Modeling the 3D Geometry of the Cortical Surface With Genetic Ancestry

Summary

Knowing how the human brain is shaped by migration and admixture is a critical step in studying human evolution [1, 2], as well as preventing the bias of hidden population structure in brain research [3, 4]. Yet the neuroanatomical differences engendered by population history are still poorly understood. Most of the inference relies on craniometric measurements, because morphology of the brain is presumed to be the neurocranium’s main shaping force before bones are fused and ossified [5]. Although studies have shown that the shape variations of cranial bones are consistent with population history [6–8], it is unknown how much human ancestry information is retained by the human cortical surface. In our group’s previous study, we found that the area measures of cortical surface and total brain volumes of European descendants in the United States correlate significantly with their ancestral geographic locations in Europe [9]. Here, we demonstrate that the 3-dimensional geometry of cortical surface is highly predictive of individuals’ genetic ancestry in West Africa, Europe, East Asia, and America, even though their genetic background has been shaped by multiple waves of migratory and admixture events. The geometry of the cortical surface contains richer information about ancestry than the areal variability of the cortical surface, independent of total brain volumes. Besides explaining more ancestry variance than other brain imaging measurements, the 3D geometry of the cortical surface further characterizes distinct regional patterns in the folding and gyrification of the human brain associated with each ancestral lineage.

Results

The participants were recruited as part of the Pediatric Imaging, Neurocognition, and Genetics (PING) study. A detailed overview of the study can be found in previous publications (e.g., [3, 4, 10]). Research protocols and data are publicly available online [11]. Briefly, PING was a multi-site project recruiting children and adolescents from ages 3 to 21 at 10 sites in the United States. All participants were screened for history of major developmental, psychiatric, or neurological disorders; brain injury; or other medical conditions that affect development. Participants then received neurodevelopmental assessments, standardized multimodal neuroimaging, and genome-wide genotyping. The overall PING sample consisted of 1,493 participants; 1,152 individuals remained after quality control of the genotyping and neuroimaging data (for quality control processes and demographics of the participants, see Supplemental Information and Table S1). We focused our analyses on 562 individuals older than 12 years (289 males, mean age: 16.6 years, standard deviation: 2.6 years). Considering that the morphological features of cortical surface change little after age 12 [10], this stratified approach further reduced the residual confounds of developmental effects.

The proportions of genetic ancestry were estimated using principal component (PC) analysis with whole-genome single nucleotide polymorphism (SNP) reference panels for ancestry [12–14]. Four continental populations were used as ancestral references: West Africa (YRI, as Yoruba in Ibadan), Europe (CEU, as Utah residents with northern and western European ancestry), East Asia (EA), and America (NA, as America natives). The metrics for summarizing genetic ancestry in each ancestral component were standardized as proportions, ranging from 0% to 100%. These proportions represent how similar an individual is to the reference population genetically [14].

Morphological Prediction for Genetic Ancestry

We first tested whether the surface geometry of the cerebral cortex predicted the proportion of genetic ancestry among participants. To characterize variation in the geometry, we reconstructed the cortical surfaces from all individuals’ T1-weighted scans, then represented the positions of the corresponding surface vertices using standard 3-D Cartesian coordinates. The reconstruction and registration processes ensure that each vertex on the reconstructed cortical surface is located in a homologous position with respect to the curvature patterns for individuals [15, 16]. Taking the coordinates of all vertices as a whole, we then have information about shape variation of the cortical surface, including aspect ratios, sulcal depth, and gyrification. The prediction models were fit with ridge regression while treating gender, age, age squared, total brain volumes, and the scanner on which the image data were acquired, as nuisance covariates. The model performance was evaluated using leave-one-out cross-validation (LOOCV).

As Figure 1 shows, the geometry of the cortical surface has good predictive value for each of the ancestry components. The variances explained by the models are 66% for ancestry in YRI, 55% for ancestry in CEU, 49% for ancestry in EA, and 47% for ancestry in NA. To determine to what degree the geometric differences reflect variation in area expansion of cortical surface, comparable models were computed using vertex-wise surface area (Table 1). Also, to examine possible roles in the prediction of simpler morphological attributes, such as aspect ratios of the cerebrum and volumes of subcortical structures, we conducted comparable analyses predicting ancestry from these measures. None has as much information about ancestry as the geometry of cortical surface (Table 1).

Figure 1. Predicting the proportion of genetic ancestry by cortical surface geometry.

YRI: Yoruban, as the West Africa ancestry. CEU: Utah residents with northern and western European ancestry. EA: East Asia. NA: America natives. In all predictive models, the variables have been residualized with respect to the age, age squared, gender, total brain volumes, and scanner used. All models excluded individuals with a 0% proportion of genetic ancestry to that specific component. LOOCV: leave-one-out cross-validation. The colors of the data points are determined by the proportion of genetic ancestry as illustrated in the figure legend.

Table 1.

Percentage of Variance Explained in Different Predictive Models

	Cortical Surface Geometry	Cortical Surface Area	Brain Aspect Ratios	Subcortical Volumes
YRI	66%	17%	10%	5%
CEU	55%	12%	2%	2%
EA	49%	9%	6%	6%
NA	47%	9%	9%	0%

Cortical surface geometry and cortical surface area are sampled in icosahedral level 4, which contains 642 vertices in each hemisphere. All models are fit with the same setting and evaluated with leave-on-out cross validation. Nuisance covariates, including gender, age, age squared, total brain volumes, and scanner, were regressed out before calculating the variance explained in LOOCV.

Characterization of the Cortical Shape Morphs

We then reconstructed the 3D geometry of the cortical surface based on the linear relationship we observed between cortical surface geometry and the proportion of genetic ancestry. This allowed us to visualize how the geometry of the cortical surface changes as a function of increasing proportion of genetic ancestry in each ancestral component. The morphing of 3D cortical surfaces from neutral ancestry (25% of genetic ancestry in all four components) to 100% ancestry in each component is demonstrated in Figure 2 (for dynamic morphing of surface geometry, see movies S1 to S4 in the online materials). As Figure 2 illustrates, the textural contrasts between regions of the cortical surface indicate that the morphing process has complex, unique patterns for each ancestral component, while the intensity varies from region to region. For example, as the proportion of the YRI component increases, the temporal surfaces move posteriorly and inward. The proportion of the CEU component is associated with protrusion of the occipital and frontal surfaces. Increases in the proportion of the EA component are accompanied by variations in temporal-parietal regions. The NA component is associated with flattening of the frontal and occipital surfaces.

Figure 2. Color-coded morphing process of the 3D geometry of the cortical surface.

The still image illustrates how each vertex on the cortical surface morphs from an ancestry-neutral 3D cortical surface (a 25% proportion of genetic ancestry in all ancestral components) to a 3D cortical surface with a 100% proportion of genetic ancestry in a specific ancestral component. The morphing coefficients were estimated from the PING sample. Here, the colors represent the direction of the morphing process. Moving along the medial-lateral axis is coded in red, along the anterior-posterior axis in green, along the dorsal-ventral axis in blue. The final color is the combination of these three, depending on which direction the vertices move. For each viewing perspective, the coloring frame of reference is rendered on the top of each column. The length of each morphing line is the actual distance between two 3D cortical surfaces. For dynamic morphing animations, see movies S1 to S3 in the online materials.

Figure 3 summarizes the mean magnitudes and variations of the morphing in each cortical surface region defined by genetic correlations [17]. The mean magnitudes vary from cortical region to cortical region, corresponding to the description above. In addition, YRI, EA, and NA all have relatively high magnitude and variations of morphing in the posterolateral-temporal region.

Figure 3. Mean magnitude and variations of morphing across 12 regions of cortical surface.

Labeled in the topmost images, the following regions are defined in a previous publication [17]: 1. central region, 2. occipital cortex, 3. posterolateral-temporal region, 4. superiorparietal region, 5. orbitofrontal region, 6. superiotemporal region, 7. inferiorparietal region, 8. dorsomedialfrontal region, 9. anteromedial-temporal region, 10. precuneus, 11. dorsolateral-prefrontal cortex, 12. parsopercularis. The Euclidean distances between cortical surface of 100% ancestry and neutral ancestry were calculated for each vertex. Depending on the surface regions where the vertices are situated, the mean and standard deviations of the Euclidean distances are shown in the box plots.

Discussion

Our data indicate that the unique folding patterns of gyri and sulci are closely aligned with genetic ancestry. The geometry robustly predicts each individual’s genetic background even though the population has been shaped by waves of migration and admixtures [12, 18]. Previously, only modeling of facial features has achieved 64% of explained variance in the YRI ancestry among African Americans [19]. Our 3D representation of cortical surface geometry performs similarly in predicting YRI ancestry and also performs well for the other three continental ancestries. As data in Table 1 show, the explanatory power is not due to the differences in total brain volumes, nor to the differences in areal expansion of the cortical surface. Instead, regional folding patterns characterize each ancestral lineage.

On the other hand, the global shapes of the reconstructed cortical surface geometry match W. W. Howells’ description on craniometry of 2,524 ancient human crania from 28 populations [20]. Crania of African ancestry tended to have a narrower cranial base, and those of Northern European ancestry had elongated occipital and frontal regions. Crania of East Asian ancestry had a high cranial vault, and those of Native American ancestry had a flatter cranium. Regarding the morphing differences of YRI, EA, and NA, all had high magnitude and variations in the posterior-temporal regions (Figure 3).These findings are consistent with the notion that temporal bones contain more variations across ancestral groups [6].

At first glance, these results are surprising because our model is based on the contemporary United States population, which is the historical product of migrations, slave trades, and local admixture events [18, 21, 22]. Nevertheless, the coordinates of reference-inferred PC space reflect information about individuals’ ancestral origins (Figure S1) [14, 21, 23]. Our group’s previous study also showed that the individuals’ positions in the PC space are matched with their ancestral locations, rather than their current geographic locations [9]. Therefore, our 3D representation might, to a certain degree, reflect the neuroanatomical and/or neurocranial changes along the human migratory path in the dispersal from Africa [24]. Based on our current model, we simulated what might be expected from the Out-of-Africa scenario in the Supplemental Information (Figure S3). More precise characterization of an individual’s ancestral origins would require more complex estimates of ancestry based on global-scale reference panels [25]. Further understanding of neuroanatomical change along the Out-of-Africa scenario based on brain imaging data will require future studies using sampling methods similar to the Human Genome Diversity Project [26].

It is important to note that these ancestry-related geometric features of the cortical surface are not substantially attributable to variation in cortical surface area. Previous studies of ancient crania often interpreted the shape differences as evidence of relative size alterations of different cortical functional domains[5, 27]. Our results suggest that in the case of the contemporary population, the differences in cortical surface geometry might not reflect variation in the relative surface area of different functional cortical regions. In prior studies, regionalization of the cortex has been linked to cognitive differences in humans [3, 4]. Any functional significance of the cortical surface geometry, per se, remains to be established. The effects reported here might be mediated by neutral drift of the phenotypic variations [28]. They can also result from a complex interaction between the brain and neurocranium, with the former expanding while the latter acts as physical resistance. Nevertheless, the causal relationships between the observed shapes and crania are beyond the scope of our current study.

An implication of our ancestry-related 3D models is that, unless properly controlled, hidden population structures could present a challenge in brain imaging studies of admixed populations [23]. The regional differences between ancestral groups include changing sulcus depths and folding angles. This issue becomes particularly relevant in large, multi-site U.S. and international brain imaging studies [29]. With the advent of inexpensive, high-throughput genotyping, it is now possible to control for spurious effects due to ancestry admixture using genetically derived admixture factors in the statistical analysis of data [3, 4]. It is also possible that the phenomena we observed are linked with specific ancestral haplotypes. It may therefore be possible to use the ancestral information to improve statistical power for gene discovery with methods such as admixture mapping [30].