weight and ϕst the unobserved effect sizes at these ages. We used a likelihood ratio test statistic () to compare the two models and quantify the evidence for nonlinearity in effect size trends with age and assessed significance using p-values computed using a chi-squared distribution with degree-of-freedom 2.

(39) =2lnmaxL(φ)maxL(φ;φ2=φ3=0),

where (φ) is the likelihood under the model in Equation 37.

We apply this test for each of the founder effects in each of the 14 diet-independent loci (112 tests) and each of the 19 founder-x-diet interaction effects in each of the diet-dependent loci (760 tests). We assign a diet-independent trend as significantly nonlinear if we observe a p<104 and assign a diet-dependent trend as significantly nonlinear if we observe a p<105 (Bonferroni-corrected p-value thresholds).

Data and software release

Request a detailed protocol

Data used in this study, and significant QTLs, can be downloaded from the following links:

Our Python implementation of the EMMA and G×EMM models can be downloaded from https://github.com/calico/do_qtl; (Raj, 2021; copy archived at swh:1:rev:3d23d3fbd7768b74b0bba6183de1570a65a2d93d).

Data availability

All data used in the study, and results of analyses, are publicly available at https://figshare.com/projects/Age_and_diet_shape_the_genetic_architecture_of_body_weight_in_Diversity_Outbred_mice/92201.

The following data sets were generated
    1. Wright KM
    2. Deighan A
    3. Di Francesco A
    4. Freund A
    5. Jojic V
    6. Churchill G
    7. Raj A
    (2020) figshare
    Genotype data: Age and diet shape the genetic architecture of body weight in Diversity Outbred mice.
    https://doi.org/10.6084/m9.figshare.13190735
    1. Wright KM
    2. Deighan A
    3. Di Francesco A
    4. Freund A
    5. Jojic V
    6. Churchill G
    7. Raj A
    (2020) figshare
    Phenotype data: Age and diet shape the genetic architecture of body weight in Diversity Outbred mice.
    https://doi.org/10.6084/m9.figshare.13190702
    1. Wright KM
    2. Deighan A
    3. Di Francesco A
    4. Freund A
    5. Jojic V
    6. Churchill G
    7. Raj A
    (2020) figshare
    Results: Age and diet shape the genetic architecture of body weight in Diversity Outbred mice.
    https://doi.org/10.6084/m9.figshare.13190708
    1. Wright KM
    2. Deighan A
    3. Di Francesco A
    4. Freund A
    5. Jojic V
    6. Churchill G
    7. Raj A
    (2020) figshare
    Covariate data: Age and diet shape the genetic architecture of body weight in Diversity Outbred mice.
    https://doi.org/10.6084/m9.figshare.13190615

References

    1. Locke AE
    2. Kahali B
    3. Berndt SI
    4. Justice AE
    5. Pers TH
    6. Day FR
    7. Powell C
    8. Vedantam S
    9. Buchkovich ML
    10. Yang J
    11. Croteau-Chonka DC
    12. Esko T
    13. Fall T
    14. Ferreira T
    15. Gustafsson S
    16. Kutalik Z
    17. Luan J
    18. Mägi R
    19. Randall JC
    20. Winkler TW
    21. Wood AR
    22. Workalemahu T
    23. Faul JD
    24. Smith JA
    25. Zhao JH
    26. Zhao W
    27. Chen J
    28. Fehrmann R
    29. Hedman ÅK
    30. Karjalainen J
    31. Schmidt EM
    32. Absher D
    33. Amin N
    34. Anderson D
    35. Beekman M
    36. Bolton JL
    37. Bragg-Gresham JL
    38. Buyske S
    39. Demirkan A
    40. Deng G
    41. Ehret GB
    42. Feenstra B
    43. Feitosa MF
    44. Fischer K
    45. Goel A
    46. Gong J
    47. Jackson AU
    48. Kanoni S
    49. Kleber ME
    50. Kristiansson K
    51. Lim U
    52. Lotay V
    53. Mangino M
    54. Leach IM
    55. Medina-Gomez C
    56. Medland SE
    57. Nalls MA
    58. Palmer CD
    59. Pasko D
    60. Pechlivanis S
    61. Peters MJ
    62. Prokopenko I
    63. Shungin D
    64. Stančáková A
    65. Strawbridge RJ
    66. Sung YJ
    67. Tanaka T
    68. Teumer A
    69. Trompet S
    70. van der Laan SW
    71. van Setten J
    72. Van Vliet-Ostaptchouk JV
    73. Wang Z
    74. Yengo L
    75. Zhang W
    76. Isaacs A
    77. Albrecht E
    78. Ärnlöv J
    79. Arscott GM
    80. Attwood AP
    81. Bandinelli S
    82. Barrett A
    83. Bas IN
    84. Bellis C
    85. Bennett AJ
    86. Berne C
    87. Blagieva R
    88. Blüher M
    89. Böhringer S
    90. Bonnycastle LL
    91. Böttcher Y
    92. Boyd HA
    93. Bruinenberg M
    94. Caspersen IH
    95. Chen Y-DI
    96. Clarke R
    97. Daw EW
    98. de Craen AJM
    99. Delgado G
    100. Dimitriou M
    101. Doney ASF
    102. Eklund N
    103. Estrada K
    104. Eury E
    105. Folkersen L
    106. Fraser RM
    107. Garcia ME
    108. Geller F
    109. Giedraitis V
    110. Gigante B
    111. Go AS
    112. Golay A
    113. Goodall AH
    114. Gordon SD
    115. Gorski M
    116. Grabe H-J
    117. Grallert H
    118. Grammer TB
    119. Gräßler J
    120. Grönberg H
    121. Groves CJ
    122. Gusto G
    123. Haessler J
    124. Hall P
    125. Haller T
    126. Hallmans G
    127. Hartman CA
    128. Hassinen M
    129. Hayward C
    130. Heard-Costa NL
    131. Helmer Q
    132. Hengstenberg C
    133. Holmen O
    134. Hottenga J-J
    135. James AL
    136. Jeff JM
    137. Johansson Å
    138. Jolley J
    139. Juliusdottir T
    140. Kinnunen L
    141. Koenig W
    142. Koskenvuo M
    143. Kratzer W
    144. Laitinen J
    145. Lamina C
    146. Leander K
    147. Lee NR
    148. Lichtner P
    149. Lind L
    150. Lindström J
    151. Lo KS
    152. Lobbens S
    153. Lorbeer R
    154. Lu Y
    155. Mach F
    156. Magnusson PKE
    157. Mahajan A
    158. McArdle WL
    159. McLachlan S
    160. Menni C
    161. Merger S
    162. Mihailov E
    163. Milani L
    164. Moayyeri A
    165. Monda KL
    166. Morken MA
    167. Mulas A
    168. Müller G
    169. Müller-Nurasyid M
    170. Musk AW
    171. Nagaraja R
    172. Nöthen MM
    173. Nolte IM
    174. Pilz S
    175. Rayner NW
    176. Renstrom F
    177. Rettig R
    178. Ried JS
    179. Ripke S
    180. Robertson NR
    181. Rose LM
    182. Sanna S
    183. Scharnagl H
    184. Scholtens S
    185. Schumacher FR
    186. Scott WR
    187. Seufferlein T
    188. Shi J
    189. Smith AV
    190. Smolonska J
    191. Stanton AV
    192. Steinthorsdottir V
    193. Stirrups K
    194. Stringham HM
    195. Sundström J
    196. Swertz MA
    197. Swift AJ
    198. Syvänen A-C
    199. Tan S-T
    200. Tayo BO
    201. Thorand B
    202. Thorleifsson G
    203. Tyrer JP
    204. Uh H-W
    205. Vandenput L
    206. Verhulst FC
    207. Vermeulen SH
    208. Verweij N
    209. Vonk JM
    210. Waite LL
    211. Warren HR
    212. Waterworth D
    213. Weedon MN
    214. Wilkens LR
    215. Willenborg C
    216. Wilsgaard T
    217. Wojczynski MK
    218. Wong A
    219. Wright AF
    220. Zhang Q
    221. Brennan EP
    222. Choi M
    223. Dastani Z
    224. Drong AW
    225. Eriksson P
    226. Franco-Cereceda A
    227. Gådin JR
    228. Gharavi AG
    229. Goddard ME
    230. Handsaker RE
    231. Huang J
    232. Karpe F
    233. Kathiresan S
    234. Keildson S
    235. Kiryluk K
    236. Kubo M
    237. Lee J-Y
    238. Liang L
    239. Lifton RP
    240. Ma B
    241. McCarroll SA
    242. McKnight AJ
    243. Min JL
    244. Moffatt MF
    245. Montgomery GW
    246. Murabito JM
    247. Nicholson G
    248. Nyholt DR
    249. Okada Y
    250. Perry JRB
    251. Dorajoo R
    252. Reinmaa E
    253. Salem RM
    254. Sandholm N
    255. Scott RA
    256. Stolk L
    257. Takahashi A
    258. Tanaka T
    259. van ’t Hooft FM
    260. Vinkhuyzen AAE
    261. Westra H-J
    262. Zheng W
    263. Zondervan KT
    264. Heath AC
    265. Arveiler D
    266. Bakker SJL
    267. Beilby J
    268. Bergman RN
    269. Blangero J
    270. Bovet P
    271. Campbell H
    272. Caulfield MJ
    273. Cesana G
    274. Chakravarti A
    275. Chasman DI
    276. Chines PS
    277. Collins FS
    278. Crawford DC
    279. Cupples LA
    280. Cusi D
    281. Danesh J
    282. de Faire U
    283. den Ruijter HM
    284. Dominiczak AF
    285. Erbel R
    286. Erdmann J
    287. Eriksson JG
    288. Farrall M
    289. Felix SB
    290. Ferrannini E
    291. Ferrières J
    292. Ford I
    293. Forouhi NG
    294. Forrester T
    295. Franco OH
    296. Gansevoort RT
    297. Gejman PV
    298. Gieger C
    299. Gottesman O
    300. Gudnason V
    301. Gyllensten U
    302. Hall AS
    303. Harris TB
    304. Hattersley AT
    305. Hicks AA
    306. Hindorff LA
    307. Hingorani AD
    308. Hofman A
    309. Homuth G
    310. Hovingh GK
    311. Humphries SE
    312. Hunt SC
    313. Hyppönen E
    314. Illig T
    315. Jacobs KB
    316. Jarvelin M-R
    317. Jöckel K-H
    318. Johansen B
    319. Jousilahti P
    320. Jukema JW
    321. Jula AM
    322. Kaprio J
    323. Kastelein JJP
    324. Keinanen-Kiukaanniemi SM
    325. Kiemeney LA
    326. Knekt P
    327. Kooner JS
    328. Kooperberg C
    329. Kovacs P
    330. Kraja AT
    331. Kumari M
    332. Kuusisto J
    333. Lakka TA
    334. Langenberg C
    335. Marchand LL
    336. Lehtimäki T
    337. Lyssenko V
    338. Männistö S
    339. Marette A
    340. Matise TC
    341. McKenzie CA
    342. McKnight B
    343. Moll FL
    344. Morris AD
    345. Morris AP
    346. Murray JC
    347. Nelis M
    348. Ohlsson C
    349. Oldehinkel AJ
    350. Ong KK
    351. Madden PAF
    352. Pasterkamp G
    353. Peden JF
    354. Peters A
    355. Postma DS
    356. Pramstaller PP
    357. Price JF
    358. Qi L
    359. Raitakari OT
    360. Rankinen T
    361. Rao DC
    362. Rice TK
    363. Ridker PM
    364. Rioux JD
    365. Ritchie MD
    366. Rudan I
    367. Salomaa V
    368. Samani NJ
    369. Saramies J
    370. Sarzynski MA
    371. Schunkert H
    372. Schwarz PEH
    373. Sever P
    374. Shuldiner AR
    375. Sinisalo J
    376. Stolk RP
    377. Strauch K
    378. Tönjes A
    379. Trégouët D-A
    380. Tremblay A
    381. Tremoli E
    382. Virtamo J
    383. Vohl M-C
    384. Völker U
    385. Waeber G
    386. Willemsen G
    387. Witteman JC
    388. Zillikens MC
    389. Adair LS
    390. Amouyel P
    391. Asselbergs FW
    392. Assimes TL
    393. Bochud M
    394. Boehm BO
    395. Boerwinkle E
    396. Bornstein SR
    397. Bottinger EP
    398. Bouchard C
    399. Cauchi S
    400. Chambers JC
    401. Chanock SJ
    402. Cooper RS
    403. de Bakker PIW
    404. Dedoussis G
    405. Ferrucci L
    406. Franks PW
    407. Froguel P
    408. Groop LC
    409. Haiman CA
    410. Hamsten A
    411. Hui J
    412. Hunter DJ
    413. Hveem K
    414. Kaplan RC
    415. Kivimaki M
    416. Kuh D
    417. Laakso M
    418. Liu Y
    419. Martin NG
    420. März W
    421. Melbye M
    422. Metspalu A
    423. Moebus S
    424. Munroe PB
    425. Njølstad I
    426. Oostra BA
    427. Palmer CNA
    428. Pedersen NL
    429. Perola M
    430. Pérusse L
    431. Peters U
    432. Power C
    433. Quertermous T
    434. Rauramaa R
    435. Rivadeneira F
    436. Saaristo TE
    437. Saleheen D
    438. Sattar N
    439. Schadt EE
    440. Schlessinger D
    441. Slagboom PE
    442. Snieder H
    443. Spector TD
    444. Thorsteinsdottir U
    445. Stumvoll M
    446. Tuomilehto J
    447. Uitterlinden AG
    448. Uusitupa M
    449. van der Harst P
    450. Walker M
    451. Wallaschofski H
    452. Wareham NJ
    453. Watkins H
    454. Weir DR
    455. Wichmann H-E
    456. Wilson JF
    457. Zanen P
    458. Borecki IB
    459. Deloukas P
    460. Fox CS
    461. Heid IM
    462. O’Connell JR
    463. Strachan DP
    464. Stefansson K
    465. van Duijn CM
    466. Abecasis GR
    467. Franke L
    468. Frayling TM
    469. McCarthy MI
    470. Visscher PM
    471. Scherag A
    472. Willer CJ
    473. Boehnke M
    474. Mohlke KL
    475. Lindgren CM
    476. Beckmann JS
    477. Barroso I
    478. North KE
    479. Ingelsson E
    480. Hirschhorn JN
    481. Loos RJF
    482. Speliotes EK
    483. LifeLines Cohort Study
    484. ADIPOGen Consortium
    485. AGEN-BMI Working Group
    486. CARDIOGRAMplusC4D Consortium
    487. CKDGen Consortium
    488. GLGC
    489. ICBP
    490. MAGIC Investigators
    491. MuTHER Consortium
    492. MIGen Consortium
    493. PAGE Consortium
    494. ReproGen Consortium
    495. GENIE Consortium
    496. International Endogene Consortium
    (2015) Genetic studies of body mass index yield new insights for obesity biology
    Nature 518:197–206.
    https://doi.org/10.1038/nature14177
  1. Book
    1. Singh A
    (2013) Allele Heterogeneity
    In: Gellman MD, Turner JR, editors. Encyclopedia of Behavioral Medicine. New York: Springer. pp. 66–67.
    https://doi.org/10.1007/978-1-4419-1005-9_677
    1. Speliotes EK
    2. Willer CJ
    3. Berndt SI
    4. Monda KL
    5. Thorleifsson G
    6. Jackson AU
    7. Lango Allen H
    8. Lindgren CM
    9. Luan J
    10. Mägi R
    11. Randall JC
    12. Vedantam S
    13. Winkler TW
    14. Qi L
    15. Workalemahu T
    16. Heid IM
    17. Steinthorsdottir V
    18. Stringham HM
    19. Weedon MN
    20. Wheeler E
    21. Wood AR
    22. Ferreira T
    23. Weyant RJ
    24. Segrè AV
    25. Estrada K
    26. Liang L
    27. Nemesh J
    28. Park J-H
    29. Gustafsson S
    30. Kilpeläinen TO
    31. Yang J
    32. Bouatia-Naji N
    33. Esko T
    34. Feitosa MF
    35. Kutalik Z
    36. Mangino M
    37. Raychaudhuri S
    38. Scherag A
    39. Smith AV
    40. Welch R
    41. Zhao JH
    42. Aben KK
    43. Absher DM
    44. Amin N
    45. Dixon AL
    46. Fisher E
    47. Glazer NL
    48. Goddard ME
    49. Heard-Costa NL
    50. Hoesel V
    51. Hottenga J-J
    52. Johansson A
    53. Johnson T
    54. Ketkar S
    55. Lamina C
    56. Li S
    57. Moffatt MF
    58. Myers RH
    59. Narisu N
    60. Perry JRB
    61. Peters MJ
    62. Preuss M
    63. Ripatti S
    64. Rivadeneira F
    65. Sandholt C
    66. Scott LJ
    67. Timpson NJ
    68. Tyrer JP
    69. van Wingerden S
    70. Watanabe RM
    71. White CC
    72. Wiklund F
    73. Barlassina C
    74. Chasman DI
    75. Cooper MN
    76. Jansson J-O
    77. Lawrence RW
    78. Pellikka N
    79. Prokopenko I
    80. Shi J
    81. Thiering E
    82. Alavere H
    83. Alibrandi MTS
    84. Almgren P
    85. Arnold AM
    86. Aspelund T
    87. Atwood LD
    88. Balkau B
    89. Balmforth AJ
    90. Bennett AJ
    91. Ben-Shlomo Y
    92. Bergman RN
    93. Bergmann S
    94. Biebermann H
    95. Blakemore AIF
    96. Boes T
    97. Bonnycastle LL
    98. Bornstein SR
    99. Brown MJ
    100. Buchanan TA
    101. Busonero F
    102. Campbell H
    103. Cappuccio FP
    104. Cavalcanti-Proença C
    105. Chen Y-DI
    106. Chen C-M
    107. Chines PS
    108. Clarke R
    109. Coin L
    110. Connell J
    111. Day INM
    112. den Heijer M
    113. Duan J
    114. Ebrahim S
    115. Elliott P
    116. Elosua R
    117. Eiriksdottir G
    118. Erdos MR
    119. Eriksson JG
    120. Facheris MF
    121. Felix SB
    122. Fischer-Posovszky P
    123. Folsom AR
    124. Friedrich N
    125. Freimer NB
    126. Fu M
    127. Gaget S
    128. Gejman PV
    129. Geus EJC
    130. Gieger C
    131. Gjesing AP
    132. Goel A
    133. Goyette P
    134. Grallert H
    135. Grässler J
    136. Greenawalt DM
    137. Groves CJ
    138. Gudnason V
    139. Guiducci C
    140. Hartikainen A-L
    141. Hassanali N
    142. Hall AS
    143. Havulinna AS
    144. Hayward C
    145. Heath AC
    146. Hengstenberg C
    147. Hicks AA
    148. Hinney A
    149. Hofman A
    150. Homuth G
    151. Hui J
    152. Igl W
    153. Iribarren C
    154. Isomaa B
    155. Jacobs KB
    156. Jarick I
    157. Jewell E
    158. John U
    159. Jørgensen T
    160. Jousilahti P
    161. Jula A
    162. Kaakinen M
    163. Kajantie E
    164. Kaplan LM
    165. Kathiresan S
    166. Kettunen J
    167. Kinnunen L
    168. Knowles JW
    169. Kolcic I
    170. König IR
    171. Koskinen S
    172. Kovacs P
    173. Kuusisto J
    174. Kraft P
    175. Kvaløy K
    176. Laitinen J
    177. Lantieri O
    178. Lanzani C
    179. Launer LJ
    180. Lecoeur C
    181. Lehtimäki T
    182. Lettre G
    183. Liu J
    184. Lokki M-L
    185. Lorentzon M
    186. Luben RN
    187. Ludwig B
    188. MAGIC
    189. Manunta P
    190. Marek D
    191. Marre M
    192. Martin NG
    193. McArdle WL
    194. McCarthy A
    195. McKnight B
    196. Meitinger T
    197. Melander O
    198. Meyre D
    199. Midthjell K
    200. Montgomery GW
    201. Morken MA
    202. Morris AP
    203. Mulic R
    204. Ngwa JS
    205. Nelis M
    206. Neville MJ
    207. Nyholt DR
    208. O’Donnell CJ
    209. O’Rahilly S
    210. Ong KK
    211. Oostra B
    212. Paré G
    213. Parker AN
    214. Perola M
    215. Pichler I
    216. Pietiläinen KH
    217. Platou CGP
    218. Polasek O
    219. Pouta A
    220. Rafelt S
    221. Raitakari O
    222. Rayner NW
    223. Ridderstråle M
    224. Rief W
    225. Ruokonen A
    226. Robertson NR
    227. Rzehak P
    228. Salomaa V
    229. Sanders AR
    230. Sandhu MS
    231. Sanna S
    232. Saramies J
    233. Savolainen MJ
    234. Scherag S
    235. Schipf S
    236. Schreiber S
    237. Schunkert H
    238. Silander K
    239. Sinisalo J
    240. Siscovick DS
    241. Smit JH
    242. Soranzo N
    243. Sovio U
    244. Stephens J
    245. Surakka I
    246. Swift AJ
    247. Tammesoo M-L
    248. Tardif J-C
    249. Teder-Laving M
    250. Teslovich TM
    251. Thompson JR
    252. Thomson B
    253. Tönjes A
    254. Tuomi T
    255. van Meurs JBJ
    256. van Ommen G-J
    257. Vatin V
    258. Viikari J
    259. Visvikis-Siest S
    260. Vitart V
    261. Vogel CIG
    262. Voight BF
    263. Waite LL
    264. Wallaschofski H
    265. Walters GB
    266. Widen E
    267. Wiegand S
    268. Wild SH
    269. Willemsen G
    270. Witte DR
    271. Witteman JC
    272. Xu J
    273. Zhang Q
    274. Zgaga L
    275. Ziegler A
    276. Zitting P
    277. Beilby JP
    278. Farooqi IS
    279. Hebebrand J
    280. Huikuri HV
    281. James AL
    282. Kähönen M
    283. Levinson DF
    284. Macciardi F
    285. Nieminen MS
    286. Ohlsson C
    287. Palmer LJ
    288. Ridker PM
    289. Stumvoll M
    290. Beckmann JS
    291. Boeing H
    292. Boerwinkle E
    293. Boomsma DI
    294. Caulfield MJ
    295. Chanock SJ
    296. Collins FS
    297. Cupples LA
    298. Smith GD
    299. Erdmann J
    300. Froguel P
    301. Grönberg H
    302. Gyllensten U
    303. Hall P
    304. Hansen T
    305. Harris TB
    306. Hattersley AT
    307. Hayes RB
    308. Heinrich J
    309. Hu FB
    310. Hveem K
    311. Illig T
    312. Jarvelin M-R
    313. Kaprio J
    314. Karpe F
    315. Khaw K-T
    316. Kiemeney LA
    317. Krude H
    318. Laakso M
    319. Lawlor DA
    320. Metspalu A
    321. Munroe PB
    322. Ouwehand WH
    323. Pedersen O
    324. Penninx BW
    325. Peters A
    326. Pramstaller PP
    327. Quertermous T
    328. Reinehr T
    329. Rissanen A
    330. Rudan I
    331. Samani NJ
    332. Schwarz PEH
    333. Shuldiner AR
    334. Spector TD
    335. Tuomilehto J
    336. Uda M
    337. Uitterlinden A
    338. Valle TT
    339. Wabitsch M
    340. Waeber G
    341. Wareham NJ
    342. Watkins H
    343. Procardis Consortium
    344. Wilson JF
    345. Wright AF
    346. Zillikens MC
    347. Chatterjee N
    348. McCarroll SA
    349. Purcell S
    350. Schadt EE
    351. Visscher PM
    352. Assimes TL
    353. Borecki IB
    354. Deloukas P
    355. Fox CS
    356. Groop LC
    357. Haritunians T
    358. Hunter DJ
    359. Kaplan RC
    360. Mohlke KL
    361. O’Connell JR
    362. Peltonen L
    363. Schlessinger D
    364. Strachan DP
    365. van Duijn CM
    366. Wichmann H-E
    367. Frayling TM
    368. Thorsteinsdottir U
    369. Abecasis GR
    370. Barroso I
    371. Boehnke M
    372. Stefansson K
    373. North KE
    374. McCarthy MI
    375. Hirschhorn JN
    376. Ingelsson E
    377. Loos RJF
    (2010) Association analyses of 249,796 individuals reveal 18 new loci associated with body mass index
    Nature Genetics 42:937–948.
    https://doi.org/10.1038/ng.686

Decision letter

  1. Sara Hägg
    Reviewing Editor; Karolinska Institutet, Sweden
  2. Naama Barkai
    Senior Editor; Weizmann Institute of Science, Israel
  3. Andrew Dahl
    Reviewer; University of Chicago, United States
  4. Amelie Boud
    Reviewer; Centre for Genomic Regulation, Spain

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Decision letter after peer review:

Thank you for submitting your article "Age and diet shape the genetic architecture of body weight in Diversity Outbred mice" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Matt Kaeberlein as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Andrew Dahl (Reviewer #2); Amelie Baud (Reviewer #3).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

Wright et al., have empirically studied the subtleties of the genetic architecture of a classic quantitative trait, body weight, providing evidence of gene-by-age and gene-by-diet effects, both at the level of the entire genome and at the level of individual genetic loci. The authors also identified likely causal variants at individual loci. To do so, they used a mouse population descended from multiple, known founders and reconstructed the founder haplotypes at each locus, and also used genomic annotations. Finally, the authors also explored pleiotropy, allelic heterogeneity and locus heterogeneity at the associated loci.

This is an outstanding dissection of the genetic architecture of a complex trait at the genome-wide level, at the locus level, across time and across environments. The experimental design is excellent, with the use of a multiparental mouse population that also permits high-resolution mapping. The statistical analyses are good and advanced, leveraging the GxEMM model recently published by Dahl et al., as well as tools developed specifically for this mouse population to trace back the origin of each QTL to the progenitors of the population; the corresponding results are presented in very informative figures. Finally, the paper is well structured and written, however, additional clarifications, justifications and possibly small additional analyses are requested to make the paper a better fit to the eLife community.

Major Comments:

1. Please justify the p-value thresholds:

– Lines 306-7 say that a 10fold weaker threshold is used for the interaction test--why? Also, where do these thresholds come from?

– Is it sound to take this two-step approach of first finding loci and then 'fine mapping' variants nearby? It is definitely fine if the thresholds are valid genome-wide--but then why not test genome-wide?

2. Please eliminate nonnegativity constraints on VCs in software:

– Are you truncating estimates to be nonnegative? If so, this will cause bias, which I believe is visible in your simulations. This is essential to fix for testing, and also for unbiased VC estimates.

– Users will test random effects, so it is necessary to validate (or remove) this part of your software.

– I don't see an assessment of tests or ses for the VCs. A quick supp figure corresponding to the current simulations would suffice, I think.

3. Please assess standard errors/false positive rates/power for testing VCs:

– Are there evidence that GxEMM improves power for fixed effect interaction tests? I think just commenting on this is enough, but it does seem important if you want others to use your approach. Comparing power (or some measure of false positives) for EMMA vs GxEMM in real data and/or simulation would work.

4. Mouse phenotyping:

– the text needs to say upfront that the mice included in this study were subject to a battery of phenotyping beyond weighing, including metabolic cage phenotyping, blood drwas, and "challenge-based phenotyping procedures". Currently this information comes late in the text (L287), and the title of Figure 1 Supplement 1 reads "Raw phenotype measurements", which does not alert the reader to these additional phenotypes. Some details about this phenotyping are needed, in particular how many blood draws were taken and which specific procedures are included under "challenge-based phenotyping".

This extra phenotyping means that the body weight results presented in this study could be study-specific. In particular, the decrease in PVE observed with age in this study, which contrasts with increased heritability with age observed in previous studies (see Introduction), could be the result of increased noise due to extra phenotyping. The implications of this extra phenotyping need to be discussed.

Importantly, it does not mean that this study is any less relevant than mouse studies without challenging life events, humans also experience challenging life events

5. Clarifications on Figure 4:

– What do PVE_tot and PVE_e represent (more on this below)? In particular, L291 "total genetic variance for the 40% CR group" is non-sensical to me because I thought PVE_tot was across all mice – as suggested by the fact that there is a single dark grey line in Figure 4 showing PVE_tot. Also, I was expecting PVE_tot (dark grey line in Figure 4) to be the sum of non diet-dependent genetic effects + all 5 PVE_e (colored lines). Yet the dark grey line is below the coloured lines. Hence I don't have the right intuition for what PVE_tot and PVE_e are. Looking at the equations didn't help: shouldn't it be PVE_e in Equation (7) and PVE_tot in Equation (13)?

Clarifications are needed.

– Figure 4: I love this figure, but worry a bit about the role of GxGeneration:

– The legend gave me a strong (and wrong) sense that the GxE was all about diet. If possible, disentangling the two GxE contributions visually would be helpful.

– I don't understand why generation is included (eg you could instead include cage effects, or neither).

– Why not have an apples-to-apples comparison using GxGeneration as the baseline for questions about GxDiet?

– What is dark grey? The hom estimate from GxEMM? The hom+GxGeneration (but then shouldn't there be one estimate for each generation)?

– More generally, I don't follow how generation is being used in this paper. Is it just always in the background as a random effect, and if so, why? Why don't you test for locus-x-generation effects--if they're not expected to be interesting, why do they explain so much variance? If they are absorbing confounding, why not also adjust for cage (perhaps with the GxEMM 'IID' model to save degrees of freedom)? I suspect whatever you're doing is reasonable, but maybe just explain a bit more clearly.

Figure 4—figure supplement 2, (b), right – isn't it concerning that the permuted phenotypes are heritable? How can this be? I recognize you don't use this kinship in your real analyses (though believe you do use the genotypes from which the kinship is built), but some explanation here would increase confidence in the overall approach.

I'd like to see versions of Figure 4 that: overlay the interventions, as in supplement Figure 1; only show the E-specific genetic variances (ie substracting the sig2g_hom terms); and show the heritabilities (ie normalize by total variance at each time point). These would make nice supplement Figures if easy to generate.

6. GWAS-related comments:

– The GWAS section and the corresponding figures are overwhelming. I think it would benefit from shortening and using sub-headings. I think it would be good to show first how likely causal variants can be identified (including FAP groups, functional annotations etc. and have one figure for that), with the chr6 and chr 12 loci as examples, and THEN show these loci across age and diets to discuss pleiotropy and heterogeneity.

– except if we consider triallelic SNPs, the distinction between allelic heterogeneity ("a single locus harboring multiple functional alleles each with distinct phenotypic effects") and (what I call locus heterogeneity) "a single genomic region contains multiple functional body weight loci that are only revealed with sufficient fine-mapping resolution" (L346-349) is difficult to make because noise can blur genetic associations, making it impossible to know the exact position of the causal variant(s). Consistent with the difficulty I just mentioned, in the section L419-443 and corresponding Supplementary Figures, I find that the assignments of the loci to either allelic or locus heterogeneity not obvious/robust.

I think it would be best to focus on distinguishing between pleiotropy (same causal variant(s) affecting body weight at different ages or in different diets) and non-pleiotropy (be it allelic or locus heterogeneity), rather than trying to distinguish between allelic and locus heterogeneity.

Statistical tests exist to test the null hypothesis {same locus affects both traits} in multiparental populations (doi.org/10.1534/g3.119.400098). In addition, looking at the FAPs of the lead variants, as the authors do here, would be a great way to complement this statistical test. Statistical evidence of different loci + differing top FAPS would provide strong evidence for some form of heterogeneity.

If you are going to focus on pleiotropy yes/no, it would be good to show a locus with likely pleiotropic effects (with the caveat that this would be the null hypothesis so not really evidence, rather suggestive).

An idea now, not a request: It would be interesting to know whether the loci for which different causal variants are predicted to affect different traits (e.g. different ages) are also loci for which multiple causal variants are predicted to affect either trait (see doi:10.1038/ng.2644).

7. Gene enrichment:

– is there a significant enrichment in genes affecting neurological behavior in this study, to support the claim that the fine-mapped genes implicate neurological and metabolic processes? If not, probably best to move the corresponding section of the text to the discussion. Also discuss whether the associations between neurological genes and body weight could arise due to the stressful extra phenotyping of the mice (are those associations mostly for ages after the challenge-based phenotyping for example, or are they seen even for ages before that stressful phenotyping?)

[Editors’ note: further revisions were suggested prior to acceptance, as described below.]

Thank you for submitting a revised version of your article "Age and diet shape the genetic architecture of body weight in Diversity Outbred mice" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Matt Kaeberlein as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Andrew Dahl (Reviewer #2); Amelie Baud (Reviewer #3).

The manuscript has improved since last submission, however, few comments many for clarity need to be addressed before the manuscript can be accepted for publication.

Essential Revisions:

1) The methods used to calculate the genome-wide significance threshold for the GWAS are now explained with sufficient detail but I believe the thresholds used are much less stringent than commonly accepted (commonly accepted = some estimate of a genome-wide threshold of 0.05). Two lines of evidence for this:

– The authors "used the number of top singular values that explain at least 90% of variation across markers (500), as a proxy for the number of independent QTL signals,", from which they calculated the Bonferroni corrected threshold. They mention in their response to the reviewers that this approach is similar to what is used in human GWAS. However, based on the results presented in Table 1 of https://onlinelibrary.wiley.com/doi/epdf/10.1002/gepi.20430 for example, I believe a value much greater than 90% of variation needs is used when applying this approach to human GWAS data to get close to the commonly accepted threshold of 5.10-8. SimpleM (presented in Table 1 of the paper above) uses a cutoff of 99.5% and still yields a more lenient threshold than 5.10-8.

– Other GWAS in DO mice have used much more stringent thresholds than p-value of 10-4: for example, https://link.springer.com/article/10.1007/s00335-018-9745-8#Sec11 used a P value threshold of 7.3.10-7 to claim genome-wide significance at 0.05; https://reader.elsevier.com/reader/sd/pii/S0168952519300654?token=DEFE149F2C71C5E4EA274AB6AE3638597EEACA678E9451C00AEE8CCD806BDA0E20ACE4BAA8DEF5D6BE37F274132E9F61&originRegion=eu-west-1&originCreation=20210602135101 used an even more stringent threshold – see Figure 3C.

As a consequence of the likely very lenient significance thresholds used in GWAS, all the claims from GWAS and subsequent analyses are supported by limited statistical evidence. In particular, it is difficult to be confident that the apparent lack of pleiotropic signals observed across loci is a real feature of the architecture of body weight: the different FAPs observed etc. could be the result of noise. The genes identified in Table 1 are also supported by limited statistical evidence.

2) I find the GWAS-related text still poorly organised and overwhelming. For example, effects as a function of age are discussed in three different sections (L404, L454, and L483). Similarly pleiotropy is discussed in three different sections, and functional variants are discussed in different sections too. Perhaps group everything related to pleiotropy in one section, effects as a function of age in another, and candidate variants/genes in a third?

The non-linearity in diet is hard to understand and appreciate in the current manuscript. It is quite interesting (surprising), if true. Perhaps give it its own section?

3) Using a diagonal omega matrix is problematic not only for pairs of diets but also for pairs of generations I think, as interaction effect sizes could well be correlated in pairs of generations. The response from the authors to this issue is not entirely satisfactory to me, as they acknowledge that biases are possible but suggest biases will be small without really explaining why they will be small.

– Lines 291-5 – I don't understand why permutations don't fully break the correspondence between both sources of relatedness and phenotype? The kinship is just a matrix, and I don't see how substructures in this matrix could be invariant to random permutation. (You are mean centering such that row/column sums are 0, right? this is required for REML). In other words, what null are you testing if the permutations are also significant? Why doesn't your claim imply that a pure noise phenotype would have significant GxE?

4) For negative VC estimates -- the new supp Figure is great, as is the new flag, but don't see these mentioned in the text? I suggest you acknowledge your constrained approach causes bias (cf your simulations), but that you address it in principle with the new flag and also there is no reason to worry about your real data analyses (because you have such strong signals). I think it's important to have unbiased inference for GxE h2 (much more than for additive h2).

https://doi.org/10.7554/eLife.64329.sa1

Author response

Major Comments:

1. Please justify the p-value thresholds:

– Lines 306-7 say that a 10fold weaker threshold is used for the interaction test--why? Also, where do these thresholds come from?

The Bonferroni corrected p-value threshold used in the manuscript is a genome-wide threshold and was computed based on the expected number of linkage disequilibrium (LD) blocks (independent signals) in our study (similar to the strategy used in human GWAS studies). If the LD structure is perfectly block-like, the number of LD blocks can be computed from the number of non-zero singular values of the marker genotype matrix. In practice, we used the number of top singular values that explain at least 90% of variation across markers (500), as a proxy for the number of independent QTL signals, resulting in a p-value threshold of 10-4. This same threshold is reasonable for our interaction tests since we have only one test per marker (similar to the additive tests); however, we use a weaker threshold of 10-3 for interaction tests to explore suggestive GxE interaction QTLs. Our choice of thresholds, although arbitrary, are intended to help us focus our attention to the strongest QTLs with sufficient statistical support and are not intended to quantify the exact number of QTLs linked to body weight. We have updated the manuscript to better explain our choice of p-value thresholds.

– Is it sound to take this two-step approach of first finding loci and then 'fine mapping' variants nearby? It is definitely fine if the thresholds are valid genome-wide--but then why not test genome-wide?

The p-value thresholds we used are valid genome-wide and it is certainly reasonable to simply test genome-wide. We chose to take the two-step approach for two reasons:

1. At the genotyped markers, testing for association between founder-of-origin and phenotype has higher power than testing for association between allele count and phenotype (PMC4169154, PMC3149489). Notably, we only have informative founder-of-origin inferences at genotyped markers.

2. The two-step approach is much more computationally efficient. While this two-step approach certainly has an increased false-negative rate, evaluating the tests longitudinally through age should mitigate this substantially.

2. Please eliminate nonnegativity constraints on VCs in software:

– Are you truncating estimates to be nonnegative? If so, this will cause bias, which I believe is visible in your simulations. This is essential to fix for testing, and also for unbiased VC estimates.

– Users will test random effects, so it is necessary to validate (or remove) this part of your software.

– I don't see an assessment of tests or ses for the VCs. A quick supp figure corresponding to the current simulations would suffice, I think.

Thank you for bringing to our attention the possibility that variance components do not need to be constrained to be positive, in order to have valid estimates of heritability. We re-computed the PVE estimates for all ages and diets without truncating the estimates to be non-negative, and found very little difference in the magnitude and trends of PVE with age (see Figure 4 —figure supplement 4). We chose to keep the estimates with non-negativity constraints as the main result, since these variance components are more interpretable in the context of our study. We have, however, included an option in the software to relax the non-negativity constraints on the variance components.

3. Please assess standard errors/false positive rates/power for testing VCs:

– Are there evidence that GxEMM improves power for fixed effect interaction tests? I think just commenting on this is enough, but it does seem important if you want others to use your approach. Comparing power (or some measure of false positives) for EMMA vs GxEMM in real data and/or simulation would work.

To address the question of power for both fixed effect additive and interaction tests, we performed QTL mapping using the EMMA model, at the same set of ages as for the GxEMM model. In order to compare the GxEMM results to the standard linear mixed model, we reran the body weight association analyses at all ages using the EMMA model. We found that for both the additive and interaction models, at p-value thresholds of 10-3 and 10-4, the EMMA model identified more loci associated with body weight, quantified both by the number of independent genomic regions and average length of time in days per locus. We have updated the Results section, included Figure 5 —figure supplement 2 , and added a table in Supplementary file 3 to highlight this point.

We, however, do not claim the difference in QTLs to be due to a difference in power nor label any model-specific QTLs as false positives. Comparing Figure 5 —figure supplements 1 and 2, we observe that it is rare for a strong association under one model to be completely ablated under the other model; instead we observe subtle differences in likelihood ratio and p-values. Given the high temporal resolution in phenotype measurements, we feel it is more informative to focus on the effect size estimates (and longitudinal trends therein) of reasonably significant QTLs rather than focusing on the (admittedly arbitrarily and binary) problem of testing for QTLs. We relied on GxEMM since it is a more expressive model and EMMA is a special case of GxEMM.

4. Mouse phenotyping:

– The text needs to say upfront that the mice included in this study were subject to a battery of phenotyping beyond weighing, including metabolic cage phenotyping, blood drwas, and "challenge-based phenotyping procedures". Currently this information comes late in the text (L287), and the title of Figure 1 Supplement 1 reads "Raw phenotype measurements", which does not alert the reader to these additional phenotypes. Some details about this phenotyping are needed, in particular how many blood draws were taken and which specific procedures are included under "challenge-based phenotyping".

This is a good point and we have updated the ‘Data’ section to describe the challenge-based phenotyping procedures (L129 – L134).

This extra phenotyping means that the body weight results presented in this study could be study-specific. In particular, the decrease in PVE observed with age in this study, which contrasts with increased heritability with age observed in previous studies (see Introduction), could be the result of increased noise due to extra phenotyping. The implications of this extra phenotyping need to be discussed.

Importantly, it does not mean that this study is any less relevant than mouse studies without challenging life events, humans also experience challenging life events

Agreed; for this reason – and for all the unmeasured factors that we are unaware of – some of these results may be specific to this study. With data from another dietary intervention study in aged, outbred mice, we could assess whether this is the case. We have added text to the ‘Discussion’ to highlight this point (L557 – L559).

5. Clarifications on Figure 4:

– What do PVE_tot and PVE_e represent (more on this below)? In particular, L291 "total genetic variance for the 40% CR group" is non-sensical to me because I thought PVE_tot was across all mice – as suggested by the fact that there is a single dark grey line in Figure 4 showing PVE_tot. Also, I was expecting PVE_tot (dark grey line in Figure 4) to be the sum of non diet-dependent genetic effects + all 5 PVE_e (colored lines). Yet the dark grey line is below the coloured lines. Hence I don't have the right intuition for what PVE_tot and PVE_e are. Looking at the equations didn't help: shouldn't it be PVE_e in Equation (7) and PVE_tot in Equation (13)?

Clarifications are needed.

PVEtot is computed using all mice (dark grey line in Figure 4) and PVEe is computed using mice belonging to diet ‘e’ (colored lines in Figure 4). Since PVEtot and PVEe (for each diet) are computed on different sets of mice, we can’t expect PVEtot to be the sum of all PVEe (i.e., the denominators are different). This is clear from the equations: the denominator of Equation 7 for PVEtot is VarY , the phenotypic variance computed using all the samples, while the denominator of Equation 13 for PVEe is VarY|e , the variance computed using only samples belonging to diet ‘e’. Crucially, PVEe is not the fraction of PVEtot contributed by diet ‘e’.

Thank you for pointing out the error on L291; we have removed the term “total” from that sentence.

– Figure 4: I love this figure, but worry a bit about the role of GxGeneration:

– The legend gave me a strong (and wrong) sense that the GxE was all about diet. If possible, disentangling the two GxE contributions visually would be helpful.

– I don't understand why generation is included (eg you could instead include cage effects, or neither).

– Why not have an apples-to-apples comparison using GxGeneration as the baseline for questions about GxDiet?

– More generally, I don't follow how generation is being used in this paper. Is it just always in the background as a random effect, and if so, why? Why don't you test for locus-x-generation effects--if they're not expected to be interesting, why do they explain so much variance? If they are absorbing confounding, why not also adjust for cage (perhaps with the GxEMM 'IID' model to save degrees of freedom)? I suspect whatever you're doing is reasonable, but maybe just explain a bit more clearly.

The GxEMM model includes variance components for interaction terms with ‘diet’ and ‘generation’. In our study design, ‘generation’ is a proxy for the shared environment that the mice are exposed to throughout the study; in particular, mice of the same generation are subject to various phenotyping pipelines together. In this sense, ‘generation’ effects are essentially batch effects. We chose to always control for GxGeneration random effects in the model since this helped give us better and more precise estimates for PVEe. We don’t test for GxGeneration fixed effects, both because they are not expected to be interesting and because their contribution to diet-dependent PVE values are relatively small (see Figure 4 —figure supplement 5).

PVEe is computed using mice on diet ‘e’ and includes contributions from the diet ‘e’ term and from all the ‘generation’ terms. Thus, PVEe for two diets does have some shared components coming from the generation variance components; however, it is difficult to decompose which parts are truly shared and which are diet-specific, since the PVE are computed on different sets of mice, and thus the contribution of kinship is slightly different (see Equation 14). We have added Figure 4 —figure supplement 4 highlighting the contribution of diet-specific variance components and generation-specific variance components for each PVEe.

We could certainly control for GxCage effects, either assuming the same variance component for all cages as in the GxEMM ‘IID’ model or using a hierarchical model for the cage variance components. Assuming that all cages shared the same variance component parameter, we observed that the estimate for the cage-specific variance components was negligibly small, suggesting that either cage effects were negligible or the ‘IID’ model is too restrictive. A hierarchical model for cage effects is an interesting avenue for future work, but is out of the scope for our current study.

– What is dark grey? The hom estimate from GxEMM? The hom+GxGeneration (but then shouldn't there be one estimate for each generation)?

Dark grey in Figure 4 is the PVE computed using all mice. This includes the contribution from all the diet-specific and generation-specific variance components. We have included Figure 4 —figure supplement 5 showing the contribution of each of these components to the dark grey line.

Figure 4—figure supplement 2, (b), right – isn't it concerning that the permuted phenotypes are heritable? How can this be? I recognize you don't use this kinship in your real analyses (though believe you do use the genotypes from which the kinship is built), but some explanation here would increase confidence in the overall approach.

The SNP-based kinship can be decomposed into two components: the genetic sharing between the founder strains (the 8 founders have some degree of genetic similarity amongst themselves at the SNP level) and the genetic sharing arising from the breeding strategy used to develop the DO panel. In other words, SNP-based kinship accounts for sharing due to identity-by-state while FAP-based kinship accounts for sharing due to identity-by-descent. Randomly permuting the phenotype labels breaks the link between phenotypic similarity and the latter component of SNP-based kinship, but does not effectively break the link with the former component. Computing kinship using the founder allele probability effectively controls for any genetic similarity between the founder strains (i.e., it explicitly only includes the latter component). We believe this explains the inflation of PVE after permuting phenotypes with the SNP-based kinship, observed both for the EMMA and GxEMM models. Based on this reasoning, we hypothesize that the observed inflation would have been worse at earlier generations of DO mice and predict that this inflation will continue to decrease with future generations of DO mice. We have added text in the manuscript elaborating on this observation.

I'd like to see versions of Figure 4 that: overlay the interventions, as in supplement Figure 1; only show the E-specific genetic variances (ie substracting the sig2g_hom terms); and show the heritabilities (ie normalize by total variance at each time point). These would make nice supplement Figuresif easy to generate.

We have added Figure 4 —figure supplement 2, overlaying the interventions onto the PVE trends. We have also added Figure 4 —figure supplement 5 breaking down each PVE (total and diet-dependent) into contributions from the diet-specific terms and the generation-specific terms.

6. GWAS-related comments:

– The GWAS section and the corresponding figures are overwhelming. I think it would benefit from shortening and using sub-headings. I think it would be good to show first how likely causal variants can be identified (including FAP groups, functional annotations etc. and have one figure for that), with the chr6 and chr 12 loci as examples, and THEN show these loci across age and diets to discuss pleiotropy and heterogeneity.

We appreciate the comments from this reviewer and agree that the GWAS section as originally constructed was overwhelming. To address this, we have now broken this section up into two subsections and eliminated the confusing discussion about distinguishing between pleiotropy, allelic heterogeneity and tightly linked loci (see below). We believe these changes will greatly increase reader comprehension.

In regards to figure 5, we agree there is a lot of information in each subpanel. We have re-organized these results into two figures to first discuss the identification of focal variants and second the effect of a focal variant on body weight at different ages.

– except if we consider triallelic SNPs, the distinction between allelic heterogeneity ("a single locus harboring multiple functional alleles each with distinct phenotypic effects") and (what I call locus heterogeneity) "a single genomic region contains multiple functional body weight loci that are only revealed with sufficient fine-mapping resolution" (L346-349) is difficult to make because noise can blur genetic associations, making it impossible to know the exact position of the causal variant(s). Consistent with the difficulty I just mentioned, in the section L419-443 and corresponding Supplementary Figures, I find that the assignments of the loci to either allelic or locus heterogeneity not obvious/robust.

I think it would be best to focus on distinguishing between pleiotropy (same causal variant(s) affecting body weight at different ages or in different diets) and non-pleiotropy (be it allelic or locus heterogeneity), rather than trying to distinguish between allelic and locus heterogeneity.

Yes, this is a good point. We agree that in general we do not have a large enough mapping population for the fine-mapping analysis to distinguish between allelic and locus heterogeneity. We have updated the language in the manuscript to state that these results are not consistent with a single pleiotropic locus.

Statistical tests exist to test the null hypothesis {same locus affects both traits} in multiparental populations (doi.org/10.1534/g3.119.400098). In addition, looking at the FAPs of the lead variants, as the authors do here, would be a great way to complement this statistical test. Statistical evidence of different loci + differing top FAPS would provide strong evidence for some form of heterogeneity.

If you are going to focus on pleiotropy yes/no, it would be good to show a locus with likely pleiotropic effects (with the caveat that this would be the null hypothesis so not really evidence, rather suggestive).

This is a great point. We do not apply this formal test of pleiotropy. We have updated the manuscript to state that when we identify a single FAP associated with body weight at multiple ages (e.g. Figures 6E,F; 7E) that this is suggestive of pleiotropy.

Additionally, we interpret the identification of distinct FAPs at the same genomic region (e.g. chromosome 6 locus with AJ/NOD and B6/CAST/PWK FAPs highlighted in Figures 5 and 6) as strong evidence against a single pleiotropic locus.

An idea now, not a request: It would be interesting to know whether the loci for which different causal variants are predicted to affect different traits (e.g. different ages) are also loci for which multiple causal variants are predicted to affect either trait (see doi:10.1038/ng.2644).

Yes, we agree this is an interesting idea to pursue in follow-up analyses.

7. Gene enrichment:

– is there a significant enrichment in genes affecting neurological behavior in this study, to support the claim that the fine-mapped genes implicate neurological and metabolic processes? If not, probably best to move the corresponding section of the text to the discussion. Also discuss whether the associations between neurological genes and body weight could arise due to the stressful extra phenotyping of the mice (are those associations mostly for ages after the challenge-based phenotyping for example, or are they seen even for ages before that stressful phenotyping?)

We have not tested for a specific GO term enrichment of neurological and metabolic processes for the candidate genes. In response to this comment, we have edited this section of the results to make clear that we do not claim any specific enrichment of these processes. In keeping with comments from another reviewer, we feel that highlighting these six specific candidate genes (Negr1, Gphn, Trm1-like, Creb5, PI3K-C2.γ, and Edem3) adds significant value to the paper. Moreover, we believe that highlighting these six genes in the Results section immediately following our discussion of the fine-mapping data is the clearest way to convey this information to the reader and we have kept this in the Results section.

We agree with this comment (and previous comments) that this particular result, and nearly any other result we present in the paper could in fact be caused by peculiarities of this experiment rather than the ‘true’ genetic basis of variation in body weight and how that is affected by age and diet. We have emphasized this caveat in the Discussion. Future studies that incorporate longitudinal measurements of body weight in both mice and humans should help identify how much of our results are robust to differences in experimental design.

[Editors’ note: further revisions were suggested prior to acceptance, as described below.]

Essential Revisions:

1) The methods used to calculate the genome-wide significance threshold for the GWAS are now explained with sufficient detail but I believe the thresholds used are much less stringent than commonly accepted (commonly accepted = some estimate of a genome-wide threshold of 0.05). Two lines of evidence for this:

– The authors "used the number of top singular values that explain at least 90% of variation across markers (500), as a proxy for the number of independent QTL signals,", from which they calculated the Bonferroni corrected threshold. They mention in their response to the reviewers that this approach is similar to what is used in human GWAS. However, based on the results presented in Table 1 of https://onlinelibrary.wiley.com/doi/epdf/10.1002/gepi.20430 for example, I believe a value much greater than 90% of variation needs is used when applying this approach to human GWAS data to get close to the commonly accepted threshold of 5.10-8. SimpleM (presented in Table 1 of the paper above) uses a cutoff of 99.5% and still yields a more lenient threshold than 5.10-8.

– Other GWAS in DO mice have used much more stringent thresholds than p-value of 10-4: for example, https://link.springer.com/article/10.1007/s00335-018-9745-8#Sec11 used a P value threshold of 7.3.10-7 to claim genome-wide significance at 0.05; https://reader.elsevier.com/reader/sd/pii/S0168952519300654?token=DEFE149F2C71C5E4EA274AB6AE3638597EEACA678E9451C00AEE8CCD806BDA0E20ACE4BAA8DEF5D6BE37F274132E9F61&originRegion=eu-west-1&originCreation=20210602135101 used an even more stringent threshold – see Figure 3C.

As a consequence of the likely very lenient significance thresholds used in GWAS, all the claims from GWAS and subsequent analyses are supported by limited statistical evidence. In particular, it is difficult to be confident that the apparent lack of pleiotropic signals observed across loci is a real feature of the architecture of body weight: the different FAPs observed etc. could be the result of noise. The genes identified in Table 1 are also supported by limited statistical evidence.

We appreciate the reviewers pointing out evidence from both human GWAS and mouse QTL studies suggesting the need for a more stringent p-value cutoff for claiming significance. While the two papers on GWAS in DO mice used a more stringent p-value threshold to claim genome-wide significance at 0.05, the Yuan et al., paper used a different permutation scheme to generate their null distribution and compute their p-value threshold and the Saul et al., paper does not describe how their threshold was selected.

Following the approach in the Gao et al., paper on Bonferroni correction in human GWAS studies, we computed the eigenvalues of the full linkage disequilibrium matrix among genotyped variants and used the top K eigenvalues that explain Y% of total genetic variation as a proxy for the number of independent tests. For Y = 90, we obtained K = 500, leading to a p-value threshold of 1e-4. For Y = 99.5, we obtained K = 4455, leading to a p-value threshold of 1.1e-5. A Bonferroni correction for multiple hypotheses is considered to be highly conservative, particularly for traits like Body Mass that are known to have a strong genetic component. An alternate and less conservative approach would be to use a p-value threshold corresponding to a fixed, expected number of false discoveries. Using the Benjamini-Hochberg method, a procedure that is expected to be robust to correlations due to linkage disequilibrium, we computed a p-value threshold of 1.2e-4 corresponding to a 10% expected false discovery rate.

Given the high heritability of Body Mass and the availability of longitudinal Body Mass data, we reasoned that it was excessively conservative to rely on a single Bonferroni-based p-value threshold as the only source of statistical evidence. Instead, in Table 1, we listed loci that (i) passed a more liberal Bonferroni threshold (matching the 10% FDR threshold), (ii) were identified as significant at multiple ages, and (iii) were found to be significant using fine-mapping analysis. In order to clarify the strength of evidence based on p-values alone, we marked in Table 1 loci with p-value less than 1.1e-5 with *** (denoting them as highly significant), loci with p-value less than 1e-4 with ** (denoting them as significant), and p-value less than 1e-3 with * (denoting them as suggestive). Multiple sources of evidence supporting these loci suggest that the observed associations and related aspects of the architecture of body weight are not the result of noise.

2) I find the GWAS-related text still poorly organised and overwhelming. For example, effects as a function of age are discussed in three different sections (L404, L454, and L483). Similarly pleiotropy is discussed in three different sections, and functional variants are discussed in different sections too. Perhaps group everything related to pleiotropy in one section, effects as a function of age in another, and candidate variants/genes in a third?

The non-linearity in diet is hard to understand and appreciate in the current manuscript. It is quite interesting (surprising), if true. Perhaps give it its own section?

We have restructured the GWAS-related text to improve readability and organization of information. First, we discussed the use of founder allele patterns, along with functional variants and candidate genes at the two loci we’ve highlighted, then we discussed pleiotropy, and, finally, we discussed effects as a function of age. We have retained a separate, broader discussion on candidate genes that we’ve identified in this study. We have also separated the paragraph on non-linearity in effects with diet into its own section.

3) Using a diagonal omega matrix is problematic not only for pairs of diets but also for pairs of generations I think, as interaction effect sizes could well be correlated in pairs of generations. The response from the authors to this issue is not entirely satisfactory to me, as they acknowledge that biases are possible but suggest biases will be small without really explaining why they will be small.

We acknowledge the lack of explanation behind our intuition that off-diagonal terms in the omega matrix are expected to be small, leading to small biases in estimates of heritability. To address this, we modified our code to estimate a full omega matrix and computed heritability estimates at each age using the full matrix. We observed that off-diagonal entries of the full omega matrix were often orders of magnitude smaller than the diagonal entries, leading to very small changes in heritability estimates using the full matrix. We have added Figure 4 —figure supplement 5 comparing the heritability estimates between a diagonal omega matrix and a full omega matrix.

– Lines 291-5 – I don't understand why permutations don't fully break the correspondence between both sources of relatedness and phenotype? The kinship is just a matrix, and I don't see how substructures in this matrix could be invariant to random permutation. (You are mean centering such that row/column sums are 0, right? this is required for REML). In other words, what null are you testing if the permutations are also significant? Why doesn't your claim imply that a pure noise phenotype would have significant GxE?

The kinship is a matrix of relatedness between pairs of individuals; in this sense, substructures in this matrix (or even the entire kinship matrix) could be invariant to certain permutations of samples depending on whether kinship is computed using genotypes or using founder allele probabilities. This invariance is much more readily observed in the DO mouse population (or other lab-outbred populations) than in natural populations of mice or humans, particularly because the DO population is currently only ~30 generations removed from the inbred Collaborative Cross mouse lines from which they were derived.

As an extreme example, if the mice in our study were composed of just “copies” of the 8 founder strains, then the genotype-based kinship matrix would have eight blocks of matrix-of-ones along the diagonal (since the 8 founders are completely homozygous), with off-diagonal blocks capturing the kinship between pairs of founders. Such a kinship matrix is invariant under certain permutations of sample ids; these permutations would give significant heritability estimates and, thus, straightforward permutations would not be a valid approach to construct a null of no association between genotype and phenotype.

The current DO population matches a less extreme version of the example above, due to ~30 generations of outbreeding. Thus, permuting the genotype-based kinship matrix does not fully break the association between genotype and phenotype. We believe this effect would have been more extreme in earlier DO generations, and will continue to weaken in later DO generations.

4) For negative VC estimates -- the new supp Figure is great, as is the new flag, but don't see these mentioned in the text? I suggest you acknowledge your constrained approach causes bias (cf your simulations), but that you address it in principle with the new flag and also there is no reason to worry about your real data analyses (because you have such strong signals). I think it's important to have unbiased inference for GxE h2 (much more than for additive h2).

We have now addressed in the manuscript the possible bias in our constrained approach, shown that we observed little difference with and without the constraint, and highlighted that our software allows for relaxing this constraint via a flag.

https://doi.org/10.7554/eLife.64329.sa2

Article and author information

Author details

  1. Kevin M Wright

    Calico Life Sciences LLC, South San Francisco, United States
    Contribution
    Conceptualization, Formal analysis, Investigation, Methodology, Validation, Visualization, Writing – original draft
    Contributed equally with
    Anil Raj
    Competing interests
    Kevin is an employee of Calico Life Sciences LLC
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-8772-2687
  2. Andrew G Deighan

    The Jackson Laboratory, Bar Harbor, United States
    Contribution
    Data curation, Resources
    Competing interests
    No competing interests declared
  3. Andrea Di Francesco

    Calico Life Sciences LLC, South San Francisco, United States
    Contribution
    Validation, Writing - review and editing
    Competing interests
    Andrea is an employee of Calico Life Sciences LLC
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-6867-8203
  4. Adam Freund

    Calico Life Sciences LLC, South San Francisco, United States
    Contribution
    Funding acquisition, Project administration, Supervision
    Competing interests
    Adam is an employee of Calico Life Sciences LLC
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7956-5332
  5. Vladimir Jojic

    Calico Life Sciences LLC, South San Francisco, United States
    Contribution
    Methodology, Software, Writing - review and editing
    Competing interests
    Vladimir is an employee of Calico Life Sciences LLC
  6. Gary A Churchill

    The Jackson Laboratory, Bar Harbor, United States
    Contribution
    Conceptualization, Funding acquisition, Project administration, Resources, Writing - review and editing
    For correspondence
    gary.churchill@jax.org
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9190-9284
  7. Anil Raj

    Calico Life Sciences LLC, South San Francisco, United States
    Contribution
    Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Supervision, Validation, Visualization, Writing – original draft
    Contributed equally with
    Kevin M Wright
    For correspondence
    anil@calicolabs.com
    Competing interests
    Anil is an employee of Calico Life Sciences LLC
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-4412-0883

Funding

Calico Life Sciences LLC

  • Kevin M Wright
  • Andrew G Deighan
  • Andrea Di Francesco
  • Adam Freund
  • Vladimir Jojic
  • Gary A Churchill
  • Anil Raj

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

The authors acknowledge Madeleine Cule, Anastasia Baryshnikova, Dale Zhang, and Nicholas Bernstein for their comments on the article and reviewing the software developed for all analyses, and Adam Baker for helping with the design of Table 1. The authors would also like to acknowledge Amelie Baud, Andrew Dahl, and an anonymous reviewer for their insightful comments and suggestions that helped greatly improve our article.

Ethics

All animal procedures were reviewed and approved by the Jackson Laboratory Animal Care and Use Committee (summary # 06005).

Senior Editor

  1. Naama Barkai, Weizmann Institute of Science, Israel

Reviewing Editor

  1. Sara Hägg, Karolinska Institutet, Sweden

Reviewers

  1. Andrew Dahl, University of Chicago, United States
  2. Amelie Boud, Centre for Genomic Regulation, Spain

Publication history

  1. Received: October 26, 2020
  2. Preprint posted: November 5, 2020 (view preprint)
  3. Accepted: May 20, 2022
  4. Version of Record published: July 15, 2022 (version 1)

Copyright

© 2022, Wright et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 1,094
    Page views
  • 217
    Downloads
  • 1
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Kevin M Wright
  2. Andrew G Deighan
  3. Andrea Di Francesco
  4. Adam Freund
  5. Vladimir Jojic
  6. Gary A Churchill
  7. Anil Raj
(2022)
Age and diet shape the genetic architecture of body weight in diversity outbred mice
eLife 11:e64329.
https://doi.org/10.7554/eLife.64329
  1. Further reading

Further reading

    1. Chromosomes and Gene Expression
    2. Genetics and Genomics
    Joseph V Geisberg, Zarmik Moqtaderi ... Kevin Struhl
    Research Advance Updated

    Alternative polyadenylation yields many mRNA isoforms whose 3’ termini occur disproportionately in clusters within 3’ untranslated regions. Previously, we showed that profiles of poly(A) site usage are regulated by the rate of transcriptional elongation by RNA polymerase (Pol) II (Geisberg et al., 2020). Pol II derivatives with slow elongation rates confer an upstream-shifted poly(A) profile, whereas fast Pol II strains confer a downstream-shifted poly(A) profile. Within yeast isoform clusters, these shifts occur steadily from one isoform to the next across nucleotide distances. In contrast, the shift between clusters – from the last isoform of one cluster to the first isoform of the next – is much less pronounced, even over large distances. GC content in a region 13–30 nt downstream from isoform clusters correlates with their sensitivity to Pol II elongation rate. In human cells, the upstream shift caused by a slow Pol II mutant also occurs continuously at single nucleotide resolution within clusters but not between them. Pol II occupancy increases just downstream of poly(A) sites, suggesting a linkage between reduced elongation rate and cluster formation. These observations suggest that (1) Pol II elongation speed affects the nucleotide-level dwell time allowing polyadenylation to occur, (2) poly(A) site clusters are linked to the local elongation rate, and hence do not arise simply by intrinsically imprecise cleavage and polyadenylation of the RNA substrate, (3) DNA sequence elements can affect Pol II elongation and poly(A) profiles, and (4) the cleavage/polyadenylation and Pol II elongation complexes are spatially, and perhaps physically, coupled so that polyadenylation occurs rapidly upon emergence of the nascent RNA from the Pol II elongation complex.

    1. Genetics and Genomics
    Christin S Kuo, Spyros Darmanis ... Mark A Krasnow
    Research Article

    Pulmonary neuroendocrine cells (PNECs) are sensory epithelial cells that transmit airway status to the brain via sensory neurons and locally via calcitonin gene-related peptide (CGRP) and γ- aminobutyric acid (GABA). Several other neuropeptides and neurotransmitters have been detected in various species, but the number, targets, functions, and conservation of PNEC signals are largely unknown. We used scRNAseq to profile hundreds of the rare mouse and human PNECs. This revealed over 40 PNEC neuropeptide and peptide hormone genes, most cells expressing unique combinations of 5–18 genes. Peptides are packaged in separate vesicles, their release presumably regulated by the distinct, multimodal combinations of sensors we show are expressed by each PNEC. Expression of the peptide receptors predicts an array of local cell targets, and we show the new PNEC signal angiotensin directly activates one subtype of innervating sensory neuron. Many signals lack lung targets so may have endocrine activity like those of PNEC-derived carcinoid tumors. PNECs are an extraordinarily rich and diverse signaling hub rivaling the enteroendocrine system.