Birthweight phenotype genotype mismatch in cardiometabolic programming

Abstract

Background

The Barker hypothesis, that poor fetal growth can program adult cardiometabolic disease, has been contradicted by genetic investigations that demonstrate direct pleiotropic effects from birthweight genetics on cardiometabolic risk. We investigated if the cardiometabolic effects of birthweight genetics were dependent on the birthweight phenotype of the offspring.

Methods

Repeated measures of adult body mass index, insulin resistance, low-density lipoprotein cholesterol and systolic blood pressure were modeled in a cohort of 1167 offspring followed from mid-pregnancy through to age 27. In linear mixed-effects models, we tested the associations of birthweight polygenic scores in the presence of a high vs low birthweight phenotype.

Results

Here we show that a birthweight phenotype-genotype mismatch associates with central adiposity and insulin resistance in young adults. In fetuses that fail to reach their growth potential (i.e. low phenotypic birthweight with high birthweight polygenic score) we find increased central adiposity and insulin resistance. In those that exceed their growth potential (high phenotypic birthweight with low birthweight polygenic score), data suggests increased insulin resistance. We find no evidence of phenotype-genotype mismatch influence on low-density lipoprotein cholesterol and systolic blood pressure.

Conclusion

These data support the Barker hypothesis, and we suggest it is the deviation from fetal growth potential rather than crude BW that underpins the relationship with adult disease risk. Identification of individuals with a birthweight phenotype-genotype mismatch could allow early identification of children at future risk.

Plain language summary

Individuals born at low birthweight have increased cardiometabolic risk in adulthood. Recent studies have suggested that this is not caused by poor fetal growth but rather by birthweight-determining genes directly affecting cardiometabolic risk. Whether these genetic effects depend on the environmentally achieved birthweight is unknown. By examining 1167 individuals followed from birth to young adulthood, we show that a mismatch between genetically predicted birthweight and environmentally achieved birthweight influences body mass index and insulin resistance. These findings provide evidence that deviating from the genetic growth potential influences cardiometabolic risk. In the future, examining birthweight genes in the context of the achieved birthweight could help identify individuals at risk of cardiometabolic disease.

Byg et al., examine whether polygenic scores for birthweight on adult cardiometabolic risk depend on the phenotypic birthweight. The effects of birthweight genes on insulin resistance and BMI changed according to birthweight phenotype, suggesting an increased cardiometabolic risk when there is a birthweight phenotype genotype mismatch.

Introduction

A robust association between birthweight (BW) and cardiovascular disease (CVD) has been replicated many times¹ (Fig. 1A) and has been considered secondary to phenotypic programming from poor fetal growth^2–4. This theoretical framework is known as the Barker hypothesis (Fig. 1B)³. Potential confounding of this association has hindered causal inference, and more recent causal investigations using BW genetics were challenged because the exposure of “poor fetal growth”, due to an “adverse intrauterine environment”, was not captured in classic genome-wide association studies (GWAS) of BW⁵. The alternative explanations for the association between BW and CVD include direct pleiotropy (Fig. 1C), where fetal genes that affect BW are independently associated with CVD, or dynastic effects related to confounding from maternal genes associated with both BW and CVD (Fig. 1D). One elegant analytic solution to investigate the Barker hypothesis used genes with maternal influences on offspring BW as proxies of an adverse intrauterine environment^5,6 while controlling for fetal BW genes to avoid direct pleiotropic effects (Fig. 1E). Despite sufficient power, such studies found no effect from maternal BW genes, whereas fetal BW genes had associations with cardiometabolic outcomes⁷. The authors concluded that the Barker hypothesis was an unlikely explanation for the BW-CVD association compared to the horizontal pleiotropic effects of fetal BW genes.

Fig. 1. Examining the association between birthweight and cardiovascular disease.

a The association between birthweight and adult disease, as originally proposed by Barker, here exemplified with cardiovascular disease (CVD). b The theory of a thrifty phenotype proposes that poor fetal growth signals a harsh environment, resulting in changes that program short-term survival at the cost of long-term health. Although we are interested in the arrow between poor fetal growth and CVD, since we cannot measure fetal growth directly, we rely on birthweight as a proxy, which means that confounding is possible. c The theory of genetic pleiotropy underlying the association between birthweight and adult outcomes. For instance, the fetal insulin hypothesis suggests that impaired insulin signaling has the dual effect of lowering birthweight and increasing future risk of diabetes mellitus type 2. d The theory of maternal genetic confounding or reverse causality, as exemplified by maternal hypertension genotype/phenotype affecting birthweight and fetal hypertension genotype/phenotype. e Teasing out maternal and fetal genetic contributions to avoid genetic correlation means using genetic variants for the maternal influences on birthweight to test if the intrauterine environment exerts programming effects while controlling for fetal genotype. f A DAG incorporating all potential explanations demonstrates that an unaffected phenotype-genotype interaction is obtained when adjusting for fetal birthweight genotype and phenotype*. Phenotype* signifies the residuals from regressing measured birthweight on birthweight polygenic score. Green = Exposure; Blue = outcome or ancestor of outcome; Pink = Confounder; Dark gray = not on causal pathway; light gray = unobserved.

This conclusion conflicts with two parallel lines of natural experiments. One is the association between famine exposure during gestation and low birthweight, where poor growth conditions are associated with adult adiposity and insulin resistance⁸. Secondly, the study of within-pair BW differences in monozygotic twins demonstrated an association between reduced fetal growth and insulin resistance^9–14. In these settings, the BW changes without an apparent genetic cause and the associations with factors related to adult cardiovascular risk cannot be explained by genetic pleiotropy. It is possible that developmental programming is only present in specific conditions, such as severe nutritional deprivation or twin pregnancy. Alternatively, the genetic approaches have been inadequate to capture the effects of poor fetal growth related to an adverse intrauterine environment.

In obstetrics, fetal growth restriction (FGR) is defined as an infant that “appears to have failed to grow to its expected size—based on its genetically pre-determined potential”¹⁵. This occurs across the entire birthweight range, with the highest proportion of FGR seen in babies with a low BW phenotype. Fifty percent of cases of FGR are seen in babies that are small for gestational age, that is, their BW is less than the 10th percentile. The remaining 50% of cases are seen in babies with a higher BW phenotype (Fig. 2A)¹⁶. As such, the discrepancy between the genetically determined and phenotypically achieved BW could identify fetuses with poor fetal growth.

Fig. 2. Framework of the current analysis.

a Graphical illustration of the non-linear increase in poor fetal growth at the low end of the phenotypic birthweight spectrum, and an indication of our cut points for phenotypic groupings (smallest, middle, and largest). Modified from ref. ¹⁶. b Directed acyclic graph of the phenotype-genotype interaction analysis between the lowest and highest birthweight phenotypes in predicting cardiometabolic outcomes. The polygenic score for birthweight was used as the birthweight genotype. c Flowchart of cohort selection. *To avoid collinearity in the interaction, the residual variance from a regression of birthweight phenotype on birthweight genotype was used to categorize participants into low- (first decile), middle- (deciles two through nine), and high (tenth decile) birthweight phenotype. FGR fetal growth restriction, BMI body mass index, HOMA-IR: homeostatic model assessment for insulin resistance, LDL-c low-density lipoprotein cholesterol, SBP systolic blood pressure.

To this end, offspring polygenic scores for birthweight (BW-PGS) could be used as a proxy for genetically determined birthweight, and the interaction with environmentally achieved birthweight could provide an analytic strategy to gauge poor fetal growth. Focusing on the interaction as the exposure of interest would remove bias from confounding, genetic dynastic effects, and horizontal pleiotropy (Fig. 1F); this can be seen in a causal diagram from the absence of any back door path between the phenotype-genotype interaction term and the outcome of CVD (when adjusting for BW phenotype and genotype)¹⁷. Using this framework, we hypothesized that a mismatch between BW phenotype and genotype, modeled as an interaction, would be associated with cardiometabolic risk in adulthood.

We find that small babies have more insulin resistance and adiposity in adulthood if polygenic scores predicted higher BW, whereas large babies had more insulin resistance if polygenic scores predicted smaller birthweight.

Methods

Participants

We used data from the Raine Study, a cohort of 2900 women from Western Australia recruited during pregnancy (https://www.rainestudy.org.au/)^18,19. Recruitment took place at the King Edward Memorial Hospital and nearby private practices in Perth. For the current study, we used the sub-cohort of offspring (Generation 2; Gen2) with available BW, gestation duration, genomics data, following quality control and follow-up with data for one of the primary outcomes of interest at ages 20, 22, and 27. Every eligible participant was included to maximize the study size. Ethics approval for the original pregnancy cohort and subsequent follow-ups were granted by the Human Research Ethics Committee of King Edward Memorial Hospital, Princess Margaret Hospital, the University of Western Australia, and the Health Department of Western Australia. Parents, guardians and young adult participants provided written informed consent either before enrollment or at data collection at each follow-up.

Exposures and covariables

BW was collected from hospital records and used to create phenotypic groups. Gestational duration was assessed with an ultrasound examination at 18 weeks of pregnancy or from the date of the last menstrual period. In case of a discrepancy >7 days, the ultrasound-estimated date of delivery was used. Maternal baseline variables were collected by questionnaire at gestational week 16–18.

Genetics data

The Raine Study Gen2 participants were genotyped at the Center for Applied Genomics, Toronto, Canada. Quality control (QC) of the genome-wide association study (GWAS) genotyped data were performed as per standard protocol. In brief, a total of 1593 Raine Study Gen2 participants were genotyped on an Illumina 660 Quad Array, which included 657,366 genetic variants, consisting of ~560,000 single-nucleotide-polymorphisms (SNPs) and ~95,000 copy number variants (CNVs). Plate controls and replicates with a higher proportion of missing data were excluded before individuals were assessed for low genotyping success (>3% missing), excessive heterozygosity, gender discrepancies between the core data and genotyped data, and cryptic relatedness (π > 0.1875, in between second- and third-degree relatives—e.g., between half-siblings and cousins). At the SNP level, the SNP data were cleaned using plink²⁰ following the Wellcome Trust Case—Control Consortium protocol²¹. The exclusion criteria for SNPs included: Hardy–Weinberg equilibrium p < 1E-06; call-rate <95%; minor-allele frequency <1%; and SNPs of possible strand ambiguity (i.e., A/T and C/G SNPs). The cleaned GWAS data were phased and imputed using shapeit²² and Minimac3²³ on the Michigan Imputation Server²³ across the 22 autosomes and X chromosome against the Haplotype Reference Consortium reference panel²⁴. A total of 1494 individuals with 518,288 SNPs remained after genotype QC; imputation resulted in 39,131,578 and 1,273,927 SNPs across the 22 autosomes and X chromosomes. Principal components (PCs) analysis was carried out, using SMARTPCA from v.3.0 of EIGENSOFT²⁵, on the cleaned genotyped data, where PCs were generated for purposes of adjusting for population stratification.

Polygenic score

As reviewed elsewhere⁵, using maternal BW genes with effects through the mother could potentially reflect an adverse intrauterine environment, and to our knowledge, no GWAS of gestational age-corrected birthweight centile exists; therefore, we used offspring SNPs influencing own BW derived from a healthy reference population to proxy optimal BW⁶. As part of a meta-analysis within a large international consortium (n ~321,000), we identified 146 SNPs in the offspring associated with their own BW⁶. SNP data were extracted from the Raine Study participants, re-coded to correspond with increasing BW, weighted using the beta-coefficients reported in the meta-analysis, summed, and re-scaled. The BW-PGS was calculated for each study participant and normalized (subtracted mean and divided by standard deviation) to ease the interpretation of results.

Outcomes

The four primary outcomes were related to common clinical measures of cardiometabolic health: adiposity (assessed by BMI), insulin resistance (assessed by homeostatic model assessment for insulin resistance (HOMA-IR)), dyslipidemia (assessed by low-density cholesterol (LDL-c)) and systolic blood pressure (SBP). Only data from participants with an overnight fast were used to calculate HOMA-IR and LDL-c.

Prespecified secondary outcomes included fat-to-lean mass ratio, insulin concentration, glucose concentration, high-density lipoprotein concentration (HDL-c) and diastolic blood pressure (DBP). Post hoc analyses explored triglyceride concentration, non-HDL cholesterol and waist circumference. Trained healthcare staff collected venous samples.

$$\mathrm{BMI}\mathrm{was}\mathrm{calculated}\mathrm{based}\mathrm{on}\mathrm{the}\mathrm{formula}:BMI=\frac{weight\left(kg\right)}{heigh{\mathrm{t}\left(m\right)}^2}$$

$$\mathrm{HOMA}-\mathrm{IR}\mathrm{was}\mathrm{calculated}\mathrm{with}\mathrm{the}\mathrm{formula}:HOMA-IR=\frac{Insulin\left(\frac{mIU}{L}\right)*glucose\left(\frac{mmol}{L}\right)}{22.5}$$

LDL- was derived using Friedewalds formula $\left[\mathrm{LDL}\right]\left(\frac{\mathrm{mg}}{\mathrm{dL}}\right)=\frac{\left[\mathrm{TC}\right]\left(\frac{\mathrm{mg}}{\mathrm{dL}}\right)-\left[\mathrm{HDL}\right]\left(\frac{\mathrm{mg}}{\mathrm{dL}}\right)-\mathrm{TG}\left(\frac{\mathrm{mg}}{\mathrm{dL}}\right)}{5}$ with exclusion of triglycerides > 4.53 mmol/L.

BP was recorded in a seated position by trained nursing staff after 5 min of rest in a quiet room at 25 °C²⁴. SBP and DBP were measured from the right arm using an oscillometric sphygmomanometer (Dinamap Procare100; GE Healthcare, Chicago, Illinois, USA) with an appropriate cuff size for arm circumference. The Dinamap was set to record BP every 2 min. We discarded the first reading and calculated the average blood pressure from the two subsequent measurements. BP was recorded in the morning at 20 and 27 years and late afternoon/evening at 22 years. Height was measured using a Stadiometer (to the nearest 0.1 cm) and body weight using Wedderburn Chair Scales (to the nearest 100 g), with participants dressed in light clothes.

Fat- and lean mass (g) was measured using a Norland XR-36 densitometer (Norland Medical Systems, Inc., Fort Atkinson, WI, USA) and the ratio was then derived. Biochemistry analyses were carried out in the PathWest Laboratory at Royal Perth Hospital. Plasma glucose (mmol/L) was measured using an automated Technicon Axon analyser (Bayer Diagnostics, Sydney, Australia) using a hexokinase method. Serum insulin (mU/ml) was measured using an automated radioimmunoassay (Tosoh, Tokyo, Japan). Total cholesterol (mmol/L) and triglycerides (mmol/L) were determined enzymatically on the Cobas MIRA analyzer (Roche Diagnostics) with reagents from Trace Scientific (Melbourne, Australia), while HDL-C (mmol/L) was determined on a heparin-manganese supernatant. Waist circumference was measured with a measuring tape (nearest 0.1 cm) at the halfway point of the lowest rib and the iliac crest.

Variable distributions were inspected (Fig. S1), and extreme outliers (two HOMA-IR measures >70, one total cholesterol of 73 mmol/L, and a triglyceride of 25 mmol/L) were excluded.

Statistics and reproducibility

PGS are imperfect single-value estimates of SNP influences on phenotypic traits used in causal epidemiology as instrumental variables (IV)²⁶; however, the horizontal pleiotropy between BW genes and CVD invalidates this approach⁵. Rather than an IV, we treated offspring BW-PGS as a measure of optimal BW and examined the cardiometabolic effects in a high versus low BW phenotype setting where poor fetal growth would be virtually absent vs very prevalent (Fig. 2A). As BW is in part determined by BW genes, a grouping based on measured BW would skew the distribution of the BW-PGS between groups, introducing bias; therefore, we split the analytic cohort by using the residual variance from a regression of BW phenotype on BW-PGS (Fig. S2A–C). We chose a sex-specific cut-off at the 10^th and 90^th centile to parallel the cut-offs used when grouping small and large for gestational age in obstetrics (Fig. 2A). To avoid bias from fetal growth restriction in the middle eight deciles, we used the top decile as the reference and the interaction with the lowest decile as the primary outcome^16,27.

Linear mixed-effects models with clustering at participant ID level and random intercepts were used to include those with one or two missing data-points for cardiometabolic outcomes (Fig. 2B). We added fixed effects for commonly used and known determinants of BW-PGS (Genetic PC’s), BW-phenotype (gestational duration) and cardiometabolic outcomes (sex and age at assessment)⁵. We first examined the individual effect estimates of the phenotypic groups and BW-PGS regressed on cardiometabolic outcomes. We then progressed to the primary aim and included an interaction between BW-PGS and the phenotypic group. Inspection of residuals revealed non-normality and skew; therefore, standard errors (SE) were derived using a non-parametric bootstrap at the ID level with 5000 simulations. Two-tailed p-values were derived from z-statistics, and with four primary outcomes, a Bonferroni correction of 0.05/4 = 0.0125 was used for statistical significance.

Secondary outcomes were analysed using a similar approach to the primary outcomes, but were not subjected to the sensitivity analyses described below. Because of the skewed residuals, we reran our models with log-transformed outcomes. We explored a 20th and 80th centile cut-off and the interaction with continuous BW residuals. We explored the impact of gestational duration by excluding it as a covariable and by excluding preterm births with redrawing of the 10th and 90th centile cut-offs as described above. We also reanalyzed models based on 10th and 90th centile cut-offs with crude BW.

Statistical analysis was done in “R” version 4.3.1 with software packages, “tidyverse”, “lme4”, “lmeresampler”, and “sjPlot”²⁸. We followed STROBE guidelines in reporting results²⁹

Results

Sample description, birthweight polygenic scores, and phenotypic groupings

The study sample consisted of 1167 participants (78.2% of the genotyped Raine Study participants), with varying data availability for the four primary analyses (Fig. 2C). Participants were divided into low-, middle- and high BW phenotypes by splitting at the 10th and 90th centiles (Fig. 2A). Using crude BW would introduce collinearity with BW-PGS wherefore groups were made with residuals from regressing BW-PGS on BW (henceforth “phenotypic groups”) (Fig. S1A–C). The largest decile provided the reference to avoid bias from poor fetal growth in the middle deciles, and the interaction with the lowest decile was the primary outcome (Fig. 2A)^16,27.

At ages 20, 22, and 27, there were 1097, 893, and 856 participants, respectively, with 118 in both the largest and smallest phenotypic groupings. These participants were pooled and included in adjusted mixed effect models of repeated cardiometabolic measures from ages 20–27 years. When including prespecified covariables, the offspring BW-PGS was associated with measured BW (Table S1), and a 1 SD increase in BW-PGS was associated with an 86 g increase in BW with no significant effect difference between phenotypic groups (p-interaction >0.18, Fig. S1C). The BW-PGS scores were equally distributed between phenotypic groups, as was fetal sex. Gestational age at birth, HOMA-IR and LDL-c concentration differed between phenotypic groups; there was a trend to higher SBP in the lower phenotypic groups (See Table 1). Maternal baseline variables were unevenly distributed between groups, with lower BMI and more smoking in the low phenotypic group (Table S2). Secondary outcome variables (insulin, glucose, HDL-c, non-HDL-c, triglycerides, DEXA fat-to-lean mass, waist circumference, and diastolic blood pressure) and histograms are presented in Table S3 and Fig. S2.

Table 1.

Baseline demographics

	≤10th phenotypic centile (n = 118)	>10th to <90th phenotypic centile (n = 931)	≥ 90th phenotypic centile (n = 118)
Birthweight (g)	Perinatal exposure variables
Median (IQR)	2427.5 (2171.2−2573.8)	3390.0 (3135.0−3630.0)	4157.5 (4055.0−4350.0)
Offspring sex (% Females)	58 (49.2%)	459 (49.3%)	58 (49.2%)
Standardized BW-PGS median (IQR)	0.0 (−0.8 to 0.7)	0.0 (−0.7 to 0.7)	0.0 (−0.5 to 0.6)
Gestational age in weeks Median (IQR)	36.5 (34.2–38.1)	39.7 (38.7–40.7)	40.4 (39.7–41.3)
Age at 20-year wave	Adult cardiometabolic variables
Median (IQR)	20.0 (19.7–20.3)	19.9 (19.7–20.3)	19.9 (19.6–20.3)
Missing	9 (7.6%)	50 (5.4%)	11 (9.3%)
Age at 22-year wave
Median (IQR)	22.2 (21.8 − 22.7)	22.0 (21.7 − 22.4)	22.2 (21.8 − 22.7)
Missing	27 (22.9%)	216 (23.2%)	31 (26.3%)
Age at 27-year wave
Median (IQR)	26.6 (26.5–27.0)	26.6 (26.5–26.9)	26.6 (26.5–26.9)
Missing	29 (24.6%)	258 (27.7%)	24 (20.3%)
BMI age 20 (kg/m2)
Median (IQR)	23.5 (21.1–26.2)	23.4 (21.2–26.1)	23.8 (21.5–26.8)
Missing	17 (14.4%)	102 (11.0%)	17 (14.4%)
BMI age 22 (kg/m2)
Median (IQR)	24.2 (21.8–27.0)	23.9 (21.8–27.5)	24.5 (22.3–27.2)
Missing	35 (29.7%)	269 (28.9%)	40 (33.9%)
BMI age 27 (kg/m2)
Median (IQR)	24.2 (21.9–28.0)	24.3 (22.2–27.9)	24.4 (22.4–28.3)
Missing	32 (27.1%)	286 (30.7%)	28 (23.7%)
HOMA-IR age 20
Median (IQR)	0.7 (0.4–1.2)	0.5 (0.4–1.2)	0.5 (0.4–1.2)
Missing	32 (27.1%)	183 (19.7%)	31 (26.3%)
HOMA-IR age 22
Median (IQR)	1.5 (1.3–2.1)	1.6 (1.2–2.2)	1.4 (1.0–1.9)
Missing	40 (33.9%)	325 (34.9%)	45 (38.1%)
HOMA-IR age 27
Median (IQR)	1.2 (0.8–1.8)	1.1 (0.8–1.7)	1.1 (0.8–1.5)
Missing	32 (27.1%)	294 (31.6%)	31 (26.3%)
LDL-c age 20 (mmol/L)
Median (IQR)	2.6 (2.1–3.1)	2.4 (2.0–2.9)	2.4 (2.1–2.8)
Missing	32 (27.1%)	186 (20.0%)	31 (26.3%)
LDL-c age 22 (mmol/L)
Median (IQR)	2.8 (2.5–3.5)	2.7 (2.3–3.1)	2.6 (2.2–3.1)
Missing	40 (33.9%)	326 (35.0%)	45 (38.1%)
LDL-c age 27 (mmol/L)
Median (IQR)	3.0 (2.5–3.7)	2.8 (2.3–3.3)	2.8 (2.3–3.3)
Missing	32 (27.1%)	295 (31.7%)	31 (26.3%)
SBP age 20 (mmHg)
Median (IQR)	117.0 (110.0−127.0)	116.0 (108.0−127.0)	115.0 (109.0−122.0)
Missing	16 (13.6%)	94 (10.1%)	16 (13.6%)
SBP age 22 (mmHg)
Median (IQR)	119.0 (111.0−128.2)	118.5 (111.0−127.0)	117.5 (111.5−124.9)
Missing	35 (29.7%)	269 (28.9%)	40 (33.9%)
SBP age 27 (mmHg)
Median (IQR)	116.0 (108.0−125.0)	117.0 (110.0−124.0)	115.0 (108.2−123.0)
Missing	32 (27.1%)	286 (30.7%)	28 (23.7%)

BMI body mass index, HOMA-IR homeostatic model assessment of insulin resistance, LDL low-density lipoprotein cholesterol, SBP systolic blood pressure, IQR interquartile range, BW-PGS birthweight polygenic score.

Cardiometabolic correlates of birthweight phenotype and genotype

In models without a phenotype–genotype interaction, there were no significant differences between birthweight phenotype groups for BMI, HOMA-IR, LDL-c, and SBP (Table 2). There was, however, a significant association between BW-PGS and adult BMI. Every 1 SD increase of BW-PGS was associated with a 0.38 kg/m² (CI[0.14, 0.62] p = 0.002) higher adult BMI (Table 2). The BW-PGS was not associated with the remaining three primary outcome variables (p value >0.2 for all).

Table 2.

Output from models of BW phenotypic grouping and BW genotype treated separately

	Body mass index (n = 1150)	HOMA-IR (n = 1112)	LDL -c (n = 1111)	SBP (n = 1153)
Largest phenotypic decile (model intercept)*	β: 22.32 SE: 3.39 95% CI: [15.7, 29.0]	β: 1.21 SE: 0.657 95% CI: [−0.0458, 2.53]	β: 2.47 SE: 0.369 95% CI: [1.73, 3.18]	β: 120.7 SE: 5.71 95% CI: [110, 133]
Middle phenotypic deciles*	β: −0.46 SE: 0.388 95% CI: [−1.22, 0.30] P value: 0.23	β: 0.15 SE: 0.062 95% CI: [0.03, 0.28] P value: 0.015	β: 0.035 SE: 0.051 95% CI: [−0.07, 0.13] P value: 0.49	β: 1.68 SE: 0.686 95% CI: [0.32, 3.00] P value: 0.014
Smallest phenotypic decile*	β: −0.269 SE: 0.647 95% CI: [−1.54, 1.00] P value: 0.68	β: 0.20 SE: 0.139 95% CI: [−0.08, 0.47] P value: 0.16	β: 0.012 SE: 0.015 95% CI: [−0.024, 0.290] P value: 0.421651	β: 1.49 SE: 1.11 95% CI: [−0.74, 3.62] P alue: 0.18
Birthweight PGS effect*	β: 0.38 SE: 0.123 95% CI: [0.14, 0.62] P value: 0.002	β: 0.03 SE: 0.027 95% CI: [−0.03, 0.08] P value: 0.31	β: 0.012 SE: 0.015 95% CI: [−0.02, 0.04] P value: 0.42	β: −0.25 SE: 0.231 95% CI: [−0.69, 0.21] P value: 0.27

Bold indicates significant effects.

HOMA-IR homeostatic model assessment for insulin resistance, LDL-c low-density lipoprotein cholesterol, SBP systolic blood pressure, PGS polygenic score.

*Linear mixed model of cardiometabolic parameters adjusted for fetal sex, gestational duration, age at assessment and genetic principal components. The estimated SE was derived using 5000 repeats of a non-parametric bootstrap at the participant ID level.

We then moved on to the primary research question: the phenotype-genotype interaction between the largest and smallest phenotypic groups (primary outcomes: Table 3 and Fig. 3, secondary outcomes: Table S4 and Fig. S3).

Table 3.

Examining a birthweight phenotype-genotype interaction in predicting cardiometabolic outcomes

	PGS effect, largest phenotypic decile (reference)	Interaction, middle deciles	Interaction, smallest phenotypic decile (primary outcome)
	Primary outcomes
BMI (kg/m2) (n = 1150, o = 2675)	β: −0.40 SE: 0.380 95% CI: [−1.14, 0.35] P value: 0.30	β: 0.72 SE: 0.406 95% CI: [−0.07, 1.52] P value: 0.07	β: 1.65 SE: 0.512 95% CI: [0.63, 2.64] P value: 0.001
HOMA-IR (n = 1112, o = 2486)	β: −0.11 SE: 0.058 95% CI: [−0.23, −0.01] P value: 0.05	β: 0.13 SE: 0.066 95% CI: [0.01, 0.27] P value: 0.05	β: 0.29 SE: 0.099 95% CI: [0.10, 0.48] P value: 0.003
LDL-c (mmol/L) (n = 1111, o = 2483)	β: 0.04 SE: 0.049 95% CI: [−0.05, 0.14] P value: 0.41	β: −0.02 SE: 0.052 95% CI: [−0.12, 0.08] P value: 0.67	β: −0.09 SE: 0.066 95% CI: [−0.22, 0.03] P value: 0.16
SBP (mmHg) (n = 1153, o = 2685)	β: −1.37 SE: 0.656 95% CI: [−2.65, −0.08] P value: 0.04	β: 1.20 SE: 0.711 95% CI: [−0.17, 2.61] P value: 0.09	β: 1.39 SE: 0.927 95% CI: [−0.47, 3.16] P value: 0.14
	Notable phenotype-genotype interactions in the secondary analysis
Fat-to-lean mass ratio (n = 1065, o = 1644)	β: −0.04 SE: 0.019 95% CI: [−0.07,−0.00] P value: 0.04	β: 0.04 SE: 0.020 95% CI: [0.00, 0.08] P value: 0.03	β: 0.07 SE: 0.024 95% CI: [0.02, 0.12] P value: 0.003
Waist circumference (cm) (n = 1150, o = 2680)	β: −0.66 SE: 0.856 95% CI: [−2.33, 1.03] P value: 0.44	β: 1.53 SE: 0.918 95% CI: [−0.28, 3.32] P value: 0.10	β: 3.51 SE: 1.195 95% CI: [1.12, 5.80] P value: 0.003
Insulin level (mIU/L) (n = 1112, o = 2490)	β: −0.46 SE: 0.235 95% CI: [−0.94, −0.02] P value: 0.05	β: 0.51 SE: 0.267 95% CI: [0.01, 1.06] P value: 0.06	β: 1.18 SE: 0.394 95% CI: [0.39, 1.93] P value: 0.003
Triglyceride level (mmol/L) (n = 1111, o = 2487)	β: −0.07 SE: 0.030 95% CI: [−0.14, −0.02] P value: 0.01	β: 0.10 SE: 0.033 95% CI: [0.04, 0.17] P value: 0.004	β: 0.12 SE: 0.044 95% CI: [0.04, 0.21] P value: 0.005

BMI body mass index, HOMA-IR homeostatic model assessment for insulin resistance, LDL-c low-density lipoprotein cholesterol, SBP systolic blood pressure, PGS: polygenic score.

*Linear mixed model of cardiometabolic parameters adjusted for fetal sex, gestational duration, genetic principal components and age at assessment. The estimated SE was derived using 5000 repeats of a non-parametric bootstrap at the participant ID level. Bold indicates the significant primary results.

Fig. 3. Primary phenotype–genotype interactions for cardiometabolic outcomes.

Graphical representation of the model output in Table 3a, b) The smallest BW phenotype group (Blue graph - Bottom 10 percent) had higher adult BMI and more insulin resistance only in the presence of a higher genetically predicted BW. This effect diminished and reversed in the middle (Black graph – Middle 80 percent) and large (Green graph – Top 10 percent) BW phenotype groups. c, d No significant interaction between BW phenotypic group and BW polygenic score was found when modeling systolic blood pressure and low-density lipoprotein cholesterol, HOMA-IR homeostatic model assessment for insulin resistance, BW birthweight. Shaded area: 95% CI

For BMI, we found a significant phenotype-genotype interaction (β: 1.65, 95% CI: [0.63, 2.64] P value: 0.001). In the smallest phenotypic decile, a 1 SD increase in fetal BW-PGS resulted in a 1.25 kg/m² (95% CI: [0.587, 1.88] P value: 0.0002) increase in adult BMI, while there was no effect of BW-PGS on BMI in the largest phenotypic group (β: −0.40, 95% CI: [0.0212, 0.0667] P value: 0.30). Similar interactions were seen in the smallest phenotypic decile group for DEXA measured fat-to-lean mass ratio (interaction-p = 0.003) and waist circumference (interaction-p = 0.003) in adults.

For HOMA-IR, we found a significant phenotype-genotype interaction (β: 0.29, 95% CI: [0.10, 0.48] P value: 0.003). In the smallest phenotypic decile, a 1 SD increase in fetal BW-PGS was associated with a 0.18 increase in adult HOMA-IR (CI: [0.01, 0.32] P value: 0.02). In the largest phenotypic decile, a 1 SD increase in BW-PGS was associated with a 0.11 decrease in HOMA-IR (95% CI: [−0.23, −0.01] P value: 0.05). This was paralleled by an interaction for insulin concentration (interaction p = 0.003) but not for glucose concentration (interaction p = 0.10).

There were no significant phenotype-genotype interactions for LDL-c or non-HDL-c (interaction p = 0.16 and p = 0.61, respectively) or systolic or diastolic blood pressure (interaction p = 0.14 and p = 0.54, respectively). In the analysis of secondary outcomes, data suggested phenotype-genotype interactions for HDL-c and triglycerides (interaction p = 0.03 and 0.005, respectively).

Sensitivity analyses

Log transformation due to skewness of data and potential outliers did not change the statistical inferences (Table S5). Using 20th and 80th centile cuts diminished effect sizes but agreed with the primary analysis, as did a continuous interaction. (Table S6). Excluding gestational duration at delivery from the model did not affect estimates (Table S7). Excluding preterm births generated phenotypic groups with 1070 participants, with 108 in the low- and high-phenotypic BW groups. The interaction terms had modest effect-size reductions, but overall effects were consistent with our primary analysis (Table S8). Sensitivity analyses using the measured BW (not removing BW-PGS variance) to make 10th (2710 g) and 90th (4020 g) centile cuts reduced interaction estimates but yielded results consistent with the primary analysis (Table S9).

Discussion

We have used a phenotype-genotype interaction approach to demonstrate that the effects of offspring BW determining genes on cardiometabolic risk factors depend on the setting in which they occur. A phenotype-genotype mismatch had the most pronounced effects on cardiometabolic risk factors at the lower end of the BW spectrum; the greater the mismatch, the greater the adverse outcomes.

Our results are consistent with famine- and twin studies of adverse early exposures but disagree with previous studies in Norwegian populations using maternal BW genes as proxies for an adverse intrauterine environment^7,27. Given the demographic and ancestral similarities in our own and the Norwegian studies, neither genetics nor age-dependency in the cardiometabolic risk manifestation offers a reasonable explanation for the different results. We propose that the maternal genetic instrument used in previous studies⁶ may primarily represent a physiological rather than an adverse intrauterine environment. Indeed, the genome-wide association study separating BW genes with maternal and fetal effects was not characterized by adversity or low BW, it excluded preterm births (often comorbid with poor fetal growth) and, in some cases, also low and high BW explicitly (Supplementary Table 1 in ref. ⁶). Our study uses the offspring BW-PGS as a proxy for optimal BW, and our data supports the hypothesis that neither the BW-genotype nor BW-phenotype is the exposure of interest for the study of poor fetal growth; rather, it is the mismatch between the two. The interaction between BW and BW-PGS is not subject to traditional confounding, genetic pleiotropy or maternal genetic confounding; because of the closure of back door paths, the interaction estimate is free from such bias, which is important when interpreting effect estimates from our model output¹⁷.

For BMI, an overall effect of the BW-PGS suggested the presence of pleiotropy; however, the effects on BMI were also highly dependent on the phenotypic group. A phenotype-genotype interaction suggested that a higher genetically predicted BW was associated with a larger adult BMI only in the presence of a small phenotypically achieved BW. Our secondary analysis suggested that central adiposity drove the increase in adult BMI. The effect size of the BW-PGS in the low-phenotype group was considerable; as the BW-PGS increased, BMI changed from normal to overweight.

For insulin resistance, we did not see a significant effect in the overall cohort to suggest pleiotropy, but saw a strong interaction between high- and low-phenotypic BW. Unlike BMI, our data suggest that a phenotype-genotype mismatch had bimodal effects on HOMA-IR. In the smallest phenotypic group, babies that failed to reach their growth potential (low phenotypic BW and high BW-PGS) developed increased insulin resistance as adults. Adversity was also seen with a reduced BW-PGS in the largest phenotypic group, i.e., babies whose phenotypic BW exceeded their genetically predicted BW, demonstrated increased insulin resistance. This result questions our a priori assumption that programming is a phenomenon related exclusively to FGR and highlights fetal growth excess as a potential programming event. Our secondary analysis suggested that increased insulin levels in the offspring drove this association, consistent with the natural history of type 2 diabetes, where hyperinsulinemia precedes later hyperglycemia³⁰.

We saw mixed results for dyslipidemia with no significant effects on the primary outcome of LDL-c or non-HDL cholesterol. Our data suggest a reduction in HDL-c and an increase in triglycerides with a phenotype-genotype mismatch, but the secondary and borderline nature of the estimates weakens our confidence in the association. The literature on associations between measured BW and dyslipidemia is similarly mixed, with some studies reporting no associations for any lipids³¹, while others report adverse outcomes for total cholesterol in low BW³² and others report quadratic associations for triglyceride and HDL-c^7,33.

For SBP, there was no significant phenotype-genotype interaction to suggest developmental programming, while the highest phenotypic BW group had nominally lower blood pressure values consistent with a previous meta-analysis³⁴. This could support maternal genetic confounding through dynastic effects⁶, that is, a maternal hypertension phenotype causes lower BW and subsequent offspring hypertension through the inheritance of the predisposing hypertension genotype.

Our sensitivity analysis underscored the separation of effects stemming from fetal growth and gestation duration at birth. Although poor fetal growth is often comorbid with preterm birth (either spontaneous or iatrogenic), the exclusion of gestation duration from our model had no impact on the phenotype-genotype interaction estimate. Although excluding preterm births markedly changed our low phenotypic BW group, we saw only modest changes in the interaction effect sizes. Using continuous interaction and quintile groupings diminished the strength of the observed associations, offering support that these effects are seen primarily at the extremes of BW. We analysed phenotypic groupings based on crude measured BW, as this cut-off may be more intuitive. Bias may be introduced in these models due to co-linearity between BW-PGS and crude BW, and the cut-off at 2710 g was arbitrarily chosen based on our sample distribution of BW; nevertheless, this analysis suggested that BW-PGS could have a programming role in neonates with a BW of 2710 or less. Finally, compared to the original Generation 2 Raine-study cohort, we saw substantial attrition, which could introduce some selection bias and influence our estimates; however, we did not see signs of selective participant dropout in the BW residuals (Fig. S1A), and previous studies have not found significant selection bias in the Raine Study adult follow-up³⁵.

Our study has several strengths. Our hypothesis-driven approach, selective testing and conservative significance threshold support the validity of our results; furthermore, the focus on the interaction as the exposure of interest should minimize bias and confounding (Fig. 1F) and, if replicated, provide causal insight. We had excellent data quality for the parameters of interest and included repeated measures to increase the power and precision of our estimates. Within the genotyped sub-cohort of the Raine Study, we had reasonable follow-up and sample size.

Our study has important limitations. The assessment in young adulthood is not necessarily generalizable to other phases of life; for instance, hypertensive consequences of a phenotype-genotype mismatch may develop later in life. We did not explore potential sex interactions, and our modest power would make it difficult to detect three-way interactions (sex-by-phenotype-by-genotype). Given the substantial literature on sexual dimorphism in developmental programming³⁶, this should be explored in future work. The present investigation was spurred by a phenotype-genotype interaction for childhood aggressive behavior in the Raine Study found post hoc³⁷. The individuals with a phenotype-genotype mismatch in that study will overlap with those in the current study, introducing the risk of a type 1 error. As such, replication in other cohorts is necessary before robust conclusions can be made. Another limitation in interpreting the phenotype-genotype interaction is that the BW-PGS will not include all the genetic variants that affect BW; therefore, it could represent a genotype-genotype interaction with the non-included genes. Given the composite nature of the PGS, we deem it unlikely that there would be a systematic interaction with the non-included genes. The 10th and 90th centile cut-offs in our study are universally used in clinical settings to define small- and large for gestational age but similar cut-offs of BW residuals may not be optimal in a clinical context. Future studies using larger cohorts may be able to determine more optimal cut-offs for the BW residuals. It is also likely that gradual improvements in the genetic predictors of BW will change the optimal cut-off. For these reasons, we also believe it is premature to speculate on the clinical relevance of birthweight phenotype-genotype mismatches. Additionally, our cohort was primarily White, and both environmental exposures and ancestral differences in BW genetics mean our results are not necessarily generalizable to other populations and settings.

Finally, studying phenotype-genotype mismatch could also be used in other settings where the deviation from a genetic setpoint is of more interest than the absolute level. One such example could be the study of homeostasis for electrolytes and hormones. The normal variation within and beyond the reference range for blood biochemistry is well-recognized in the clinic³⁸. We speculate that GWAS in cohorts of healthy individuals could be used to provide SNPs that influence such normal variation (i.e., individual optimal level). If incorporated in a phenotype-genotype interaction, we would have an unbiased approach to studying the effects of phenotypic deviation from a genetic setpoint. A specific example could be the debate on causality regarding the association between reduced serum sodium (hyponatremia) and osteoporosis³⁹. In conventional genetic epidemiology, a large-scale case-control GWAS of hyponatremia could yield a genetic instrument to be used; however, akin to “poor fetal growth, “ the variable of interest is not low sodium per se but reduced sodium, away from the optimal genetic setpoint. We hope that a phenotype-genotype interaction could provide an effective and cheap alternative to gauge the unconfounded effect estimates of deviating from the genetically determined setpoints. Importantly, there is still a possibility of reverse causality if phenotype and outcome are measured simultaneously; consequently, with a cross-sectional data collection used in a phenotype-genotype study (e.g. simultaneous collection of serum sodium and osteoporosis) no inference about causality can be made, as reverse causation is possible.

Conclusion

We demonstrate that a BW phenotype-genotype mismatch predicts insulin resistance and central adiposity, supporting the Barker hypothesis. This finding requires replication in larger cohorts prior to clinical use for identifying at-risk individuals. For researchers, we believe a phenotype-genotype interaction could have multiple potential applications in studying deviations from a genetic setpoint.

Author contributions

Conceptualization and methodology: L.M.B., J.A., C.A.W. and C.E.P., Investigation: L.M.B. and C.A.W., Visualization: L.M.B. and C.E.P., Funding acquisition: C.E.P., C.A.W., L.M.B. and T.A.M., Project administration: C.E.P., R.S. and C.A.W., Supervision: C.E.P., C.A.W., R.S. and J.A., Writing—original draft: L.M.B., C.E.P., J.A., C.A.W., L.M.B., T.A.M. and R.S.

Peer review

Peer review information

Communications Medicine thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The individual-level data from the Raine Study cannot be made available or stored on another, non-Raine Study public repository due to legal and ethical constraints. The Raine Study is a publicly available data provider and data can be accessed by interested researchers upon reasonable request. To request access, please contact the Raine Study team at rainestudyscience@uwa.edu.au. The GWAS used for generating polygenic scores for birthweight was published by Warrington NM, Beaumont RN, Horikoshi M, Day FR, Helgeland O, et al. Nature Genetics 2019 (doi:10.1038/s41588-019-0403-1). Data on birthweight has been contributed by the EGG Consortium using the UK Biobank Resource and has been downloaded from www.egg-consortium.org.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s43856-026-01725-0.