1. Neuroscience
Download icon

The heritability of multi-modal connectivity in human brain activity

  1. Giles L Colclough  Is a corresponding author
  2. Stephen M Smith
  3. Thomas E Nichols
  4. Anderson M Winkler
  5. Stamatios N Sotiropoulos
  6. Matthew F Glasser
  7. David C Van Essen
  8. Mark W Woolrich  Is a corresponding author
  1. University of Oxford, United Kingdom
  2. University of Warwick, United Kingdom
  3. University of Nottingham, United Kingdom
  4. Washington University, United States
Research Article
  • Cited 5
  • Views 1,677
  • Annotations
Cite as: eLife 2017;6:e20178 doi: 10.7554/eLife.20178

Abstract

Patterns of intrinsic human brain activity exhibit a profile of functional connectivity that is associated with behaviour and cognitive performance, and deteriorates with disease. This paper investigates the relative importance of genetic factors and the common environment between twins in determining this functional connectivity profile. Using functional magnetic resonance imaging (fMRI) on 820 subjects from the Human Connectome Project, and magnetoencephalographic (MEG) recordings from a subset, the heritability of connectivity among 39 cortical regions was estimated. On average over all connections, genes account for about 15% of the observed variance in fMRI connectivity (and about 10% in alpha-band and 20% in beta-band oscillatory power synchronisation), which substantially exceeds the contribution from the environment shared between twins. Therefore, insofar as twins share a common upbringing, it appears that genes, rather than the developmental environment, have the dominant role in determining the coupling of neuronal activity.

https://doi.org/10.7554/eLife.20178.001

Introduction

Intrinsic human brain activity enables inference about the pathways and processes of information transfer in the brain. Studying intrinsic activity, when the brain is in a resting state, has given insights into many aspects of healthy and diseased brain function. Resting-state function is characterised by spatially separated regions organised into networks of strongly correlated activity (Beckmann et al., 2005; Smith et al., 2012). These networks represent both local connectivity and longer-range communication. Importantly, the strength of resting-state functional connectivity reflects many aspects of cognitive function. Connectivity in the brain changes throughout the life cycle: networks associated with attention and control may continue to develop in late adolescence (Barnes et al., 2016) and network integrity degrades during the ageing process (Dennis and Thompson, 2014). Many neurological diseases, including schizophrenia and Alzheimer’s disease (Sheline and Raichle, 2013; Greicius, 2008), have been associated with major alterations to the strength and organisation of functional connectivity (Stam and van Straaten, 2012; Stam, 2014; van Straaten and Stam, 2013). In healthy subjects, intrinsic brain activity can predict not only performance in a task, but also the specific regions which will show increased activity during that task (Sala-Llonch et al., 2012; Yamashita et al., 2015; Zou et al., 2013; Smith et al., 2013; Tavor et al., 2016). Furthermore, recent evidence suggests that the organisation of human brain function at rest is associated with a broad range of behavioural and life-style traits, and reflects a generalised measure of intelligence (Smith et al., 2015).

Here, we set out to identify the relative importance of genetic and shared environmental factors in determining these fundamental patterns of cortical communication. We employ functional magnetic resonance imaging (fMRI) recordings to perform a heritability analysis on the strength of functional connectivity within a network of 39 regions. Additionally, we perform the same heritability analysis on source-localised magnetoencephalographic (MEG) recordings (Van Veen et al., 1997) from a subset of the subjects. This complementary analysis allows us to focus specifically on communication mediated by the coupling (correlation) of oscillatory amplitudes, within particular frequency bands, using MEG as a more direct measure of neuronal activity that is unaffected by vascular confounds.

We present separate analyses of functional networks estimated from the fMRI response (Smith et al., 2013), and of MEG-derived networks in the theta (4–8 Hz), alpha (8–13 Hz) and beta (13–30 Hz) oscillatory bands. These bands span the frequency range within which the most convincing patterns of resting-state MEG connectivity have been shown to be expressed (Hipp et al., 2012; Baker et al., 2014; Hillebrand et al., 2012; Brookes et al., 2011; Mantini et al., 2007; Marzetti et al., 2013; de Pasquale et al., 2012; de Pasquale et al., 2016). We employ resting-state recordings from 820 subjects from the Human Connectome Project (HCP; WU-Minn HCP Consortium et al., 2013); these data are publicly available, and have undergone a standardised pre-processing procedure. The HCP is a twin study, and the subjects from the S900 data release with all fMRI (MEG) resting-state scans comprise 103 (19) monozygotic twin pairs, who share 100% of their genetic structure and a common environment, and 54 (13) dizygotic twin pairs, who share 50% of their genetic structure and a common environment. Each subject provided three MEG scans (of 6 minutes each), and four fMRI scans (of 15 minutes each). This stratified sample allows estimation of the relative effects of genetic influence and environmental factors on the variability observed in connectivity structure.

We compare the similarity in overall network structure between pairs of subjects, assessing the extent to which the functional connectivity of two subjects becomes more alike as their proportion of shared background and genetics increases. We fit variance-component models on each network edge, using recently developed permutation and bootstrap-based methods for fast, accurate, non-parametric statistical inference with family-wise control of type I errors (Chen, 2014). This provides estimates of the mean genetic and shared environmental influences on the observed phenotypic variation in cortical connectivity. We run similar heritability analyses on the oscillatory power and BOLD variance within the regions of interest (ROIs) which constitute the network nodes, to determine whether the observed effects of genes and developmental environment on cortical connectivity could be attributable to simple differences in signal strength. Additionally, we analyse the heritability of cortical folding patterns in each ROI, to investigate whether genetic control of anatomical structure contributes to the heritability of connectivity.

Results

The functional connectivity measured with fMRI was estimated among 39 functionally defined cortical regions of interest (network nodes) by computing partial correlations between BOLD time courses representing each region. ROIs were spatially contiguous, covered both hemispheres, and were generated using fMRI from HCP subjects—see Materials and methods. The functional connectivity measured with MEG, corresponding to amplitude coupling of oscillations in each frequency band, was assessed between the same set of regions by correlating fluctuations in oscillatory power (Engel et al., 2013; O'Neill et al., 2015). Potential confounds to these MEG functional connectivity estimates induced by the leakage of source-reconstructed signals were reduced using a multivariate orthogonalisation technique (Colclough et al., 2015). These methods represent established practice for determining networks of functional connectivity in the two modalities (Varoquaux and Craddock, 2013; Smith et al., 2011; Smith et al., 2013; Brookes et al., 2012; Hipp et al., 2012; Colclough et al., 2016; de Pasquale et al., 2012; Hillebrand et al., 2012). The group-averaged functional networks are shown in Figure 1A and rendered in 3D in the supplementary videos.

Figure 1 with 6 supplements see all
Contribution of genetic factors to functional connectivity outweighs that of the environment shared between twins.

A. Grand average functional connectome for fMRI and for the theta, alpha and beta MEG oscillatory bands. The coloured edge maps show group-average network matrices for correlations in oscillatory amplitude in each band, and partial correlations in BOLD response, thresholded for visualisation. Nodes are annotated by cortical region, and labelled in Supplementary file 1. 3D renderings of these connectomes are shown in Figure 1—videos 1 to 4. B. Similarity of pairs of network matrices, separated by the relationship status of each pair. Subjects with a shared environment, and a greater proportion of shared genetics, have more similar organisation of neuronal coupling. Violin plots show distributions of distance values between pairs of network matrices estimated from single resting-state recording sessions, on a logarithmic scale relative to the mean network separation over all pairings, for pairs of unrelated subjects (UN), siblings (SI), dizygotic twin pairs (DZ), monozygotic twin pairs (MZ) and for repeated sessions with the same subject (SAME). C. Average genetic and shared environmental contributions to the variability of functional connectivity. The bar charts show the mean genetic component (heritability, h2) and mean shared environment component (c2) from a variance decomposition model fitted on each network edge, with 95% bootstrapped confidence intervals on the mean. Annotations indicate the difference in contribution between genes and the shared or developmental environment. The values of h2 and c2 are expressed as proportions of the total variance. The difference between their total and unity is e2, the remaining environmental and measurement noise component. D. Average genetic and shared environmental contributions to the variability of oscillatory power, or BOLD response, in each ROI. Stars indicate significant differences in mean value for each of the displayed comparisons: ***p<0.001; **p<0.01; *p<0.05; n.s., not significant. The non-parametric p-values were computed by permutation and corrected for multiple comparisons over the 21 tests performed.

https://doi.org/10.7554/eLife.20178.002

The fMRI connectome shows strong bilateral connectivity, together with fronto-parietal correlations and coupling between the anterior cingulate and posterior cingulate cortices, reflecting well-known patterns of resting-state networks in fMRI (Beckmann et al., 2005; Smith et al., 2012). In the MEG data, the alpha band shows a strongly connected visual system, with strong connectivity from the visual cortex extending out to the temporal and parietal lobes, and a highly coupled posterior cingulate. The beta-band exhibits strong bilateral coupling across the sensorimotor cortices, with connectivity continuing through the superior parietal lobes and down to the occipital cortex. The theta band also exhibits visual and motor connectivity. These patterns of connectivity are in general agreement with previously reported results in these frequency bands (Hillebrand et al., 2012; Brookes et al., 2011; Baker et al., 2014; de Pasquale et al., 2012; de Pasquale et al., 2016; Hipp et al., 2012; Marzetti et al., 2013; Mantini et al., 2007; Colclough et al., 2015).

Figure 1B (left-hand plot, middle row) shows that the structure of fMRI and MEG networks of functional activity are progressively more similar as the strength of relationship is increased, from unrelated subjects, through siblings and dizygotic twins to monozygotic twins (p<10-3 in each case, except for pMZ<DZα=0.003). (Although siblings and dizygotic twins share the same proportion of genetic material, siblings have different ages and inevitably share less similar environments than do twins. This may account for the significant increase in network similarity from siblings to dizygotic twins.) Also displayed, for comparison, is the distribution of network similarity over repeated recording sessions within the same subject. Tests for significance were performed by assessing the difference in mean of the logarithm of network separation, relative to the mean separation, using a non-parametric permutation-based t-test. All 21 tests performed for this paper are quoted after a false discovery rate correction for multiple comparisons.

A three-component variance model was fitted for the variability observed in the strength of coupling in individual edges. This model ascribes proportions of the variance in a phenotype, σ2, either to additive shared genetics (A), shared environmental factors (C), or to measurement error and other known sources of variance (E),

σ2=A+C+E.

We estimate each of these components as a fraction of the total variance, giving h2=A/σ2, known as the heritability of a trait, c2=C/σ2 and c2=E/σ2. These components are related by the identity

h2+c2+e2=1.

Estimates of the mean heritability and contribution from common environment over these connections were constructed with bootstrapped confidence intervals, and the significance of the mean heritability was tested using non-parametric permutation-based methods. Although there is a large amount of variability in the data which is not attributable to either of these factors, we found that additive genetics accounted for 17%, on average, of the functional connectivity measured with fMRI. In MEG, 19% of the variability in oscillatory coupling strengths in the beta band, and 8% in the alpha band, was determined by additive genetics. These estimates and comparisons are presented graphically in Figure 1C, and tabulated with confidence intervals and p-values in Supplementary files 2 and 3.

In the absence of a permutation scheme which is invariant under the null hypothesis of equal genetic and shared environmental factors, we additionally estimate h2-c2 with bootstrapped confidence intervals to investigate the relative importance of these two components. In the fMRI data and in beta-band oscillations, where we also find significant genetic influences, we identify a differential effect of heritability and twins’ shared environment (estimates with 95% confidence intervals, h2-c2=0.15[0.13,0.16] for the fMRI and h2-c2=0.13[0.04,0.22] for the beta-band results). Permutation tests for a significant contribution of heritability on each individual functional connection were performed, but no edges were significant after correcting for family-wise error at α=0.05. Parameter estimates for individual fMRI connections are shown in Figure 1—figure supplement 1.

To investigate whether our results were particular to our choice of parcellation, we re-ran our heritability analyses using the 15-dimensional ICA decomposition of resting-state fMRI recordings released by the HCP. This parcellation contains entire resting-state networks as each node of the decomposition, yielding a functional connectivity matrix that describes inter-network, rather than intra-network relationships. We found significant genetic influences in the fMRI data (h2=0.29[0.26,0.33], p=1×10-4, uncorrected), but not for any of the activity recorded with MEG (uncorrected permutation-based p-values for h2>0 were 0.69, 0.06 and 0.14 for the theta, alpha and beta bands respectively). The parameter estimates and confidence intervals are presented in Supplementary file 4.

The three-component model was also fitted to the variability in the logarithm of the signal power in each ROI, in both the fMRI and MEG data. Significant influences of genetic factors, on average over the nodes, were found only in the fMRI data (Figure 1D).

Finally, we fitted the heritability model to the cortical surface curvature of each subject, computing average h2 within each ROI (Figure 1—figure supplement 2). Cortical curvature exhibits the highest heritability around the sensorimotor areas and insula (as has been reported previously, in an analysis using a different sample of the HCP subjects; Sotiropoulos et al., 2015). We compared this spatial profile of the heritability of curvature to the average heritability of network connections for each ROI within each modality (that is, for each ROI, the average h2 of all connections involved with that ROI). In general, no significant positive correlations were found between the spatial arrangement of cortical curvature heritability and the heritability of network connections in particular ROIs. The exception was moderate correlation (ρ=0.39) with the MEG theta band connectivity pattern, for which there was no significant heritability of connection strength in the first instance. (These data are presented in Supplementary file 5.)

Discussion

Using resting-state fMRI and MEG recordings released as part of the HCP, we have constructed the functional network structure that expresses the coupling in MEG oscillatory power within three frequency bands, and fMRI partial correlation networks, among 39 ROIs. Based on these analyses, we make two key claims. First, genetic factors help to determine the strength and form of cortical oscillatory communication and functional connectivity. We have shown that the entire functional network structure is found to be more similar for two subjects the more closely they are related, and we estimate that the average heritability of individual connection strengths is about 15–18% for BOLD correlations, and between 13% and 26% (1% and 25%) for correlations in beta-band (alpha-band) power fluctuations. Second, genetic make-up is more important, on average, than the shared environment among twins when determining the strength of these couplings (with a difference of 13–16 percentage points in fMRI and 4–22% in the MEG beta-band).

Our results are drawn from two imaging modalities, fMRI and MEG, with functional connectivities estimated from the same set of ROIs. Our analyses with these technologies produce complementary assessments of functional connectivity: slow time scale couplings of functional activation indexed by BOLD response, and faster co-ordinations in oscillatory amplitude measured with MEG. (MEG is a more direct measure of neuronal oscillatory activity, unaffected by vascular confounds.) We found, over both modalities (although, for the MEG results, only in the beta band), the same differential pattern of influence on functional connectivity from additive genetic factors and twins’ shared developmental environments. Given the range of connectivity structures that are expressed in beta-band oscillations and in the slow time scale couplings measured in fMRI, these results suggest that genes have broad control over functional connectivity in the cortex, with a contribution that outweighs shared environmental factors.

These complementary results in fMRI and MEG alpha- and beta-band oscillations, drawn from the same parcellation, provide strong new support for a neural basis of the genetic influences on BOLD connectivities. The heritability of functional connectivity measured with electrophysiology has been reported before (Posthuma et al., 2005; Schutte et al., 2013; recording heritabilities of 20–75%, predominantly in the alpha and beta bands). However, these analyses were performed between EEG sensors, using synchronisation likelihood for network estimation, a measure that is sensitive to volume conduction artefacts. We use a connectivity metric that explicitly suppresses these artefacts (Colclough et al., 2015; Brookes et al., 2012; Hipp et al., 2012; O'Neill et al., 2015), reducing the influence of heritable anatomical features and cortical folding patterns on our results. Additionally, by working with cortical reconstructions of the oscillatory sources, we gain not only substantial artefact rejection (Schoffelen and Gross, 2009; Hillebrand et al., 2005) but also interpretability; the functional networks that we identify in the beta band and which drive our genetic analyses are mostly in motor and posterior regions, in correspondence with previous findings (Hipp et al., 2012; Baker et al., 2014; Hillebrand et al., 2012; Brookes et al., 2011; Brookes et al., 2012; Mantini et al., 2007; Marzetti et al., 2013; de Pasquale et al., 2012; de Pasquale et al., 2016). Our results therefore advance the strength of the evidence for a neural mechanism that mediates the genetic control over functional connectivity. Current data cannot, however, reveal the mechanistic details of this control. We can speculate that genes will control both network dynamics (in terms of synaptic strengths or conductance delays) as well as biophysical properties (such as the distribution of local neuron populations and their immediate feedback systems)—but in the future it may be possible to combine neuroimaging results with large-scale biophysical models (Deco et al., 2008; Cabral et al., 2014) to better understand these influences.

We did not find significant genetic control of functional connectivity in the theta frequency band. This may be in part because of the difficulty in cleanly estimating functional connectivity in MEG, particularly outside the alpha and beta bands (Colclough et al., 2016), but may also be because the 89 HCP subjects with resting-state MEG scans present a much smaller sample for study than the rfMRI HCP dataset. Even the conclusions from our alpha- and beta-band analyses, while coincident with the results in fMRI, are necessarily summary characterisations of the role of genes and environment on cortical oscillatory coupling. Our claims are based on significant results from robust non-parametric tests, but the confidence intervals on our parameter estimates are inevitably broad.

The primary parcellation that we employed consists of focal, contiguous regions. Our heritability analyses therefore reflect the genetic and shared environmental influences on the strengths of connections between the component nodes of the established resting-state networks (such as the default mode, motor, visual and dorsal attention networks). In an additional analysis, we found significant effects of additive genetic factors on the strength of inter-network connections in fMRI, using an alternative lower-dimensional parcellation based on 15 entire networks as nodes for the connectome. This suggests that genetic control of functional connectivity extends across multiple spatial scales of synchronisation. For oscillatory activity measured with MEG, it is not clear that it is meaningful to summarise, in the same manner as for an fMRI inter-network analysis, the collective behaviour of these gross network structures using a single time course: there is growing evidence from MEG that the extended networks from the fMRI literature are composed of smaller sub-networks that synchronise and de-synchronise on relatively fast time scales (Baker et al., 2014; de Pasquale et al., 2012; de Pasquale et al., 2016). That we found no significant genetic effects on the connections among this additional set of 15 networks in our MEG data may simply reflect the unsuitability of static whole-network parcellations for this modality.

We have taken a comprehensive approach to our heritability analyses by controlling for a wide range of potential confounds. Of the heritable effects of physiology, anatomy and noise on functional connectivity, the influences of signal power, cortical folding patterns and subject motion are perhaps most worthy of discussion.

Higher coupling strengths are commonly observed between regions with higher signal power in both fMRI and MEG estimates of functional connectivity. This effect can be caused just by the increased signal to noise ratio in the observed regions, rather than any difference in underlying coupling strength (Friston, 2011). We fitted a genetic and environmental factors model to the power in each ROI of our analysis. In our MEG data, we found no influence of these factors on the measured signal strength, and so can exclude this signal-to-noise effect as a likely confound in the beta-band MEG results. There was a significantly heritable component to signal power in the fMRI data—but this is not surprising in and of itself. To control for its influence on our results, we have included measures of the signal power in each node as confound regressors in all of our genetic analyses of functional connectivity. (Signal power is not a perfect proxy for the signal to noise ratio, as the measured power will be affected by fluctuations in both the noise and the signal.)

Cortical folding patterns have been shown to have modest heritability (Sotiropoulos et al., 2015; Botteron et al., 2008; van Essen et al., 2014); although the strength of genetic control and the mechanism of influence are not yet certain (Tallinen et al., 2016; Gómez-Robles et al., 2015; Ronan and Fletcher, 2015). This has particular implications for the MEG measurements, where the arrangement of cortical folds will have a strong impact on the signals measured outside the scalp, and on the leakage (or volume conduction) of signals between different sensors. Our connectivity estimates use methods that are robust to source leakage artefacts (Colclough et al., 2015; Brookes et al., 2012; Hipp et al., 2012; O'Neill et al., 2015), and the lack of strong correlations between the spatial profile of the heritability of cortical curvature and the heritability of edges associated with each ROI (as shown in Figure 1—figure supplement 2 and Supplementary file 5) is good evidence, we believe, that this confound has not seriously impacted our findings.

Subjects’ motion inside a scanner during resting-state recordings is heritable (Couvy-Duchesne et al., 2014), and motion is a known resting-state confound for functional connectivity analyses (Siegel et al., 2016). While our pre-processing steps include appropriate registration of images into static reference frames, and the removal of components from the data that are associated with motion, we have additionally included in our genetic analyses a summary measure of participants’ movements within their fMRI scans, to reduce our co-measurement of heritable motion traits. Lastly, we are unable to specifically control for eye movement in our genetic analyses, as no tracked eye recordings are available for the HCP resting-state scans. (Independent components of the sensor data that correspond to eye movements are, however, removed during pre-processing.) However, it is worth noting that although saccades and micro-saccades are thought to be linked to gamma-band oscillatory patterns (Muthukumaraswamy, 2013; Orekhova et al., 2015), which we did not study, there is no strong evidence-base that they influence resting-state recordings at lower frequencies.

We used variance-component models to represent the observed variability in functional connectivity as a set of contributions from additive genetic factors and from the shared developmental environment in twin pairs. Our conclusions rest on a number of assumptions implicit in this model (Boomsma et al., 2002). These include an assumption that people choose their partners randomly; that the relevant genetic mechanisms are additive; and that there is no significant interaction between genes and the shared environment—an effect which has been reported to exist in various aspects of cognition (Nisbett et al., 2012). Most importantly, we assume that monozygotic and dizygotic twin pairs will equally share exposure to environmental factors in their upbringing. If this is an accurate assumption, then estimates of c2 and h2 are informative about the relative roles of developmental environment and additive genetics in the variation of functional connectivity phenotypes. While there is evidence in support of the equal-environment assumption (Felson, 2014; Conley et al., 2013), it has been challenged on the basis that monozygotic twins look more alike, behave more alike, and are treated more similarly than dizygotic twins (Joseph, 1998). Without entering into the discussion on this point, we note that if this assumption fails in our sample, it would lead to an overestimation of genetic heritability and an underestimation of the impact of developmental environment, therefore potentially weakening our conclusions on the differential effect of these two influences. Lastly, the c2 term in our model cannot be solely identified with developmental environment. It encompasses all effects which are common between twins, which may include intra-uterine and mitochondrial influences, in addition to the shared environmental factors.

Our results add substantially to the evidence for significant genetic control of human cortical connectivity. Structural connectivity, the distribution of the white matter tracts across the brain that enable communication, is known to be highly heritable (Zhu et al., 2015; Jahanshad et al., 2012; Kochunov et al., 2015). (A study of the heritability of structural connections with the HCP dataset estimated a mean heritability of 25%; Sotiropoulos et al., 2015) Previous studies of the heritability of functional connectivity with fMRI (which, in general, fail to control for subject motion, node power or brain size, and some of which focus instead on network topology), do also find significant genetic effects with heritability estimates lying between 20% and 60% (Smit et al., 2008; Fornito et al., 2011; van den Heuvel et al., 2013; Thompson et al., 2013; Jansen et al., 2015; Glahn et al., 2010). However, connectivity estimation is noisy, and as heritability is expressed as a fraction of the observed phenotypic variance (including measurement noise), we can expect the parameter estimates of heritability to change as network estimation methods improve. The important comparison, therefore, is of the relative importance of genes to the developmental and environmental factors shared by twins. Our observation of the stronger influence of the former over the latter on functional connectivity is a relationship shared in several other cognitive, behavioural and physiological traits measured in adults (Polderman et al., 2015).

Twin studies cannot probe the precise genetic mechanisms of these influences. Recent work (Hawrylycz et al., 2015; IMAGEN consortium et al., 2015) has correlated the profiles of gene expression in different regions of cortex, and compared these patterns of gene transcription to resting-state fMRI connectivity profiles, identifying sets of genes with co-expression patterns that reflect networks of functional connectivity. However, detailed investigations into the specific genes affecting healthy resting-state connectivity must await large datasets with coincident resting-state functional imaging and genotyping. (Genetic data will soon be available both for the HCP and for the UK Biobank imaging data.)

Taken together, our results provide strong additional evidence for a neural basis of the heritability of functional connectivity in the human brain. We identify genetic influences both in fMRI datasets and in (alpha- and beta-band) electrophysiological recordings, using functional network analyses on the same cortical parcellation. Our results more comprehensively control for a wide range of important confounds than previous work in this area. We have made particular emphasis of the increased importance of genetics over the developmental environment in determining cortical functional connectivity. The relevance and implications of this finding are widespread. Functional connectivity profiles are associated with intelligence, which is well-known to be heritable (Bartels et al., 2002; Neisser et al., 1996). But connectivity is also implicated with a very broad range of behavioural and life-style factors, including earning power, measures of health, various assessments of cognitive performance, and self-reported life satisfaction (Smith et al., 2015). As the findings of the Moving To Opportunity experiment in the United States make clear (Ludwig et al., 2013; Chetty et al., 2015), together with results from studies on the effects of schooling, socio-economic environment and adoption (Nisbett et al., 2012), the environment in which individuals develop can still make a significant impact on their performance and long-term outcomes. However, while genes are not the only important factor, results such as these presented here suggest that nature, more so than nurture, is a major force in determining integrated cognitive function and, by extension, cognitive capability.

Materials and methods

Subjects

Four 15-minute resting-state fMRI recordings from 820 subjects were collected as part of the HCP900 data release by the Human Connectome Project (WU-Minn HCP Consortium et al., 2013; WU-Minn HCP Consortium et al., 2013). Additionally, 89 of these subjects provided three 6-minute resting-state MEG recordings. All subjects are young adults (22–35 years of age) and healthy. Of the fMRI (MEG) subjects, there are 103 (19) monozygotic and 54 (13) dizygotic complete twin pairs. Zygosity was determined from subjects’ genotypes when available, and otherwise by self-report.

HCP data were acquired using protocols approved by the Washington University institutional review board. Informed consent was obtained from subjects. Anonymised data are publicly available from ConnectomeDB (db.humanconnectome.org; Hodge et al., 2016). Certain parts of the dataset used in this study, such as the family structures of the subjects, are available subject to restricted data usage terms, requiring researchers to ensure that the anonymity of subjects is protected (WU-Minn HCP Consortium et al., 2013).

fMRI analysis

Resting-state fMRI data were acquired with 2 mm isotropic spatial resolution and a temporal resolution of 0.72 s. The HCP provides comprehensively pre-processed data (WU-Minn HCP Consortium et al., 2013) that are mapped to a standard cortical surface using a multi-modal registration algorithm, MSMAll (Robinson et al., 2014; Glasser et al., 2016), and for which structured artefacts (with origins including motion, heartbeat and cerebro-spinal fluid) have been removed by a combination of ICA and FIX (Salimi-Khorshidi et al., 2014), FSL’s automated noise component classifier.

We estimated functional connectivity between 39 fMRI-derived cortical ROIs. We used the parcellation employed in Colclough et al. (Colclough et al., 2015), which contains contiguous regions identified from a resting-state 100-dimensional group-ICA decomposition of fMRI data from the first 200 subjects of the HCP project. Most ROIs have a symmetric counterpart in the opposite hemisphere, except those on the midline such as the posterior cingulate cortex. The parcellation is shown in Figure 2A. This parcellation was chosen because the use of contiguous, focal nodes creates a comprehensible network model between individual regions. The low dimensionality of MEG data restricts us to parcellations of about this number of ROIs or less: as there are only a few hundred sensors, reconstructing time courses for more than approximately 60 different regions would be noisy or nonsensical. This upper bound is supported by models that categorise signals as originating within or without the MEG dewar, which suggest a limit of about 60 measurable cortical sources (Taulu et al., 2005), and by recently developed data-driven parcellations that only find around 70 unique identifiable parcels using combined MEG and EEG data (Farahibozorg et al., 2017). To investigate whether our results were particular to our choice of parcellation, we additionally ran our heritability analyses on a lower-dimensional decomposition: the 15-dimensional group-ICA map computed from the HCP900 data. The nodes in this parcellation represent entire (non-contiguous) functional networks (such as the default mode or motor networks), and the connectivity matrix therefore represents inter-network dependencies. The parcellation is shown in Figure 2B.

Regions of interest used for functional connectivity estimation.

(A) Primary parcellation of 39 contiguous clusters, identified from a resting-state 100-dimensional group-ICA decomposition of fMRI data from the first 200 subjects of the HCP project. (B) The 15-dimensional fMRI ICA parcellation computed by the HCP as part of the S900 data release.

https://doi.org/10.7554/eLife.20178.009

A single BOLD time course to represent each ROI was constructed using multiple spatial regression in FSLnets. Partial correlation matrices were constructed using the tools in FSLnets, using mild Tikhonov regularisation (λ=0.01) and a conversion to Z-values with Fisher’s transform, referenced to the standard deviation of correlations of null data. Group-level networks were estimated as the mean Z-transformed correlation matrices over all sessions.

MEG analysis

Resting-state MEG data were acquired on a whole-head Magnes 3600 scanner (4D Neuroimaging, San Diego, CA, USA). The data have been pre-processed to compensate for head movement, to remove artefactual segments of time from the recordings (which might relate to head or eye movement), identify recording channels which are faulty, and to regress out artefacts with clear temporal signatures (such as eye-blinks, saccades, muscle artefacts or cardiac interference) using ICA (WU-Minn HCP Consortium et al., 2013). Sensor-space data were down-sampled from 509 Hz to 300 Hz, with the application of an anti-aliasing filter.

MEG data from each session were source-reconstructed using a scalar beamformer (Van Veen et al., 1997; Robinson and Vrba, 1999; Woolrich et al., 2011). Pre-computed single-shell source models are provided by the HCP at multiple resolutions, registered into the standard co-ordinate space of the Montreal Neuroimaging Institute (MNI). Data were filtered into the 1–30 Hz band before being beamformed onto a 6 mm grid using normalised lead fields. Covariance estimation was regularised using PCA rank reduction. The rank was conservatively reduced by five more than the number of ICA components removed during preprocessing. Source estimates were normalised by the power of the projected sensor noise. Source-space data were filtered into theta (4–8 Hz), alpha (8–13 Hz) and beta (13–30 Hz) bands.

The same parcellations were employed for the MEG analysis as for the fMRI. As the MEG source reconstruction was performed over a volumetric grid in the MNI’s standard space, rather than on the cortical surface, we used the volumetric versions of the fMRI ICA decompositions to form the MEG parcellations. A single time-course was constructed to represent each node as the first principal component of the ROI, after weighting the PCA over voxels by the strength of the ICA spatial map. This analysis yielded 39 time-courses for each frequency band and session for our principal parcellation (and 15 for the second).

One major confound when estimating connectivity in source-localised MEG is the spatially local bleeding of estimated sources from their true location into neighbouring regions. We compensate for these spatial leakage confounds, which can induce spurious connectivity estimates, using a symmetric orthogonalisation procedure (Colclough et al., 2015) to remove all shared signal at zero lag between the network nodes. This procedure is a multivariate extension of the orthogonalisation principle proposed in Brookes et al., 2012; Hipp et al., 2012; and O'Neill et al., 2015. It identifies the set of ROI time courses least displaced from the initial, uncorrected set, while enforcing mutual orthogonality between them, with no bias related to any reordering of the nodes. This approach suppresses any artificial correlations induced by spatial leakage, at the expense of any zero-lag connections of true neuronal origin. Lastly, power envelopes of the leakage-corrected ROI time-courses were computed by taking the absolute value of the Hilbert transform of the signals, low-pass filtering with a cut-off of 1 Hz, and down-sampling to 2 Hz (Luckhoo et al., 2012). We illustrate this stage in Figure 3.

Illustration of pipeline for MEG functional connectivity estimation.

From a whole-brain source-reconstruction, a single time-course is extracted to represent each ROI. These time-courses are bandpass filtered, then orthogonalised to remove shared signal which is potentially attributable to spatial leakage effects. The power envelope of each time-course is computed, then correlated to form the network matrix.

https://doi.org/10.7554/eLife.20178.010

Correlations between the power time-courses in each band were computed, converted to Z-values using Fisher’s transform, and standardised with reference to the standard deviation of an empirical null distribution of correlations generated from time-courses with the same temporal properties as the data under test (see Colclough et al., 2015, for a detailed description). As for the fMRI networks, group-level networks were estimated as the mean Z-transformed correlation matrices over all sessions.

We use slightly different network estimation methods in the two modalities. Both are correlation measures, to maintain the similarity in analysis between the fMRI and MEG data. Partial correlation, used in the fMRI data, is one of the most accurate and robust measures in this modality (Smith et al., 2011; Smith et al., 2013; Marrelec et al., 2006; Varoquaux and Craddock, 2013). However, partial correlation can be difficult to estimate, even with regularisation, and it would not necessarily be the optimal choice for many smaller or lower quality fMRI datasets. An assessment of the repeatability of the most common network estimation measures in MEG showed that regularised partial correlation is relatively unreliable in this modality, and that the full correlation of the power envelopes of oscillatory activity (as we use here) was the most repeatable approach (Colclough et al., 2016). We have therefore attempted to use the best-practice inference methods in each modality. We did repeat our fMRI analysis using full correlation, rather than partial correlations, and found qualitatively similar results, with no difference in the significance of the claims we would make.

Comparison of entire network structure

Similarity between pairs of functional networks was measured as the inverse Euclidean distance between the correlation matrices, after a logarithmic projection onto a Euclidean plane locally tangent to the Riemann manifold of positive semi-definite matrices (Barachant et al., 2013Ng et al., 2014). This process can decouple the inter-relations between the elements within each positive-definite network matrix, and can improve the performance of classification algorithms when network matrices are used as discriminative features. It also provides for the definition of a true distance metric, where more similar matrices are projected to more proximate locations on the plane. A suitable Euclidean space for each modality and frequency band was found as the Euclidean plane tangent to the cone of positive semi-definite matrices, taking the geometric mean of the networks computed from all sessions as the tangent point. The separation between each pair of network matrices in each band was computed, forming different distributions of network similarity for pairs of subjects, split by the shared genetics of each pair. To assess whether shared environmental factors and shared genetics were associated with global network structure, we tested the difference in mean between the distributions of the logarithm of network separation for pairs of unrelated subjects and pairs of dizygotic twins, and between pairs of dizygotic twins and monozygotic twins, using a non-parametric t-test based on 20,000 permutations of the group labels. In the fMRI data, we were also able to compare unrelated subjects to siblings, and siblings to dizygotic twins.

Three-component variance models

An ACE model for the subject-to-subject variability in connection strength was fitted for each network edge. This model splits the observed phenotypic variability into three factors: (A), additive genetics or heritability (h2), (C), common environment (c2), and (E), measurement error or external sources of variability. The three factors h2, c2 and e2 are proportions of the total variance, and therefore normalised such that they sum to one. The APACE system of permutation inference (Chen, 2014) was employed, using the mean network matrix (over resting-state recording sessions) for each subject.

On each edge, we regressed out the effect of age, the square of age, sex, an age and sex interaction, an interaction between sex and the square of age, the cube root of intra-cranial volume and of cortical volume (both estimated with FreeSurfer), a measure of subject motion in the scanner (fMRI_motion) and, for the fMRI data, the MR image reconstruction software version. Summary motion estimates are only available for the fMRI recordings, but we use these values as a proxy measure of subjects’ movements in both modalities. On each network edge, we included a regressor to account for heritable changes in node power, computed as the geometric mean of the power in each of the two nodes that each connection joins. We computed power simply as the standard deviation of the signal (fMRI time course or MEG power envelope) in each node. In the MEG data, where the standard deviation of the power of a virtual sensor is often correlated with its mean, we additionally include, in the same manner, a regressor formed from the ratio of the standard deviation to the mean of the power time course.

Finally, we also regressed out a measure of the noise passed by the beamformer for each subject. Our beamformer is described in Woolrich et al. (2011), and the noise it passes at each voxel is the denominator of equation 4 in that paper,

(HT(ri)I/σe2H(ri)),

where H is the projection of the N×3 lead field matrix (for N sensors) that maximises power at location ri, and σe2I is the noise covariance matrix. The scale of the noise is estimated from the smallest eigenvalue of the data covariance matrix. These voxel-wise noise estimates are scaled by the same weightings used to compute the ROI time courses, and averaged over ROIs to create a regressor estimating the noisiness of the beamformer for each subject.

Estimates of heritability were computed on each edge, with family-wise error corrected p-values computed by permutation, randomly shuffling monozygotic and dizygotic twin statuses 15,000 times. It is useful to consider a single, summary measure of functional connectivity for the entire functional networks. The mean values of h2 and c2 were computed over all functional connections; the p-value for mean heritability was computed by permutation, as above; confidence intervals for h2, c2 and h2-c2 were computed by bootstrap re-sampling twin-pairs with replacement, 15,000 times.

The ACE model was also fitted to the logarithm of the variance of the corrected power envelopes (variance of the BOLD time course) in each ROI. This provided confidence that the conclusions of the ACE model for functional connectivity were not being driven by heritable effects in power or SNR within the network nodes. On each ROI, we removed the same set of regressors as for the functional connectivity analyses (save the power variables) before fitting the model.

Lastly, the ACE model was fitted within each ROI to estimates of cortical curvature for each of the subjects in the fMRI sample. We regressed out the effect of age, the square of age, sex, an age and sex interaction, the cube root of intra-cranial volume and of cortical volume, subject motion and the MR image reconstruction software version, before computing the mean heritability (h2) over all points within each ROI. To compare spatial profiles of cortical folding with connectivity, we averaged the heritabilities of each network connection onto their constituent nodes, to create a spatial map suggesting the heritability of connectivity by ROI. These maps were correlated with the maps of heritability of cortical curvature in each ROI, and significance assessed by permuting ROIs 5000 times.

Methodological notes

We applied a false discovery rate correction to the 21 principal statistical tests conducted for this paper to compensate for the multiple comparisons we perform (Benjamini and Hochberg, 1995). Uncorrected and corrected p-values are available in Supplementary files 3 and 5.

All analyses were performed in Matlab. MEG network analyses were performed with the MEG-nets software (github.com/OHBA-analysis/MEG-ROI-nets), fMRI network analyses with the FSL-nets software (fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSLNets), and heritability analyses using APACE (the Advanced Permutation inference for ACE models (APACE) software is available at warwick.ac.uk/tenichols/apace).

References

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
    The 14 Th Annual Meeting of the Organizationfor Human Brain Mapping
    1. KN Botteron
    2. DL Dierker
    3. R Todd
    4. J Alexopolous
    5. D Seung
    6. K Han
    7. T Nishino
    8. E Reid
    9. A Todorov
    10. DC Van Essen
    (2008)
    Human vs computer algorithm choices in identifying identical twin pairs based on cortical shape characteristics–who’s better?, The 14 Th Annual Meeting of the Organizationfor Human Brain Mapping.
  9. 9
  10. 10
  11. 11
  12. 12
    Accelerated estimation and inference for heritability of fMRI data
    1. X Chen
    (2014)
    University of Warwick.
  13. 13
  14. 14
  15. 15
  16. 16
  17. 17
  18. 18
  19. 19
  20. 20
  21. 21
  22. 22
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
  28. 28
  29. 29
  30. 30
  31. 31
  32. 32
  33. 33
  34. 34
  35. 35
  36. 36
  37. 37
  38. 38
    The equal environment assumption of the classical twin method: a critical analysis
    1. J Joseph
    (1998)
    Journal of Mind and Behaviour 19:325–258.
  39. 39
  40. 40
  41. 41
  42. 42
  43. 43
  44. 44
  45. 45
  46. 46
  47. 47
    Transport on Riemannian manifold for functional connectivity-based classification
    1. B Ng
    2. M Dressler
    3. G Varoquaux
    4. JB Poline
    5. M Greicius
    6. B Thirion
    (2014)
    Medical Image Computing and Computer-Assisted Intervention : MICCAI ... International Conference on Medical Image Computing and Computer-Assisted Intervention 17:405–417.
    https://doi.org/10.1007/978-3-319-10470-6_51
  48. 48
  49. 49
  50. 50
  51. 51
  52. 52
  53. 53
  54. 54
  55. 55
    Recent Advances in Biomagnetism
    1. SE Robinson
    2. J Vrba
    (1999)
    302–305, Functional neuroimaging by synthetic aperture magnetometry (SAM), Recent Advances in Biomagnetism, Sendai, Japan.
  56. 56
  57. 57
  58. 58
  59. 59
  60. 60
  61. 61
  62. 62
    Data quality influences observed links between functional connectivity and behaviour
    1. JS Siegel
    2. A Mitra
    3. TO Laumann
    4. BA Seitzman
    5. M Raichle
    6. M Corbetta
    7. AZ Snyder
    (2016)
    Cerebral Cortex, 10.1093/cercor/bhw253, 27550863.
  63. 63
  64. 64
  65. 65
  66. 66
  67. 67
  68. 68
    Heritability of structural connections from HCP diffusion MRI data
    1. SN Sotiropoulos
    2. X Chen
    3. SM Smith
    4. DC Van Essen
    5. TE Behrens
    6. TE Nichols
    7. S Jbabdi
    (2015)
    Proceedings of the 23 Rd Annual Conference of the International Society for Magnetic Resonance inMedicine. 721.
  69. 69
  70. 70
  71. 71
  72. 72
  73. 73
  74. 74
  75. 75
  76. 76
    Heritability of brain structure, function, and connectivity in human connectome project data
    1. DC van Essen
    2. MF Glasser
    3. E Robinson
    4. X Chen
    5. M Jenkinson
    6. DL Dierker
    7. TE Nichols
    8. SM Smith
    (2014)
    The 20 Th Annual Meeting of the Organization for Human Brain Mapping.
  77. 77
  78. 78
  79. 79
  80. 80
  81. 81
  82. 82
  83. 83
  84. 84
  85. 85
    Genetic analysis of structural brain connectivity using DICCOL models of diffusion MRI in 522 twins
    1. D Zhu
    2. L Zhan
    3. J Faskowitz
    4. M Daianu
    5. N Jahanshad
    6. G de Zubicaray
    7. KL McMahon
    8. NG Martin
    9. MJ Wright
    10. PM Thompson
    (2015)
    Proceedings of the IEEE International Symposium on Biomedical Imaging. pp. 1167–1171.
  86. 86

Decision letter

  1. Jack L Gallant
    Reviewing Editor; University of California, Berkeley, United States

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your article "The heritability of multi-modal connectivity in human brain activity" for consideration by eLife. Your article has been reviewed by three peer reviewers, one of whom, Jack Gallant (Reviewer #1), is a member of our Board or Reviewing Editors and the evaluation has been overseen by Sabine Kastner as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: David Glahn (Reviewer #2).

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

Although the Discussion highlighted some hesitation about the publication of this work in eLife, the reviewers have decided to ask for a revision, which should deal with (1) additional analyses, (2) novelty issues, (3) other smaller issues. Of particular concern is that we are not yet convinced that this work is sufficiently novel; the authors will have to make the case for that.

Summary:

This study examines genetic influences on resting state MRI and MEG connectivity. The article is clearly written and the data analysis procedures are reasonable as far as they go. However, the genetic influences on resting state have been reported several times using similar methods, and at this point it is unclear whether this paper is appropriate for publication in eLife, or rather whether it merely reflects an incremental increase in scientific knowledge. Furthermore, the current analyses appear to be insufficient. The MEG results are rather underpowered and many essential analytical controls are missing. After consultation the reviewers decided that the paper should be returned to the authors for revisions focusing on (1) elucidating the novel contributions of the work, (2) performing additional analysis as suggested and (3) dealing with a variety of smaller issues.

Essential revisions:

1) The genetic influences on resting state have been reported several times using similar methods. However, the current paper does not really include any detailed comparison of the current results with those reported previously, so it is difficult to judge the novelty of this contribution. In revision the authors should include a detailed discussion of precisely which aspects of the paper are novel. This may require further data analysis apart from what is requested as controls below. It would also be helpful if the authors would make more of an effort to explain some potential causal mechanisms that might underlie their reported relationships.

2) Additional analyses should be included to account for potential contamination of the FC data by confounding factors having nothing to do with cortical connectivity or communication, but which are nevertheless heritable. The obvious candidates here are body motion, head motion, eye movements, or other physiological factors.

3) The results are all based on one particular parcellation. Some evidence should be provided that the results do not depend on this particular choice. The optimal way to address this problem would be to rerun the analysis pipeline using several different parcellation schemes.

4) There appear to be several differences between the way that the fMRI and the MEG data were processed. These should be justified and explained, or a more consistent approach should be used.

Reviewer #1:

This is a fine study as far as it goes and it includes several good controls for potential contaminating factors (though these could be substantially improved). However, the paper is going to require revision and additional data analysis before it is suitable for publication.

Although the authors show THAT genetics influences FC, they provide no information about WHY genetics influences FC. Which of the many mechanisms that contribute to observed BOLD FC are influenced by genetics? As the authors note, several of the results reported here (such as the heritability of FC from EEG data) have been reported already in previous studies. It is claimed here that these previous studies are less interpretable than the current study. That may be true, but then the authors need to provide the interpretation. If not then this paper loses a lot of its potential novelty and impact and I am not sure that it will be suitable for publication in this venue. Some effort should be made to explain several of the more interesting and unusual effects as well. For example, it is stated that there is significant heritability in the alpha and beta bands of the MEG data but not in the theta band. Why? What is a plausible mechanism that would generate this pattern of results?

There is one major class of potential confounds which appear to be given short shrift here: potential contamination of the FC data by confounding factors having nothing to do with cortical connectivity or communication, but which are nevertheless heritable. The problem is that genetic factors might influence processes that are well known to influence correlations in MRI data (and likely MEG data as well). For example, if genetics causes some people to wiggle more in the magnet, that is obviously going to influence FC. (In fact I am fairly certain that I saw another study recently that reported just that effect, though unfortunately I am unable to find it now.)

To take just one specific example of the above concern, it is reported that the visual system has a high degree of FC in the MEG data. The most likely candidate here would be eye movements, which will clearly affect MEG correlations, and which may also be influenced by genetics. However, I didn't see any report here that the eyes were tracked properly, or that the eye tracking data were regressed out when calculating FC. It is precisely these sorts of potential confounds (i.e., genetic influences operating on behavior that in turn influences FC rather than operating on FC directly) that must be scrupulously accounted for before publication.

Perhaps one good way to get an intuitive handle on these sorts of potential indirect genetic influences would be to run separate analyses to determine the heritability of the various confounds that are known to affect FC, such as motion, eye movements, beamforming errors (in the MEG data). For that matter other factors that might affect FC indirectly might also be heritable. BOLD data taken directly off the scanner are non-Gaussian, and these are Gausianized during pre-processing by a necessarily imprecise procedure. Is the distribution of raw BOLD signals heritable? All these seem plausible, and all could potentially affect FC. A systematic analysis of these factors would seem to be critical for making any strong claims about direct genetic influences on FC.

Smaller issues:

The claims in the Abstract and Discussion are a bit over-stated in several places given the results because they do not make any reference to places where no heritability is found. For example, while effects in the MEG alpha and beta frequency bands are observed, this is not true for the theta frequency band.

The figures and the supplementary movies in this paper are rather poor, especially the connectivity figures. There are many excellent choices for visualization software these days.

Issues related to data analysis:

The pre-processing and data analysis procedures are necessarily complicated, and of course it is always possible that some decision that was made during those procedures might have biased the results. On the other hand, if we started requiring every single step to be addressed in multiple ways to ensure against this sort of bias then none of us would get anywhere! So in the next few paragraphs I only ask for further work on data analysis steps that I think could be particularly problematic, or where additional analysis might provide worthwhile enlightenment.

The entire data analysis procedure is based on the HCP ICA parcellation. It would inspire more confidence in the conclusions if some analysis was provided that showed whether the choice of parcellation scheme makes any difference on the results or the conclusions. I don't think that it would make much difference in the results if one compared the HCP ICA parcellation to, say, the Glasser HCP parcellation, because these are both fine-scale parcellations that are likely to include more spatial information than can really be supported by genetics data anyway. (In the case of the MEG data, the HCP ICA parcellation may very well have higher resolution than can be supported by either the genetics or the MEG.) It would be immensely helpful to know whether a coarser parcellation schemewould change the results, and I think that knowing the answer to this question might also help us address the question of why these apparent genetic influences are found.

It looks as if subject motion was removed during pre-processing in both the fMRI and the MEG data, but subject motion was still used as a regressor during ACE modeling in the fMRI data. Subject motion is a problem in both fMRI and MEG, so if it was a problem in the fMRI data even after pre-processing why was pre-processing sufficient for the MEG data? At a minimum it seems that it should also be included in the MEG modeling analysis.

Very small confusing issues:

This is a side point, but I find it surprising that it is useful/helpful to regress out the MR image reconstruction software version in ACE modeling. This suggests that these data really do live at the very limit of sensitivity, and that confounds and bias really can creep into the data even after pre-processing. It seems like the best practice would be to use the same version of the software to process all the data and I find it a bit worrisome that this was not done.

In Subsection “E. Three-component variance models” it is stated that the variance partitioning analysis addresses non-negative portions of the variance. Why do these procedures produce negative variance estimates at all? Is this just statistical error?

Reviewer #2:

Colclough and colleagues examine the genetic influences on resting state MRI and MEG connectivity in "The heritability of multi-modal connectivity in human brain activity." Subjects included 461 individuals with MRI data and a group of 61 primarily overlapping individuals with MEG data all from the Human Connectome Project. Subjects were from extended twin pairs or unrelated individuals. The goals of the article are to estimate the heritability and common genetic influences on resting state and MEG measures of connectivity. The article is clearly written and the analytic plan is both complex and reasonable. Unfortunately, the findings represent only an incremental increase in scientific knowledge, as the genetic influences on resting state have been reported several times using conceptually similar methods. While the MEG results are novel, that analysis is rather underpowered, as the authors appropriately note in the Discussion. Thus, while I believe the findings should be reported, the authors should either describe the work as replication or conduct additional analyses that will increase novelty.

Reviewer #3:

This is an excellent article, which combines state-or-the-art resting-state MEG analysis methods with new approaches to assessing heritability and environmental facts in this type of electrophysiological dataset. As such it both helps to drive the field forward and will be of general interest to the eLife readership. I recommend publication after the following issues are addressed:

1) I think the authors should clarify why 39 cortical regions are chosen, as opposed to other common parcellation schemes, which often have 50-100 regions.

2) I'm surprised that the Authors did not at least attempt to reconstruct networks in the higher gamma band. Why only theta/alpha/beta?

3) One of my main concerns is the use of partial correlations in the FMRI analysis and simple correlations in the MEG network analysis. Why this difference? The Authors cite two of their own papers as justification but this not particularly convincing. Could we have a short justification here?

4) The authors state that the beta-band "exhibits broad connectivity over the whole cortex". That doesn't really appear to be the case looking at Figure 1. In ay case, what does that statement actually mean – it's quite a woolly phrase that is not backed up by any kind of quantification.

5) The authors look at both genetic and environmental factors in terms of contribution to signal power (in both FMRI and MEG), partly as a control for these as confounds in the connectivity analyses, but also as interesting exploratory analyses in their own right. I have two concerns here. 1) Can the authors please describe how "power" was assessed in the FMRI signal? I assume this is some measure of temporal variation around the mean of the voxel time series? 2) For the MEG, the power passed by a beamformer in an RSN analysis can be quite dependent on geometric effects (even after weights normalisation). Wouldn't a better measure of 'activity" in the MEG amplitude envelopes be some measure of temporal variability? The standard deviation (SD) is often used, but the SD of the virtual-sensor and the Mean are often correlated in beamformer reconstructions, so a better measure might be some proportional change (e.g. SD/Mean). "Activity" in both the FMRI and MEG time series could thus be assessed using the same (or very similar) metric.

6) In the ACE model, quite a few nuisance covariates are regressed out of the model before heritability/environmental effects were assessed. How were these parameters (and their second-order versions) chosen? In addition, when regressing out some many parameters, is there an issue with statistical power in the MEG analyses as some of the groups only have relatively few subjects (i.e. 11 monozygotic twin pairs)?

7) Finally, with these "nuisance" regressors Is there a potential problem with interaction with heritability? For example age is 100% matched in the twin-pairs, but is presumably not matched for the non-related pairings.

https://doi.org/10.7554/eLife.20178.020

Author response

Essential revisions:

1) The genetic influences on resting state have been reported several times using similar methods. However, the current paper does not really include any detailed comparison of the current results with those reported previously, so it is difficult to judge the novelty of this contribution. In revision the authors should include a detailed discussion of precisely which aspects of the paper are novel. This may require further data analysis apart from what is requested as controls below. It would also be helpful if the authors would make more of an effort to explain some potential causal mechanisms that might underlie their reported relationships.

In summary, the novel contributions of this paper are:

a) Evidence not just of the heritability of functional connectivity, but that the genetic influences outweigh those of the shared developmental environment between twins.

b) The first reporting of the heritability of functional connectivity measured with

source-reconstructed electrophysiology.

c) Evidence that the genetic control of functional connectivity has an

electrophysiological basis, through the use of MEG functional connectivity analyses on the same cortical parcellation as the fMRI results. This analysis increases confidence that the genetic influence on functional connectivity is neuronal in origin, compared to previous EEG studies, by controlling for the heritable confounds associated with those approaches.

d) An analysis of the genetic and shared environmental influences of functional

connectivity that for the first time offers comprehensive control for a wide range of potential confounds.

In our revised submission, we have extended the comparison of our work to previous papers, and added discussion of potential causal mechanisms. Before highlighting these changes, we expand on these novel contributions above.

a) Estimates of heritability are highly contingent on the measuring apparatus. The noisier the measure of the phenotype, the lower the estimate of heritability, h2, because the three components of the analysis (genes, shared environment and noise) must sum to one. The important question, then, is what the sensible scale of comparison should be for heritability. Arguably, the most appropriate choice is c2, which captures (among some other effects) the environmental and developmental influences that are shared between twins. Investigating the relative importance of these two factors allows us to make important claims about the genetic control of functional connectivity, despite the noisiness of the measure. This is a perspective which until now has been absent from the literature, but which makes the results comparable to analyses of heritability in other fields relevant to investigations of functional connectivity, and of interest to a wide audience of the general public.

c) The evidence for genetic control of functional connectivity that we present from source-space MEG analyses is a critical advance on the evidence from sensor space EEG recordings. Sensor space connectivity estimation is a long way from the best practice that can be achieved in MEG (and potentially in high density EEG) because, in addition to the greater susceptibility of sensorspace data to artefacts, much of the connectivity may be driven purely by volume conduction effects, which are likely to be highly heritable, reflecting heritable anatomical and cortical folding traits. By comparison, functional connectivity in source space is much more interpretable as being neuronal in origin – both from the reduction in biological noise and the suppression of spatial leakage effects. Furthermore, the specific network edges that carry functional connectivity and drive the genetic effect can be inspected to see if they are in plausible brain areas. In our case, the alpha-band networks are in occipital areas, and the beta-band in motor and posterior regions, in line with previous findings (Hipp et al., 2012; Brookes et al., 2011, 2012; Baker et al., 2014; Mantini et al., 2007).

d) The reviewers had a strong focus on the need to account for a wide range of potentially heritable confound factors that are being captured within the measures of functional connectivity, in addition to the concerns over volume conduction or source leakage discussed above. We agree with the reviewers, and after extending our approach in line with their suggestions, we are confident that our heritability analysis controls for these confounds more comprehensively than any previous study, in any imaging modality. This represents a major strengthening of our understanding of the genetic control of functional connectivity.

We now highlight the specific changes we have made to the paper to address the requests of the reviewers in their concerns over novelty. This includes a detailed comparison of the current results with those reported previously; a discussion of which aspects of the paper are novel; and discussion of potential causal mechanisms.

First, we have expanded our Discussion with a new paragraph to be more explicit about the contribution of our electrophysiology results, and to provide a more detailed exploration of possible mechanisms for genetic control, together with a more explicit comparison to the existing literature, “These complementary results in fMRI and MEG alpha- and beta-band oscillations, drawn from the same parcellation, provide strong new support for a neural basis of the genetic influences on BOLD connectivities. […] We can speculate that genes will control both network dynamics (in terms of synaptic strengths or conductance delays) as well as biophysical properties (such as the distribution of local neuron populations and their immediate feedback systems)—but in the future it may be possible to combine neuroimaging results with large-scale biophysical models (Cabral et al., 2014; Deco et al., 2008) to understand these influences.”

We have altered later parts of the Discussion to contextualize our results, and to be more explicit about our contributions in comparison to the fMRI and diffusion MRI literature, “Our results add substantially to the evidence for significant genetic control on human cortical connectivity. Structural connectivity, […] Our observation of the stronger influence of the former over the latter on functional connectivity is a relationship shared in several other cognitive, behavioural and physiological traits measured in adults (Polderman et al., 2015).”

We conclude with an overview of the key contributions of the paper, “Taken together, our results provide strong additional evidence for a neural basis of the heritability of functional connectivity in the human brain. […] We have made particular emphasis of the increased importance of genetics over developmental environment in determining cortical functional connectivity. The relevance and implications of this finding are widespread”

2) Additional analyses should be included to account for potential contamination of the FC data by confounding factors having nothing to do with cortical connectivity or communication, but which are nevertheless heritable. The obvious candidates here are body motion, head motion, eye movements, or other physiological factors.

We agree with the reviewers that appropriate control of possible confounds to the heritability analysis is important, if we are to interpret the results as reflecting the genetic and environmental influences on functional connectivity alone.

Most of the literature we surveyed on the genetic influences of connectivity control for demographics in their genetic analyses (age and sex, for example). Some (for example, Fu et al., 2005, and Sinclair et al., 2015 – both fMRI studies) also control for participants’ motion in the scanners.

In our initial submission, we felt that these controls did not go far enough. On top of age and sex demographics, we included measures of brain volume, and of head motion (in fMRI). We were particularly concerned that our results might simply reflect heritable changes in anatomy that influence how our signal is measured (such as cortical folding patterns), or reflect heritable patterns in signal power (if, for example, higher connectivities are observed between two nodes simply because those nodes have higher SNRs). We therefore regressed out measures of node power from our analysis, and furthermore performed control analyses for the heritabilities of power and of cortical curvature. Finally, previous electrophysiological studies of the heritability of connectivity, which have tended to measure the synchronisation likelihood between sensors, will be sensitive to any heritable anatomical differences that create similar profiles of volume conduction. We instead used an approach for connectivity analysis in MEG that was defined in source-space and removed any confounding effects attributable to magnetic field spread (or source leakage).

The reviewers have suggested several additional measures that are important to account for. We have addressed their concerns to the extent to which it is possible within the scope of the available dataset, by now including the age2 * sex interaction, a measure of head motion, a measure of beamformer noise, and an additional measure of signal power in the MEG analyses. We have also expanded our discussion of the pre-processing pipelines used to remove structured noise and physiological artefacts from the data.

Eye movement data were unfortunately not collected as part of the HCP imaging protocol. We have expanded our Discussion accordingly, where we also note that independent components of the sensor data that correspond to eye movements are, however, removed during preprocessing. The paragraphs in our Discussion that focus on potential confounds now read, “We have taken a comprehensive approach to our heritability analyses by controlling for a wide range of potential confounds. Of the heritable effects of physiology, anatomy and noise on functional connectivity, the influences of signal power, cortical folding patterns and subject motion are perhaps most worthy of discussion.”

“Higher coupling strengths are commonly observed, in both fMRI and MEG estimates of functional connectivity, between regions with higher signal power. […] To control for its influence on our results, we have included measures of the signal power in each node as confound regressors in all of our genetic analyses of functional connectivity, although it is not a perfect proxy for the signal to noise ratio, as the measured power will be affected by fluctuations in both the noise and the signal.”

“Cortical folding patterns have been shown to have modest heritability (Botteron et al., 2008; Sotiropoulos et al., 2015; Van Essen et al., 2014); although the strength of genetic control and the mechanism of its influence are not yet certain (Gomez-Robles et al., 2015; Ronan & Fletcher, 2015; Tallinen et al., 2016). [...] Our connectivity estimates use methods that are robust to source leakage artefacts (Brookes et al., 2012; Colclough et al, 2015; Hipp et al., 2012; O’Neill et al., 2015), and the lack of strong correlations between the spatial profile of the heritability of cortical curvature and the heritability of edges associated with each ROI (as shown in supplementary figure 3 and supplementary table 5) is good evidence, we believe, that this confound has not seriously impacted our findings.”

“Subjects’ motion inside a scanner during resting-state recordings is heritable (Couvy- Duchesne et al., 2014), and motion is a known resting-state confound for functional connectivity analyses (Siegel et al., 2016). […] On the other hand, it is worth noting that although saccades and microsaccades are thought to be linked to gamma-band oscillatory patterns (Muthukumaraswamy, 2013; Orekhova et al., 2015), which we are not studying, there is no strong evidence-base that they influence resting-state recordings at lower frequencies.”

Our Material and methods section has been adjusted to read, “On each edge, we regressed out the effect of age, the square of age, sex, an age and sex interaction, an interaction between sex and the square of age, the cube root of intracranial volume and of cortical volume (both estimated with FreeSurfer), a measure of subject motion in the scanner (fMRI_motion) and, for the fMRI data, the MR image reconstruction software version. […] In the MEG data, where the standard deviation of the power of a virtual sensor is often correlated with its mean, we additionally include, in the same manner, a regressor formed from the ratio of the standard deviation to the mean of the power time course.”

“Finally, we also regressed out a measure of the noise passed by the beamformer for each subject. […] These voxel-wise noise estimates are scaled by the same weightings used to compute the ROI time courses, and averaged over ROIs to create a regressor estimating the ‘noisiness’ of the beamformer for each subject.”

3) The results are all based on one particular parcellation. Some evidence should be provided that the results do not depend on this particular choice. The optimal way to address this problem would be to rerun the analysis pipeline using several different parcellation schemes.

We have re-run our analysis using an additional parcellation, the 15-dimensional ICA decomposition from the HCP data. This choice was motivated in part by the limitations of MEG data, whose effective spatial resolution cannot support high-dimensional parcellations, and in part by the reviewers’ suggestions that we investigate the heritability of inter-network connections rather than intra-network connections. The 15-dimensional parcellation is well-suited for this, as it provides entire RSNs as the ‘nodes’ of the functional network analysis (including the DMN, DAN, motor network, visual networks). To our knowledge, such a parcellation has not previously been used in the literature for functional network analyses in MEG.

In the fMRI data, we find a higher mean heritability (29%) with the lower-dimensional parcellation than our original choice, and the difference of influence between genes and the environment shared between twins is also retained.

In the smaller MEG dataset, we find no significant heritability on average over the network connections, in any frequency band, using the 15-dimensional parcellation. This result is consistent with the fact that conventional (fMRI) RSNs appear only partially in resting MEG data, or are comprised of many sub-networks that synchronise at relatively fast timescales. This makes the use of this particular parcellation inappropriate for MEG. Building reliable and informative data-driven (e.g. lower-dimensional) parcellations in MEG is still very much an open, non-trivial research problem, and investigation of a low-dimensional parcellation that would yield positive results in MEG is beyond the scope of the current work.

In summary, we find the fMRI result encouraging, and suggest that the lack of a result in the MEG data highlights issues in finding appropriate parcellations for this modality.

We have included this paragraph in our Material and methods section, “To investigate whether our results were particular to our choice of parcellation, we additionally ran our analysis on a lower-dimensional decomposition: the 15-dimensional group-ICA map computed from the HCP900 data. The nodes in this parcellation represent entire functional networks (such as the default mode or motor networks), and the connectivity matrix therefore represents inter-network dependencies.”

Altered our Results section, “To investigate whether our results were particular to our choice of parcellation, we reran our heritability analyses using the 15-dimensional ICA decomposition of restingstate fMRI recordings released by the HCP..[…] The parameter estimates and confidence intervals are presented in Supplementary file 3.”

And added to our Discussion as follows, “The primary parcellation that we employed consists of focal, contiguous regions. […] That we found no significant genetic effects on the connections between this additional set of 15 networks in our MEG data may simply reflect the unsuitability of static whole-network parcellations for this modality.”

4) There appear to be several differences between the way that the fMRI and the MEG data were processed. These should be justified and explained, or a more consistent approach should be used.

The three reviewers have touched upon three differences in the way the fMRI and MEG data are processed: in the pre-processing steps, in the network estimation method, and in the correction for confounds within the heritability analyses.

We have discussed the improvements we have made to our confound controls in point 2) above, which now include using the same regressors in each modality, apart from those that are modality-specific.

The pre-processing steps are necessarily slightly different in detail between the two modalities, but generally take the same approach of: movement compensation – registration – removal of artefactual channels and epochs – ICA decomposition and removal of noise components – source reconstruction. The registration algorithms are obviously different between modalities, and the ICA noise removal process is based on an automated classification algorithm for the fMRI data only, because it has only been trained on these data. The pre-processing stages for each modality have all been performed by the HCP, and we believe it is fair to say that these pre-processing pipelines reflect the current state of the art in the pre-processing of resting-state data.

The principal difference between the processing of the modalities raised by reviewers was our approach for constructing the network matrices: we employed regularised partial correlations for the fMRI data, but ordinary (full) correlation for the MEG data. Different modalities do require different approaches, but the objective of taking these different courses is to identify bottom-line measures that are accurate and comparable across modalities. In short, the consistent approach we have taken is to use the best-practice inference methods in each modality.

We have adjusted our Material and methods section as follows, “We use slightly different network estimation methods in the two modalities. […] We did repeat our fMRI analysis using full correlation, rather than partial correlations, and found qualitatively similar results, with no difference in the significance of the claims we would make.”

Additional remarks on this revision:

Our revised submission uses the additional data available in the HCP900 data release. These new data (28 additional MEG subjects and 359 additional MRI subjects), together with the additional control variables included in our heritability analyses, and updated twin zygosities derived from genotyping data (rather than exclusively self-reported statuses) have led to slight changes in our results and parameter estimates, although not in any substantive way to our conclusions. We estimate lower values for heritability in all modalities (although within the previous confidence intervals), which may in part be related to the tighter control over noise confounds. However, the principal claim of our paper, that genetic effects are more important than the environmental factors that twins share, is now supported in the MEG beta-band, as well as in the fMRI analysis.

Overall, as a result of the input from reviewers, we now have even more confidence in the strength of the results we are presenting, in their generalisability and in their relevance. We thank the reviewers for their contribution, and set out below any remaining comments relevant to their more specific, individual points.

Reviewer #1:

Although the authors show THAT genetics influences FC, they provide no information about WHY genetics influences FC. Which of the many mechanisms that contribute to observed BOLD FC are influenced by genetics?

Thank you for prompting us to include more comment about the plausible mechanisms mediating the genetic influence. We highlighted the extension of our discussion in part 1) of the responses above.

As the authors note, several of the results reported here (such as the heritability of FC from EEG data) have been reported already in previous studies. It is claimed here that these previous studies are less interpretable than the current study. That may be true, but then the authors need to provide the interpretation. If not then this paper loses a lot of its potential novelty and impact and I am not sure that it will be suitable for publication in this venue.

Thank you for highlighting the need to expand upon this point, as we feel that it is an important one. The relevant changes we made to our manuscript are described above in our response to the reviewers’ novelty concerns (1, above).

Some effort should be made to explain several of the more interesting and unusual effects as well. For example, it is stated that there is significant heritability in the alpha and beta bands of the MEG data but not in the theta band. Why? What is a plausible mechanism that would generate this pattern of results?

On this point, we would suggest that the lack of significant results in the theta band is not a reflection on the influence of genetics, but instead a product of the limited statistical power caused by the noisiness of estimating functional connectivity matrices in MEG. This is consistent with our past work (Colclough et al., 2016, NeuroImage), which demonstrated that all common methods for estimating resting-state connectivity in MEG show low repeatability, even over consecutive recording sessions from the same subject. The same paper showed a breakdown of this performance by frequency band, and revealed that network estimation was particularly noisy for all frequency bands other than alpha and beta. Additionally, the generators of theta power are more frontal in cortex than alpha and beta (Hipp et al.,2012, Nature Neuroscience), and the MEG scanner used by the HCP seems to have poor frontal sensitivity (and poorer sensor coverage in frontal regions), at least in comparison to the CTF MEG scanner. As a final note in this vein, our 2016 NeuroImage paper compared a wide variety of methods for inferring functional connectivity in MEG. The approach we use in this paper was chosen because it demonstrated the greatest repeatability in our method comparison.

We have slightly re-arranged the relevant part of our Discussion to make our thoughts on this point explicit, “We do not find significant genetic control of functional connectivity in the theta frequency band. […] Our claims are based on significant results from robust nonparametric tests, but the confidence intervals on our parameter estimates are inevitably broad.”

There is one major class of potential confounds which appear to be given short shrift here: potential contamination of the FC data by confounding factors having nothing to do with cortical connectivity or communication, but which are nevertheless heritable. The problem is that genetic factors might influence processes that are well known to influence correlations in MRI data (and likely MEG data as well). For example, if genetics causes some people to wiggle more in the magnet, that is obviously going to influence FC. (In fact I am fairly certain that I saw another study recently that reported just that effect, though unfortunately I am unable to find it now.)

We completely agree with this – that a major concern for genetic analyses of heritability is the wide array of possible co-varying confounds. Please see above for a more detailed response to these points (item 2).

It is possible that the paper being alluded to by the reviewer was Couvy-Duchesne et al., 2014, this seemed the most relevant authority, and we now cite it in our Discussion.

To take just one specific example of the above concern, it is reported that the visual system has a high degree of FC in the MEG data. The most likely candidate here would be eye movements, which will clearly affect MEG correlations, and which may also be influenced by genetics. However, I didn't see any report here that the eyes were tracked properly, or that the eye tracking data were regressed out when calculating FC. It is precisely these sorts of potential confounds (i.e., genetic influences operating on behavior that in turn influences FC rather than operating on FC directly) that must be scrupulously accounted for before publication.

Unfortunately, eye motion was not tracked as part of the HCP-protocol for the MEG resting-state scans, so we are unable to control for this directly. However, any independent components of the MEG sensor data that correspond to eye movements are removed during pre-processing. Further, while there is evidence that saccades and micro-saccades affect gamma-band activity, we are not actually measuring gamma-band functional connectivity, and there is not a strong evidence base for their impact on lower frequency recordings. Please see item (3) at the start of our response for how we have altered our discussion to reflect these points.

We did find strong connectivity in the visual system, particularly in the alpha-band – but this is well-reported in the MEG resting-state literature, and not thought to be just a product of eye movements (see, for example, Brookes et al., 2011, Baker et al,. 2014, Colclough et al.,2015, Hipp et al., 2012, Mantini et al., 2007, Marzetti et al., 2013, de Pasquale et al., 2012.

Perhaps one good way to get an intuitive handle on these sorts of potential indirect genetic influences would be to run separate analyses to determine the heritability of the various confounds that are known to affect FC, such as motion, eye movements, beamforming errors (in the MEG data). For that matter other factors that might affect FC indirectly might also be heritable. BOLD data taken directly off the scanner are non-Gaussian, and these are Gausianized during pre-processing by a necessarily imprecise procedure. Is the distribution of raw BOLD signals heritable? All these seem plausible, and all could potentially affect FC. A systematic analysis of these factors would seem to be critical for making any strong claims about direct genetic influences on FC.

The difficulty with performing separate analyses is that we expect many of these factors to be heritable in and of themselves. The real question is whether or not our functional connectivity estimates contain residual elements of these effects, or whether we are assessing the heritability of FC only, as we desire. For example, we have conducted additional heritability analyses on the power in each network node. In the fMRI data, we find significant heritability, on average, of these signal strengths (which is not necessarily surprising), and we must rely on our use of power as a confound regressor in the FC analysis to be confident in our claims. We believe we are the first authors to measure the heritability of FC that have tried to account for the effects of signal power in this way.

We have taken your suggestions about possible additional confounds for the MEG connectivity estimates, and included additional regressors for beamforming noise and for subject motion in our analysis. Please see the discussion on item 2) at the start of this response for details.

The claims in the Abstract and Discussion are a bit over-stated in several places given the results because they do not make any reference to places where no heritability is found. For example, while effects in the MEG alpha and beta frequency bands are observed, this is not true for the theta frequency band.

Thank you for highlighting this. We have explicitly discussed the lack of findings in theta (see above, in response to your third question), and altered our Discussion and Abstract to be specific about the bands where we find significant results.

The relevant part of the Abstract now reads, “On average over all connections, genes account for about 15% of the observed variance in fMRI connectivity (and about 10% in alpha-band and 20% in beta-band oscillatory power synchronisation),”

And that of the Discussion reads, “we estimate that the average heritability of individual connection strengths is about 15- 18% for BOLD correlations, and between 13% and 26% (1% and 25%) for correlations in beta-band (alpha-band) power fluctuations.”

The figures and the supplementary movies in this paper are rather poor, especially the connectivity figures. There are many excellent choices for visualization software these days.

We have replaced the connectivity figures with improved versions, but have distinguished between the principal figure of the paper (Figure 1), and the supplementary figures.

Particularly, we have altered the connectivity plots in Figure 1 to an alternative that has more visual impact and clear separation of the network nodes. We have retained the glass-brain style figures in the supplementary material, as we find it aids our understanding of the networks to see the connections in the context of the cortical locations of the nodes. This compromise offers a balance between visual aesthetics and ease of comprehension.

The entire data analysis procedure is based on the HCP ICA parcellation. It would inspire more confidence in the conclusions if some analysis was provided that showed whether the choice of parcellation scheme makes any difference on the results or the conclusions. I don't think that it would make much difference in the results if one compared the HCP ICA parcellation to, say, the Glasser HCP parcellation, because these are both fine-scale parcellations that are likely to include more spatial information than can really be supported by genetics data anyway. (In the case of the MEG data, the HCP ICA parcellation may very well have higher resolution than can be supported by either the genetics or the MEG.) It would be immensely helpful to know whether a coarser parcellation schemewould change the results, and I think that knowing the answer to this question might also help us address the question of why these apparent genetic influences are found.

Thank you for this discussion. We agree that an understanding of how parcellations may affect the heritability of connectivity is important, and past work has not touched upon this point. We also agree that a denser parcellation scheme would not be a sensible comparator, for the reasons you mention. We repeated our analysis with the 15-dimensional HCP ICA parcellation, which has a substantially different structure to the 39-dimensional parcellation we originally used. As your point was raised by several reviewers, we discussed our approach and findings in item 3) at the start of this response.

It looks as if subject motion was removed during pre-processing in both the fMRI and the MEG data, but subject motion was still used as a regressor during ACE modeling in the fMRI data. Subject motion is a problem in both fMRI and MEG, so if it was a problem in the fMRI data even after pre-processing why was pre-processing sufficient for the MEG data? At a minimum it seems that it should also be included in the MEG modeling analysis.

We take your point that motion is an important confound to treat carefully.

It is worth noting that the pre-processing for motion is slightly different between the modalities. The fMRI pipeline registers functional volumes into the same static space, then uses an automated classification system to remove ICA components from the data associated with motion (and other structured noise) artefacts. Additionally, a variable summarising the motion of the subject during the recording can be used as a regressor in further analyses. On the other hand, the MEG pre-processing registers the signals in the sensors into a common, static reference frame before source reconstruction, followed by an ICA decomposition process for further noise removal. No summary measure of motion in the scanner is available (though see below). These approaches represent the current best practice in pre-processing functional images for motion in these modalities.

In the absence of a MEG-specific motion confound for the resting-state scans, we have now used the summary measure from each subject’s fMRI scans as a proxy motion variable for the MEG analysis. To the extent that motion is heritable, this proxy measure is the best available approach for removing any residual motion confounds in this dataset. The changes to our text were highlighted at the start of our response, in part 2).

This is a side point, but I find it surprising that it is useful/helpful to regress out the MR image reconstruction software version in ACE modeling. This suggests that these data really do live at the very limit of sensitivity, and that confounds and bias really can creep into the data even after pre-processing. It seems like the best practice would be to use the same version of the software to process all the data and I find it a bit worrisome that this was not done.

The reconstruction of raw (complex, multi-coil, multiband, k-space) fMRI data, was carried out by the HCP. This is outside our control – and indeed at this point theirs – as it was not possible for practical reasons to store these raw data long-term. However, we have no reason to think that, after inclusion of this confound variable, any signature of this change remained in the data. A full explanation of this change is given on the HCP website: https://wiki.humanconnectome.org/display/PublicData/Ramifications+of+Image+Reconstructio n+Version+Differences

In Subsection “E. Three-component variance models” it is stated that the variance partitioning analysis addresses non-negative portions of the variance. Why do these procedures produce negative variance estimates at all? Is this just statistical error?

We apologise for the potentially confusing wording; we had written, "This model ascribes non- negative proportions of the variance either to additive shared genetics […]” Of course, variance is always non-negative, and the model appropriately constrains the variance parameters. This wording was to reference that we allow h2 and c2 to be exactly zero. On reflection, this is self- evident, and so we have simply dropped the word “non-negative.”

Reviewer #2:

Colclough and colleagues examine the genetic influences on resting state MRI and MEG connectivity in "The heritability of multi-modal connectivity in human brain activity." Subjects included 461 individuals with MRI data and a group of 61 primarily overlapping individuals with MEG data all from the Human Connectome Project. Subjects were from extended twin pairs or unrelated individuals. The goals of the article are to estimate the heritability and common genetic influences on resting state and MEG measures of connectivity. The article is clearly written and the analytic plan is both complex and reasonable. Unfortunately, the findings represent only an incremental increase in scientific knowledge, as the genetic influences on resting state have been reported several times using conceptually similar methods. While the MEG results are novel, that analysis is rather underpowered, as the authors appropriately note in the Discussion. Thus, while I believe the findings should be reported, the authors should either describe the work as replication or conduct additional analyses that will increase novelty.

Thank you for this overview. We have discussed all reviewers’ concerns over novelty at the start of this response, in item 1). This includes changes to better contextualise our findings and the way in which they advance the state of knowledge.

Reviewer #3:

1) I think the authors should clarify why 39 cortical regions are chosen, as opposed to other common parcellation schemes, which often have 50-100 regions.

Larger parcellation schemes (>60 brain regions) are unfortunately impractical for MEG: the scanner only records from several hundred sensors, and the effective dimensionality of the data, once it is localised within the cortex using established methods, is on the order of 60. Models that separate sources that are internal and external to the MEG dewar tend to provide support for 60 sources or fewer (Taulu et al., 2005), and recently developed adaptive data driven parcellations that combine MEG and EEG data only find that there are on the order of 70 unique parcels identifiable [http://biorxiv.org/content/early/2017/01/04/097774]. In our experience, ~40 regions is a good compromise for a parcellation that can represent contiguous nodes, but does not stretch beyond the support of the spatial information available in the MEG data. Note that we have, however, included an additional analysis on a lower, 15-dimensional cortical parcellation; please see item (3) at the start of our response for a more extended discussion of parcellation schemes and the results of this analysis.

We have included this paragraph in the Material and methods, to provide additional explanation, “This parcellation was chosen because the use of contiguous, focal nodes creates a comprehensible network model between individual regions. […] This upper bound is supported by models that categorise signals as originating within or without the MEG dewar, which suggest a limit of about 60 measurable cortical sources (Taulu et al., 2005), and by recently developed data-driven parcellations that only find around 70 unique identifiable parcels using combined MEG and EEG data (Farahibozorg et al., 2017).”

2) I'm surprised that the Authors did not at least attempt to reconstruct networks in the higher gamma band. Why only theta/alpha/beta?

For two reasons. First, in our previous work (Colclough et al., 2016) we assessed the reproducibility of network estimation in resting-state MEG. We found that by and large, reproducibility was low, and that network estimation is therefore noisy. The results were bad for all bands other than alpha and beta, where most structured signal originates in the resting state. Secondly, and this is the reason that the theta band was included at all, because the evidence in the literature is that resting-state networks present in MEG primarily in the theta, alpha and beta bands, but less clearly in lower or higher frequencies [Hipp et al., 2012]. This does not necessarily mean that there is not structured connectivity between oscillations in these other bands (indeed, ECoG studies find strong relationships between gamma oscillations and fMRI resting-state networks; see for example Nir et al., 2008), but simply that these relationships do not emerge in MEG data using conventional functional connectivity techniques for MEG. We now refer to this in a sentence in the Introduction, “These bands span the frequency range within which the most convincing patterns of resting-state MEG connectivity have been shown to be expressed (Baker et al., 2014; Brookes et al., 2011a; Hipp et al., 2012; Mantini et al., 2007; Marzetti et al., 2013; de Pasquale et al., 2012, 2015).”

3) One of my main concerns is the use of partial correlations in the FMRI analysis and simple correlations in the MEG network analysis. Why this difference? The Authors cite two of their own papers as justification but this not particularly convincing. Could we have a short justification here?

We agree that the original presentation of our methodology should have included more explicit discussion of some of these choices. We have addressed the reviewer’s concerns on this point at the start of our response, under item 4).

4) The authors state that the beta-band "exhibits broad connectivity over the whole cortex". That doesn't really appear to be the case looking at Figure 1. In ay case, what does that statement actually mean – it's quite a woolly phrase that is not backed up by any kind of quantification.

Thank you for highlighting this. We have removed the poor phrasing, and now describe the group-mean beta-band connectivity matrices as, “The beta-band exhibits strong bilateral coupling across the sensorimotor cortices, with connectivity continuing through the superior parietal lobes and down to occipital cortex.”

5) The authors look at both genetic and environmental factors in terms of contribution to signal power (in both FMRI and MEG), partly as a control for these as confounds in the connectivity analyses, but also as interesting exploratory analyses in their own right. I have two concerns here. 1) Can the Authors please describe how "power" was assessed in the FMRI signal? I assume this is some measure of temporal variation around the mean of the voxel time series? 2) For the MEG, the power passed by a beamformer in an RSN analysis can be quite dependent on geometric effects (even after weights normalisation). Wouldn't a better measure of 'activity" in the MEG amplitude envelopes be some measure of temporal variability? The standard deviation (SD) is often used, but the SD of the virtual-sensor and the Mean are often correlated in beamformer reconstructions, so a better measure might be some proportional change (e.g. SD/Mean). "Activity" in both the FMRI and MEG time series could thus be assessed using the same (or very similar) metric.

We agree that we had neglected to mention how this was calculated. We had in fact used the standard deviation of the node time courses (fMRI signal or MEG power envelope) as our measure of “power,” as suggested. We have now explained our approach in the text. We have also followed the reviewer’s advice, and included an additional power confound in our MEG genetic analyses that is the coefficient of variation (SD/mean) for each node. We have highlighted the altered text in item 2) near the top of this response, where we address our use of additional confounds.

6) In the ACE model, quite a few nuisance covariates are regressed out of the model before heritability/environmental effects were assessed. How were these parameters (and their second-order versions) chosen? In addition, when regressing out some many parameters, is there an issue with statistical power in the MEG analyses as some of the groups only have relatively few subjects (i.e. 11 monozygotic twin pairs)?

As gross brain size is known to vary greatly by gender and age, it is routine in brain imaging studies to remove effects of age, gender and, when the sample size supports it, age2 and interactions between these variables. Most genetic analyses will also include regressors of this form to control for differences in basic demographics between unrelated subject pairings. The MEG sample we originally used totaled 44 twins; while of course a larger sample is desirable (and we have now increased it to 64), inaccurate fitting of age2 and age*gender interactions will only degrade our residuals and reduce our ability to detect heritability. Hence, we see this as a conservative approach (and it retains symmetry with the use of these regressors in the larger fMRI sample).

7) Finally, with these "nuisance" regressors Is there a potential problem with interaction with heritability? For example age is 100% matched in the twin-pairs, but is presumably not matched for the non-related pairings.

It is precisely because the lack of age matching between siblings that these nuisance regressors are essential. We have considered this carefully, and have not been able to construct a setting where use of covariates (which remove variance) can artificially inflate heritability.

https://doi.org/10.7554/eLife.20178.021

Article and author information

Author details

  1. Giles L Colclough

    1. Oxford Centre for Human Brain Activity (OHBA), Wellcome Centre for Integrative Neuroimaging, Department of Psychiatry, University of Oxford, Oxford, United Kingdom
    2. Oxford Centre for Functional MRI of the Brain (FMRIB), Wellcome Centre for Integrative Neuroimaging, Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom
    3. Department of Engineering Science, University of Oxford, Oxford, United Kingdom
    Contribution
    Conceptualization, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing—original draft, Project administration, Writing—review and editing
    For correspondence
    giles.colclough@ohba.ox.ac.uk
    Competing interests
    No competing interests declared
    ORCID icon 0000-0003-1074-7186
  2. Stephen M Smith

    Oxford Centre for Functional MRI of the Brain (FMRIB), Wellcome Centre for Integrative Neuroimaging, Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom
    Contribution
    Conceptualization, Data curation, Software, Supervision, Funding acquisition, Investigation, Visualization, Methodology, Project administration, Writing—review and editing
    Competing interests
    No competing interests declared
  3. Thomas E Nichols

    1. Department of Statistics, University of Warwick, Coventry, United Kingdom
    2. Warwick Manufacturing Group, International Manufacturing Centre, University of Warwick, Coventry, United Kingdom
    Contribution
    Software, Investigation, Methodology, Writing—review and editing
    Competing interests
    No competing interests declared
  4. Anderson M Winkler

    Oxford Centre for Functional MRI of the Brain (FMRIB), Wellcome Centre for Integrative Neuroimaging, Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom
    Contribution
    Investigation, Methodology, Writing—review and editing
    Competing interests
    No competing interests declared
  5. Stamatios N Sotiropoulos

    1. Oxford Centre for Functional MRI of the Brain (FMRIB), Wellcome Centre for Integrative Neuroimaging, Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom
    2. Sir Peter Mansfield Imaging Centre, School of Medicine, University of Nottingham, Nottingham, United Kingdom
    Contribution
    Software, Investigation, Methodology, Writing—review and editing
    Competing interests
    No competing interests declared
  6. Matthew F Glasser

    School of Medicine, Washington University, St. Louis, United States
    Contribution
    Data curation, Writing—review and editing
    Competing interests
    No competing interests declared
  7. David C Van Essen

    School of Medicine, Washington University, St. Louis, United States
    Contribution
    Data curation, Funding acquisition, Writing—review and editing
    Competing interests
    Senior editor, eLife
    ORCID icon 0000-0001-7044-4721
  8. Mark W Woolrich

    1. Oxford Centre for Human Brain Activity (OHBA), Wellcome Centre for Integrative Neuroimaging, Department of Psychiatry, University of Oxford, Oxford, United Kingdom
    2. Oxford Centre for Functional MRI of the Brain (FMRIB), Wellcome Centre for Integrative Neuroimaging, Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom
    3. Department of Engineering Science, University of Oxford, Oxford, United Kingdom
    Contribution
    Conceptualization, Software, Supervision, Funding acquisition, Investigation, Methodology, Project administration, Writing—review and editing
    For correspondence
    mark.woolrich@ohba.ox.ac.uk
    Competing interests
    No competing interests declared

Funding

Research Councils UK (Digital Economy Programme (EP/G036861/1, Centre for Doctoral Training in Healthcare Innovation))

  • Giles L Colclough

Medical Research Council (MRC UK MEG Partnership Grant (MR/K005464/1))

  • Giles L Colclough
  • Mark W Woolrich

Wellcome Trust (098369/Z/12/Z)

  • Giles L Colclough
  • Stephen M Smith
  • Thomas E Nichols
  • Anderson M Winkler
  • Stamatios N Sotiropoulos
  • Mark W Woolrich

Wellcome Trust (100309/Z/12/Z)

  • Giles L Colclough
  • Stephen M Smith
  • Thomas E Nichols
  • Anderson M Winkler
  • Stamatios N Sotiropoulos
  • Mark W Woolrich

Wellcome Trust (106183/Z/14/Z)

  • Giles L Colclough
  • Stephen M Smith
  • Thomas E Nichols
  • Anderson M Winkler
  • Stamatios N Sotiropoulos
  • Mark W Woolrich

Wellcome Trust (203139/Z/16/Z)

  • Giles L Colclough
  • Stephen M Smith
  • Thomas E Nichols
  • Anderson M Winkler
  • Stamatios N Sotiropoulos
  • Mark W Woolrich

National Institutes of Health (R01EB015611-01)

  • Thomas E Nichols
  • Matthew F Glasser
  • David C Van Essen

National Institutes of Health (NRSA fellowship (F30-MH097312))

  • Thomas E Nichols
  • Matthew F Glasser
  • David C Van Essen

National Institutes of Health (1U54MH091657)

  • Thomas E Nichols
  • Matthew F Glasser
  • David C Van Essen

Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq,211534/2013-7)

  • Anderson M Winkler

Engineering and Physical Sciences Research Council (EP/L023067)

  • Stamatios N Sotiropoulos

National Institute for Health Research (NIHR Oxford Biomedical Research Centre)

  • Mark W Woolrich

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

Acknowledgements

The authors would like to thank Matthew Brookes and Andrew Quinn for helpful discussions, and Paul McCarthy for his assistance in creating figures. Functional MRI and MEG data were provided 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. GLC is funded by the Research Councils UK Digital Economy Programme (EP/G036861/1, Centre for Doctoral Training in Healthcare Innovation); SMS by a Wellcome Trust Strategic Award (098369/Z/12/Z); TEN by the Wellcome Trust (100309/Z/12/Z) and the NIH (R01EB015611-01); AMW by the National Research Council of Brazil (CNPq, 211534/2013-7); SNS by the UK Engineering and Physical Sciences Research Council (EP/L023067); MFG by an NRSA fellowship (F30-MH097312, NIH); DCVE by the NIH (1U54MH091657); and MWW by the Wellcome Trust (106183/Z/14/Z) and the MRC UK MEG Partnership Grant (MR/K005464/1). This research was supported by the NIHR Oxford Biomedical Research Centre. The Wellcome Centre for Integrative Neuroimaging is supported by core funding from the Wellcome Trust (203139/Z/16/Z).

Ethics

Human subjects: HCP data were acquired using protocols approved by the Washington University institutional review board. Informed consent was obtained from subjects. Anonymised data are publicly available from ConnectomeDB (db.humanconnectome.org; Hodge et al., 2016). Certain parts of the dataset used in this study, such as the family structures of the subjects, are available subject to restricted data usage terms, requiring researchers to ensure that the anonymity of subjects is protected (Van Essen et al., 2013).

Reviewing Editor

  1. Jack L Gallant, Reviewing Editor, University of California, Berkeley, United States

Publication history

  1. Received: August 3, 2016
  2. Accepted: July 13, 2017
  3. Accepted Manuscript published: July 26, 2017 (version 1)
  4. Accepted Manuscript updated: August 1, 2017 (version 2)
  5. Version of Record published: September 29, 2017 (version 3)

Copyright

© 2017, Colclough et al.

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

Metrics

  • 1,677
    Page views
  • 327
    Downloads
  • 5
    Citations

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

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)

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

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