The individuality of shape asymmetries of the human cerebral cortex

  1. Yu-Chi Chen  Is a corresponding author
  2. Aurina Arnatkevičiūtė
  3. Eugene McTavish
  4. James C Pang
  5. Sidhant Chopra
  6. Chao Suo
  7. Alex Fornito
  8. Kevin M Aquino
  9. for the Alzheimer's Disease Neuroimaging Initiative
  1. Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Australia
  2. Monash Biomedical Imaging, Monash University, Australia
  3. Monash Data Futures Institute, Monash University, Australia
  4. Healthy Brain and Mind Research Centre, Faculty of Health Sciences, Australian Catholic University, Australia
  5. Department of Psychology, Yale University, United States
  6. BrainPark, Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Australia
  7. School of Physics, University of Sydney, Australia
  8. Center of Excellence for Integrative Brain Function, University of Sydney, Australia
  9. BrainKey Inc, United States

Abstract

Asymmetries of the cerebral cortex are found across diverse phyla and are particularly pronounced in humans, with important implications for brain function and disease. However, many prior studies have confounded asymmetries due to size with those due to shape. Here, we introduce a novel approach to characterize asymmetries of the whole cortical shape, independent of size, across different spatial frequencies using magnetic resonance imaging data in three independent datasets. We find that cortical shape asymmetry is highly individualized and robust, akin to a cortical fingerprint, and identifies individuals more accurately than size-based descriptors, such as cortical thickness and surface area, or measures of inter-regional functional coupling of brain activity. Individual identifiability is optimal at coarse spatial scales (~37 mm wavelength), and shape asymmetries show scale-specific associations with sex and cognition, but not handedness. While unihemispheric cortical shape shows significant heritability at coarse scales (~65 mm wavelength), shape asymmetries are determined primarily by subject-specific environmental effects. Thus, coarse-scale shape asymmetries are highly personalized, sexually dimorphic, linked to individual differences in cognition, and are primarily driven by stochastic environmental influences.

Editor's evaluation

The article is of interest to scientists who study neuroanatomy or the many behavioral phenotypes that have been proposed be associated with left–right asymmetry of the human brain. The methodology is sophisticated and rigorously applied.

https://doi.org/10.7554/eLife.75056.sa0

Introduction

Asymmetries in brain structure and function are found throughout the animal kingdom (Duboc et al., 2015; Güntürkün et al., 2020; Corballis and Häberling, 2017; Güntürkün and Ocklenburg, 2017) and can be discerned at multiple spatial scales, ranging from differences in the size and shape of the cerebral hemispheres through measures of regional morphometry and connectivity to cellular and molecular organization (Güntürkün et al., 2020; Güntürkün and Ocklenburg, 2017; Esteves et al., 2021). At the coarsest scale, the most salient feature of anatomical asymmetry in the human brain is cerebral torque, in which the right hemisphere appears to be warped in the rostral direction relative to the left hemisphere (Li et al., 2018; Toga and Thompson, 2003; Zhao et al., 2021). More fine-grained asymmetries of specific sulci/gyri (Kang et al., 2015) and brain regions Kong et al., 2018; Plessen et al., 2014 have also been investigated. For example, the superior temporal sulcus, which is adjacent to the Wernicke’s area, shows a leftward asymmetry in length (Gómez-Robles et al., 2013).

Asymmetries in brain organization are often considered at an average level across a population of individuals (Toga and Thompson, 2003; Kong et al., 2018; Kong et al., 2022; Deep-Soboslay et al., 2010; Postema et al., 2019). These population-based asymmetry features have been studied extensively and are thought to have important implications for both functional lateralization and abnormal brain function in a wide range of psychiatric and neurological diseases (Esteves et al., 2021; Plessen et al., 2014; Postema et al., 2019; Cai et al., 2015; Fling et al., 2014). For example, the planum temporale of the left hemisphere, which encompasses Wernicke’s area, has been consistently shown to be larger than the right for most healthy individuals (Güntürkün et al., 2020; Toga and Thompson, 2003; Royer et al., 2015; Takao et al., 2011), and patients with schizophrenia often show reduced leftward asymmetry in planum temporale compared to healthy individuals (Clark et al., 2010; Ratnanather et al., 2013; Corballis, 2013). However, many findings with respect to asymmetries of specific brain regions have been inconsistent in terms of the directions and magnitudes of asymmetry observed (Kong et al., 2018; Plessen et al., 2014; Kong et al., 2022). The correlates of these asymmetries are also unclear (Güntürkün and Ocklenburg, 2017; Plessen et al., 2014; Kong et al., 2022; Kurth et al., 2018; Núñez et al., 2018). For example, two fundamental characteristics often examined in relation to cerebral asymmetry are sex and handedness. Some studies have found that the surface area (Kong et al., 2018), shape (Núñez et al., 2018; Wachinger et al., 2015; Kovalev et al., 2003), volume (Guadalupe et al., 2015), and torque (Zhao et al., 2021) of cortical structures in males are more asymmetric than in females, whereas other studies have found no sex differences (Takao et al., 2011; Narr et al., 2007). Similarly, some studies have found associations between cerebral asymmetry and handedness (Zhao et al., 2021; Deep-Soboslay et al., 2010; Steinmetz et al., 1991), with others reporting no such effect (Kong et al., 2018; Plessen et al., 2014; Wachinger et al., 2015; Narr et al., 2007; Good et al., 2001; Guadalupe et al., 2014; Maingault et al., 2016).

Some of these inconsistencies may arise from the disparate methodologies and the heterogeneous nature of the brain asymmetries across the population (Toga and Thompson, 2003; Kong et al., 2018; Gómez-Robles et al., 2013; Kong et al., 2022; Deep-Soboslay et al., 2010; Postema et al., 2019). Despite some consistent asymmetry features across the population (Güntürkün et al., 2020; Toga and Thompson, 2003; Royer et al., 2015; Takao et al., 2011), there is also considerable individual variability around population means, with many people often showing little or even reversed asymmetries relative to the prevalent pattern of the population (sometimes also referred to as antisymmetry) (Corballis and Häberling, 2017; Gómez-Robles et al., 2013; Gómez-Robles et al., 2016; Neubauer et al., 2020). The distinction between population-level and individual-specific asymmetries is essential as they are thought to arise from distinct mechanisms (Gómez-Robles et al., 2016; Sherwood and Gómez-Robles, 2017). Populational-level asymmetries are hypothesized to have a genetic basis (Zhao et al., 2021; Kong et al., 2018; Gómez-Robles et al., 2016; Neubauer et al., 2020; Sherwood and Gómez-Robles, 2017; Francks, 2015; de Kovel et al., 2018; Graham and Özener, 2016; Sha et al., 2021), whereas individual-specific asymmetries, which describe the way in which a given individual departs from the population mean, may reflect environmental influences, developmental plasticity, or individual-specific genetic perturbations (Gómez-Robles et al., 2016; Neubauer et al., 2020; Sherwood and Gómez-Robles, 2017; Francks, 2015; de Kovel et al., 2018; Graham and Özener, 2016; Nadig et al., 2021). Notably, cortical asymmetries of the human brain are more variable across individuals than other primates at both regional and global hemispheric levels (Gómez-Robles et al., 2013; Neubauer et al., 2020). The variability is most evident in regions of heteromodal association cortex, leading some to conclude that high levels of variability in asymmetry may have emerged in line with the evolution of human-specific cognition (Gómez-Robles et al., 2013), although the relationship between the asymmetries of the human brain and individual differences in cognition is still largely unknown.

Traditional analysis methods, which rely on standard image processing techniques such as image registration and spatial smoothing, minimize individual variation and thus have limited sensitivity for studying individual-specific asymmetries (Gomez-Robles et al., 2018; Wachinger et al., 2016). Moreover, most past studies have focused on morphological properties related to the size of specific brain regions, such as estimates of gray matter volume, cortical thickness, or surface area, often measured at fine-grained resolutions, such as individual voxels or the vertices of cortical surface mesh models (Cai et al., 2015; Takao et al., 2011; Kurth et al., 2018; Good et al., 2001; Maingault et al., 2016; Kruggel and Solodkin, 2020; Kurth et al., 2015). Many of the most obvious features of cerebral asymmetry arise from variations in brain shape, which are not captured by size-related descriptors (Wachinger et al., 2015; Reuter et al., 2009). Indeed, it is possible for two objects to have identical volume but have very different shapes (Reuter et al., 2009; Ge et al., 2016). In addition, shape variations can occur at different spatial resolution scales, from the presence and configuration of specific sulci at fine scales to more global patterns such as cerebral petalia at coarser scales. Conventional analyses only consider the finest resolvable scale (i.e., point-wise differences) and have limited sensitivity for identifying important morphological variations that occur over large swathes of cortical tissue.

A comprehensive, multiscale description of cortical shape, from the finest to coarsest scales, can be derived through a spectral analysis of cortical geometry based on solutions to the Helmholtz equation (Wachinger et al., 2015; Reuter et al., 2009; Reuter et al., 2006), which is fundamental in many branches of physics, engineering, chemistry, and biology (Lévy, 2006). The equation can be solved by formulating it as an eigenfunction–eigenvalue problem of the Laplace–Beltrami operator (LBO) (see ‘Materials and methods’). Importantly, the characteristics of the eigenfunctions and eigenvalues depend on the cortical shape for which the equation is solved (Reuter et al., 2006; Lévy, 2006), and thus, the spectral analysis provides a comprehensive description of the intrinsic geometry of a given object, akin to a ‘Shape-DNA’ (see ‘Materials and methods’; Reuter et al., 2006). The application of such Shape-DNA analysis to human magnetic resonance imaging (MRI) data has shown that shape properties of cortical and subcortical structures have superior sensitivity compared to traditional, size-based measures for identifying individual subjects (Wachinger et al., 2015), classifying and predicting the progress of psychiatric and neurological diseases (Wachinger et al., 2016; Richards et al., 2020), and detecting genetic influences on brain structure (Ge et al., 2016; Wachinger et al., 2018). However, a detailed characterization of individual-specific asymmetries in cerebral shape is lacking.

Here, we introduce methods for constructing an individual-specific measure of cortical asymmetry, called the shape asymmetry signature (SAS; see ‘Materials and methods’). The SAS characterizes pure shape asymmetries of the whole cortical surface, independent of variations in size, across a spectrum of spatial scales. We apply this methodology to three independent longitudinal datasets to test the hypothesis that cortical shape asymmetry is a highly personalized and robust feature that can identify individuals, akin to a cortical asymmetry fingerprint. We then use the identifiability values to identify optimal spatial scales at which robust individual differences are most salient. We also compare the identifiability of the SAS and shape descriptors of individual hemispheres, asymmetries in traditional size-based descriptors, or patterns of inter-regional functional connectivity (so-called connectome fingerprinting; Finn et al., 2015) to test the hypothesis that the SAS is a more individually unique property of brain organization than unihemispheric and functional properties. We further elucidate the relationships between the SAS and sex, handedness, as well as cognitive performance across multiple tasks. Finally, we test the hypothesis that individual-specific asymmetry features are largely driven by environmental influences using classical heritability modeling of twin data.

Results

Cortical shape asymmetries are individually unique

To understand how cortical shape asymmetries vary across individuals, we examined the degree to which different cortical shape descriptors (defined below) can be used to identify individual brains from a large sample of T1-weighted MRIs. We analyzed healthy subjects from three open-source datasets – the latest Open Access Series of Imaging Studies (OASIS-3; LaMontagne et al., 2019), the Human Connectome Project (HCP; Van Essen et al., 2013), and the Alzheimer’s Disease Neuroimaging Initiative (ADNI; https://ida.loni.usc.edu/) – in which individuals had at least two anatomical MRI scans acquired at different time points (separated by 1 day to several years; see ‘Materials and methods’). For each dataset, we asked whether the shape descriptors for an individual estimated from the first scan could accurately identify the same participant’s second scan. Within each dataset, the shape descriptor was calculated from the cortical surfaces at the white and gray matter boundary estimated either from FreeSurfer (Fischl et al., 2002) (OASIS-3 and ADNI) or FreeSurfer-HCP (HCP), which is a FreeSurfer (Fischl et al., 2002) pipeline with some HCP-specific enhancements (Glasser et al., 2013). Shape-DNA (Reuter et al., 2009; Reuter et al., 2006) analysis was employed to obtain multidimensional shape descriptors for each hemisphere that quantify the shape of each individual’s cortex, as defined by the eigenvalue spectrum of the LBO (Figure 1A and B; see ‘Materials and methods’). Each eigenvalue is associated with a corresponding eigenfunction, which describes shape variations at a particular spatial wavelength, ordered from coarse to fine-grained scales (Figure 1B). These eigenfunctions are orthogonal by construction and thus represent a basis set for cortical shape variations much like the sinusoidal basis used in Fourier decomposition of signals, with the corresponding eigenvalue being analogous to the wave frequency at each spatial scale. Critically, we normalized the surface area (Reuter et al., 2009) of the meshes prior to Shape-DNA analysis to ensure that the resulting eigenvalue spectra were independent of individual differences in brain size (see ‘Materials and methods’).

Figure 1 with 1 supplement see all
Schematic of our analysis workflow.

(A) The shapes of the left and right hemispheres are independently analyzed using the Laplace–Beltrami operator (LBO) via the Shape-DNA algorithm (Reuter et al., 2009; Reuter et al., 2006). (B) Eigenfunctions of the LBO are obtained by solving the Helmholtz equation on the surface, given by Δf=λf, where f corresponds to a distinct eigenfunction, and λ is the corresponding eigenvalue. Each eigenvalue λi,i=1,2,,1000, quantifies the degree to which a given eigenfunction is expressed in the intrinsic geometry of the cortex. Higher-order eigenvalues describe shape variations at finer spatial scales. (C) The shape asymmetry signature (SAS) is quantified as the difference in the left and right hemisphere eigenvalue spectra, providing a summary measure of multiscale cortical shape asymmetries. To investigate the identifiability of the SAS, we use Pearson’s correlation to calculate the similarity between the SAS vectors obtained for the time 1 (t1) and time 2 (t2) two scans from the same individuals (diagonal elements of the matrices) as well as the correlation between t1 and t2 scans between different subjects (off-diagonal elements). We estimate identifiability by first correlating the initial two eigenvalues, then the initial three eigenvalues, and so on to a maximum of 1000 eigenvalues. Here, we show examples of correlation matrices obtained when using the first 10, 50, 144, and 1000 eigenvalues, and the cortical surface reconstructions show the shape variations captured by corresponding spatial scales. (D) Repeating the identifiability analysis up to a maximum of 1000 eigenvalues yields a curve with a clear peak, representing the scale at which individual differences in cortical shape are maximal. For the SAS, this peak occurs when the first 144 eigenvalues are used (black dashed line), which offers a fairly coarse description of shape variations (see panel C). We then use a similar analysis approach to investigate associations between scale-specific shape variations and sex, handedness, cognitive functions, as well as heritability. The data in this figure are from the OASIS-3 (n = 233) cohort, and the cortical surfaces are from a population-based template (fsaverage in FreeSurfer).

To investigate the uniqueness of these shape descriptors to individual brains, we performed an identifiability analysis (Amico and Goñi, 2018; Mansour L et al., 2021), where identifiability was quantified as the degree to which the surface eigenvalue spectrum of an individual at scan time 1 was more similar to the same person’s spectrum at time 2, relative to other people’s time 2 spectra (Figure 1C; see also ‘Materials and methods). To determine whether identifiability is maximized at any specific scales, we repeated the analysis multiple times, initially by taking only the first two eigenvalues, which describe shape variations at the coarsest scale, and then incrementally adding eigenvalues representing more fine-grained features to a maximum of 1000. Plotting the identifiability score as a function of the number of eigenvalues allows us to identify characteristic spatial scales at which the identifiability score is maximized (Figure 1D). In other words, it allows us to identify the scales at which individual-specific shape features are most pronounced. We repeated this procedure using the eigenvalue spectra for the left and right hemispheres alone, the combination of both (which describes the shape of both hemispheres), and for the SAS, which quantifies shape asymmetries as the difference between the left and right hemisphere eigenvalue spectra (see Figure 2 for details). Finally, we utilized the spatial scales with maximum identifiability (Figure 1D) to examine the relationships between the SAS and sex, handedness, cognition, and heritability. In general, a brain with a higher degree of shape asymmetry has SAS values that more strongly depart from zero (Figure 1—figure supplement 1).

Figure 2 with 3 supplements see all
Identifiability of different shape descriptors at different spatial scales.

(A–C) Identifiability scores for shape features across eigenvalue indices. The identifiability scores of the shape asymmetry signature (SAS) are generally higher than the scores for shape descriptors of individual hemispheres or scores obtained when concatenating both hemispheres across three datasets (OASIS-3: n = 233; ADNI: n = 208; HCP test–retest: n = 45). The SAS scores are also much higher than the scores obtained by randomly shuffling the order of the subjects at time 2 (shaded area represents mean ± 2 SDs). (D–F) The cumulative mean identifiability scores for each eigenvalue group, derived from correspondence with spherical harmonics (Robinson et al., 2016). The peak mean identifiability occurs at the 11th eigenvalue group for the OASIS-3 (D) and ADNI data (E), representing the first 144 eigenvalues. The curve of the mean identifiability score for the HCP test–retest data (F) flattens after the 11th group and peaks at the 16th group. (G) Cortical surfaces reconstructed at different spatial scales, starting with only the first eigen-group (E1) and incrementally adding more groups to a maximum of the first 12 eigen-groups (E12). (H) Overhead view of the spatial scale corresponding to the eigen-group at which identifiability is maximal in the OASIS-3 and ADNI datasets (i.e., the first 11 eigen-groups, corresponding to the first 144 eigenvalues).

Figure 2A–C shows the identifiability scores obtained for the different shape descriptors. In all three datasets, across a broad range of spatial scales, identifiability was highest for the SAS, followed by the combination of left and right hemisphere eigenvalues, and then each hemisphere alone. This result indicates that individual variability in the asymmetry of cortical shape is greater than the variability of shape across the whole cortex or within each hemisphere alone. Figure 2A–C also shows identifiability scores obtained when trying to identify an individual’s left hemisphere using right hemisphere shape descriptors obtained at the same time point. These scores are very low, indicating that shape variations between the two hemispheres are largely independent of each other and lack a consistent pattern amongst subjects. In other words, for any given person, the shape of one hemisphere offers little individually unique information about the shape of the other hemisphere.

Individually unique variations of cortical shape asymmetry are maximal at coarse spatial scales

We next investigated the scale-specificity of SAS identifiability. Figure 2A–C shows that SAS identifiability sharply increases to a peak as we use more eigenvalues to characterize the surface shape at finer scales (i.e., as we add more shape information from finer spatial scales), before gradually falling again. This peak identifies a characteristic spatial scale in which individual differences in shape asymmetries are maximally unique (see also Figure 2—figure supplement 1).

Peak SAS identifiability was observed using the first 126 and 122 eigenvalues for the OASIS-3 (Figure 2A) and ADNI (Figure 2B) data, respectively. At these scales, the subject identifiability scores were 4.93 (p=0; estimated by permutation; see ‘Statistical analysis’ section for details) for OASIS-3 and 5.03 (p=0) for ADNI. For the HCP data, peak SAS identifiability was observed when using the first 268 eigenvalues (identifiability score = 6.74; p=0; Figure 2C), but the identification curve flattened after the first 137 eigenvalues (identifiability score = 6.56), which is closely aligned with the OASIS-3 and ADNI datasets.

In the case of a perfect sphere, the shape spectral analysis yields subsets of degenerate eigenvalues with equal magnitude (Robinson et al., 2016), within which the corresponding eigenfunctions represent orthogonal rotations of the same spatial pattern at a given scale. For example, eigenfunctions 2–4 of a sphere represent coarse-scale gradients in the anterior–posterior, inferior–superior, and left–right axes. As the cortex is topologically equivalent to a sphere, the spherical eigen-groups offer a natural way to identify characteristic spatial scales, more succinctly summarize cortical shape variations (Robinson et al., 2016), and smooth out eigenvalue-specific fluctuations at a given scale (see ‘Materials and methods’). We averaged the identifiability scores for each harmonic group and plotted these as a function of the group index in Figure 2D–F. The group mean identifiability score peaks at the 11th eigenvalue group for the OASIS-3 (mean identifiability score = 4.93) and ADNI (mean identifiability score = 5.06) datasets, which is comprised of the first 144 eigenvalues. Identifiability also reaches a near-plateau at the 11th group for the HCP data (mean identifiability score = 6.47), with an additional marginal increase observed at the 16th group (mean identifiability score = 6.69). Thus, the first 144 eigenvalues represent a stable and robust characteristic scale at which individual uniqueness in cortical shape asymmetry is strongest. The 11th group corresponds to a wavelength of approximately 37 mm in the case of the population-based template (fsaverage in FreeSurfer; Supplementary file 1 shows the corresponding wavelengths of the first 14 eigen-groups; Figure 2G shows the spatial scales corresponding to the cumulative eigen-groups).

A reconstruction of the cortical surface using the first 144 eigenfunctions is shown in Figure 2H. The reconstruction captures shape variations at a coarse scale, representing major primary and secondary sulci, but with minimal additional details. If we include additional eigenfunctions to capture more fine-scale anatomical variations, inter-session image differences increase, suggesting that finer spatial scales may be capturing dynamic aspects of brain structure that are more susceptible to increased measurement noise (Figure 2—figure supplement 2). This same characteristic scale was obtained after repeating the identifiability analysis over the longest inter-scan intervals in the ADNI and OASIS-3 datasets (Figure 2—figure supplement 3), indicating that our results are robust over time windows ranging from 1 day to more than 6 years.

Shape asymmetries are more identifiable than classical morphological and functional measures

We next compared the identifiability of the SAS to scores obtained using asymmetries in classical morphological descriptors such as regional surface area, cortical thickness, and gray matter volume, and measures of inter-regional functional connectivity (Figure 3), which have previously been shown to yield high identifiability (Finn et al., 2015; Amico and Goñi, 2018). Identifiability scores obtained with the SAS were much higher than those obtained by regional asymmetries in size-related morphological measures with the HCP-MMP1 atlas (Glasser et al., 2016; Figure 3A and B). We also found that SAS identifiability was higher when using our surface area normalization procedure compared to the SAS computed without this procedure (Figure 3—figure supplement 1; see ‘Materials and methods’). Since the normalization isolates the pure effects of shape independent of brain size, the results converge to indicate that individual variability in brain anatomy is higher when considering asymmetries in cortical shape compared to more traditional size-based morphological descriptors.

Figure 3 with 1 supplement see all
Cortical shape asymmetries are more identifiable than size-related descriptors or functional connectivity.

(A) Identifiability scores for the shape asymmetry signature (SAS) are higher than those obtained for asymmetries based on cortical surface area (identifiability score = 0.81), volume (identifiability score = 0.66), and thickness (identifiability score = 0.33) for the OASIS-3 dataset (n = 232; see ‘Materials and methods’). (B) Matrices of the Pearson correlation coefficients for SAS and size-based morphological asymmetries from MRI scans taken at different time points (t1 and t2) of the OASIS-3 subjects. (C) SAS identifiability is higher than the identifiability based on functional connectivity, assessed with parcellations at different resolution scales in the HCP test–retest dataset (n = 44). (D) Matrix of the Pearson correlation coefficients for SAS of the HCP test-retest subjects. (E) Four resolution scales of parcellations used in the functional connectivity analysis (shown on an inflated fsaverage surface in FreeSurfer). (F) Matrices of the Pearson correlation coefficients for functional connectivity using the Schaefer 100 (identifiability score = 1.57), HCP-MMP1 (identifiability score = 2.06), Schaefer 300 (identifiability score = 2.11), and Schaefer 900 (identifiability score = 2.69) parcellations.

Figure 3C–F compares the identifiability scores obtained from the SAS to those obtained using inter-regional functional connectivity (see ‘Materials and methods’), within the HCP test–retest data. Functional connectivity was quantified for the entire cortex using four different regional parcellations defined at different spatial scales (Schafer 100, Schaefer 300, Schaefer 900 [Schaefer et al., 2018], and HCP-MMP1 [Glasser et al., 2016] atlas). The SAS outperformed all functional identifiability scores, indicating that cortical shape shows greater specificity to an individual than patterns of functional connectivity.

Cortical shape asymmetries are related to sex but not handedness

Sex and handedness are two characteristics that have frequently been examined in relation to brain asymmetry (Güntürkün et al., 2020; Toga and Thompson, 2003; Kong et al., 2018; Plessen et al., 2014; Wachinger et al., 2015; Guadalupe et al., 2015; Narr et al., 2007; Good et al., 2001; Kong et al., 2021). We used a general linear model (GLM) with permutation testing and accounting for familial structure (Winkler et al., 2016; Winkler et al., 2015) of the HCP data to evaluate the association between these two characteristics and the SAS defined at each eigenvalue ranging between the 2nd and 144th. After false discovery rate (FDR) correction, males and females showed significant differences in asymmetry scores for the 2nd (PFDR = 0.037), 6th (PFDR = 0.037), 8th (PFDR = 0.039), 52nd (PFDR = 0.030), and 84th (PFDR = 0.037) eigenvalues (Figure 4A), where female brains showed more rightward asymmetry than males in these eigenvalues. These five eigenvalues come from four different eigen-groups, and the corresponding spatial scales of these eigenvalues are shown in Figure 4B. These eigenvalues relate to shape variations over coarse scales. For instance, for the 2nd eigenvalue (L = 1; see ‘Materials and methods’ for the definition of L), the wavelength is of order 300 mm, which is about half the circumference of the brain; for the most-fine grained eigenvalue, the 84th eigenvalue (L = 9), the wavelength is about 44 mm. We note however that the sex differences are small, with considerable overlap between male and female distributions (Figure 4A). No such effects of handedness on the SAS surpassed the FDR-corrected threshold. We also found that the overall asymmetry level (i.e., the sum of the SAS) was not correlated with either handedness or sex.

Sex differences in eigenvalue asymmetries.

(A) Smoothed distributions and boxplots with mean and interquartile range (Allen et al., 2019) of the eigenvalues among males (n = 504) and females (n = 602). Under these five spatial scales, female brains show a greater rightward asymmetry than males. The p-values are false discovery rate (FDR)-corrected values of the correlation between sex and shape asymmetry signature (SAS), obtained via a general linear model (GLM). The d values are effect sizes (Cohen’s d). L denotes eigen-group. (B) The corresponding eigenfunction of each eigenvalue in panel (A) that shows the gradients of spatial variation on a population-based template.

Individual differences in cortical shape asymmetry correlate with cognitive functions

We used canonical correlation analysis (CCA) (Winkler et al., 2020) to examine associations between the SAS and 13 cognitive measures from the HCP dataset (n = 1094; see ‘Materials and methods’) selected to represent a wide range of cognitive functions (Kong et al., 2019; see ‘Materials and methods’ for details). To reduce the dimensionality of the SAS measures and ensure equivalent representation of asymmetries at each spatial scale, we took the mean SAS value for each of the 1st to 11th eigen-groups, spanning the 2nd to 144th eigenvalues. To minimize collinearity of the cognitive variables, we applied principal component analysis (PCA) to the 13 cognitive measures and retained the first four principal components (PCs), which explained 80% of the variance. The analysis revealed a single statistically significant canonical mode (CCA r = 0.187; PFWER = 0.032; Figure 5A). Figure 5B shows that the mode has significant positive loadings from mean SAS scores in eigen-groups 2, 4, 5, and 11, and significant negative loadings from eigen-groups 3, 6, 7, and 10. Figure 5C indicates that 12 of the 13 cognitive measures showed significant positive correlations with the canonical variate, indicating that it captures covariance with general cognitive ability. Thus, our findings identify strong scale-specificity of associations between cortical shape asymmetry and cognition, with a greater leftward asymmetry in scales captured by eigen-groups 2 (~170 mm wavelength), 4 (~95 mm wavelength), 5 (~75 mm wavelength), and 11 (~37 mm wavelength) being associated with better performance across most cognitive measures, and a greater leftward asymmetry in scales captured by eigen-groups 3 (~120 mm wavelength), 6 (~65 mm wavelength), 7 (~55 mm wavelength), and 10 (~40 mm wavelength) being associated with poorer cognitive performance.

Individual differences in cortical shape asymmetry correlate with general cognitive ability.

(A) Scatterplot of the association between the cognitive and shape asymmetry signature (SAS) canonical variates with the corresponding least-squares regression line in black. (B) Canonical variate loadings of each eigen-group. (C) Correlations between the original cognitive measures and the cognitive canonical variate. Error bars show ±2 bootstrapped standard errors (SE). Asterisks denote bootstrapped PFDR < 0.05. The data in this figure are from the HCP dataset (n = 1106).

Cortical shape asymmetries are primarily driven by unique environmental influences

To characterize genetic and environmental effects on cortical shape and its asymmetry, we calculated the heritability of each eigenvalue within the left and right hemispheres, as well as for the SAS. We used data from 138 monozygotic (MZ) twin pairs, 79 dizygotic (DZ) twin pairs, and 160 of their non-twin siblings drawn from the HCP dataset (Van Essen et al., 2013; see ‘Materials and methods’ for details). Unihemispheric shape descriptors demonstrated strong heritability at very coarse spatial scales and moderate heritability at slightly finer scales. For instance, the heritability of the first eigen-group (second to fourth eigenvalues) of both hemispheres ranged between 0.52 < h2 < 0.69 (all PFDR < 0.05; Figure 6A and B). These eigenvalues are related to shape variations on the coarsest scale that does not include any sulcal or gyral features (the corresponding wavelength is approximately 170 mm). Beyond the second eigen-group, heritability estimates dropped to below 0.5 (PFDR < 0.05 for most eigenvalues), and beyond the fourth eigen-group they dropped below 0.3. Most eigenvalues with statistically significant heritability estimates were confined to the first six eigen-groups, which correspond to wavelengths greater than or equal to approximately 65 mm (Figure 6A and B, insets). These results indicate that genetic influences on the shape of each cortical hemisphere are expressed over very coarse scales at which only primary cortical folds such as the Sylvian and central sulci are apparent. Estimates of common environmental influence on both hemispheres were uniformly low across the 2nd to 144th eigenvalues (range 0–0.20).

Figure 6 with 2 supplements see all
Heritability of cortical shape.

(A, B) Heritability of the eigenvalues of the left (A) and right (B) hemispheres. The insets show the corresponding spatial scales by reconstructing the surfaces using the first six eigen-groups. (C) Heritability of the shape asymmetry signature (SAS). The inset shows the corresponding spatial scale with some level of genetic influence, obtained by reconstructing the surface using the first five eigen-groups. (D) Unique environmental influences to the SAS at each eigenvalue. Statistical significance is evaluated after false discovery rate (FDR) correction. Note that significance is not estimated for unique environmental effects as this represents the reference model against which other genetically informed models are compared. We use 79 same-sex dizygotic (DZ) twin pairs, 138 monozygotic (MZ) twin pairs, and 160 of their non-twin siblings.

In contrast to unihemispheric shape variations, all the heritability estimates of the SAS were low (<0.28; Figure 6C), with only four eigenvalues (2, 3, 16, and 28) showing statistically significant heritability after FDR correction (PFDR = 0.004–0.022) and no heritability estimates exceeding 0.30. Thus, at any given scale, genes account for less than one-third of the phenotypic variance in the SAS. These four eigenvalues are confined to the first five eigen-groups, with corresponding wavelengths greater than or equal to approximately 75 mm (Figure 6C, inset). Estimates of common environmental influences were uniformly low (range 0–0.14), whereas unique (subject-specific) environmental influences on the SAS were consistently high across the full range of eigenvalues considered, ranging between 0.72 and 1.00 (Figure 6D).

Notably, heritability estimates for non-surface area normalized eigenvalues of individual hemispheres, which capture variations in both shape and size, were uniformly high across all scales, and the scale-specific effects were eliminated (Figure 6—figure supplement 1), indicating that variations in cortical size are under greater genetic influence than cortical shape. These results underscore the importance of controlling for size-related variations in shape analyses.

Discussion

Asymmetries in brain anatomy are widely viewed as a critical characteristic for understanding brain function. Here, we employed a multiscale approach to quantify individualized shape asymmetries of the human cerebral cortex. We found that cortical shape asymmetries were highly personalized and robust, with shape asymmetries at coarse spatial scales being the most discriminative among individuals, showing differences between males and females, and correlating with individual differences in cognition. Heritability estimates of shape descriptors in individual hemispheres were high at very coarse scales but declined to moderate values at finer scales. By contrast, the heritability of cortical shape asymmetry was low at all scales, with such asymmetries being predominantly influenced by individual-specific environmental factors.

Identifiability of cortical shape asymmetry is maximal at coarse scales

Cortical asymmetries have traditionally been investigated at fine-scale, voxel, or vertex-level resolutions (Cai et al., 2015; Takao et al., 2011; Kurth et al., 2018; Good et al., 2001; Maingault et al., 2016; Kruggel and Solodkin, 2020; Kurth et al., 2015). These approaches may ignore meaningful effects (i.e., properties that are individually unique and correlated with cognition) at coarser spatial scales. Our SAS quantifies these underlying variations across the whole brain and along a spectrum of spatial scales. Our approach is akin to studying seismic waves of earthquakes with different wave frequencies at the global tectonic scale, instead of focusing on a particular city. The ability to assess shape along a spectrum of spatial scales is important since brain asymmetry is a multidimensional and multivariate phenotype (Corballis and Häberling, 2017; Kong et al., 2022; Kruggel and Solodkin, 2020).

Few studies have assessed individual variations in shape at coarse scales. Neubauer et al., 2020 found that individual-specific asymmetry in endocranial shape is reliable across two time points. The endocranial shape is the imprint of the cortical surface shape but contains only very coarse shape information (Neubauer et al., 2020). Moreover, levels of brain torque (both horizontal and vertical) are robust across time (Kong et al., 2021). Wachinger et al., 2015 used shape descriptors at coarse scales derived from the eigenvalues of the LBO for all brain structures to achieve accurate subject identification. Taken together with our findings, these results indicate that coarse features of cortical shape are highly personalized and unique to individuals.

It is perhaps surprising that individual differences in cortical shape are most strongly expressed at coarse scales, given the known variability of fine-grained anatomical features such as the presence and trajectories of tertiary sulci. It is possible that local changes in gray matter volume affect fine-scale geometry in such a way that it carries less identifying information, or that such fine scales carry too much measurement noise to be used for the purpose of identification. Traditional analysis methods use smoothing to address the issue of image noise (Kurth et al., 2015), but smoothing can also suppress actual variations at fine scales. Our multiscale approach affords a more comprehensive characterization of shape variations across multiple spatial scales. An important avenue of future work will involve investigating the functional consequences of these pronounced individual differences.

Cortical shape, rather than shape asymmetry, is heritable

Genetic effects on cortical thickness and surface area are generally bilateral (Kong et al., 2018; Chen et al., 2011; Chen et al., 2012; Chen et al., 2013), resulting in few lateralized effects (Kong et al., 2018; Eyler et al., 2014). Accordingly, it has been postulated that individual-specific asymmetries may be largely determined by environmental factors (Gómez-Robles et al., 2016; Sherwood and Gómez-Robles, 2017; Francks, 2015; de Kovel et al., 2018; Graham and Özener, 2016). In line with this hypothesis, we found that individualized cortical shape asymmetries were associated with low heritability and were driven mainly by unique environmental effects. These environmental effects are captured by the E parameter of the ACTE heritability model that we used in our analysis. This parameter also includes the contributions of measurement error. However, our effects cannot be explained by the SAS being a noisier descriptor of morphology as it showed the highest identifiability (Figure 2A-C). A noisy measure will not be able to discriminate between individuals in this way. Thus, taking the findings of our identifiability and heritability analyses together, we can conclude that individual differences in SAS scores are primarily driven by unique environmental influences rather than measurement error.

Previous studies have found some evidence of environmental influences on brain asymmetry (Güntürkün et al., 2020; Güntürkün and Ocklenburg, 2017; Esteves et al., 2021; Felton et al., 2017). Early in the intrauterine environment, fetal posture and light may influence brain asymmetry (Güntürkün et al., 2020; Güntürkün and Ocklenburg, 2017; Esteves et al., 2021), and during postnatal maturation, language learning has been linked to specific asymmetry features. For example, bilinguals have stronger rightward asymmetry of cortical thickness of the anterior cingulate than monolinguals (Felton et al., 2017). However, the mechanisms of how environmental effects shape brain asymmetry are largely unknown, and epigenetics may also play a role (Güntürkün et al., 2020; Güntürkün and Ocklenburg, 2017).

In contrast to shape asymmetries, the shape of individual hemispheres showed greater heritability at coarse scales, consistent with results from previous studies on other morphological measurements (Kong et al., 2018; Sha et al., 2021; Kruggel and Solodkin, 2020). The scales at which genetic effects on unihemispheric shape were observed captured variations in primary sulci, consistent with evidence that the primary folds, which develop early in life, are less variable across individuals and under greater genetic control than other folds (i.e., secondary and tertiary folds) (Kruggel and Solodkin, 2020; Ronan and Fletcher, 2015; Kruggel, 2018). Previous studies have found that genetic influences on the cerebral thickness, geodesic depth, and surface curvature generally occur along the posterior–anterior and inferior–superior axes (Kruggel and Solodkin, 2020; Valk et al., 2020). These two axes correspond to the second and third eigenvalues of the LBO, which also showed strong heritability in the shapes of both hemispheres in our results. In addition to these two axes, we found strong heritability at very coarse scales in other directions that have not been described in previous studies. Our approach thus identifies dominant spatial scales and gradients of heritability in shape.

Shape asymmetries, sex, and handedness

Using our multiscale approach, we did not find a relationship between shape asymmetry and handedness, consistent with numerous studies showing that handedness is unrelated to anatomical brain asymmetry in cortical thickness, volume, surface area, shape, and voxel-based morphometric (VBM) analysis (Kong et al., 2018; Plessen et al., 2014; Núñez et al., 2018; Wachinger et al., 2015; Narr et al., 2007; Good et al., 2001; Maingault et al., 2016).

Numerous studies, focusing primarily on size-related descriptions such as gray matter volume and cortical thickness, have found that female brains are more symmetric than male brains (Zhao et al., 2021; Kong et al., 2018; Núñez et al., 2018; Wachinger et al., 2015; Kovalev et al., 2003; Guadalupe et al., 2015). Our analysis reveals that, although the overall level of shape asymmetry did not differ between male and female brains, female brains displayed a greater rightward shape asymmetry than male brains at certain coarse spatial scales, such as along the anterior–posterior axis. The mechanisms giving rise to these scale-specific sex differences require further investigation.

Shape asymmetries are correlated with general cognitive performance

We found that individual differences in cortical shape asymmetry are correlated with cognitive performance in a scale-specific way. Specifically, we found that a greater leftward asymmetry across a wide range of spatial scales, corresponding to wavelengths of approximately 37, 75, 95, and 170 mm, and greater rightward asymmetry at wavelengths of approximately 40, 55, 65, and 120 mm, are associated with better performance across nearly all cognitive measures considered. Previous studies have found that asymmetries in cortical thickness and surface area are negatively correlated with cognition (Nadig et al., 2021; Yeo et al., 2016), but these studies only measured the level of asymmetry and did not consider the direction (i.e., leftward or rightward) of the asymmetry. The scale-specificity of the associations that we find underscores the importance of viewing brain asymmetry as a multiscale rather than a unidimensional trait.

The magnitudes of the associations are modest, but they are consistent with effect sizes reported in past research (Nadig et al., 2021; Yeo et al., 2016). These modest correlations with cognition may reflect a robustness of cognitive abilities to stochastic perturbations of brain morphology, given that our heritability analysis revealed a dominant effect of unique environmental factors in driving individual differences in cortical shape asymmetries.

Conclusion

We developed a multiscale approach and found that cortical shape asymmetries are robust and personalized neuroanatomical phenotypes, especially at coarse spatial scales. Some of these coarse scales are more strongly rightward asymmetric in females compared to males. The cortical shape asymmetries also show scale-dependent associations with cognition. Finally, individual-specific cortical shape asymmetries are driven mainly by subject-specific environmental influences rather than by genetics, contrasting with the shape of individual hemispheres, which shows strong heritability at coarse scales.

Materials and methods

Neuroimaging data

We used healthy subject data from three open-source neuroimaging datasets: the latest OASIS-3 (LaMontagne et al., 2019), the HCP (Van Essen et al., 2013), and the ADNI (https://ida.loni.usc.edu/) to develop and test our new asymmetry shape measure – the SAS (see below for details). To test for relationships of sex, handedness, and heritability, we restricted our analysis to the HCP dataset, which provides twin and non-twin sibling information and handedness measurement as a continuous variable, as the sample sizes of the left-handers in the other two datasets are too small (n = 15 in the ADNI data; n = 18 in the OASIS-3 data).

OASIS-3

Request a detailed protocol

We used 239 healthy participants with at least two longitudinal MRI sessions using 3T scanners from the latest release of the OASIS-3 (LaMontagne et al., 2019). We excluded six subjects whose SAS was an outlier in at least one of those sessions due to poor image quality and major errors in image segmentation. These subjects had more than two eigenvalues of the first 200 eigenvalues that departed from the population mean values by more than four standard deviations. The remaining 233 subjects (99 males; 134 females) were aged from 42 to 86 (mean = 66.03; standard deviation = 8.81) when they entered the study. We also repeated the analyses using all the subjects including the outliers, and the resulting number of eigenvalues with peak identifiability was identical to the initial analysis that excluded the outliers. For comparing the identifiability of the SAS and the asymmetry from traditional measurements (volume, cortical thickness, and surface area), we further excluded one subject because some of this subject’s files were corrupted and could not be segmented. For subjects with more than two MRI sessions (n = 115), our main analysis used the initial session as the time 1 (t1) session and the session closest in time to the initial session as the time 2 (t2) session. The intervals between these two sessions were one to 3151 days (mean = 2.95 years; standard deviation = 1.67 years). To ensure the robustness of our methods, we used sessions with the longest intersession interval (mean interval of 6.24 years; standard deviation of 1.88 years) to reanalyze the subject identifiability. These healthy participants had no history of neurological or psychiatric diseases. We also excluded subjects with a Mini-Mental State Examination (MMSE) score equal to or lower than 26 as this indicates that a subject is at risk of being diagnosed with dementia (O’Bryant et al., 2008).

OASIS-3 (LaMontagne et al., 2019) provides surface meshes based on the T1-weighted MRI images created by FreeSurfer version 5.3 with the cross-sectional pipeline (i.e., to treat the T1 and T2 sessions independently; Fischl et al., 2002), including the FreeSurfer patch (10 December 2012) and the HCP patch (http://surfer.nmr.mgh.harvard.edu/pub/dist/freesurfer/5.3.0-HCP; LaMontagne et al., 2019). A trained lab member of the OASIS project reviewed the image segmentation, and for the images that failed the quality control, TkMedit (http://freesurfer.net/fswiki/TkMedit), a FreeSurfer toolbox, was used to revise the images and rerun the FreeSurfer pipeline (LaMontagne et al., 2019). After the re-segmentation, the images were excluded if they still failed a quality control process (LaMontagne et al., 2019). The details of the OASIS-3 dataset can be found in LaMontagne et al., 2019 and the OASIS website (https://www.oasis-brains.org/). We used the actual output files provided by the OASIS-3 without any further corrections.

HCP

Request a detailed protocol

We used participants from the HCP (Van Essen et al., 2013) s1200 release (https://www.humanconnectome.org/), which includes 1113 subjects with T1-weighted MRI. All subjects of the s1200 release were healthy young adults (aged 22–35, mean = 28.80, standard deviation = 3.70). The structural images (T1-weighted and T2-weighted scans) of the HCP have a high isotropic resolution (0.7 mm; see Van Essen et al., 2013 for details), and all images underwent the HCP-specific minimal preprocessing pipeline (Glasser et al., 2013). We used native surface meshes created by the FreeSurfer (version 5.3)-HCP pipeline (Fischl et al., 2002; Glasser et al., 2013; Jenkinson et al., 2012; Jenkinson et al., 2002) from T1-weighted MRI images using 3T scanners. For subject identification, we employed the test–retest subsample, which consists of 45 healthy subjects (13 males, 32 females) aged from 22 to 35 (mean = 30.29; standard deviation = 3.34), including 17 pairs of MZ twins. The intervals between the test session (the t1 session in our analysis) and the retest session (t2) were between about 1 and 11 months (mean interval of 4.7 months). To compare the identifiability of the SAS and the resting-state functional connectivity, we further excluded one subject without REST1 data in one session.

For analyzing the relationships between SAS and sex as well as handedness, we excluded three subjects with unclear zygosity and four subjects with outlying SAS values (using the same criteria as used in the OASIS-3) from the s1200 release subjects, and GLM of sex and handedness effects were applied to cross-sectional data of these remaining 1106 subjects (504 males; 602 females). We further excluded 12 subjects who did not have all 13 cognitive measures analyzed in our CCA (detailed below). Among the s1200 release subjects were 79 same-sex DZ twin pairs and 138 MZ twin pairs; 160 of these twin pairs have non-twin sibling imaging data. For twin pairs with more than one non-twin sibling, we selected one sibling at random (Arnatkeviciute et al., 2021). We used the resulting twin and non-twin siblings data for the heritability analysis.

ADNI

Request a detailed protocol

The ADNI database (adni.loni.usc.edu) was launched in 2003 as a public–private partnership, led by principal investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial MRI, positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD).

Participants in the ADNI sample completed multiple MRI sessions, but the number of sessions was not consistent across subjects. We used 208 healthy control subjects from the ADNI 1 who had both the baseline MRI session (the t1 session) and a follow-up MRI session 6 months later (the t2 session). These subjects comprised 109 males and 99 females aged 60–90 (mean = 76.21; standard deviation = 5.10) upon study entry. Of these 208 subjects, 135 subjects also had an MRI session 3 years later from the initial session. To evaluate the stability of our methods, we reanalyzed these 135 subjects using data from the 3-year follow-up as the t2 session. The preprocessing procedure included gradwarping, B1 correction, and/or N3 scaling. We used the ADNI-provided surface meshes generated by the cross-sectional FreeSurfer (version 4.3) from T1-weighted MRI image. Detailed descriptions of image acquisition, quality control, and preprocessing are described at http://adni.loni.usc.edu/methods/mri-tool/mri-analysis/ and Jack et al., 2008.

Spectral shape analysis

Request a detailed protocol

We utilized the eigenvalues of the LBO applied to cortical surface mesh models generated with FreeSurfer (Fischl et al., 2002). The eigendecomposition of each individual’s cortical surface was estimated using the Shape-DNA software (Wachinger et al., 2015; Reuter et al., 2009; Reuter et al., 2006), which provides algorithms that extract and optimize the eigenvalues and eigenfunctions from the LBO based on the intrinsic shape of an object (Reuter et al., 2009; Reuter et al., 2006). The Shape-DNA software (Reuter et al., 2009; Reuter et al., 2006) uses the cubic finite element method to solve the Helmholtz equation (Equation 1), also known as the Laplacian eigenvalue problem:

(1) f=-λf

where Δ is the LBO, and f is the eigenfunction with corresponding eigenvalue λ. The eigenvalues of the Helmholtz equation are a sequence ranging from zero to infinity, that is, 0 ≤ λ1λ2 ≤…< ∞, and changes in shape result in changes in the eigenvalue spectrum (Reuter et al., 2006).

Spectral shape analysis via LBO is a departure from traditional morphological analyses that focus on either specific locations (i.e., regions defined by a cortical atlas) or global differences (such as total hemispheric volume). Spectral shape analysis focuses instead on differences in the spatial scales of variation. The decomposed spatial scales can be linearly combined to reconstruct the surface via the eigenfunctions and their corresponding coefficients (the contribution of each set of eigenfunctions to the original surface; see Figure 2G for examples of reconstructed surfaces).

Importantly, Shape-DNA achieves better results for retrieving object shapes than numerous cutting-edge shape-retrieval methods (Lian et al., 2013). Shape-DNA compresses the cortical-surface geometry from around 5 mb into only less than 3 kb, making it computationally efficient for further analysis (Wachinger et al., 2015). The code for calculating Shape-DNA is written in Python and is freely available (http://reuter.mit.edu/software/shapedna/). We applied the Shape-DNA code to the data and analyzed the resulting eigenvalues using MATLAB.

Eigenvalue normalization

Request a detailed protocol

To account for differences in brain sizes among participants, the eigenvalue spectra from Shape-DNA should be normalized (Reuter et al., 2009). Previous studies Wachinger et al., 2015; Wachinger et al., 2016; Wachinger et al., 2018 have applied volume normalization to normalize the eigenvalue spectrum to unit volume via the following equation (Wachinger et al., 2015; Wachinger et al., 2016):

(2) λ=v2/Dλ

where v is the Riemannian volume of the manifold, λ is the original eigenvalue spectrum (λ = [ λ1 , λ2 ,…]), and λ is the volume normalized eigenvalue spectrum. Although this approach has been used in the literature, it is still unable to isolate shape properties as it does not control the effect of different surface areas among objects. For example, in Figure 7, each line is the eigenvalue spectrum for the cortical surface of one subject, and these eigenvalue spectra are straight lines (although they are not straight lines if we zoom in these figures) increasing along with the indices: each eigen-spectrum line has its own slope. Specifically, slopes of the native eigenvalue spectra from each subject are different (Figure 7A) and related to the volume of the manifold. Even though volume normalization decreases the differences in the slopes of the eigenvalue spectra, the slopes remain quite different (Figure 7B) and are driven by differences in surface area (Reuter et al., 2009). More specifically,

(3) λn4πnarea(M)

where λ is the eigenvalue and n is the eigenvalue index. Hence, an appropriate surface area-based normalization is essential to isolate the effects of shape that are distinct from size, particularly given the evidence that the right hemisphere tends to have a greater cortical surface area than the left hemisphere (Kong et al., 2018). Without surface area normalization, differences between the hemispheres may be primarily driven by differences in the surface area of the two hemispheres.

Eigenvalue spectra with and without normalization.

(A) Native eigenvalue spectra. (B) Eigenvalue spectra with volume normalization. (C) Eigenvalue spectra with surface area normalization. All of these results are from the left white surfaces of 233 subjects from the OASIS- 3 data. Each line represents a subject. The slopes of the spectra in (A) and (B) differ among subjects, whereas those in (C) are almost identical.

To perform surface area normalization, we obtained the unit surface area by dividing the vertex coordinates on each axis by the square root of the total surface area (Equation 4).

(4) Vx=Vxarea(M);Vy=Vyarea(M);Vz=Vzarea(M)

where Vx, Vy, Vz are the coordinates of all vertices on the X-axis, Y-axis, and Z-axis, respectively; area (M) is the surface area of object M; Vx′, Vy′, Vz′ are the coordinates of transformed vertices on the X-axis, Y-axis, and Z-axis, respectively. Surface area normalization is stricter than volume normalization for spectral alignment, and the eigenvalue spectra with surface area normalization have a nearly identical slope (Reuter et al., 2009; Figure 7C).

The shape asymmetry signature

Request a detailed protocol

The LBO eigenvalues measure the intrinsic geometry of an object and are isometry invariant. Hence, the eigenvalue spectra are identical regardless of object positions, rotations, and mirroring (i.e., perfect projection from the brain structure of the right hemisphere to the left does not change the eigenvalue spectrum) (Wachinger et al., 2015; Reuter et al., 2006). Therefore, brain asymmetry can be calculated directly from the eigenvalue spectra of the two hemispheres (Wachinger et al., 2015; Wachinger et al., 2016) without image registration or smoothing (Wachinger et al., 2015; Reuter et al., 2006). In this study, after calculating the eigenvalues with surface area normalization, we subtracted the eigenvalue spectra of the right hemisphere from those of the left hemisphere in the same subject at each spatial scale (each eigenvalue index) to define the SAS. Formally, the SAS for subject i is given by

(5) Λi=λLiλRi

where λ is the eigenvalue spectrum λ = (λ1 , λ2 ,..., λn) from the left (L) and right (R) hemispheres, each of which represents a certain spatial scale. There are other possible asymmetry indices (Kong et al., 2018; Moodie et al., 2020), but those indices are not appropriate for a surface area-normalized eigenvalue analysis, as our normalization already accounts for size effects.

Moodie et al., 2020 proposed subtracting the mean of the asymmetry values across subjects from the individual asymmetry values to represent the asymmetry. We tested this approach with our method, and the results were generally the same, as the eigenvalues were normalized before calculating the asymmetry. For simplicity, we defined the SAS using Equation 5 to represent the individual-specific asymmetry.

To further check the possible influence of image quality on the SAS, we first took the mean of the Euler number of the left and right hemispheres using FreeSurfer, which is widely used as an index of image quality (Morgan et al., 2019; Kaufmann et al., 2019; Rosen et al., 2018), and then calculated the Pearson’s correlation between the mean Euler number and the SAS across the first 200 eigenvalues. For the HCP s1200 dataset, the correlations were all below 0.07 (PFDR > 0.05). For the OASIS-3, the correlations were all below 0.18 (PFDR > 0.05) at either time 1 or time 2 MRI session. These results indicate that image quality does not strongly influence the SAS, which is in line with past findings that the eigenvalues and eigenfunctions of the LBO are robust to image noise (Reuter, 2010).

Subject identification

Request a detailed protocol

Our first aim was to validate the SAS as a useful and robust measure of individual-specific asymmetry. We, therefore, evaluated the extent to which the SAS of each individual subject measured at time 1 (t1) could be used to identify the same person at time 2 (t2) in the longitudinal data, akin to a neuromorphological fingerprint. The t1 – t2 Pearson correlations were then estimated between all pairs of N individuals, resulting in an N×N correlation matrix.

Amico and Goñi, 2018 defined identifiability as the difference between the mean of within-subject correlations (diagonal elements of the Pearson correlation matrix in Figure 1C) and the mean of between-subject correlations (off-diagonal elements of the Pearson correlation matrix in Figure 1C). This approach allows for a more quantitative and finer-grained comparison of the identifiability of different metrics compared to other approaches that just rely on binary identification accuracy (e.g., Finn et al., 2015; Amico and Goñi, 2018; Mansour L et al., 2021). However, this approach does not take into account the variance of the observations. To examine the within- and between- subject similarities, we utilized the Glass’s ∆, which is the standardized difference between the mean values of two categories of observations, normalized by the standard deviation of a control group (Glass et al., 1981), which is the between-subject group in our case. Glass’s ∆ has been recommended when the standard deviations of the two groups are substantially different (Glass et al., 1981; Lakens, 2013), which is the case for the between- and within-subject groups. Thus, our identifiability score was given by

(6) Identifiabilityscore=mean(rii)mean(rij)SD(rij)

where SD is the standard deviation. Higher scores indicate a greater capacity to discriminate between individuals. We also tested the pooled standard deviation of the two groups (Mansour L et al., 2021), as in the estimation of Cohen’s d, and the results were generally consistent to those using the Glass’s ∆.

We also evaluated the identifiability performance of the SAS with respect to unihemispheric descriptors of either combining size and shape or shape alone: namely, the eigenvalues (native, volume-normalized, or surface area-normalized) from the same hemispheres between time 1 and time 2 follow-up; concatenating eigenvalues of both left and right hemispheres between time 1 and time 2; and identifying the shape of one hemisphere from the shape of the other hemisphere both at time 1 or both at time 2. Finally, we compared the identifiability score of the SAS to the asymmetry based on commonly used size-related measures (i.e., volume, cortical thickness, and surface area), and resting-state functional connectivity.

Identifying spatial scales for optimum subject identifiability

Request a detailed protocol

Given a surface of N vertices, spectral shape analysis yields up to N eigenvalues, raising the question of how many eigenvalues constitute a sufficient description of cortical shape. Is a full representation of the entire surface necessary for optimal subject identifiability, or can this be achieved using a more compact set of eigenvalues? If so, the specific number of eigenvalues required would define the relevant spatial scale of shape differences that characterize the individual-specific asymmetry at which individual differences are most prevalent.

To address this question, we decomposed the cortical surface and use an increasing number of eigenvalues, from the first two eigenvalues (λ1 , λ2) to the first 1000 eigenvalues (λ1 , λ2,λ3, …, λ1000), each time computing the SAS and evaluating subject identifiability. For example, we first quantified the shape of cortical surface using only λ1 and λ2 , thus capturing the coarsest scales of cortical shape. We then quantified the surface using λ1 through λ3 , then λ1 through λ4 , and so on. If there is a specific spatial scale that is optimal for this subject identifiability, we expect to see a peak in the identifiability score as a function of the truncation number, k. This peak not only defines the spatial scale at which individual variability, and thus individual-specific asymmetry, is most strongly expressed, but it also identifies a meaningful point at which to define a compressed summary of individual-specific asymmetry using the eigenvalue spectrum.

Cortical shape harmonics

Request a detailed protocol

The cerebral cortex is topologically equivalent to a sphere. Solving the Helmholtz equation for a sphere yields groups of eigenfunctions with the same eigenvalues and spatial wavelength, progressing along orthogonal axes (Robinson et al., 2016). These groups in the solutions to the idealized spherical case are known as the spherical harmonics. The zeroth group (L = 0) is comprised of the first eigenvalue; the first group (L = 1) is comprised of the second, third, and fourth eigenvalues; the second group (L = 2) is comprised of the fifth to ninth eigenvalues, and so on, with 2 (L + 1)–1 eigenvalues in the Lth group. Robinson et al., 2016 showed that while the eigenvalues between the cortical surface and sphere are different, the spherical grouping provides a rough division of the convoluted cortical surface. This is a useful grouping approach to investigate eigenfunctions and eigenvalues as the constituents of each group have roughly the same spatial wavelength. By averaging over several eigenvalues with similar spatial scales, we can also increase the stability of the truncation number across datasets. For example, the peak SAS identifiability appeared at the first 126 and 122 eigenvalues for the OASIS-3 and ADNI data, respectively, and these eigenvalues are all within the 11h eigen-group (L = 11).

To estimate the corresponding wavelength of each eigen-group, we used an approximation of the spatial wavelength in the spherical case:

(7) W=2πRsL(L+1)

where Rs is the equivalent sphere of the original object (for the fsaverage case, Rs is about 67 mm) and L is the index of the eigen-group. We used the population-based template (fsaverage) as an example to show the wavelengths of the first 14 eigen-groups in Supplementary file 1.

Cortical segmentation

Request a detailed protocol

We applied the HCP-MMP1 atlas (Glasser et al., 2016) to segment cortical regions for accessing size-related morphological asymmetry, functional connectivity, and regional volume heritability. This atlas is based on a surface alignment approach that aligns the images using cortical folding patterns and minimizes the spatial smoothness (Glasser et al., 2016; Coalson et al., 2018), thus offering more accurate inter-subject registration than volume-based registration (Glasser et al., 2016). Moreover, regions in the left and right hemispheres of the HCP-MMP1 atlas are corresponding and thus can be used for accessing cortical asymmetry. In addition to the HCP-MMP1 atlas, we also employed the Schaefer atlas (Schaefer 100, 300, and 900) (Schaefer et al., 2018) for constructing functional connectivity (FC) and regional volume heritability. The Schaefer atlas has superior functional homogeneity and has different parcellation scales (Schaefer et al., 2018); therefore, it can be used for comparing the identifiability of the FC and estimating regional volume heritability at different scales. Specifically, each hemisphere has 50 regions in the Schaefer 100 atlas, 150 regions in the Schaefer 300 atlas, and 450 regions in the Schaefer 900 atlas (Schaefer et al., 2018). However, regions in the left and right hemispheres of the Schaefer atlas are not corresponding; therefore, the atlas cannot be used for assessing brain asymmetry.

Non-shape descriptors of brain anatomical asymmetry

Request a detailed protocol

To compare identifiability scores obtained with SAS to asymmetries using size-related descriptors, including volume, cortical thickness, and surface area, we had to ensure that the asymmetry values were purely from the asymmetry effect and were not affected by the effect of total brain size. A traditional asymmetry index (Kong et al., 2022; Kurth et al., 2018; Sha et al., 2021) is

(8) AIS,i=(PLS,iPRS,i)0.5(PLS,i+PRS,i)

where AIS,i is the asymmetry index for parcellation S of subject i. PLS,i is the value of the morphological measurement from parcellation S from subject i’s left hemisphere; and PRS,i is from the right hemisphere. However, this asymmetry index is unable to entirely eliminate the effect of total brain size. Here, we use cortical volume as an example. We suppose the total brain volume effect (α) exists, and the effects of each ROI-based volume of the left (Equation 9) and right (Equation 10) hemispheres are

(9) VLF,i=δLiVLi+αi
(10) VRF,i=δRiVRi+αi

where VLi and VRi are the volumes of region i in the left and right hemispheres, respectively, δLi and δLi are the scaling coefficients, and αi is the effect of total brain volume on region i. Thus, VLF,i and VRF,i are the overall effects of volume on region i. We can apply VLF,i and VRF,i to the traditional asymmetry index as in Equation 8 to get

(11) VAIi=[(δLiVLF,i+αi)(δRiVRF,i+αi)]0.5[(δLiVLF,i+αi)+(δRiVRF,i+αi)]

By rearranging this equation, we obtain

(12) VAIi=δLiVLF,iδRiVRF,i0.5δLiVLF,i+0.5δRiVRF,i+αi

which shows that the total volume effect αi still remains in the denominator and is not removed by the traditional asymmetry index defined in Equation 8.

In this study, we adjusted the asymmetry index for the mean of each morphological measurement, such as the asymmetry of cortical thickness, volume, and surface area. Specifically, we revised the traditional asymmetry index by subtracting the mean value of the measurement across all parcellations of each subject before calculating the asymmetry index defined in Equation 8. This revised asymmetry measure RAIS,i is explicitly calculated as

(13) RAIS,i=(PLS,iMi)(PRS,iMi)0.5[(PLS,iMi)+(PRS,iMi)]

where Mi is the mean value of the measurement across all regions in parcellation of subject i. We note that this is an important point as without this correction, the asymmetry measure is dependent on the mean value.

We employed a multi-modal parcellation, HCP-MMP1 version 1.0 (Glasser et al., 2016), on the OASIS-3 subjects. We excluded one subject whose cortical surfaces could not be segmented by the HCP-MMP1 atlas. There are 180 regions in each hemisphere of the HCP-MMP1 atlas, including the hippocampus that was excluded in our analysis. We created one vector per size-related measure that quantified the asymmetry index per subject and then used these asymmetry indices in the subject identifiability analyses.

Functional connectivity

Request a detailed protocol

We used the resting-state FC from the first session (REST1) in the test sample as the first FC time point (t1) and FC from the first session in the retest session as the second FC time point (t2). We utilized the fMRI signals that were preprocessed by the HCP functional and ICA-FIX pipelines (Glasser et al., 2013). We did not apply any spatial smoothing on the signals. FC was calculated using the upper triangle entries of the Pearson correlation matrix between nodes from the atlas (Finn et al., 2015). To compare the identifiability of the SAS and the FC across different parcellation scales and atlas, we repeated the FC analysis with the Schaefer 100, 300, and 900 atlas (Schaefer et al., 2018) and the HCP-MMP1 atlas (Glasser et al., 2016) from the subjects in the HCP test–retest subsample (n = 44; we excluded one subject without REST1 data in one session).

Relationships with sex and handedness

Request a detailed protocol

Sex and handedness are two common characteristics that have been widely examined in the asymmetry literature (Corballis and Häberling, 2017; Güntürkün and Ocklenburg, 2017; Toga and Thompson, 2003; Kong et al., 2018; Plessen et al., 2014; Kong et al., 2022; Deep-Soboslay et al., 2010; Núñez et al., 2018; Wachinger et al., 2015; Narr et al., 2007; Steinmetz et al., 1991; Good et al., 2001; Guadalupe et al., 2014; Maingault et al., 2016; Kong et al., 2021). We used a GLM to analyze relationships between each eigenvalue with sex and handedness on 1106 HCP subjects (see ‘HCP’ section). The HCP dataset provides the handedness preference measured by the Edinburgh Handedness Inventory (EHI) (Oldfield, 1971). EHI is the most widely used handedness inventory (Vlachos et al., 2013; Willems et al., 2014), with resulting scores range from –100 (complete left-handedness) to 100 (complete right-handedness) (Oldfield, 1971). Handedness preference is not a bimodal phenomenon (Dragovic, 2004), and cutoff scores to categorize the handedness are still arbitrary. We therefore used the EHI score as a continuous variable in our main analysis, which is a widely used approach (Maingault et al., 2016; Kong et al., 2021). To further confirm the robustness of the relationship between handedness and the SAS, we tested two thresholds to categorize handedness. First, right-handed (EHI: 71–100), left-handed (EHI: –100 to -71), and ambidextrous (EHI: –70–70) (Deep-Soboslay et al., 2010; Narr et al., 2007; Dragovic, 2004) second, right-handed (EHI: 50–100), left-handed (EHI: –100 to -50), and ambidextrous (EHI: –49–49) (Vlachos et al., 2013; Perlaki et al., 2013). Regardless of the threshold, the categorized handedness variable was still unrelated to the SAS (2–144 eigenvalues).

Relationships with cognition

Request a detailed protocol

We followed Kong et al., 2019 and used 13 cognitive measures in the HCP data dictionary that represent a wide range of cognitive functions, namely, (1) Visual Episodic Memory (PicSeq_Unadj); (2) Cognitive Flexibility (CardSort_Unadj); (3) Inhibition (Flanker_Unadj); (4) Fluid Intelligence (PMAT24_A_CR); (5) Reading (ReadEng_Unadj); (6) Vocabulary (PicVocab_Unadj); (7) Processing Speed (ProcSpeed_Unadj); (8) Delay Discounting (DDisc_AUC_40K); (9) Spatial Orientation (VSPLOT_TC); (10) Sustained Attention – Sens (SCPT_SEN); (11) Sustained Attention – Spec (SCPT_SPEC); (12) Verbal Episodic Memory (IWRD_TOT); and (13) Working Memory (ListSort_Unadj). We used PCA to reduce dimensionality and minimize collinearity in the CCA. The first four PCs explained 80% of the variance and were retained for our primary analysis. Similarly, we reduced the dimensionality of the SAS measures and ensured equal representation across different spatial scales by taking the mean of the SAS across each eigen-group (from 1st to 11th groups). These 11 mean SAS values and the first 4 cognitive PCs were then subjected to CCA to identify linear combinations of SAS and cognitive measures that maximally covary with each other (Smith et al., 2015). Inference on the resulting canonical variates was performed using a permutation-based procedure (Winkler et al., 2020), and robust estimates of canonical loadings were obtained using bootstrapping (Dong et al., 2020), as described in the ‘Statistical analysis’ section.

Heritability of brain shape

Request a detailed protocol

We used MZ and same-sex DZ twin pairs and their non-twin siblings to calculate the heritability of brain shape and cortical volume. For twin pairs with more than one non-twin sibling, we selected one sibling at random. We estimated the heritability of each eigenvalue from individual hemispheres and the SAS. To emphasize the importance of properly controlling surface area, we show the heritability of eigenvalues with volume normalization (but without surface area normalization; Figure 6—figure supplement 1). We also calculated the heritability from ROI-based volumes of individual hemispheres (Figure 6—figure supplement 2). Regional cortical volumes of individual hemispheres were estimated for each region of the Schaefer 100, 300, and 900 atlas (Schaefer et al., 2018), as well as the HCP-MMP1 atlas (Glasser et al., 2016).

Under the assumption that MZ twins are genetically identical whereas DZ twins on average share half of their DNA, structural equation modeling (SEM) can be used to decompose the phenotypic variance of a trait into additive genetic (A), common environmental (C), and unique (subject-specific) environmental (E) factors (Arnatkeviciute et al., 2021). Twins raised together are likely to share a more common environment compared to their non-twin siblings; therefore, including a set of non-twin siblings into the analysis allows us to additionally separate common environmental contributions into twin-specific (T) and twin non-specific common environmental factors (C). The heritability analyses of brain shape and volume were performed independently using standard SEM implemented in OpenMx software (Boker et al., 2011; Neale et al., 2016) in R.

For each eigenvalue and parcellated volume, outlying values were first excluded using the boxplot function in R keeping data points (v) in a range Q1−1.5 × IQR < v < Q3 + 1.5 × IQR, where Q1 and Q3 are the first and third quartiles respectively, and IQR is the interquartile range (Arnatkeviciute et al., 2021). For each phenotype, we then fitted a set of biometric models – ACTE, ACE, ATE, CTE, TE, CE, E – using age and sex as covariates, where the letters indicate the factors present in the model. The goodness of fit between the models was compared using the Akaike information criterion (AIC) (Parzen et al., 1998), and the best-fitting model for each measure was selected based on the lowest AIC value. Consequently, the heritability for each measure was derived from the best-fitting model, corresponding to the best model that balances the ability to explain data with model complexity. To ensure that the general heritability pattern was not dependent on the model selection, we also calculated the heritability estimates from the full ACTE model (without model selection) at each eigenvalue (with surface area normalization) of individual hemispheres as well as the SAS. The heritability estimates were highly correlated with those with model selection (Pearson correlation r = 0.92–0.96).

Statistical analysis

Request a detailed protocol

We applied a permutation test to evaluate the statistical significance of a given identifiability score for a given number of eigenvalues. We randomly shuffled the subject order of the SAS of the t2 session 50,000 times and then compared the original identifiability score with all the permuted peak identifiability score truncated at each iteration independently to calculate the P-value. Statistical inference for models evaluating associations between SAS and sex and handedness was also performed using a permutation test with 100,000 iterations by randomly shuffled the subjects’ sex and handedness data.

When analyzing associations between the SAS and cognition, we used a recently developed permutation-based procedure for CCA inference (Winkler et al., 2020) with 50,000 iterations. The P-values of the canonical modes were controlled over family-wise error rate (FWER; FWER corrected P-values are denoted PFWER), which is more appropriate than the FDR when measuring the significant canonical mode (Winkler et al., 2020). The results were consistent when controlling for age and sex as confounding variables. To identify reliable loadings of each SAS eigen-group on the canonical variate, we used bootstrapping with 1000 iterations of the correlation between each SAS eigen-group and the SAS canonical variate. The resulting standard errors were used to estimate z-scores for each loading by dividing the original correlation by the standard errors, and then the z-scores were used to compute two-tailed p-values. We then used FDR (q = 0.05) to correct for multiple comparisons of P-values of all the eigen-groups. We also used the same approach to identify reliable correlations of cognitive measures on the corresponding canonical variate. Due to the family structure of the HCP data, we kept the subjects’ family structures intact when shuffling or bootstrapping the subjects using the Permutation Analysis of Linear Models (PALM) software package (Winkler et al., 2016; Winkler et al., 2015).

The statistical significance of the heritability estimates was evaluated through model comparison between models with and without parameter A. For example, if the ACE model was the best-fitting model, the P-value was derived by comparing the ACE and CE models; if the best-fitting model was CE, we compared this model with the ACE model to get the P-value for the A parameter. We also used the same approach for measuring the statistical significance of the common environmental factor (C). FDR (q = 0.05) was used to correct for multiple comparisons (corrected P-values are denoted PFDR) in all analyses except for the CCA, where FWER was controlled using a permutation-based procedure (Winkler et al., 2020).

Data availability

All data generated or analysed during this study are included in the manuscript. All code and dependent toolboxes used in this study can be found at: https://github.com/cyctbdbw/Shape-Asymmetry-Signature, (copy archived at swh:1:rev:242d90d06e90e7f200182acef46d403f3419a85a). The code of shape-DNA can be found at: http://reuter.mit.edu/software/shapedna/. The OASIS-3 dataset is available under https://www.oasis-brains.org/. The ADNI dataset is available under https://adni.loni.usc.edu. The HCP dataset is available under https://db.humanconnectome.org/.

References

  1. Book
    1. Glass GV
    2. McGaw B
    3. Smith ML
    (1981)
    Meta-Analysis in Social Research
    SAGE Publications.
  2. Book
    1. Gomez-Robles A
    2. Reyes LD
    3. Sherwood CC
    (2018) Landmarking brains
    In: Bruner E, Ogihara N, editors. Digital Endocasts. Springer. pp. 115–126.
    https://doi.org/10.1007/978-4-431-56582-6
  3. Conference
    1. Lévy B
    (2006) Laplace-Beltrami eigenfunctions towards an algorithm that “understands” geometry
    IEEE International Conference on Shape Modeling and Applications - SMI’06.
    https://doi.org/10.1109/SMI.2006.21
  4. Book
    1. Parzen E
    2. Tanabe K
    3. Kitagawa G
    (1998) Selected papers of hirotugu akaike
    In: Petrov BN, editors. Information Theory and an Extension of the Maximum Likelihood Principle 949. New York, NY: Springer. pp. 267–281.
    https://doi.org/10.1007/978-1-4612-1694-0
    1. Wachinger C
    2. Golland P
    3. Kremen W
    4. Fischl B
    5. Reuter M
    (2015)
    Alzheimer’s disease neuroimaging I. brainprint: a discriminative characterization of brain morphology
    NeuroImage 109:232–248.

Decision letter

  1. Ingrid S Johnsrude
    Reviewing Editor; University of Western Ontario, Canada
  2. Christian Büchel
    Senior Editor; University Medical Center Hamburg-Eppendorf, Germany
  3. Antoine Balzeau
    Reviewer

Our editorial process produces two outputs: (i) public reviews designed to be posted alongside the preprint for the benefit of readers; (ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Decision letter after peer review:

Thank you for submitting your article "The individuality of shape asymmetries of the human cerebral cortex" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Christian Büchel as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Antoine Balzeau (Reviewer #2).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission. The theoretical notions of asymmetry being explored should be more carefully defined and related to past literature, and alternative explanations of some of the results need to be considered. The reviews follow below.

Reviewer #1 (Recommendations for the authors):

– It would be very helpful to have improved anatomical representations of the inter-individual differences. For instance, could it be shown what an individual with an extreme eigenvalue at a specific scale would look like, for some of the scales that are argued to be biologically important?

– The results would be strengthened if a measure of bihemispheric (symmetric shape signature) were included, rather than or in addition to unihemispheric shape of the left and right hemisphere separately, to strengthen the conclusions about asymmetry.

– An analysis of a quantitative QC metric could strengthen the paper. For instance, looking at associations of the Euler index from FreeSurfer with SAS across scales may help clarify if there is an association between image quality and these metrics.

– Can the authors clarify that the calculation of SAS is not affected by whether there is a 1:1 correspondence (homology) between left and right hemisphere vertices, which varies between processing pipelines.

Reviewer #2 (Recommendations for the authors):

I am pleased to identify myself as Antoine Balzeau. I am a paleoanthropologist, working on the evolution of the human brain. I also work with more "classic" neuroscientists. Please, excuse me if some of my comments or questions appear to be strange for you. I am not able to understand all the mathematic behind your methodology. However, my expertise allows me to comment some of your approaches and ideas, and to ask some questions about the way you deal with bilateral variation. I detail those points below, following the order of the manuscript.

Abstract. Here and later, you mention a lot the context of studies of brain asymmetries. But your study is not really dealing with those aspects. You develop a new tool and find an original application, I agree with this. However, I think that you should improve the way you present the results about bilateral variation and develop implications for morphological and functional aspects. Or only focus on your specific approach and results, that could be enough.

Line 48. You mention DA. However, statement lines 47 and 48 is not completely true. Fluctuating asymmetry is also analysed within a group or a population. DA is only a part of the information obtained within studies of bilateral variation.

Line 51, 52 "regions", "directions", "magnitudes" are not "inconsistent". Different methods give different results about DA and FA in studies of brain bilateral variations. There are consistent results on many specific traits. You cannot compare as a whole the literature and say there is not info from all this amount of research. (I agree with lines 59-61, but it does not justify what you have written above).

Line 62: "little…. Average", this general statement is strange. Again, there are some results showing consistent DA, high absolute asymmetry, antisymmetry….

Line 65-66 FA is not only "the degree to which a given individual departs from the mean". FA is also studied among and between samples, that is the most informative way to discuss developmental instability.

Lines 94-99. I do not understand well this section. I do not see the link with your following sentence. Moreover, I do not think that your study is "a detailed characterization of FA in cerebral shape". At least, I think that you could have said much more on the topic.

Line 101, individual-specific measure of FA. You have to explain this. Why would it be FA? In fact, that is only individual bilateral variation.

Line 106. How can you say that it is done with greater accuracy? You have to demonstrate this.

The end of the introduction is a kind of abstract of what has been done. There are no H0 to be tested, no justification of the approach, no scientific question.

Line 124-125, it is not clear what are the shape descriptors your analyse, nor which aspects of cortical shape asymmetries you study.

Lines 132-145, 150-161, 181-186 should be in the method section.

Line 146-150: methods too, and not clear here why it relates to studies of bilateral variation.

Line 190-191. How do you reach this conclusion? Should be true if DA is present, but how can you compare differences in R-L shapes and shape of an object?

Line 194-197: the conclusion is that one hemisphere does not predict the shape of the other. OK. But why do not you quantify those bilateral variations? Moreover, there is no quantification of absolute asymmetry, DA, or FA. Would not it be possible to extract information about those parameters with you approach? That would be much more informative.

Figures 1 and 2 are more about the methods that true results.

Line 219-223. What is the implication of this? Isn't it related to the fact that the approximation for lower values is too large to extrapolate any information?

Line 226. I do not understand what are identifiability scores, a p = 0, and do not have any idea about the meaning in the variation of the scores between samples. Moreover, why do you do comparison between sample and not for each individual? Finally, why do you have differences between the samples, one with a flat area, differences in shape and values between samples? It seems that you describe global results by sample but never extract something from the descriptions of general results.

Line 230-238: move to the method section.

Line 238-242: what is the implication of this? What does it mean?

243. Why 144 would be a "stable and robust scale"? how can you decide it is 144.

251-253. OK, it works for individual identification. But is it better than any simple comparison of brain shape? Why is it interesting to have tool that allows to identify an individual with 2 mri dataset? Moreover, I would prefer to see more results about bilateral variation thanks to your approach instead of this possibility of identification.

Lines 264-266. You do not demonstrate this supposed higher ability.

Lins 271-272. OK. But what is the implication of this?

Lines 291-295: move to methods. And why did you decide to do this way?

Figure 4: that is the first presentation of bilateral variation. But not test, no description, no quantification of DA, FA, absolute asym…. Why?

Figure 4B, what is the color code? lines 317-324: methods.

Lines 33-335: you mention direction of asymmetries, but there is no test for DA for those parameters.

Figure s5. What are the results for absolute asymmetry?

Lines 344-349: methods.

Line 352-3. Ok, but why is it nteresting to exclude sulcal and gyral features? Isn’t it a too large approximation of the shape? Moreover, you never mention any attempts to test measurement errors, that may have a large impact on FA analyses.

Lines 357-9. Why?

Lines 378-379: what is the evidence that proves this 1/3 relationship?

Discussion

391, in my opinion you have not examined FA.

393: highly personalized, ok, but it is not the most interesting aspect of FA in theory.

Line 394 “most discriminative”, it has not been demonstrated.

Lines 390-398 are an abstract of the paper, not the beginning of a discussion.

Line 401 which “meaningful effects”?

Line 403-404 this image is not appropriate. Line 405-6: ok, but it has to give results that are interpretable, it is not really the case here.

Lines 407 and following: there are several papers on endocranial casts with detailed shape information.

410 which time?

411-412, what is the implication?

416: why are not they “captured reliably”?

419 explain in which aspects it is a “more comprehensive characterization”.

Lines 434-440: No link with your study and results.

Line 463: ok, but please elaborate a bit.

Lines 474-475: this is a rather naïve statement. Please explain or delete.

Conclusion: to be removed, it is an abstract.

Methods

503: how do you determine those errors?

On the reverse, you do not quantify your own methodological error. Have you tried to repeat your methods? What is the effect of imaging resolution (always important when working on bilateral variation)?

Line 759 what are the “morphological measurements” that you mention here?

Line 770: what are the asym-index measurements?

Line 798: what is reduced?

https://doi.org/10.7554/eLife.75056.sa1

Author response

Reviewer #1 (Recommendations for the authors):

1. It would be very helpful to have improved anatomical representations of the inter-individual differences. For instance, could it be shown what an individual with an extreme eigenvalue at a specific scale would look like, for some of the scales that are argued to be biologically important?

Thank you for this valuable suggestion. We have included an additional figure (Figure 1—figure supplement 1) that shows cortical surface reconstructions at different scales for varying values of the SAS.

2. The results would be strengthened if a measure of bihemispheric (symmetric shape signature) were included, rather than or in addition to unihemispheric shape of the left and right hemisphere separately, to strengthen the conclusions about asymmetry.

Thank you for this comment. In Figure 2 (A) to (C), we showed that the subject identifiability of the SAS is not only higher than unihemispheric shape (either left or right hemisphere) but also higher than combining the shapes of both hemispheres (concatenating the left and right hemispheres). For this reason, the subsequent analyses focus on the SAS. We also note that the SAS can readily accommodate cases of pure symmetry, and in such cases, the SAS would be equal to zero. We now include a demonstration of the point in the new Figure 2—figure supplement 2.

3. An analysis of a quantitative QC metric could strengthen the paper. For instance, looking at associations of the Euler index from FreeSurfer with SAS across scales may help clarify if there is an association between image quality and these metrics.

Thank you for raising this important issue. We have checked that the Euler number of the cortical surfaces, which is recommended as a quantitative index of image quality (1), is unrelated to the SAS, indicating that image quality has a limited influence on our measure of shape asymmetry.

4. Can the authors clarify that the calculation of SAS is not affected by whether there is a 1:1 correspondence (homology) between left and right hemisphere vertices, which varies between processing pipelines.

Thank you for raising this important point. Before the calculation of SAS, we calculated the eigenvalues of the Laplace-Beltrami Operator (LBO) from the left and right hemispheres independently. The LBO is not directly computed on the vertices of surface mesh but on the underlying manifold of the surface (14). In other words, the LBO quantifies the shape of the entire hemisphere, and thus, the SAS is not affected by the 1:1 correspondence between the left and right hemisphere vertices.

Reviewer #2 (Recommendations for the authors):

I am pleased to identify myself as Antoine Balzeau. I am a paleoanthropologist, working on the evolution of the human brain. I also work with more “classic” neuroscientists. Please, excuse me if some of my comments or questions appear to be strange for you. I am not able to understand all the mathematic behind your methodology. However, my expertise allows me to comment some of your approaches and ideas, and to ask some questions about the way you deal with bilateral variation. I detail those points below, following the order of the manuscript.

Thank you for your detailed feedback and suggestions.

1. Abstract. Here and later, you mention a lot the context of studies of brain asymmetries. But your study is not really dealing with those aspects. You develop a new tool and find an original application, I agree with this. However, I think that you should improve the way you present the results about bilateral variation and develop implications for morphological and functional aspects. Or only focus on your specific approach and results, that could be enough.

We interpret this comment to be focused on the Abstract. We note that studies on asymmetry provide an important context and motivation for our approach; namely, to develop a method that can disentangle contributions from size and shape across multiple spatial resolution scales. To better develop the background in the Abstract, we have now amended the text so that it reads:

“Asymmetries of the cerebral cortex are found across diverse phyla and are particularly pronounced in humans, with important implications for brain function and disease. However, many prior studies have confounded asymmetries due to size with those due to shape. Here, we introduce a novel approach to characterize asymmetries of the whole cortical shape, independent of size, across different spatial frequencies using magnetic resonance imaging data in three independent datasets…”.

We have also made several further amendments to other sections of the text to clarify our reporting on prior results. These are detailed in our responses to subsequent comments.

2. Line 48. You mention DA. However, statement lines 47 and 48 is not completely true. Fluctuating asymmetry is also analysed within a group or a population. DA is only a part of the information obtained within studies of bilateral variation.

To the best of our knowledge, fluctuating asymmetry (FA) and directional asymmetry (DA) are not consistently defined across studies and across disciplines. In our original manuscript, we followed the definitions of FA and DA in references (15-17). However, we agree that using these terms may confuse some readers with different backgrounds. We have therefore refrained from using these terms in our revised manuscript and instead used population-level/population-based asymmetry and individual-specific asymmetry. These two terms more clearly articulate the type of asymmetry being considered.

3. Line 51, 52 “regions”, “directions”, “magnitudes” are not “inconsistent”. Different methods give different results about DA and FA in studies of brain bilateral variations. There are consistent results on many specific traits. You cannot compare as a whole the literature and say there is not info from all this amount of research. (I agree with lines 59-61, but it does not justify what you have written above).

Line 62: “little…. Average”, this general statement is strange. Again, there are some results showing consistent DA, high absolute asymmetry, antisymmetry….

We agree. We now cite examples of consistent results and have revised Lines 57 to 66 of the revised manuscript:

“Asymmetries in brain organization are often considered at an average level across a population of individuals (7, 18-21). These population-based asymmetry features have been studied extensively and are thought to have important implications for both functional lateralization and abnormal brain function in a wide range of psychiatric and neurological diseases (20, 22-25). For example, the planum temporale of the left hemisphere, which encompasses Wernicke’s area, has been consistently shown to be larger than the right for most healthy individuals (21, 26-28), and patients with schizophrenia often show reduced leftward asymmetry in planum temporale compared to healthy individuals (29-31). However, many findings with respect to asymmetries of specific brain regions have been inconsistent in terms of the directions and magnitudes of asymmetry observed (18, 19, 23).”

We have also revised Lines 74 to 78 of the revised manuscript:

“Despite some consistent asymmetry features across the population (21, 26-28), there is also considerable individual variability around population means, with many people often showing little or even reversed asymmetries relative to the prevalent pattern of the population (sometimes also referred to as anti-symmetry) (16, 32-34).”

4. Line 65-66 FA is not only “the degree to which a given individual departs from the mean”. FA is also studied among and between samples, that is the most informative way to discuss developmental instability.

As per our response to Comment 2, we deleted all references to FA and DA, and amended the following sentences in Lines 80 to 85 of the revised manuscript:

“Populational-level asymmetries are hypothesized to have a genetic basis (16, 18, 33, 35-40), whereas individual-specific asymmetries, which describe the way in which a given individual departs from the population mean, may reflect environmental influences, developmental plasticity, or individual-specific genetic perturbations (16, 17, 33, 35-38). Notably, cortical asymmetries of the human brain are more variable across individuals than other primates, at both regional and global hemispheric levels (16, 32).”

5 Lines 94-99. I do not understand well this section. I do not see the link with your following sentence. Moreover, I do not think that your study is “a detailed characterization of FA in cerebral shape”. At least, I think that you could have said much more on the topic.

The introduction outlines the general principles of shape-DNA and our spectral approach, which provides a context for our aims and for the results that follow. The Materials and methods section outline in detail how the approach is implemented. We now include in Lines 108 to 115 of the revised manuscript an indication to readers that further details are provided in the Materials and methods section. We have also amended the statement about FA so that it now reads:

“The spectral analysis provides a comprehensive description of the intrinsic geometry of a given object, akin to a “Shape-DNA” (see Materials and methods). (14). The application of such Shape-DNA analysis to human magnetic resonance imaging (MRI) data has shown that shape properties of cortical and subcortical structures have superior sensitivity compared to traditional, size-based measures for identifying individual subjects (41), for classifying and predicting the progress of psychiatric and neurological diseases (42, 43), and for detecting genetic influences on brain structure (44, 45). However, a detailed characterization of individual-specific asymmetries in cerebral shape is lacking.”

6. Line 101, individual-specific measure of FA. You have to explain this. Why would it be FA? In fact, that is only individual bilateral variation.

As per our responses to Comments 1 and 2, we no longer use the term FA and instead use individual-specific asymmetry, which we define in Lines 655 to 673 of the revised manuscript. Moreover, as per our response to Reviewer 1’s Comment 2, we have provided an additional figure (Figure 1—figure supplement 1) to show that the SAS value of a perfectly symmetric brain is zero, and it increases for more asymmetric brains.

7. Line 106. How can you say that it is done with greater accuracy? You have to demonstrate this.

This is one of our key findings. In Figure 3, we show that the SAS is more accurate at identifying individuals than shape descriptors of individual hemispheres, asymmetries in traditional size-based descriptors, or patterns of inter-regional functional connectivity.

8. The end of the introduction is a kind of abstract of what has been done. There are no H0 to be tested, no justification of the approach, no scientific question.

We apologize for the lack of clarity. Our motivation follows from the idea that the eigen decomposition of the LBO can be used as a form of shape-DNA (14); i.e., it can be used to identify individual brains (41). As individual-specific asymmetries appear to be highly variable in humans (16, 32), our primary hypothesis was that the SAS would be associated with greater identifiability than measures of unihemispheric cerebral shape, volume, or function. The null hypothesis then is that the SAS does not show greater identifiability than unihemispheric measures. We secondarily explored associations with gender, handedness and cognition, given the extant literature linking various forms of cerebral asymmetry to these factors. Finally, we tested whether the SAS is more strongly influenced by environmental factors, given arguments that individual-specific asymmetries may be primarily driven by environmental or developmental plasticity (16, 17, 33, 35-38). We have amended Lines 116 to 130 of the revised manuscript to more clearly outline our aims and hypotheses. We underline our hypotheses here for convenience:

“Here, we introduce methods for constructing an individual-specific measure of cortical asymmetry, called the shape asymmetry signature (SAS; see Materials and methods). The SAS characterizes pure shape asymmetries of the whole cortical surface, independent of variations in size, across a spectrum of spatial scales. We apply this methodology to three independent longitudinal datasets to test the hypothesis that cortical shape asymmetry is a highly personalized and robust feature that can identify individuals, akin to a cortical asymmetry fingerprint. We then use the identifiability values to identify optimal spatial scales at which robust individual differences are most salient. We also compare the identifiability of the SAS and shape descriptors of individual hemispheres, asymmetries in traditional size-based descriptors, or patterns of inter-regional functional connectivity (so-called connectome fingerprinting (46)) to test the hypothesis that the SAS is a more individually unique property of brain organization than unihemispheric and functional properties. We further elucidate the relationships between the SAS and sex, handedness, as well as cognitive performance across multiple tasks. Finally, we test the hypothesis that individual-specific asymmetry features are largely driven by environmental influences using classical heritability modelling of twin data.”

9. Line 124-125, it is not clear what are the shape descriptors your analyse, nor which aspects of cortical shape asymmetries you study.

Lines 124 to 125 are just the first sentence of the first paragraph of the Results section and are intended to orient readers. The descriptions come below. We have now amended the Lines 133 to 135 of the revised manuscript so that it reads:

“To understand how cortical shape asymmetries vary across individuals, we examined the degree to which different cortical shape descriptors (defined below) can be used to identify individual brains from a large sample of T1-weighted magnetic resonance images (MRIs).”

10. Lines 132-145, 150-161, 181-186 should be in the method section.

We agree that these lines contain methodological detail. However, given the eLife format, which places Methods at the end of the paper, we contend that it is essential to include this level of methodological detail in the Results section to provide sufficient context for interpreting the Results.

11. Line 146-150: methods too, and not clear here why it relates to studies of bilateral variation.

We performed the identifiability analysis to investigate whether the cortical shape asymmetry is an individualized feature of human brain. This is an excellent and widely-used (47, 48) method for determining just how individually specific such asymmetries are. As per our response to Comment 10, we include this level of methodological detail here to provide context for interpreting the Results section.

12. Line 190-191. How do you reach this conclusion? Should be true if DA is present, but how can you compare differences in R-L shapes and shape of an object?

Identifiability quantifies person-specific variability. If a phenotype has high identifiability, then it is highly specific to individuals. Given that the SAS is associated with higher identifiability than shape descriptors of unilateral hemispheres and descriptors obtained from combining both left and right hemispheres, we can conclude that the SAS captures greater individual variability. The eigen decomposition of the LBO allows us to characterize and compare shapes, as detailed in the Materials and methods section.

13. Line 194-197: the conclusion is that one hemisphere does not predict the shape of the other. OK. But why do not you quantify those bilateral variations? Moreover, there is no quantification of absolute asymmetry, DA, or FA. Would not it be possible to extract information about those parameters with you approach? That would be much more informative.

The SAS is precisely designed to quantify hemispheric differences in shape. It is possible To compute an absolute version of the SAS, but this removes specific individually unique information. We confirm this in the analysis in Author response image 1, which shows that the absolute value of SAS is associated with lower identifiability than our original SAS.

Author response image 1

14. Figures 1 and 2 are more about the methods that true results.

Figure 1 provides a visual schematic of our methodological approach and is essential for contextualizing the Results section. It is common to include such figures in our field. Figure 2 presents some of our key findings, comparing the identifiability of the SAS with other shape measures and truncating the spatial scales to represent the maximal identifiability.

15. Line 219-223. What is the implication of this? Isn’t it related to the fact that the approximation for lower values is too large to extrapolate any information?

The implications of the maximal subject identifiability at coarse scales are discussed in the Identifiability of cortical shape asymmetry is maximal at coarse scales section of the Discussion. We repeatedly evaluated the identifiability using up to 1000 eigenvalues. There is no reason to think that this value is too large to extrapolate useful information. For example, Germanaud et al. (49) used 5000 eigenfunctions (corresponding to 5000 eigenvalues) to examine cortical curvature. As per our response to Comment 1 of Reviewers 1’s Public Review, we have now included details about the potential role of measurement error at finer spatial scales in Lines 266 to 269, 437 to 439, and Figure 2—figure supplement 2.

16. Line 226. I do not understand what are identifiability scores, a p = 0, and do not have any idea about the meaning in the variation of the scores between samples. Moreover, why do you do comparison between sample and not for each individual? Finally, why do you have differences between the samples, one with a flat area, differences in shape and values between samples? It seems that you describe global results by sample but never extract something from the descriptions of general results.

Thank you for your questions. We have amended this sentence in Lines 240 to 242 of the revised manuscript:

“At these scales, the subject identifiability scores were 4.93 (P = 0; estimated by permutation; see Statistical analysis section in Materials and methods for details) for OASIS-3 and 5.03 (P = 0) for ADNI.”

To compute the identifiability score, we estimate correlations between all possible pairs of time 1 and time 2 phenotypes (e.g., SAS scores). The identifiability score is essentially the difference between the correlations observed for the same person at both time points and different people at the two time points. In other words, identifiability is high if a person i’s time 1 scan is more correlated with his/her scan at time 2 than with someone else’s scan. The formal definition of identifiability is provided in the Subject Identification section of the Materials and methods.

We perform statistical inference on the identifiability scores by permuting the participant identities. This gives us a distribution of identifiability scores under the null hypothesis. The p-value is then given by the fraction of null values that exceed the observed identifiability score. A p = 0 means that no null values exceeded the observed score. This approach is necessary as no parametric statistic is available for inference on identifiability scores, which is commonly used in the literature (e.g., (46, 48)). This procedure is outlined in detail in the Materials and methods section.

We examine three datasets to test the robustness of our findings; i.e., we replicate our results in independent samples. It does not really make sense to combine the datasets since they represent different cohorts measured with different scanners. The fact that we observe such consistent findings across all cohorts speaks to the robustness of our results. We focus on the general, consistent trends across the three cohorts in the Discussion section.

17. Line 230-238: move to the method section.

This description is essential for readers to understand the full context of the results that follow, since the Materials and methods are presented at the end of the paper.

18. Line 238-242: what is the implication of this? What does it mean?

The eigen-groups provide a way to group eigenvalues into spatial variations of similar wavelengths, in direct analogy to spherical harmonics, as detailed in Lines 730 to 750 of the revised manuscript. We also explain the implication and meaning of eigen-groups in Lines 259 to 266 of the revised manuscript:

“Thus, the first 144 eigenvalues represent a stable and robust characteristic scale at which individual uniqueness in cortical shape asymmetry is strongest. The 11th group corresponds to a wavelength of approximately 37 mm in the case of the population-based template (fsaverage in FreeSurfer; Appendix 1—Table 1 shows the corresponding wavelengths of the first 14 eigen–groups; Figure 2G shows the spatial scales corresponding to the cumulative eigen-groups). A reconstruction of the cortical surface using the first 144 eigenfunctions is shown in Figure 2H.

The reconstruction captures shape variations at a coarse scale, representing major primary and secondary sulci, but with minimal additional details.”

19. 243. Why 144 would be a "stable and robust scale"? how can you decide it is 144.

251-253. OK, it works for individual identification. But is it better than any simple comparison of brain shape? Why is it interesting to have tool that allows to identify an individual with 2 mri dataset? Moreover, I would prefer to see more results about bilateral variation thanks to your approach instead of this possibility of identification.

We choose 144 because the first 144 eigenvalues correspond to the first 11 eigen-groups, and across the three datasets peak or near-peak identifiability is observed within the 11th eigen-group. It is stable and robust because we obtain the same result across three independent datasets. We show in Figures 2 and 3 that the SAS offers stronger identifiability than unihemispheric shape descriptors and asymmetries of cortical volume, cortical thickness, or surface area, so it is better than simpler shape and size-related descriptors. We focus on identifiability as a way of quantifying how individually unique shape asymmetries are. This question speaks to how variable such asymmetries are across the population. Note that the eigenvalues of the LBO provide a compact representation of shape variation at different scales, and the corresponding eigenfunctions (examples of which are shown in Figures 2 (G) and (H)) represent the basis set of spatial patterns. The SAS thus offers an efficient summary of cortical asymmetries. In the revised manuscript, we now include Figure 1—figure supplement 1, which more clearly illustrates the correspondence between different SAS values and cortical shape at different spatial scales. After determining the spatial scales with the optimal subject identifiability, we linked the SAS to sex and cognitive function, as discussed in Lines 309 to 361 of the revised manuscript.

20. Lines 264-266. You do not demonstrate this supposed higher ability.

We have demonstrated that “individual variability in brain anatomy is higher when considering asymmetries in cortical shape compared to more traditional size-based morphological descriptors” in Figure 3 (A) to (B) and Figure 3—figure supplement 1. Specifically, Figure 3 (A) shows that identifiability scores for the SAS are higher than those obtained for asymmetries based on cortical surface area (identifiability score = 0.81), volume (identifiability score = 0.66), and thickness (identifiability score = 0.33). Figure 3—figure supplement 1 demonstrates that the identifiability scores calculated from the surface area normalized SAS are generally higher than the scores calculated using native eigenvalues and eigenvalues with volume normalization (but without surface area normalization) for individual hemispheres, the combination of both hemispheres, and asymmetry.

21. Lines 271-272. OK. But what is the implication of this?

Functional connectivity has been shown to be a highly personalized brain phenotype, akin to a neuro-fingerprint, in numerous studies (46, 48, 50, 51). Here, we show that the SAS has an even better subject identifiability than this widely studied feature. It indicates that cortical shape asymmetries are highly specific to individual brains.

22. Lines 291-295: move to methods. And why did you decide to do this way?

As detailed above, eLife presents the Materials and methods section at the end of the article, hence this detail is required to provide sufficient context for the results that follow. Using a GLM with permutation testing in this context is standard in the field for this sort of analysis. We use permutation testing to infer statistical significance because the subjects in the HCP dataset are not independent; some of them are from the same twin pairs or non-twin siblings.

23. Figure 4: that is the first presentation of bilateral variation. But not test, no description, no quantification of DA, FA, absolute asym…. Why?

The SAS quantifies cortical shape asymmetries. Its effects are quantified extensively in Figures 2 to 4. The brain pictures in Figure 4 (B) simply show the eigenfunctions that correspond to the eigenvalues at which significant associations were identified. It is intended simply to show the gradients of spatial variation associated with the implicated eigenvalues. Detailed descriptions of eigenvalues and eigenfunctions are provided in the Spectral shape analysis section of Materials and methods. To clarify this further, we have amended Lines 332 to 334:

“(B) The corresponding eigenfunction of each eigenvalue in panel (A) that shows the gradients of spatial variation on a population-based template.”

24. Figure 4B, what is the color code?

It is the value of each cortical point for each eigenfunction. We have now added colorbars in Figure 4 (B):

25. Lines 317-324: methods.

As detailed above, it is difficult for readers to understand this section without briefly describing the methods first.

26. Lines 33-335: you mention direction of asymmetries, but there is no test for DA for those parameters.

Thank you for your question. To avoid confusion, we have changed the terms in the revised manuscript, replacing DA with population-based asymmetry to refer to the asymmetry map across a population. The direction of the asymmetry is captured by the sign of the SAS, such that negative values indicate rightward asymmetry and positive values indicate leftward asymmetry. The direction of asymmetry is specific to individuals. Since the canonical correlation analysis measures individual variations in both the shape asymmetry and cognitive function, our interpretation is correct.

27. Figure s5. What are the results for absolute asymmetry?

We find that when using absolute asymmetry (i.e., the eigen-groups from the SAS with absolute value), the canonical correlation analysis (CCA) results are not significant (PFWER > 0.6). As mentioned in our response to Comment 13, absolute asymmetry also shows inferior identifiability to our SAS measure.

28. Lines 344-349: methods.

As per above, this description is necessary for context.

29. Line 352-3. Ok, but why is it interesting to exclude sulcal and gyral features? Isn't it a too large approximation of the shape? Moreover, you never mention any attempts to test measurement errors, that may have a large impact on FA analyses.

To clarify, we do not explicitly exclude sulcal and gyral features. Our findings tell us that identifiability is optimal at spatial scales that are coarser than fine-grained anatomical features such as some specific sulci and gyri. This is interesting because it indicates that highly unique and individual properties of brain shape asymmetry are found at coarse scales. As indicated in our response to Comments 1 and 3 of the Public Review of Reviewer 1, we revised the manuscript to now include a characterization of measurement error across scales and show that the SAS is robust to variations in scan quality, but that fine spatial scales may be too noisy to inform identifiability.

30. Lines 357-9. Why?

As shown in Figure 6, strong heritability is only found within the first six eigen-groups, corresponding to wavelengths greater than or equal to ~65 mm. The insets show the corresponding spatial scales of the reconstructions.

31. Lines 378-379: what is the evidence that proves this 1/3 relationship?

Figure 6 (C) shows that the of the SAS never exceeds 0.30.

32. Discussion

391, in my opinion you have not examined FA.

393: highly personalized, ok, but it is not the most interesting aspect of FA in theory.

We have revised Lines 410 to 414 to remove all the FA terms in the revised manuscript:

“Here, we employed a multiscale approach to quantify individualized shape asymmetries of the human cerebral cortex. We found that cortical shape asymmetries were highly personalized and robust, with shape asymmetries at coarse spatial scales being the most discriminative among individuals, showing differences between males and females, and correlating with individual differences in cognition.”

33. Line 394 "most discriminative", it has not been demonstrated.

We have demonstrated this in Figure 2 (A) to (C) and Lines 201 to 211 of the revised manuscript by showing that the SAS has the highest identifiability.

34. Lines 390-398 are an abstract of the paper, not the beginning of a discussion.

In our field it is common to begin the Discussion section with a brief summary of findings.

35. Line 401 which "meaningful effects"?

We have revised this sentence in Lines 420 to 421 of the revised manuscript so that it now reads:

“These approaches may ignore meaningful effects (i.e., properties that are individually unique and correlated with cognition) at coarser spatial scales.”

36. Line 403-404 this image is not appropriate. Line 405-6: ok, but it has to give results that are interpretable, it is not really the case here.

Lines 403 to 404 of the original manuscript state that:

“Our approach is akin to studying seismic waves of earthquakes with different wave frequencies at the global tectonic scale, instead of focusing on a particular city.”

This sentence provides an intuitive understanding for our multiscale analysis compared to traditional point-wise analyses. In fact, our approach draws on mathematical tools established for over a century and used in diverse fields of physics and engineering to characterize spatial properties of different systems. We contend that our results are indeed interpretable, as we outlined in the manuscript, but the interpretation is just different to that offered by traditional neuroscientific analyses. As per our response to Reviewer 1’s Comment 2, the revised manuscript now includes Figure 2—figure supplement 2 to aid interpretation.

37. Lines 407 and following: there are several papers on endocranial casts with detailed shape information.

We agree that there are several papers on endocranial casts, but most of those papers simply quantified the linear dimensions of Broca’s cap or petalia and did not characterize the actual 3D shape of the entire endocranial casts or brain surface (16). Many of those papers also ignored individual variations in actual shapes, focusing only on average patterns across a population (16). We have revised the sentence in Line 427 of the revised manuscript:

“Few studies have assessed individual variations in shape at coarse scales.”

38. 410 which time?

Kong et al. (6) used the HCP test-retest dataset, which is one of the datasets we used in this study, and the intervals between time 1 and time 2 were about one to 11 months.

39. 411-412, what is the implication?

We have added an additional sentence in Lines 432 to 434 of the revised manuscript:

“Wachinger et al. (21) used shape descriptors at coarse scales derived from the eigenvalues of the LBO for all brain structures to achieve accurate subject identification. Taken together with our findings, these results indicate that coarse features of cortical shape are highly personalized and unique to individuals.”

40. 416: why are not they "captured reliably"?

Line 416 states that:

“It is possible that these fine-grained features are not captured reliably, meaning that they contribute noise to identifiability analyses.”

In brief, we provide an additional figure (Figure 2—figure supplement 2), which used one subject’s images from two MRI sessions. The figure shows that more inter-sessional differences (quantified by the Euclidean distances between the vertices) between time 1 and time 2 images are more observable at finer spatial scales. In other words, information at finer scales was more vulnerable to measurement noise or subtle fluctuations in volume affecting shape.

41. 419 explain in which aspects it is a "more comprehensive characterization".

We have revised the sentence in Lines 441 to 442 of the revised manuscript:

“Our multiscale approach affords a more comprehensive characterization of shape variations across multiple spatial scales.”

42. Lines 434-440: No link with your study and results.

We found that cortical shape asymmetry is mainly driven by unique environmental effects. This section outlines potential ways in which this may occur.

43. Line 463: ok, but please elaborate a bit.

Thank you for your suggestions, but our scale-specific asymmetry analysis is a new approach with no existing references to cite and the specific mechanism is beyond the scope of the present study. For this reason, we say that it requires further investigation.

44. Lines 474-475: this is a rather naïve statement. Please explain or delete.

This sentence briefly summarizes the preceding paragraph and is merely based on our results. Specifically, at the beginning of this paragraph (Lines 490 to 493 of the revised manuscript), we mentioned:

“We found that a greater leftward asymmetry across a wide range of spatial scales, corresponding to wavelengths of approximately 37, 75, 95, and 170 mm, and greater rightward asymmetry at wavelengths of approximately 40, 55, 65, and 120 mm, are associated with better performance across nearly all cognitive measures considered.

Based on our result, we concluded that different scales may have different relationships with cognition. For this reason, a multiscale approach can reveal associations between brain and behavior that may not be detected with single-scale approaches.”

45. Conclusion: to be removed, it is an abstract.

We believe that it is useful to end the paper with a brief summary of findings. We have renamed this section as Summary instead of Conclusions.

46. Methods

503: how do you determine those errors?

On the reverse, you do not quantify your own methodological error. Have you tried to repeat your methods? What is the effect of imaging resolution (always important when working on bilateral variation)?

Line 503 reads:

“We excluded six subjects whose SAS was an outlier in at least one of those sessions due to poor image quality and major errors in image segmentation. These subjects had more than two eigenvalues of the first 200 eigenvalues that departed from the population mean values by more than four standard deviations.”

Segmentation errors are determined via manual inspection of all FreeSurfer outputs. FreeSurfer is a widely validated technique (52-54), and errors are typically determined by visual inspection. Moreover, the eigen decomposition of the LBO is robust to variations in mesh resolution and image quality (4). As outlined in our responses to Reviewer 1’s Recommendations for the authors and Public Review Comment 3, we now also show that the Euler number from FreeSurfer (1), a quantitative metric of image quality, was not associated with the SAS. We have repeated the SAS on three independent datasets to show that it quantified individualized brain features.

47. Line 759 what are the "morphological measurements" that you mention here?

We now explain "morphological measurements" in Lines 789 to 790 of the revised manuscript:

“In this study, we adjusted the asymmetry index for the mean of each morphological measurement, such as the asymmetry of cortical thickness, volume, and surface area.”

48. Line 770: what are the asym-index measurements?

The asymmetry-index measurements correspond to the asymmetry of cortical thickness, surface area, and volume, all of which were quantified via the asymmetry index. We have revised this sentence in Lines 801 to 803 of the revised manuscript:

“We created one vector per size-related measure that quantified the asymmetry index per subject and then used these asymmetry indices in the subject identifiability analyses.”

49. Line 798: what is reduced?

As indicated in Lines 837 to 841 of the revised manuscript, we used PCA to reduce the dimensionality of the cognitive data, replacing 13 different cognitive test scores with 4 principal components explaining 80% of the test variance. An advantage of this approach is that the components are orthogonal, which mitigates biases in the CCA. We reduced the dimensionality of the SAS values from 143 eigenvalues to 11 by taking the mean of the first 11 eigen-groups. Our rationale for using eigen-groups is outlined in detail in Lines 730 to 750 of the revised manuscript.

References

1. Rosen AFG, Roalf DR, Ruparel K, Blake J, Seelaus K, Villa LP, et al. Quantitative assessment of structural image quality. Neuroimage. 2018;169:407-18.

2. Morgan SE, Seidlitz J, Whitaker KJ, Romero-Garcia R, Clifton NE, Scarpazza C, et al. Cortical patterning of abnormal morphometric similarity in psychosis is associated with brain expression of schizophrenia-related genes. Proc Natl Acad Sci U S A. 2019;116(19):9604-9.

3. Kaufmann T, van der Meer D, Doan NT, Schwarz E, Lund MJ, Agartz I, et al. Common brain disorders are associated with heritable patterns of apparent aging of the brain. Nat Neurosci. 2019;22(10):1617-23.

4. Reuter M. Hierarchical Shape Segmentation and Registration via Topological Features of Laplace-Beltrami Eigenfunctions. International Journal of Computer Vision. 2009;89(2-3):287-308.

5. Maingault S, Tzourio-Mazoyer N, Mazoyer B, Crivello F. Regional correlations between cortical thickness and surface area asymmetries: A surface-based morphometry study of 250 adults. Neuropsychologia. 2016;93(Pt B):350-64.

6. Kong XZ, Postema M, Schijven D, Castillo AC, Pepe A, Crivello F, et al. Large-Scale Phenomic and Genomic Analysis of Brain Asymmetrical Skew. Cereb Cortex. 2021.

7. Deep-Soboslay A, Hyde TM, Callicott JP, Lener MS, Verchinski BA, Apud JA, et al. Handedness, heritability, neurocognition and brain asymmetry in schizophrenia. Brain. 2010;133(10):3113-22.

8. Dragovic M. Categorization and validation of handedness using latent class analysis. Acta Neuropsychiatr. 2004;16(4):212-8.

9. Narr KL, Bilder RM, Luders E, Thompson PM, Woods RP, Robinson D, et al. Asymmetries of cortical shape: Effects of handedness, sex and schizophrenia. Neuroimage. 2007;34(3):939-48.

10. Vlachos F, Avramidis E, Dedousis G, Katsigianni E, Ntalla I, Giannakopoulou M, et al. Incidence and Gender Differences for Handedness among Greek Adolescents and Its Association with Familial History and Brain Injury Research in Psychology and Behavioral Sciences. 2013;1(1):6-10.

11. Perlaki G, Horvath R, Orsi G, Aradi M, Auer T, Varga E, et al. White-matter microstructure and language lateralization in left-handers: a whole-brain MRI analysis. Brain Cogn. 2013;82(3):319-28.

12. Oldfield RC. The assessment and analysis of handedness: the Edinburgh inventory. Neuropsychologia. 1971;9(1):97-113.

13. Willems RM, Van der Haegen L, Fisher SE, Francks C. On the other hand: including left-handers in cognitive neuroscience and neurogenetics. Nat Rev Neurosci. 2014;15(3):193-201.

14. Reuter M, Wolter FE, Peinecke N. Laplace–Beltrami spectra as ‘Shape-DNA’ of surfaces and solids. Computer-Aided Design. 2006;38(4):342-66.

15. Yeo RA, Ryman SG, Pommy J, Thoma RJ, Jung RE. General cognitive ability and fluctuating asymmetry of brain surface area. Intelligence. 2016;56:93-8.

16. Neubauer S, Gunz P, Scott NA, Hublin JJ, Mitteroecker P. Evolution of brain lateralization: A shared hominid pattern of endocranial asymmetry is much more variable in humans than in great apes. Sci Adv. 2020;6(7):eaax9935.

17. Nadig A, Seidlitz J, McDermott CL, Liu S, Bethlehem R, Moore TM, et al. Morphological integration of the human brain across adolescence and adulthood. Proc Natl Acad Sci U S A. 2021;118(14).

18. Kong XZ, Mathias SR, Guadalupe T, Group ELW, Glahn DC, Franke B, et al. Mapping cortical brain asymmetry in 17,141 healthy individuals worldwide via the ENIGMA Consortium. Proc Natl Acad Sci U S A. 2018;115(22):E5154-E63.

19. Kong XZ, Postema MC, Guadalupe T, de Kovel C, Boedhoe PSW, Hoogman M, et al. Mapping brain asymmetry in health and disease through the ENIGMA consortium. Hum Brain Mapp. 2020.

20. Postema MC, van Rooij D, Anagnostou E, Arango C, Auzias G, Behrmann M, et al. Altered structural brain asymmetry in autism spectrum disorder in a study of 54 datasets. Nat Commun. 2019;10(1):4958.

21. Toga AW, Thompson PM. Mapping brain asymmetry. Nat Rev Neurosci. 2003;4(1):37-48.

22. Esteves M, Ganz E, Sousa N, Leite-Almeida H. Asymmetrical Brain Plasticity: Physiology and Pathology. Neuroscience. 2020.

23. Plessen KJ, Hugdahl K, Bansal R, Hao X, Peterson BS. Sex, age, and cognitive correlates of asymmetries in thickness of the cortical mantle across the life span. J Neurosci. 2014;34(18):6294-302.

24. Cai Y, Liu J, Zhang L, Liao M, Zhang Y, Wang L, et al. Grey matter volume abnormalities in patients with bipolar I depressive disorder and unipolar depressive disorder: a voxel-based morphometry study. Neurosci Bull. 2015;31(1):4-12.

25. Fling BW, Dutta GG, Schlueter H, Cameron MH, Horak FB. Associations between Proprioceptive Neural Pathway Structural Connectivity and Balance in People with Multiple Sclerosis. Front Hum Neurosci. 2014;8:814.

26. Gunturkun O, Strockens F, Ocklenburg S. Brain Lateralization: A Comparative Perspective. Physiol Rev. 2020;100(3):1019-63.

27. Royer C, Delcroix N, Leroux E, Alary M, Razafimandimby A, Brazo P, et al. Functional and structural brain asymmetries in patients with schizophrenia and bipolar disorders. Schizophr Res. 2015;161(2-3):210-4.

28. Takao H, Abe O, Yamasue H, Aoki S, Sasaki H, Kasai K, et al. Gray and white matter asymmetries in healthy individuals aged 21-29 years: a voxel-based morphometry and diffusion tensor imaging study. Hum Brain Mapp. 2011;32(10):1762-73.

29. Clark GM, Crow TJ, Barrick TR, Collinson SL, James AC, Roberts N, et al. Asymmetry loss is local rather than global in adolescent onset schizophrenia. Schizophr Res. 2010;120(1-3):84-6.

30. Ratnanather JT, Poynton CB, Pisano DV, Crocker B, Postell E, Cebron S, et al. Morphometry of superior temporal gyrus and planum temporale in schizophrenia and psychotic bipolar disorder. Schizophr Res. 2013;150(2-3):476-83.

31. Corballis MC. Early signs of brain asymmetry. Trends Cogn Sci. 2013;17(11):554-5.

32. Gomez-Robles A, Hopkins WD, Sherwood CC. Increased morphological asymmetry, evolvability and plasticity in human brain evolution. Proc Biol Sci. 2013;280(1761): 115-26.

33. Gomez-Robles A, Hopkins WD, Schapiro SJ, Sherwood CC. The heritability of chimpanzee and human brain asymmetry. Proc Biol Sci. 2016;283(1845).

34. Corballis MC, Haberling IS. The Many Sides of Hemispheric Asymmetry: A Selective Review and Outlook. J Int Neuropsychol Soc. 2017;23(9-10):710-8.

35. Sherwood CC, Gómez-Robles A. Brain Plasticity and Human Evolution. Annual Review of Anthropology. 2017;46(1):399-419.

36. Francks C. Exploring human brain lateralization with molecular genetics and genomics. Ann N Y Acad Sci. 2015;1359:1-13.

37. de Kovel CGF, Lisgo SN, Fisher SE, Francks C. Subtle left-right asymmetry of gene expression profiles in embryonic and foetal human brains. Sci Rep. 2018;8(1):12606.

38. Graham J, Özener B. Fluctuating Asymmetry of Human Populations: A Review. Symmetry. 2016;8(12).

39. Zhao L, Matloff W, Shi Y, Cabeen RP, Toga AW. Mapping Complex Brain Torque Components and Their Genetic and Phenomic Architecture in 24,112 healthy individuals. 2021.

40. Sha Z, Schijven D, Carrion-Castillo A, Joliot M, Mazoyer B, Fisher SE, et al. The genetic architecture of structural left-right asymmetry of the human brain. Nat Hum Behav. 2021.

41. Wachinger C, Golland P, Kremen W, Fischl B, Reuter M, Alzheimer's Disease Neuroimaging I. BrainPrint: a discriminative characterization of brain morphology. Neuroimage. 2015;109:232-48.

42. Wachinger C, Salat DH, Weiner M, Reuter M, Alzheimer's Disease Neuroimaging I. Whole-brain analysis reveals increased neuroanatomical asymmetries in dementia for hippocampus and amygdala. Brain. 2016;139(Pt 12):3253-66.

43. Richards R, Greimel E, Kliemann D, Koerte IK, Schulte-Korne G, Reuter M, et al. Increased hippocampal shape asymmetry and volumetric ventricular asymmetry in autism spectrum disorder. Neuroimage Clin. 2020;26:102207.

44. Ge T, Reuter M, Winkler AM, Holmes AJ, Lee PH, Tirrell LS, et al. Multidimensional heritability analysis of neuroanatomical shape. Nat Commun. 2016;7:13291.

45. Wachinger C, Nho K, Saykin AJ, Reuter M, Rieckmann A, Alzheimer's Disease Neuroimaging I. A Longitudinal Imaging Genetics Study of Neuroanatomical Asymmetry in Alzheimer's Disease. Biol Psychiatry. 2018;84(7):522-30.

46. Finn ES, Shen X, Scheinost D, Rosenberg MD, Huang J, Chun MM, et al. Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity. Nat Neurosci. 2015;18(11):1664-71.

47. Mansour LS, Tian Y, Yeo BTT, Cropley V, Zalesky A. High-resolution connectomic fingerprints: Mapping neural identity and behavior. Neuroimage. 2021;229:117695.

48. Amico E, Goni J. The quest for identifiability in human functional connectomes. Sci Rep. 2018;8(1):8254.

49. Germanaud D, Lefevre J, Toro R, Fischer C, Dubois J, Hertz-Pannier L, et al. Larger is twistier: spectral analysis of gyrification (SPANGY) applied to adult brain size polymorphism. Neuroimage. 2012;63(3):1257-72.

50. Mars RB, Passingham RE, Jbabdi S. Connectivity Fingerprints: From Areal Descriptions to Abstract Spaces. Trends Cogn Sci. 2018;22(11):1026-37.

51. Noble S, Spann MN, Tokoglu F, Shen X, Constable RT, Scheinost D. Influences on the Test-Retest Reliability of Functional Connectivity MRI and its Relationship with Behavioral Utility. Cereb Cortex. 2017;27(11):5415-29.

52. Fischl B, Salat DH, Busa E, Albert M, Dieterich M, Haselgrove C, et al. Whole Brain Segmentation Automated Labeling of Neuroanatomical Structures in the Human Brain. Neuron. 2002;33(3):341–55.

53. Fischl B. FreeSurfer. Neuroimage. 2012;62(2):774-81.

54. Cardinale F, Chinnici G, Bramerio M, Mai R, Sartori I, Cossu M, et al. Validation of FreeSurfer-estimated brain cortical thickness: comparison with histologic measurements. Neuroinformatics. 2014;12(4):535-42.

https://doi.org/10.7554/eLife.75056.sa2

Article and author information

Author details

  1. Yu-Chi Chen

    1. Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Melbourne, Australia
    2. Monash Biomedical Imaging, Monash University, Melbourne, Australia
    3. Monash Data Futures Institute, Monash University, Melbourne, Australia
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing
    For correspondence
    yu-chi.chen@monash.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9167-6417
  2. Aurina Arnatkevičiūtė

    Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Melbourne, Australia
    Contribution
    Software, Visualization, Methodology, Writing – review and editing
    Competing interests
    No competing interests declared
  3. Eugene McTavish

    1. Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Melbourne, Australia
    2. Monash Biomedical Imaging, Monash University, Melbourne, Australia
    3. Healthy Brain and Mind Research Centre, Faculty of Health Sciences, Australian Catholic University, Fitzroy, Australia
    Contribution
    Software, Methodology
    Competing interests
    No competing interests declared
  4. James C Pang

    1. Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Melbourne, Australia
    2. Monash Biomedical Imaging, Monash University, Melbourne, Australia
    Contribution
    Software, Supervision, Methodology, Writing – review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-2461-2760
  5. Sidhant Chopra

    1. Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Melbourne, Australia
    2. Monash Biomedical Imaging, Monash University, Melbourne, Australia
    3. Department of Psychology, Yale University, New Haven, United States
    Contribution
    Software, Methodology, Visualization
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-0866-3477
  6. Chao Suo

    1. Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Melbourne, Australia
    2. Monash Biomedical Imaging, Monash University, Melbourne, Australia
    3. BrainPark, Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Melbourne, Australia
    Contribution
    Resources, Data curation, Software, Methodology
    Competing interests
    No competing interests declared
  7. Alex Fornito

    1. Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Melbourne, Australia
    2. Monash Biomedical Imaging, Monash University, Melbourne, Australia
    Contribution
    Conceptualization, Resources, Supervision, Funding acquisition, Investigation, Visualization, Methodology, Project administration, Writing – review and editing
    Contributed equally with
    Kevin M Aquino
    Competing interests
    Reviewing editor, eLife
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9134-480X
  8. Kevin M Aquino

    1. Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Melbourne, Australia
    2. Monash Biomedical Imaging, Monash University, Melbourne, Australia
    3. School of Physics, University of Sydney, Sydney, Australia
    4. Center of Excellence for Integrative Brain Function, University of Sydney, Sydney, Australia
    5. BrainKey Inc, San Francisco, United States
    Contribution
    Conceptualization, Resources, Software, Supervision, Validation, Investigation, Visualization, Methodology, Project administration, Writing – review and editing
    Contributed equally with
    Alex Fornito
    Competing interests
    is a scientific advisor and shareholder in BrainKey Inc, a medical image analysis software company
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7435-0236

Funding

Sylvia and Charles Viertel Charitable Foundation (Senior Medical Research Fellowship)

  • Alex Fornito

National Health and Medical Research Council (1197431)

  • Alex Fornito

National Health and Medical Research Council (1146292)

  • Alex Fornito

Australian Research Council (DP200103509)

  • Alex Fornito

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

Acknowledgements

AF was supported by the Sylvia and Charles Viertel Foundation, National Health and Medical Research Council (IDs: 1197431 and 1146292), and Australian Research Council (ID: DP200103509). Data were provided in part by OASIS-3: Principal Investigators: T Benzinger, D Marcus, J Morris; NIH P50 AG00561, P30 NS09857781, P01 AG026276, P01 AG003991, R01 AG043434, UL1 TR000448, R01 EB009352. AV-45 doses were provided by Avid Radiopharmaceuticals, a wholly owned subsidiary of Eli Lilly. Data were provided in part by the Human Connectome Project, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University. Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc; Cogstate; Eisai Inc; Elan Pharmaceuticals, Inc; Eli Lilly and Company; EuroImmun; F Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (https://www.fnih.org/). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California.

Ethics

Human subjects: This study only involved subjects from the open-sourced datasets, and all subjects were de-identified by the datasets. Each dataset was approved by its relevant ethics committee and obtained written informed consent from each participant.

Senior Editor

  1. Christian Büchel, University Medical Center Hamburg-Eppendorf, Germany

Reviewing Editor

  1. Ingrid S Johnsrude, University of Western Ontario, Canada

Reviewer

  1. Antoine Balzeau

Publication history

  1. Preprint posted: September 15, 2021 (view preprint)
  2. Received: October 28, 2021
  3. Accepted: October 4, 2022
  4. Accepted Manuscript published: October 5, 2022 (version 1)
  5. Accepted Manuscript updated: October 6, 2022 (version 2)
  6. Version of Record published: November 16, 2022 (version 3)

Copyright

© 2022, Chen et al.

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

Metrics

  • 843
    Page views
  • 239
    Downloads
  • 0
    Citations

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

Download links

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

Downloads (link to download the article as PDF)

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

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

  1. Yu-Chi Chen
  2. Aurina Arnatkevičiūtė
  3. Eugene McTavish
  4. James C Pang
  5. Sidhant Chopra
  6. Chao Suo
  7. Alex Fornito
  8. Kevin M Aquino
  9. for the Alzheimer's Disease Neuroimaging Initiative
(2022)
The individuality of shape asymmetries of the human cerebral cortex
eLife 11:e75056.
https://doi.org/10.7554/eLife.75056

Further reading

    1. Neuroscience
    Kevin Vinberg, Jörgen Rosén ... Fredrik Ahs
    Research Article Updated

    Understanding the neural basis for individual differences in the skin conductance response (SCR) during discriminative fear conditioning may inform on our understanding of autonomic regulation in fear-related psychopathology. Previous region-of-interest (ROI) analyses have implicated the amygdala in regulating conditioned SCR, but whole brain analyses are lacking. This study examined correlations between individual differences in SCR during discriminative fear conditioning to social stimuli and neural activity throughout the brain, by using data from a large functional magnetic resonance imaging study of twins (N = 285 individuals). Results show that conditioned SCR correlates with activity in the dorsal anterior cingulate cortex/anterior midcingulate cortex, anterior insula, bilateral temporoparietal junction, right frontal operculum, bilateral dorsal premotor cortex, right superior parietal lobe, and midbrain. A ROI analysis additionally showed a positive correlation between amygdala activity and conditioned SCR in line with previous reports. We suggest that the observed whole brain correlates of SCR belong to a large-scale midcingulo-insular network related to salience detection and autonomic-interoceptive processing. Altered activity within this network may underlie individual differences in conditioned SCR and autonomic aspects of psychopathology.

    1. Neuroscience
    Rong Zhao, Stacy D Grunke ... Joanna L Jankowsky
    Research Article

    Neurodegenerative diseases are characterized by selective vulnerability of distinct cell populations; however, the cause for this specificity remains elusive. Here we show that entorhinal cortex layer 2 (EC2) neurons are unusually vulnerable to prolonged neuronal inactivity compared with neighboring regions of the temporal lobe, and that reelin+ stellate cells connecting EC with the hippocampus are preferentially susceptible within the EC2 population. We demonstrate that neuronal death after silencing can be elicited through multiple independent means of activity inhibition, and that preventing synaptic release, either alone or in combination with electrical shunting, is sufficient to elicit silencing-induced degeneration. Finally, we discovered that degeneration following synaptic silencing is governed by competition between active and inactive cells, which is a circuit refinement process traditionally thought to end early in postnatal life. Our data suggests that the developmental window for wholesale circuit plasticity may extend into adulthood for specific brain regions. We speculate that this sustained potential for remodeling by entorhinal neurons may support lifelong memory but renders them vulnerable to prolonged activity changes in disease.