Shared genetic influences on adolescent body mass index and brain structure: A voxel-based morphometry study in twins

Abstract

Background:

Previous research has demonstrated significant relationships between obesity and brain structure. Both phenotypes are heritable, but it is not known whether they are influenced by common genetic factors. We investigated the genetic etiology of the relationship between individual variability in brain morphology and BMIz using structural MRI in adolescent twins.

Method:

The sample (n=258) consisted of 54 monozygotic and 75 dizygotic twin pairs (mean(SD) age = 13.61(0.505), BMIz = 0.608(1.013). Brain structure (volume and density of gray and white matter) was assessed using VBM. Significant voxelwise heritability of brain structure was established using the Accelerated Permutation inference for ACE models (APACE) program, with structural heritability varying from 15 to 97%, depending on region. Bivariate heritability analyses were carried out comparing additive genetic and unique environment models with and without shared genetics on BMIz and the voxels showing significant heritability in the APACE analyses.

Results:

BMIz was positively related to gray matter volume in the brainstem and thalamus and negatively related to gray matter volume in the bilateral uncus and medial orbitofrontal cortex, gray matter density in the cerebellum, prefrontal lobe, temporal lobe, and limbic system, and white matter density in the brainstem. Bivariate heritability analyses showed that BMIz and brain structure share ~1/3 of their genes and that ~95% of the phenotypic correlation between BMIz and brain structure is due to shared additive genetic influences. These regions included areas related to decision-making, motivation, liking vs. wanting, taste, interoception, reward processing/learning, caloric evaluation, and inhibition.

Conclusion:

These results suggested genetic factors are responsible for the relationship between BMIz and heritable BMIz related brain structure in areas related to eating behavior.

1. Introduction

Obesity has become a major health concern worldwide (James et al., 2001). Child and adolescent obesity rates have significantly increased since the 1970s (Sassi et al., 2009), with worldwide obesity rates in children and adolescents having risen from 0.7% to 5.6% in girls and 0.9% to 7.8% in boys in this time period (NCD-RisC, 2017). Early life obesity is typically measured through Body Mass Index Z score (BMIz) or percentile, a measure examining where a person’s weight falls on a distribution for people their age, sex, and height with a Z score above 1.645 (the 95^th percentile) considered obese. Early life obesity increases the risk of later health problems, even if the person is no longer obese (Reilly and Kelly, 2011). A better understanding of etiological mechanisms of obesity and its neural correlates can facilitate prevention and intervention efforts. Obesity is a complex phenotype influenced by both genetic and environmental factors. Twin studies can estimate the heritability of a trait (e.g. BMI or brain structure), which is a proportion of the total phenotypic (observed) inter-individual variation that is attributable to genetic differences, and genetic correlation between traits (rG), i.e. the proportion of genes two variables share by comparing the covariances of a phenotype between people with differing degrees of genetic (e.g. 100% for identical twins, on average 50% for fraternal twins, etc.) and environmental similarity (e.g. 100% for twins raised together, 0% for unrelated individuals raised apart; Neale and Maes, 2004). Twin and family studies have shown that the risk for obesity is strongly influenced by genetic factors, with ~80% of adolescent BMIz variance being genetic (Elks et al., 2012), though the exact mechanisms by which genetic factors contribute to BMIz are not well understood. Research has found that eating behaviors predict weight and weight gain (Gallant et al., 2010; Pothos et al., 2009; Sung et al., 2009), are significantly heritable (Carnell et al., 2008; Keskitalo et al., 2008; Sung et al., 2010; Tholin et al., 2005), and share genetic influences with obesity (Keskitalo et al., 2008). Both eating behavior and obesity have been linked to brain structure in regions related to reward, caloric evaluation, taste, decision making, and inhibition, reflecting that brain structure and BMIz are linked through eating behavior, as either a cause or a consequence of overeating behaviors (Kennedy et al., 2016; Maayan et al., 2011; Su et al., 2017; van der Laan et al., 2016). Given that BMIz, eating behavior, and brain structure are heritable and phenotypically related (Blokland et al., 2012; Carnell et al., 2008; Elks et al., 2012; Jansen et al., 2015; Kennedy et al., 2016; Keskitalo et al., 2008; Maayan et al., 2011; Su et al., 2017; Sung et al., 2010; Tholin et al., 2005; van der Laan et al., 2016; van der Lee et al., 2017), it is reasonable to expect that there are shared genetic factors between BMIz and brain structure, possibly mediated by variance in eating behavior related to heritable brain structure. It is important to note that both heritability and genetic correlations between traits can change with age, particularly during development.

Understanding the heritable relationship between BMIz and brain structure during adolescence is particularly important as adolescence is a significant period in the development of the frontostriatal reward and inhibition regions (DePasque and Galvan, 2017; Galvan 2005; Sowell 2001) which are related to the overeating behaviors associated with obesity (Maayan et al., 2011; Su et al., 2017; van der Laan et al., 2016). Obesity is related to striatal volumes and response to food rewards (Kennedy et al., 2016; Stice et al., 2008) in a manner consistent with a reward deficit hypothesis that states individuals with a weaker reward response tend to overeat in order to reach a reward threshold (Blum 2014). Prefrontal regions involved in executive function and inhibition and insular regions involved in interoception have also been implicated in obesity and eating behavior (Kennedy et al., 2016; Kishinevsky et al., 2012; Le et al., 2007; Maayan et al., 2011; Pannacciulli et al., 2006; Tuulari et al., 2016). In particular, Kennedy et al (2016) found that BMIz was negatively related to gray matter volume in the medial prefrontal cortex/anterior cingulate, frontal pole, caudate, and uncus, as well as white matter volume in the anterior limb of the internal capsule, adjacent to the caudate. These regions are involved in decision-making, motivation, conflict monitoring, emotion processing, and reward processing (Gilbert et al., 2006; Kerns et al., 2004; Kim, 2013; Rushworth, 2008; Stice et al., 2008), suggesting that atypical forms of related behaviors may be involved in the overeating that leads to increased BMIz. Examining the relationship between these regions and obesity using a twin dataset can help inform to what extent BMIz and the brain structures linked to these behaviors are genetically related.

Another line of research into the etiology of obesity established heritable relationships between brain structure and obesity. In a diffusion tensor imaging (DTI) study of adult obesity in a sample of Mexican-American families, Spieker et al. (2015) found genetic correlations between average fractional anisotropy and obesity in the corpus callosum, internal capsule, thalamic radiations, and superior fronto-occipital fasciculus, ranging from −0.25 to −0.39. Surface and subcortical volume analyses found genetic correlations with cortical surface area in most lobes of the brain (range −0.013 to −0.385) and with regional subcortical volume (range −0.085 to −0.274). Curiously, the phenotypic correlations (i.e. Pearson’s) between obesity and brain structure were small (all below +/−0.1; Curran et al., 2013). Significance values were not reported for the surface and subcortical analyses, making interpretations difficult. These analyses were carried out in adults with a very wide age range (18-81 years old) and a lower obesity heritability (58%) than reported in adolescents. They also used summary measures (e.g. average surface area of the superior frontal gyrus), which are insensitive to local variation in what are often large structures encompassing multiple functionally distinct regions. It is important to note that estimates of heritability and genetic correlations in the above studies were based on family pedigree data, where genetic and common environmental factors are difficult to disentangle because the degree of gene sharing and environment sharing is strongly correlated within extended family structures (except special cases such as adoption).

Two major gaps in knowledge remain. First, while brain structure heritability, obesity heritability, and the association between brain structure and obesity have been separately examined (Blokland et al., 2012; Elks et al., 2012; Janowitz et al., 2015; Kennedy et al., 2016; Lenroot et al., 2009; Swagerman et al., 2014), it remains unclear whether genetic influences on brain volume and density significantly overlap with genetic influences on BMIz. Here we address this gap in knowledge by conducting a series of bivariate genetic analyses to estimate the degree of genetic overlap between BMIz and brain structure. Second, most of the previous research has been conducted in adult samples, whereas the frontostriatal BMIz-related brain structures involved in reward, decision-making, evaluation, and inhibition are largely shaped during adolescence, making adolescence a critical developmental stage for the study of the relationships between structural variation in the brain and BMIz. The present study addressed this problem by utilizing a sample of adolescent twins, allowing the estimation of shared genetics in a cohort where obesity is most heritable (Elks et al., 2012). It is not clear whether genetic influences on BMIz and brain structure are independent or shared, and whether phenotypic correlations between them can be attributed, at least in part, to overlapping genetic influences. Accordingly, the aim of this study was to address these questions through the assessment of genetic correlations between brain structure and BMIz using bivariate heritability analyses in a sample of adolescent twins.

We hypothesized that there are shared genetic influences for BMIz and brain regions implicated in eating behavior that account for a significant portion of the phenotypic relationship between BMIz and brain structure. Based on our previous study using similar methods in a sample of unrelated adolescents (Kennedy et al., 2016), we hypothesized that the volume of brain regions involved in regulation of eating behavior, including striatal, medial temporal, medial prefrontal, frontal pole, and anterior cingulate gray matter and peristriatal white matter volumes, would be negatively related to BMIz. Based on previous twin research, we expected that BMIz would have a high heritability (~80%; Elks et al., 2012). While we were not aware of previous voxel-based morphometry (VBM) heritability studies in adolescents, meta-analyses of adult regional structural volumes (as derived from lobular volume and FreeSurfer subcortical segmentations) suggested that the hypothesized regions should have univariate heritabilities in the 65-75% range (Blokland et al., 2012).

2. Materials and Methods

2.1. Participants

Monozygotic (MZ) and dizygotic (DZ) twin pairs were recruited through the Missouri Family registry. Participants were excluded based on history of neurological or psychiatric illness, physical or intellectual impairment, history of substance use, pregnancy, or non-removable metal in their body. The study protocol was approved by Washington University School of Medicine’s Human Research Protection Office. Informed consent was obtained from parents or legal guardians and an informed assent was obtained from child participants. The final sample used for these analyses consisted of 129 twin pairs (54 monozygotic and 75 dizygotic), ages 12 to 14 years old (mean(SD) 13.61(0.505) years old). Zygosity was determined using a modified version of the twin zygosity questionnaire (TZQ; Goldsmith, 1991) completed by parents of twins and included ratings of physical similarity and differences between twins, difficulty of telling the twins apart by family members, friends, and other parties. When TZQ scores were unavailable for both twins (two pairs), zygosity determination was based on an analogue scale measures of facial similarity completed by the study coordinator and research assistants who interacted with both twins during their laboratory visit. All twin pairs were the same sex, with 63 male (25 MZ) and 66 female (29 MZ) twin pairs.

2.2. BMI measurement

BMI z-score (BMIz) was calculated based on height, weight, and age obtained during a laboratory visit, typically the same day as the scan, using formulas derived from the 2000 Center for Disease Control growth charts (Kuczmarski et al., 2002). As child and adolescent obesity rates have risen since 2000 (Sassi et al., 2009), BMIz for this sample did not have a mean of 0. Z-scores derived by standardizing our sample’s BMIz values were used to identify outliers. Individuals with a standardized Z outside +/− 3 were considered outliers and they and their twin were removed from the dataset, resulting in two twin pairs being excluded. Post-hoc analyses including outliers did not change the phenotypic or genetic relationship between BMIz and brain structure. The mean(SD) BMIz was 0.608(1.013) and had a range of −2.36 to 2.76.

2.3. MRI Acquisition and Preprocessing

All participants were scanned on a Siemens 3T Prisma Fit at Washington University Medical School’s Neuroimaging Laboratories. A new navigator-guided T1 MPRAGE scan, developed to compensate for movement (Tisdall et al., 2012), was obtained using an axial acquisition in a 32-channel head coil array with a TR of 2500ms, TE of 2.88ms, TI of 1060ms, 176 slices, no gap, and field of view of 256mm. VBM preprocessing was carried out using Statistical Parametric Mapping 12’s Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (Ashburner, 2007; Friston et al., 2007) package to warp segmented whole brain gray and white matter volumes into 1mm³ Montreal Neurological Institute space. Spatial smoothing was carried out using an 8mm full width at half maximum. Images were inspected for failed warping, ultimately no one was excluded. Brain masks were created by averaging all subject’s tissue maps and filtering out voxels with values under 0.3; the gray matter mask was then manually edited to exclude some para-orbitofrontal non-gray matter.

2.4. Statistical Analyses

Preliminary statistical analyses were carried out using the Permutation Analysis of Linear Models (PALM; Winkler et al., 2014) package as it can control for familial non-independence. The relationship between BMIz and gray and white matter volume and density was evaluated while controlling for age, sex, ethnicity (Caucasian/non-Caucasian), the interaction between BMIz and age, and total tissue volume. Clusters surviving a Threshold Free Cluster Enhancement family wise error rate of 0.05 (1000 permutations) were used for further analyses examining univariate brain structural heritability.

Univariate heritability of BMIz (controlling for age, sex, ethnicity and the interaction between BMIz and age) was carried out using OpenMx (Neale et al., 2016) implemented in R-3.4.4. Models tested were additive genetics (A), non-additive genetics (D), and unique environment (E; ADE), A, shared environment (C), and E (ACE), A and E, C and E, and only E. Model comparisons were performed to identify the best fitting model. This was repeated for the total gray and white matter volumes. Voxelwise brain structure heritability was examined using the Accelerate Permutation inference for ACE models (APACE; Chen et al., 2014) script on the significant clusters from the PALM analyses. APACE takes the voxelwise heritability and compares it to a null created by randomizing zygosity 1000 times. All voxels surviving a False Discovery Rate of 0.05 were used for bivariate analyses.

Each significantly heritable voxel was extracted and entered into a bivariate heritability analysis with BMIz in a lower Cholesky model (see Figure 1) in OpenMx. This was done after controlling for the effects of age, sex, ethnicity and the interaction of BMIz and age on BMIz and the effects of age, sex, ethnicity, the interaction between BMIz and age, and total tissue volume on brain structure. Three models were tested: A and E with paths allowing common and unique genetics and error, A and E with no common genetics between BMIz and brain structure (AEdA21), and an E only model. The last two models were tested against the AE model in likelihood ratio tests. A false discovery rate of 0.05 was applied to the voxelwise AE vs. AEdA21 model to control for multiple comparisons when testing if there were significant shared genetics between BMIz and brain structure. Clusters greater than 100 voxels were reported here. For summary purposes, this was repeated on average voxel values for each cluster. This was repeated (save for multiple comparison corrections) to examine the shared genetics of BMIz, total gray matter volume, and total white matter volume.

Figure 1:

Bivariate additive genetic non-shared environment heritability model. MZ = monozygotic, DZ = Dizygotic. A/E11/22 = Unique genetics/non-shared environment on BMIz/VBM. A/E21 = Shared genetics/non-shared environment between BMIz and VBM.

Data from this ongoing project can be made available by request of the principal investigator (A.P.A.) after de-identification and documentation is completed. Upon the completion of the study, the data will be shared with the scientific community, pending availability of appropriate resources. All code can be supplied upon request and can be freely shared or reused. This is in compliance with the funding body and institutional review board.

3. Results

3.1. Regression Results

3.1.1. Volume

BMIz was positively related to gray matter volume in the brainstem and in the bilateral thalamus. BMIz was negatively related to gray matter volumes in the bilateral uncus, extending into the orbitofrontal cortex (OFC). BMIz did not show a significant relationship with white matter volume. See Table 1 and Figure 2 for more detail.

Table 1:

TFCE results for the association of BMIz with gray and white matter volume and density

Volume
Gray Matter
BMIz+
Area	k	X	Y	Z	TFCE p	Sig h2/TFCE	AEpD21/TFCE	AEpD21/Sig h2
Brainstem	1399	−4	−35	−13	0.019	0.77	0.77	1.00
L Thalamus	1152	−16	−28	2	0.037	1.00	1.00	1.00
R Thalamus	1919	9	−31	−3	0.029	0.71	0.71	1.00
BMIz−
Area	k	X	Y	Z	TFCE p	Sig h2/TFCE	AEpD21/TFCE	AEpD21/Sig h2
L Uncus/medial OFC	5897	−17	6	−32	0.001	0.30	0.30	1.00
R Uncus/medial OFC	5727	19	8	−35	0.001	0.17	0.17	1.00
Density
Gray Matter
BMIz−
Area	k	X	Y	Z	TFCE p	Sig h2/TFCE	AEpD21/TFCE	AEpD21/Sig h2
L Cerebellum	6448	−32	−58	−51	0.025	0.10	0.09	0.99
R Cerebellum	8392	32	−55	−51	0.024	0.20	0.20	1.00
R Uncus/OFC/inferior temporal/fusiform/ amygdala/hippocampus/parahippocampal/ inferior, middle and superior frontal/insula/ B cingulate/medial frontal/frontal pole	127507	15	48	−22	0.012	0.12	0.08	0.73
L Uncus/OFC/inferior temporal/insula/ supramarginal/pallidum/amygdala/ hippocampus	29978	−27	9	−35	0.023	0.20	0.16	0.82
R Inferior occipital	127	46	−80	2	0.046	0.00	0.00	-
White Matter
BMIz−
Area	k	X	Y	Z	TFCE p	Sig h2/TFCE	AEpD21/TFCE	AEpD21/Sig h2
Brainstem	1976	−2	−30	−15	0.019	0.55	0.22	0.40

Table 1: Regions significantly associated with BMIz. TFCE = Threshold Free Cluster Enhancement analyses. BMIz+/− = Structures positively/negatively associated with BMIz. k = Volume, XYZ = MNI peak coordinate, TFCE p = peak TFCE significance, Sig h2/TFCE = proportion of TFCE cluster voxels significantly heritable, AEpD21/TFCE = proportion of TFCE cluster voxels with significant shared genetics, AEpD21/Sig h2 = proportion of significantly heritable voxels with significant shared genetics. L = Left, R = Right, B = Bilateral. / = cluster covers multiple regions. OFC = Orbitofrontal cortex.

Figure 2:

Brain regions significantly related with BMIz. A: Gray and white matter density. B: Gray matter volume. Color scales indicate the significance size of the correlations. Yellow-red indicates significant positive correlations, whereas light blue to dark blue indicate negative correlations. Lighter colors indicate higher significance level (lower p-value).

3.1.2. Density

BMIz was negatively related to gray matter density in the cerebellum, scattered clusters in all lobes, the insula, cingulate, and some subcortical structures. BMIz was negatively related to white matter density in the brainstem. See Table 1 and Figure 2 for a full list of regions.

3.2. Univariate Heritability

The AE model had the best fit for BMIz with a heritability of 89%. For total gray matter volume, an ACE model showed the best fit (A² = 0.43, C² = 0.47, E² = 0.1, probability ACE is not a better fit than AE (p ACE < AE) = 0.002, p ACE < CE < 0.001, p ACE < E < 0.001). For total white matter volume, the AE model had the best fit (A² = 0.93, E² = 0.07, p ACE < AE = 0.074, p ACE < CE < 0.001, p ACE < E < 0.001). See Table 2 for more information. APACE analyses of clusters significantly associated with BMIz found significantly heritable brain gray matter volumes (36 to 89%) in parts of all regions associated with BMIz. Significantly heritable gray matter density (15 to 97%) was observed in most regions where a relationship between BMIz and brain structure were observed. White matter densities were heritable (56 to 97%) in the brainstem in the superior cerebellar peduncle and spinothalamic tract. See Table 3 and Figure 3 for the full list of significant regions.

Table 2:

Univariate heritability analyses

Analysis	ADE A2	ADE D2	ADE E2	ACE A2	ACE C2	ACE E2	AE A2	AE E2	CE C2	CE E2	rMZ	rDZ	LL_ADE	LL_ACE	LL_AE	LL_CE	LL_E
BMIz	0.66	0.23	0.11	0.90	0.00	0.10	0.90	0.10	0.60	0.40	0.89	0.40	640.67	641.08	641.08	682.12	738.68
Gray Volume	0.90	0.00	0.10	0.43	0.47	0.10	0.90	0.10	0.78	0.22	0.90	0.70	−1165.56	−1175.63	−1165.56	−1156.74	−1038.28
White Volume	0.93	0.00	0.07	0.65	0.28	0.07	0.93	0.07	0.74	0.26	0.92	0.62	−1194.88	−1198.07	−1194.88	−1159.99	−1058.87

Table 2: Standardized univariate heritability values for ADE, ACE, AE, and CE values. rMZ = monozygotic twin correlations, rDZ = dizygotic twin correlations. LL_ = −2 log likelihood values.

Table 3:

Significantly heritable brain regions

Volume
Gray Matter
BMI+
Area	k	X	Y	Z
R Thalamus	1365	9	−31	−3
L Thalamus	1152	−16	−28	2
Brainstem	1073	−4	−36	−16
BMI−
Area	k	X	Y	Z
L Uncus/medial OFC	1775	−18	7	−32
R Uncus/medial OFC	428	19	8	−32
R Temporal pole/OFC	320	28	14	−28
R Inferior temporal	224	23	−7	−49
Density
Gray Matter
BMI−
Area	k	X	Y	Z
B Anterior cingulate/medial prefrontal	3753	−2	28	−5
L Hippocampus/pallidum/posterior insula	3100	−16	−12	−14
R OFC/insula/temporal pole	2974	44	25	−4
L Insula	1533	−45	5	−13
R medial OFC/medial prefrontal/frontal pole	1408	13	49	−21
R Insula	1360	34	15	4
R Cerebellum	1267	31	−54	−49
R Lingual	971	27	−39	−12
R Fusiform/inferior temporal	879	24	−7	−49
L Cerebellum	618	−35	−55	−49
R OFC	587	33	39	−12
R Fusiform/inferior temporal	516	−27	−11	−48
R Middle cingulate	470	3	3	31
R Cerebellum	417	27	−58	−58
B Middle cingulate	395	3	−27	32
B Posterior cingulate	372	1	−39	27
L Anterior insula	336	−31	25	−1
R Fusiform	306	35	−55	−18
L Posterior cingulate	286	−10	−56	28
R Posterior operculum	272	54	12	13
L Inferior temporal	271	−44	−14	−38
L Medial superior frontal	190	−5	49	27
L OFC	166	−23	33	−14
R Hippocampus/parahippocampal gyrus	166	27	−26	−14
L Parietal operculum	149	−38	−27	21
R Medial superior frontal	138	10	40	41
R Medial superior frontal	135	13	41	40
White Matter
BMI−
Area	k	X	Y	Z
R Superior cerebellar peduncle	303	8	−35	−25
L Superior cerebellar peduncle	243	−3	−33	−13
L Superior cerebellar peduncle	208	−1	−32	−13
L Superior cerebellar peduncle	176	−5	−46	−24
R Spinothalamic tract	150	15	−22	−19

Table 3: Significantly heritable regions associated with BMIz. BMIz+/− = Structures positively/negatively associated with BMIz. k = Volume, XYZ = MNI peak coordinate. L = Left, R = Right, B = Bilateral. / = cluster covers multiple regions. OFC = Orbitofrontal cortex.

Figure 3:

Voxelwise heritability of brain structure related to BMIz. Color scales indicate the proportion of phenotypic variance attributable to genetic factors. The yellow-red scale indicates heritability of density, and the blue scale indicates heritability of volume. Lighter colors represent higher heritability.

3.3. Bivariate Heritability and Genetic Correlations

Bivariate heritability analyses found significant shared genetic influences between BMIz and brain structure. For the gray matter analyses, significant genetic correlations were observed in clusters in most regions that were both related to BMIz and heritable. The mean(SD) genetic correlation (rG) was 0.33(0.04) for gray matter volumes positively related to BMIz, −0.42(0.09) for volumes negatively related to BMIz, and was −0.3(0.07) for densities negatively related to BMIz. The mean(SD) percentage of the phenotypic correlation explained by shared genetics (%A) was 99(5.5)% for gray matter volumes positively related to BMIz, 94(6.8)% for volumes negatively related to BMIz, and 101(12)% for density (%A can exceed 100% if the variance due to error reduces the relationship between two variables). For white matter density, shared genetic influences were observed in the superior cerebellar peduncle with a mean(SD) rG of −0.25(0.02) and %A of 94(7.3)%. There were no voxels where the model with no genetics fit best. See Table 4 and Figure 4 for the full list of regions. BMIz was not related to total gray matter volume (probability the AE model (p AE) is a better fit than the AE model with no shared genetics (AE dA21) = 0.81) but was a trend for total white matter volume (p AE better fit than AE dA21 = 0.086). Total gray and white matter volume were significantly related (p AE better fit than AE dA21 < 0.001) with an rG of 0.71 and %A of 87.9%.

Table 4:

Regions with significant shared genetics between BMIz and gray or white matter volume or density

Volume
Gray Matter
BMI+
Area	k	x	y	z	SA2	LL_AE	LL_AEdA21	LL_E	pAEdA21	PE	rG	rE	%A	%E	SrMZ	SrDZ
R Thalamus	1365	9	−31	−3	0.67	−359.04	−345.99	−228.06	<0.001	<0.001	0.34	0.04	97.26	2.74	0.63	0.39
L Thalamus	1152	−16	−28	2	0.73	−345.87	−333.46	−214.69	<0.001	<0.001	0.33	0.00	99.94	0.06	0.68	0.31
PAG	1073	−4	−36	−16	0.66	−619.57	−594.39	−486.03	<0.001	<0.001	0.46	0.00	99.74	0.26	0.66	0.55
BMI−
Area	k	x	y	z	SA2	LL_AE	LL_AEdA21	LL_E	pAEdA21	pE	rG	rE	%A	%E	SrMZ	SrDZ
L Uncus/temporal pole	1775	−18	7	−32	0.64	−310.63	−288.03	−190.76	<0.001	<0.001	−0.47	−0.17	91.48	8.52	0.65	0.36
R Uncus/temporal pole	428	19	8	−32	0.61	−378.36	−352.15	−260.34	<0.001	<0.001	−0.51	−0.13	93.33	6.67	0.65	0.31
R Temporal pole	320	28	14	−28	0.62	−338.56	−320.58	−218.83	<0.001	<0.001	−0.43	−0.02	99.02	0.98	0.72	0.22
R Fusiform	224	23	−7	−49	0.57	−324.31	−306.27	−206.45	<0.001	<0.001	−0.44	0.00	99.88	0.12	0.58	0.29
Density
Gray Matter
BMI−
Area	k	x	y	z	SA2	LL_AE	LL_AEdA21	LL_E	pAEdA21	pE	rG	rE	%A	%E	SrMZ	SrDZ
L Hippocampus/pallidum/insula	2672	−16	−12	−14	0.66	−401.13	−388.95	−268.28	<0.001	<0.001	−0.33	0.07	105.21	−5.21	0.71	0.21
B Anterior cingulate/L medial prefrontal	2554	−2	28	−5	0.83	−802.19	−780.33	−635.36	<0.001	<0.001	−0.39	0.19	108.36	−8.36	0.85	0.25
R Insula/uncus	2250	44	25	−4	0.53	−321.02	−313.55	−209.55	0.006	<0.001	−0.31	0.06	106.21	−6.21	0.60	0.00
R Cerebellum	1262	31	−54	−49	0.73	−644.71	−637.46	−498.63	0.007	<0.001	−0.25	−0.12	90.49	9.51	0.78	0.24
R Medial OFC/frontal pole	1156	13	49	−21	0.66	−540.79	−534.50	−412.27	0.012	<0.001	−0.25	0.06	106.25	−6.25	0.68	0.23
L Insula/temporal pole	982	−45	5	−13	0.69	−539.70	−530.14	−404.75	0.002	<0.001	−0.29	−0.02	98.54	1.46	0.71	0.28
R Fusiform/inferior temporal	860	24	−7	−49	0.57	−513.65	−502.82	−392.49	0.001	<0.001	−0.35	0.11	110.97	−10.97	0.67	0.01
R Lingual/parahippocampal	859	27	−39	−12	0.67	−429.45	−421.37	−292.86	0.004	<0.001	−0.27	0.05	104.80	−4.80	0.72	0.30
L Cerebellum	612	−35	−55	−49	0.72	−315.82	−307.66	−175.57	0.004	<0.001	−0.27	−0.12	91.02	8.98	0.77	0.10
L Medial prefrontal	596	−4	56	−12	0.62	−655.29	−646.17	−531.51	0.003	<0.001	−0.31	0.02	101.62	−1.62	0.68	0.18
L Fusiform/inferior temporal	464	−27	−11	−48	0.64	−767.09	−744.27	−643.80	<0.001	<0.001	−0.47	0.04	102.33	−2.33	0.72	0.01
R Anterior insula	447	35	15	4	0.63	−580.61	−567.99	−459.81	<0.001	<0.001	−0.36	0.13	111.05	−11.05	0.63	0.19
R Cerebellum	417	27	−58	−58	0.62	−450.65	−441.78	−324.63	0.003	<0.001	−0.30	0.03	102.61	−2.61	0.63	0.19
B Middle cingulate	377	3	3	31	0.75	−596.90	−578.41	−456.11	<0.001	<0.001	−0.38	−0.02	98.87	1.13	0.75	0.29
L Anterior insula	336	−31	25	−1	0.74	−647.77	−633.27	−504.18	<0.001	<0.001	−0.34	−0.09	94.68	5.32	0.80	0.28
R Posterior insula	328	40	−18	18	0.68	−451.83	−444.71	−324.83	0.008	<0.001	−0.26	0.15	115.72	−15.72	0.73	0.11
R Fusiform	292	35	−55	−18	0.69	−600.26	−591.89	−458.94	0.004	<0.001	−0.27	−0.03	97.79	2.21	0.72	0.37
B Posterior cingulate	280	3	−27	32	0.80	−471.84	−459.20	−315.67	<0.001	<0.001	−0.31	−0.03	98.20	1.80	0.82	0.27
L Inferior temporal	267	−44	−14	−38	0.80	−692.90	−678.58	−528.03	<0.001	<0.001	−0.32	−0.15	92.71	7.29	0.84	0.34
B Posterior cingulate	262	1	−37	27	0.74	−670.98	−651.56	−523.86	<0.001	<0.001	−0.39	−0.01	99.61	0.39	0.79	0.32
R Inferior frontal	168	45	43	−12	0.54	−489.75	−479.10	−367.94	0.001	<0.001	−0.35	0.13	113.95	−13.95	0.62	0.14
L Posterior cingulate	144	−10	−56	28	0.83	−616.39	−598.11	−443.33	<0.001	<0.001	−0.35	0.04	101.78	−1.78	0.83	0.45
L Post Insula	143	−38	−27	21	0.71	−511.30	−502.23	−368.40	0.003	<0.001	−0.28	−0.02	98.42	1.58	0.76	0.27
R Hippocampus/parahippocampal	140	28	−29	−22	0.53	−634.55	−625.37	−520.62	0.002	<0.001	−0.33	−0.10	91.13	8.87	0.57	0.22
L Medial OFC	123	−23	33	−14	0.75	−713.37	−694.35	−569.74	<0.001	<0.001	−0.39	0.05	102.52	−2.52	0.79	0.26
White Matter
BMI-
Area	k	x	y	z	SA2	LL_AE	LL_AEdA21	LL_E	pAEdA21	pE	rG	rE	%A	%E	SrMZ	SrDZ
L Superior cerebellar peduncle	163	−3	−33	−13	0.67	−470.25	−462.92	−328.18	0.007	<0.001	−0.25	−0.12	89.84	10.16	0.72	0.34
R Superior cerebellar peduncle	161	8	−35	−25	0.59	−631.26	−624.33	−503.69	0.008	<0.001	−0.27	−0.08	91.79	8.21	0.65	0.18
L Superior cerebellar peduncle	112	−5	−46	−24	0.67	−620.60	−613.22	−483.09	0.007	<0.001	−0.26	0.01	100.81	−0.81	0.70	0.27

Table 4: Regions with significant shared genetics. BMIz+/− = Structures positively/negatively associated with BMIz. k = Volume, XYZ = MNI peak coordinate. SA2 = standardized heritability of the selected structure. LL_ = −2 log likelihood. pAEdA21 = probability the AE model fits better than the model with no shared genetics. pE = probability the AE model is better than the E model. rG = genetic correlation. rE = error correlation. SrMZ = Structure’s monozygotic twin correlation. SrDZ = Structure’s dizygotic twin correlation. L = Left, R = Right, B = Bilateral. / = cluster covers multiple regions. PAG = Periaqueductal gray, OFC = Orbitofrontal cortex.

Figure 4:

Genetic correlations between BMIz and brain structure. A: Gray and white matter density. B: Gray matter volume. Yellow-red = positive correlations, light blue-blue = negative correlations.

4. Discussion

4.1. Accuracy of Hypotheses and Replication of Previous Research

Our hypotheses that BMIz would be heritable as well as phenotypically and genetically related to brain structure in adolescents were partially supported. BMIz heritability was high, as were gray and white matter volumes and densities in many of the areas associated with BMIz. BMIz was significantly associated with gray matter volume in the uncus/orbitofrontal cortex and with gray matter density in the medial prefrontal, frontal pole, and anterior cingulate. While the expected caudate volume differences were not observed, BMIz was associated with pallidum density, another region involved in reward processing. Most heritable clusters shared genetic influences with BMIz. Our findings suggest that the observed phenotypic correlations between BMI and heritable brain structure can be attributed almost entirely (>90%) to common genetic influences.

Brain structural correlates of BMIz were largely consistent with previous research. Structural BMI analyses typically reported reduced volume or density in subcortical, insula, cerebellar, and prefrontal structures that were observed in our density analyses (Kennedy et al., 2016; Pannacciulli et al., 2006; Willette and Kapogiannis, 2015). The positive correlations with gray matter volume and negative with white matter density measures appeared to be novel. This may be due to the use of the permutation based Threshold Free Cluster Enhancement significance approach, as previous research used the random field theory based clustering approach which is less accurate (Eklund et al., 2016) and less sensitive to small, highly significant regions (Hayasaka and Nichols, 2003). VBM based volume and density heritabilities appeared similar or higher than volume heritabilities previously reported (van der Lee et al., 2017; Weise et al., 2017). This may be because previous VBM heritability research was based on adult samples while ours was from a younger age group and may be subject to developmental effects on heritability (Peper et al., 2009; van den Berg 2006). Genetic correlations between brain structure and BMIz were higher in our analyses than those previously reported (Curran et al., 2013; Spieker et al., 2015), possibly due to the previous study’s use of pedigree data which conflates genetic and common environmental effects and use of summary measures from different structural modalities.

The direction of causation in the relationship between brain structure and BMIz is difficult to resolve in a cross-sectional analysis. One plausible hypothesis is that genetically determined differences in brain structure contribute to individual differences in the regulation of eating behavior which in turn influences BMIz. This causal interpretation is indirectly supported by the fact that the majority of brain regions associated with BMIz in the present study have been previously linked to eating behaviors (Kishinevsky et al., 2012; Maayan et al., 2011; van der Laan et al., 2016). An alternative model would be that differences in BMIz lead to differences in brain structure, which is supported by evidence that weight loss surgery resulted in density recovery (Tuulari et al., 2016). The issue of causal relationships between BMIz and brain structure can be clarified by longitudinal or co-twin control studies.

4.2. Functional Associations of Structures Phenotypically Correlated with BMIz

Gray matter volume or density reductions were observed in structures in some way related to food consumption, including systems related to sensation, reward evaluation, and decision-making. As we did not evaluate eating behavior directly any inference on the behavioral significance of structural variances related to BMIz in these analyses is speculative. Cingulate and prefrontal volume reductions have been linked to increased self-reported dietary restraint, though restrainers generally do not eat less than non-restrainers and are likely to be overweight, suggesting self-reported restraint is more linked to intention than behavior (van der Laan et al., 2016). Reduced orbitofrontal cortex (OFC) volume has been linked to disinhibited eating behavior, where someone eats impulsively to environmental food cues (Maayan et al., 2011). Anterior cingulate cortex activity has also been negatively linked to food related disinhibition, but only in the obese (Martin et al., 2010). In cases where someone has both high dietary restraint and high disinhibition, people do not update their eating behaviors based on earlier food exposure, causing them to overeat when new eating opportunities appear (Westenhoefer et al., 1994). The volume and density reductions observed in our analyses in the prefrontal and cingulate cortices support the high food disinhibition/restraint model of overeating in obesity. We also observed structural decreases in the insula, hippocampus, and spinothalamic tract, which are involved in interoceptive awareness (Davidson et al., 2007; Kennedy and Dimitropoulos, 2014; Krames and Foreman, 2007; Rolls, 2016). This may result in reduced awareness of bodily signals of hunger and satiety, leading to overeating and disinhibition related eating to external food cues. The reduced pallidum and posterior cingulate density may also contribute to overeating as the pallidum is involved in controlling reward based on interoceptive need (Nangunoori et al., 2016, Smith and Berridge, 2005, Tindell et al., 2006) and the posterior cingulate has been linked to craving (Brewer et al., 2013) and emotional salience of food in a sensory specific satiety task (Small et al., 2001). Structural reductions here may inhibit the updating of reward values as people reach satiety. Observed reductions in the primary taste and smell cortices of the brain, namely the insula (Rolls, 2016) and uncus (Kiernan, 2012), may reflect reduced sensory input from consumption, which may lead to less sensory reward and overeating to reach a reward threshold. Uncus volume, obesity, and agentic positive emotionality, a personality measure linked to long-term goal pursuit have all been found to be significantly related to one another (Kennedy et al., 2016). Uncus volume and density reductions may represent problems maintaining the long-term restrictive behaviors necessary to maintain a successful diet.

Structural differences related to BMIz were also found in some regions not usually linked to eating behavior or where the observed relationship isn’t well understood. The frontal pole is involved in understanding emotional states (Gilbert et al., 2006). Alexithymia, or difficulty recognizing internal emotional states, is associated with problems regulating emotion (Pandey et al., 2011) and predicts emotional eating in obese women with a binge-eating disorder (Pinaquy et al., 2012). Eating is one method of emotional regulation (Macht, 2008); difficulty processing emotional states related to atypical frontal pole structure may lead to problems in emotion regulation, in turn resulting in overuse of food as a regulatory tool. The parahippocampus is functionally associated with food anticipation, with hunger level correlating with the strength of activation (Stice et al., 2013). The cerebellum is primarily known for its role in movement (Manto, 2012) but has also been linked to meal anticipation (Mendoza, 2010). Disruptions to meal anticipatory regions may result in overeating through excess snacking. The fusiform responds to satiation (Hinton, 2004) and is more active to food cues in people with the obesity linked FTO gene (Kuhn, 2016). Reduced fusiform density may inhibit satiation processing, leading to overeating. Obesity related disruptions to white matter in the cerebellar peduncle have been linked to impaired motor competence (Augustijin et al., 2017; Verstynen et al., 2012). People with poor motor competence may struggle with physical activity, creating a disincentive towards exercise, promoting a positive caloric imbalance through inactivity rather than overeating. The periaqueductal gray (PAG) works with the raphe magnus in inhibiting pain responses during ingestion (Mason, 2011). Atypical PAG activity may facilitate overeating by masking feelings of discomfort caused by overeating related gut distention. Again, as we lacked extensive cognitive and behavioral measures, these interpretations are speculative but may provide avenues for future research.

4.3. Heritability and Shared Genetics

Most brain structures phenotypically related to BMIz were significantly heritable and shared significant genetics with BMIz. Genetic correlation analyses suggest that BMIz and heritable brain volume/density associated with BMIz share around one third of their genes, and that around 95% of the observed phenotypic correlation between BMIz and brain structure can be attributed to these shared genetic influences. Our findings suggest that there is a significant common genetic underpinning between BMIz and most of the heritable brain structures related to BMIz. While structural reductions in adulthood are at least partially attributed to the deleterious effects of an obesogenic diet (Mueller et al., 2012; Walther et al., 2010), these results suggest that there are brain regions where BMIz related decreases in adolescent volume or density are genetic.

4.4. Significance and Summary of Findings

These analyses were significant in several ways. They extended previous structural BMIz research to density analyses where adolescent analyses were lacking, established structural reductions in areas important to eating behavior, established that brain structure in several of these regions is significantly heritable, and that the phenotypical relationship in these regions is largely due to genes shared by BMIz and brain structure. To our knowledge, this is the first bivariate heritability analysis of VBM derived brain structure and BMIz, and one of a very few bivariate neuroimaging analyses known to the author not examining summary values (Chen et al., 2011; Hulshoff Pol et al., 2006; Rao et al., 2018; Rimol et al., 2010). Examining neuroimaging data at the vertex or voxel level is important as summary measures ignore local variation in structure and may overlook significant associations spread out across multiple contiguous regions (Betjemann et al., 2010; Glahn et al., 2010; Glahn et al., 2013; Greenspan et al., 2016; Jahanshad et al., 2013; Karlsgodt et al., 2010; Kochunov et al., 2016; Patel et al., 2017; Pinel and Dehaene, 2013; Spieker et al., 2015); important as there is functional variance within regions and trait-associated variance does not necessarily conform to the anatomical boundaries typically used for segmentation or parcellation.

Overall, these analyses suggested there are shared genetic factors between BMIz and brain structure underlying multiple neurobehavioral domains tied to eating behavior. Heritable structural differences in areas related to sensation, caloric and reward evaluation, decision-making, and reward processing could contribute to obesogenic behaviors through impaired inhibition, eating absent hunger, or overeating due to reduced sensation related to interoception or taste. Almost all of the phenotypic relationship between BMIz and these regions is due to genetic factors. This suggests that there is a genetic component to many of the observed neurostructural differences associated with adolescent obesity, though lacking longitudinal data, we can’t tell if these genetic factors cause altered brain structure leading to obesity or cause obesity, leading to altered brain structure. As the analyses presented here were limited to the relationship between BMIz and brain structure, omitting cognitive, behavioral, etc. measures, the interpretations presented here are speculative.

Limitations and future directions

These analyses were limited in some ways. Lacking an extensive consumptive behavior/cognition battery, our interpretations of the possible ways the observed structural differences might manifest were highly speculative. We also lacked sufficient power to examine more complex bivariate ACE or ADE models that would have allowed us to explore shared environment or non-additive genetic effects; however, the univariate BMIz analyses suggested that there was not any shared environment or non-additive effects to analyze. Furthermore, we had to use a voxelwise correction for our bivariate heritability significance rather than use the permutation-null comparison significance method used in APACE as processing the 1000 minimum permutations required would have taken approximately a year.

Post-hoc Analyses

Adult BMI has been linked to increased in-scanner movement (Siegal et al., 2017), meaning structural differences attributed here to BMIz may instead reflect movement artifacts. Post-hoc analyses were conducted on the average density or volume values extracted from the clusters that were genetically linked to BMIz, adding as a covariate mean movement parameters from resting state scans obtained in the same session. Movement and BMIz were significantly related (partial r = 0.359, p < 0.001, controlling for age, sex, ethnicity, and the interaction between BMIz and age), however the relationship between BMIz and brain structure remained significant in all clusters after the addition of mean movement as a covariate. This sample included individuals who were underweight (below a Z of −1.64 or <5^th percentile) which is associated with structural differences separate from those found in obesity (Phillipou et al., 2018). All regions with a significant genetic correlation had both significant phenotypic and genetic correlations following post-hoc analyses where three twin pairs that included an underweight member were removed.

Future research using longitudinal data should clarify the direction of causality in the relationship between brain structure and BMIz and test the hypothesis that this relationship is mediated by eating behavior. One important prediction suggested by the present findings is that both brain structure (in the regions identified here) and BMIz would show significant genetic correlations with disordered eating behaviors.

5. Conclusions

These analyses suggested that both BMIz and brain structures related to eating behavior in adolescents are influenced by overlapping genetic factors. This evidence for a genetic predisposition suggests that if a child has a family history of obesity, they may benefit from interventions to compensate for the associated neurobehavioral differences, even if obesity has not manifested. This may take the form of mindfulness training to focus on the sensation of their food and interoceptive signals, interventions to focus on the long-term ramifications of their eating behavior, or digitally assisted calorie counting to compensate for poor internal caloric evaluation. Mindfulness and calorie counting have demonstrated some efficacy in altering eating behavior or promoting weight loss in people already obese (Alberts et al., 2012; Katterman et al., 2014; Mason et al., 2016; Mendes et al., 2017). As the health consequences of early life obesity can persist after weight loss (Reilly and Kelly, 2011), it is important to address these behaviors early.

Figure 5:

Literature-based eating and obesity related functions of regions phenotypically and genetically linked to BMIz in these analyses. Significant volume and density clusters in a glass brain color coded by relevant function. Striped clusters associated with multiple relevant functions, colors in striped clusters refer to cluster as a whole, not specific colored region. Black clusters (pulvinar, inferior temporal) have no known functional relevance.