1. Neuroscience
Download icon

Oscillatory hyperactivity and hyperconnectivity in young APOE-ɛ4 carriers and hypoconnectivity in Alzheimer’s disease

  1. Loes Koelewijn
  2. Thomas M Lancaster
  3. David Linden
  4. Diana C Dima
  5. Bethany C Routley
  6. Lorenzo Magazzini
  7. Kali Barawi
  8. Lisa Brindley
  9. Rachael Adams
  10. Katherine E Tansey
  11. Aline Bompas
  12. Andrea Tales
  13. Antony Bayer
  14. Krish Singh  Is a corresponding author
  1. Cardiff University, United Kingdom
  2. Neuroscience and Mental Health Research Institute, Cardiff University, United Kingdom
  3. Core Bioinformatics and Statistics Team, College of Biomedical and Life Sciences, Cardiff University, United Kingdom
  4. College of Human and Health Sciences, Swansea University, United Kingdom
  5. School of Medicine, Cardiff University, United Kingdom
Research Article
  • Cited 0
  • Views 1,333
  • Annotations
Cite this article as: eLife 2019;8:e36011 doi: 10.7554/eLife.36011

Abstract

We studied resting-state oscillatory connectivity using magnetoencephalography in healthy young humans (N = 183) genotyped for APOE-ɛ4, the greatest genetic risk for Alzheimer’s disease (AD). Connectivity across frequencies, but most prevalent in alpha/beta, was increased in APOE-ɛ4 in a set of mostly right-hemisphere connections, including lateral parietal and precuneus regions of the Default Mode Network. Similar regions also demonstrated hyperactivity, but only in gamma (40–160 Hz). In a separate study of AD patients, hypoconnectivity was seen in an extended bilateral network that partially overlapped with the hyperconnected regions seen in young APOE-ɛ4 carriers. Using machine-learning, AD patients could be distinguished from elderly controls with reasonable sensitivity and specificity, while young APOE-e4 carriers could also be distinguished from their controls with above chance performance. These results support theories of initial hyperconnectivity driving eventual profound disconnection in AD and suggest that this is present decades before the onset of AD symptomology.

https://doi.org/10.7554/eLife.36011.001

Introduction

Once established, the cognitive impairments characteristic of Alzheimer’s disease (AD) are irreversible (Herrup, 2015). To effectively identify measures to predict and prevent AD, it is vital to identify biomarkers at an early age, before the presence of any symptoms, and one increasingly popular route for this is neuroimaging of genetic risk groups. The greatest genetic risk factor for developing late-onset AD is the apolipoprotein E4 (APOE-ɛ4) allele (Corder et al., 1993; van der Flier et al., 2011).

APOE consists of 299 amino acids divided into receptor-binding and lipid-binding regions and has three isoforms (ɛ2, ɛ3 and ɛ4) that only differ at two of the amino acids (112 and 158), both of which are on the receptor-binding part of the protein. APOE has a crucial role in transporting cholesterol and other lipids, both in plasma and in the brain, but it is its role in regulating amyloid-β that has received much attention, due to the latter’s apparently crucial role in AD (Reinvang et al., 2013; Zhao et al., 2018). The pathological accumulation of amyloid-β could occur via multiple mechanisms of over-production and/or impaired clearance (Zhao et al., 2018) and it appears that the presence of the APOE-ɛ4 isoform may differentially lead to one or more of these occurring. For example, compared to APOE-ɛ3, APOE-ɛ4 stimulates increased recycling of the amyloid precursor protein (APP), hence leading to an increase in amyloid-β production (Ye et al., 2005). This mostly appears to be via modified interactions between APOE-ɛ4 and the LRP1 receptor. Similarly, it has been shown that APOE can bind to amyloid-β to form a complex that then promotes clearance via cell internalization of this complex (Zhao et al., 2018). As the APOE-ɛ4-amyloid-β complex is less stable than the other isoforms, this may make it less effective in amyloid clearance. Finally, APOE-ɛ4 may interfere with protease-mediated clearance of amyloid-β via downregulation of, for example, insulin-degrading-enzyme (IDE) (Du et al., 2009).

Whatever the precise mechanism, APOE-ɛ4 carriers have a much increased risk of developing sporadic AD (Corder et al., 1993), of progression to AD if diagnosed with mild cognitive impairment (MCI) (Petersen, 1995), and to develop AD at an earlier age (Corder et al., 1993; van der Flier et al., 2011) compared to non-carriers. Studying early effects of the APOE-ɛ4 allele on the brain may therefore reveal insights into the neural pathways leading to AD.

Within groups of AD patients and elderly controls, APOE-ɛ4 carriers show alterations in brain structure and function that follow similar patterns to those that characterise AD, but are stronger than in non-carriers. Imaging studies show that APOE-ɛ4 carriers present with more severely atrophied temporal areas (Filippini et al., 2009a; Filippini et al., 2009b; Juottonen et al., 1998), and have elevated levels in temporo-parietal areas of the two main harmful brain deposits seen in AD: the tau protein (Ossenkoppele et al., 2016), causing neurofibrillary tangles, and amyloid-β (Drzezga et al., 2009), the peptide accumulating in the pathognomonic intercellular plaques.

Functional neuroimaging during a ‘resting state’ can show intrinsic fluctuations in levels of background brain activity. Fast fluctuations of rhythmic neural activity occur spontaneously and are thought to underlie communication between brain areas (Schnitzler and Gross, 2005). These neural oscillations, particularly in frequency bands such as the alpha band, have been linked to patterns of white matter connectivity (Teipel et al., 2009). APOE plays a role in lipid transport, important for myelination of white matter (Herrup, 2015). Studying neural oscillations in conjunction with variation in this gene may therefore reveal important changes in neural connectivity. Neural oscillations can be non-invasively measured in vivo in humans using electro- or magnetoencephalography (E/MEG), which measure fluctuations in electric potentials or magnetic fields and hence directly reflect neural activity at the scalp with high temporal resolution (Baillet, 2017).

To date, M/EEG studies in elderly APOE-ɛ4 carriers have shown changes in the oscillatory alpha band (Babiloni et al., 2006; Canuet et al., 2012; de Waal et al., 2013; Jelic et al., 1997; Kramer et al., 2008; Ponomareva et al., 2008) and in the much lower-frequency delta band (Cuesta et al., 2015; de Waal et al., 2013) in regions similar to those reduced in AD and MCI. These findings suggest that the presence of an APOE-ɛ4 allele is associated with differences in oscillatory brain function at a stage preceding AD symptomology, but it is not clear how early in the life span this is established. One very small-sample study has measured MEG oscillations in healthy young genotyped individuals, and found that low-frequency (theta) oscillations were increased in the medial frontal cortex of APOE-ɛ4 carriers, but did not investigate the alpha band, nor connectivity between brain regions (Filbey et al., 2006).

Here, we used MEG to study resting-state neural activity non-invasively in a large sample of healthy young individuals who were genotyped for APOE alleles (51 ɛ4 carriers and 108 non-carriers). We assessed whether and how background oscillatory neural activity and connectivity across the whole brain in a range of frequencies is affected in young APOE-ɛ4 carriers, decades before the effects of ageing or potential onset of AD symptomology. To relate the significance of these findings to brain dysfunction in AD, we further compared the young APOE-ɛ4-related findings with resting-state MEG data from patients with established AD and healthy age-matched controls. Finally, using machine learning approaches, we assessed the diagnostic information content of these MEG-derived static connectivity matrices by quantifying their accuracy in solving three problems: 1) Whether young APOE-ε4 carriers could be distinguished from young non-carriers, 2) Whether AD patients could be distinguished from elderly controls and 3) Whether a classifier trained on identifying young E4 carriers could identify AD patients, purely from their MEG connectivity matrices.

Results

Participants

APOE-ɛ4 carriers did not significantly differ from non-carriers in age, gender, years of education, hippocampal volume, head motion during the recording or number of good-quality resting-state data epochs included in the analysis (Table 1). Performance differed on only 1 out of 8 cognitive tasks, the NAB Mazes task (reasoning and problem solving), where ɛ4 carriers performed better than non-carriers, but this result (p=0.03) did not survive a false-discovery-rate correction for multiple comparisons (Benjamini and Yekutieli, 2001). Similarly, within the 68 regions of the Desikan-Killiany gyral-based atlas, there was no significant difference found in cortical thickness, volume or area, in any region, when corrected for multiple comparisons.

Table 1
Participant demographics and statistical assessment of APOE-ɛ4 group differences (two-sample t-tests).
https://doi.org/10.7554/eLife.36011.002
All MEG datasetsAPOE-ɛ4 carriersAPOE-ɛ4 non-carriersP value
Number18351108
Male/female60/12318/3336/720.81
Allele composition41 ɛ3ɛ4
6 ɛ4ɛ4
4 ɛ2ɛ4
87 ɛ3ɛ3
17 ɛ2ɛ3
4 ɛ2ɛ2
Age*24.5 ± 5.4 years25.2 ± 6.8 years24.2 ± 4.6 years0.23
Years of education16.4 ± 0.9 (N = 15)17.0 ± 1.3 (N = 34)0.13
Cognitive tasks:
Trail making test
BACS symbol coding
HVLT-R (verbal learning)
WMS-III SS (spatial span)
Letter-number span
NAB mazes (reasoning)
BVMT-R (visual learning)
Category fluency
27.8 ± 10.1 (N = 47)
65.0 ± 8.7 (N = 48)
27.7 ± 3.9 (N = 48)
16.9 ± 3.4 (N = 47)
15.0 ± 2.9 (N = 48)
22.9 ± 3.0 (N = 48)
28.3 ± 5.5 (N = 48)
29.2 ± 6.8 (N = 47)
27.1 ± 8.8 (N = 98)
66.3 ± 9.1 (N = 100)
27.2 ± 4.3 (N = 100)
16.5 ± 3.5 (N = 99)
15.1 ± 2.8 (N = 100)
21.4 ± 4.3 (N = 100)
27.8 ± 5.1 (N = 100)
28.6 ± 6.9 (N = 100)
0.65
0.42
0.51
0.51
0.86
0.03
0.54
0.60
Hippocampal volume (L) absolute, relative4343 ± 408 mm3
2.734 ± 0.26
4362 ± 422 mm3
2.705 ± 0.24
4228 ± 394 mm3
2.746 ± 0.27
0.74
0.35
Hippocampal volume (R) absolute, relative4439 ± 368 mm3
2.797 ± 0.25
4417 ± 363 mm3
2.757 ± 0.23
4416 ± 345 mm3
2.798 ± 0.26
0.67
0.35
Total ICV1596 ± 161 cm31618 ± 156 cm3
(N = 50)
1589 ± 166 cm3
(N = 106)
0.29
MEG head motion0.28 ± 0.23 cm0.32 ± 0.26 cm0.27 ± 0.21 cm0.18
MEG number of epochs144.75 ± 7.63144.25 ± 7.02145.06 ± 7.750.53
  1. *Values are represented as group mean ±SD.

    The MCCB includes tests of processing speed (BACS symbol coding, category fluency: animal naming, trail making test), working memory (WMS-III: spatial span, letter-number span), verbal and visual learning, and reasoning and problem-solving skills (NAB mazes).

  2. Relative left (L) and right (R) hippocampal volumes are calculated as absolute hippocampal volumes divided by total intracranial volume (ICV), multiplied by 1000.

APOE-ɛ4 carrier versus non-carrier group differences in oscillatory connectivity

Figure 1 shows differences in static oscillatory network connectivity for APOE-ɛ4 carriers, compared to controls. In the top row, it is clearly evident that, for the individual frequency bands, the beta band has the greatest number of valid connections (224) and the highest number of edges (n = 13) that exceed our initial thresholding criterion of p<0.05 (uncorrected). All of these edges show increased connectivity in the beta band.

Each column shows oscillatory amplitude correlations comparing APOE-ɛ4 carriers and non-carriers for each of the five frequency ranges where valid edges were found, and combined across frequency.

The top row shows where there are ‘valid’ edge connections in either of the two groups. Colours, with a key at the top of the figure, indicate regions of brain areas that the connections originate from. The numbers on the plots show the number of valid edges found. Note that Low Gamma is not shown as no ‘valid’ edges were found. The second row shows the unpaired t-statistic for APOE-ɛ4 carriers compared to non-carriers. Increases are displayed in red and decreases in blue and only those connections significant at p<0.05 (uncorrected) are shown. The numbers on each plot (n+/n-) show the number of significant increases and decreases found in each map. Note that, across all tests, only one edge connection, in the Theta band, was found to be significantly different when corrected for multiple comparisons (p<0.036, right parietal). Using the cluster-level network-based size statistic, shown in the third row, significantly large clusters were found for the theta and Combined maps. The bottom, fourth row, shows the results of the cohort random sampling analysis. Edges are only shown if they are within the 95% confidence limit for APOE-ɛ4 -related increases (red) or decreases (blue). Note that, in general, the predominant effect is for edges to show increases in connectivity in APOE-ɛ4 carriers, compared to non-carriers and that this is most clearly shown in the Combined map (right-hand column), in which the most valid connections are identified and the most significant, and consistent, differences are seen. In all connection plots, line opacity indicates relative connection strength within the plot (most opaque = strongest). Data and scripts needed to generate this figure can be found at http://dx.doi.org/10.17605/OSF.IO/E4CJX.

https://doi.org/10.7554/eLife.36011.003

In the Combined maps, there are not only more edges than any of the individual bands (n = 282) but this synthesised vector combination also shows enhanced sensitivity to cohort effects, with more connections found to be significantly different at the edge-level (n = 31), the cluster-size level (n = 19) and after cohort resampling tests (n = 26). Again, the balance of the sign of the effect is skewed towards hyperconnectivity in the majority of comparisons, although, interestingly, there is some evidence of hypoconnectivity (blue lines) in left medial temporal regions, including the hippocampus.

What can also be observed is a bias towards effects in the right hemisphere, with many of the significantly different edges arising from the right parietal cortex, but also including medial precuneus regions. These differences therefore at least partially overlap with posterior regions of the human default-mode network (DMN). Finally, what is encouraging is that most of these significant differences are observable in the 95% confidence interval test using random sub-sampling of the cohort (bottom row), suggesting that our results may well generalise to the wider population, rather than being driven by a small number of cases, or controls, in our sample.

APOE-ɛ4 carrier versus non-carrier group differences in oscillatory activity

When we assessed activity across the brain using a voxelwise assessment of temporal variability in the amplitude envelope at each location in the brain, the only differences found using TFCE (p<0.05 corrected) were in the two gamma ranges, 40–60 Hz and 60–140 Hz and are shown in Figure 2. These were exclusively gamma hyperactivities in a right-lateralised profile, distributed over lateral parietal, temporal and frontal regions. When a more conservative voxel-level threshold (p<0.05, corrected) was applied, only one region survived in the 60–140 Hz range, localised over the right middle and inferior frontal gyri.

Spatial distribution of differences in oscillatory activity, between young APOE-ε4 carriers and their matched controls and the relationship with hyper- and hypo- connectivity.

Top panel shows differences in activity as a coloured overlay depicting unpaired t-statistics (thresholded at p<0.05, corrected using TFCE), with yellow/orange representing greater activity in the APOE-ε4 carriers. Significant effects were only found in the right hemisphere and only for the two gamma ranges tested, 40–60 Hz and 60–140 Hz. The overlays are shown depicted on a template brain surface. The far-right column shows the results for the 60–140 Hz comparison using a more stringent voxel-wise correction for multiple-comparisons. The bottom panel shows the distribution of differences in both connectivity and activity, between APOE-ε4 carriers and non-carriers, within each AAL node that shows any significant effect. For clarity the left and right hemispheres are shown separately. Red lines show the number of significant connections to/from each node that are of higher strength in APOE-ε4 carriers, compared to non-carriers. Blue lines show the number of significant connections to/from each AAL node that are weaker in APOE-ε4 carriers. For the right hemisphere only, as effects were only found on the right, the green lines show the most significant t-statistic found in each AAL region for the analysis of gamma activity. Positive values indicate that gamma activity, in either band, was higher in APOE-ε4 carriers compared to non-carriers. Note that connection numbers are the sum across all frequency ranges tested, with most of these coming from the low frequencies – only one connection (FrontalInfTriL to TemporalPoleMidR) showed hyperconnectivity in the gamma range. In contrast, only gamma showed hyperactivity, so the green lines/marker exclusively reflect right-hemisphere gamma hyperactivity. The spreadsheet used to generate the graphs on this figure can be found at http://dx.doi.org/10.17605/OSF.IO/E4CJX.

https://doi.org/10.7554/eLife.36011.004

In summary, gamma hyperactivities were found in mostly right-lateralised regions, some of which also appear to show hyperconnectivity in the lower frequency ranges (principally beta and alpha). This overlap is described graphically in the lower part of Figure 2.

Alzheimer’s disease versus elderly controls

Here, we compared our finding of a spatially-localised increase in functional brain connectivity in young APOE-ɛ4 carriers to connectivity changes in a cohort of AD patients, in which we have recently reported strongly decreased alpha and beta-band oscillatory network activity in parieto-temporal areas with MEG (Koelewijn et al., 2017), for AD patients compared to age-matched controls. In our reanalysis of a subset of the same dataset, using the current pipeline (Figure 3) to compare AD patients to age-matched controls, we show strongly decreased oscillatory connectivity in predominantly parieto-temporal connections, with the most significant, and extensive, effects visible in the Alpha and Beta ranges. Again, similar to the young APOE-ɛ4 analysis, the most valid connections, and the most extensive pattern of significant differences, was observed in the synthesised vector combination of frequency-specific effects (right hand column). We also note that, in contrast to the APOE-ɛ4 analysis, most of the observed significant cohort differences are reduced connectivity, i.e. hypoconnectivity, in AD compared to elderly controls.

Each column shows oscillatory amplitude correlations comparing AD patients and matched elderly controls for each of the six frequency ranges where valid edges were found, and combined across frequency, and the Combined map.

The top row shows where there are ‘valid’ edge connections in either of the two groups. Colours, with a key at the top of the figure, indicate regions of brain areas that the connections originate from. The numbers on the plots show the number of valid edges found. The second row shows the unpaired t-statistic for AD compared to controls - Increases are displayed in red and decreases in blue and only those connections significant at p<0.05 (uncorrected) are shown. The numbers on each plot (n+/n-) show the number of significant increases and decreases found in each map. In terms of corrected significance only one connection each reached the p<0.05 threshold in Beta, High-Gamma and Combined (not shown). Using the cluster-level network-based size statistic, shown in the third row, significantly large clusters were found for the Delta, Alpha, Beta and Combined maps. The bottom row shows the results of the cohort random sampling analysis. Edges are only shown if they are within the 95% confidence limit for AD-related increases (red) or decreases (blue). Note that, in general, the predominant effect is for edges to show decreases in connectivity in AD patients, compared to controls, and that this is most clearly shown in the Combined map (right-hand column), in which the most valid connections are identified and the most significant, and consistent, differences are seen. Data and scripts needed to generate this figure can be found at http://dx.doi.org/10.17605/OSF.IO/E4CJX.

https://doi.org/10.7554/eLife.36011.005

Finally, when using the temporal standard-deviation of the oscillatory envelope to quantify activation at each AAL node, this cohort of patients did not show any significant differences, in any frequency band, compared to their age-matched controls.

Comparing the effects in young APOE-ɛ4 carriers to elderly patients with Alzheimer’s disease

Figure 4 shows a comparison of effects, taken from the most sensitive Combined connectivity maps, in order to directly visualise the overlap of statistical differences in the two experiments. In terms of the connections identified as valid (top row of Figure 1, Figure 3 and Figure 4a), there are clear similarities but also differences, principally in terms of the balance between parieto-occipital connectivity and parieto-temporal-frontal connections. This may reflect differences due to the large discrepancy in age between the two experimental cohorts. In terms of cohort effects, it can clearly be seen that in the APOE-ɛ4 comparison, young carriers mostly demonstrate hyperconnectivity in a right dominant network, whilst in the AD comparison, more extensive bilateral decreases in connectivity are seen. In Figure 4c we show this in more detail by displaying the unthresholded t-statistics for each comparison and then comparing the magnitude of effect-sizes for the two groups in the scatter plot. There is a clear negative correlation of effect sizes between the two experiments (r = −0.18, p<3.6×10−6), confirming the fact that, on average, connections that show hyperconnectivity in APOE-ɛ4 carriers demonstrate hypoconnectivity in Alzheimer’s cases. However, there is clearly not a perfect correspondence between the pattern of effects.

We can also perform a more conservative comparison of the effects in the two experiments by restricting our analysis to those connections, and AAL regions, that show a significant effect in both experiments. This is shown in Table 2. In the Alpha band there is a single occipital connection that has reduced connectivity in both APOE-ɛ4 carriers and AD. In the Beta band there is a single connection from the medial precuneus region to right lateral parietal that shows hyperconnectivity in APOE-ɛ4 carriers and hypoconnectivity in people with AD, compared to their respective controls. Similarly, in the Combined analysis, there are three connections in the right lateral-parietal/temporal region that show this same pattern of effects.

Table 2
Connections which show a statistically significant effect in both experiments.
https://doi.org/10.7554/eLife.36011.006
Frequency RangeAAL Region 1AAL Region 2E4 carriers versus non-carriers (T-statistic)AD versus elderly controls (T-statistic)
AlphaCuneus leftLingual left−2.5−1.8
BetaParietal superior rightPrecuneus left+2.3−2.3
CombinedOccipital mid rightTemporal inferior right+2.3−2.2
CombinedSupramarginal rightHeschl right+2.5−3.2
CombinedTemporal Superior rightHeschl right+3.0−3.0
Comparison of connectivity differences in the two experiments, assessed using the Combined connectivity maps.

(a) Replotting of the results for the Combined maps, extracted from Figure 1 and Figure 2 and displayed in the same format. At the bottom of this panel, a conjunction analysis is shown in which connections that meet the 95% confidence criterion for both experiments are displayed. Three connections survive this test and are shown in blue to reflect the fact that the effects have opposite sign (see Table 2 for more details). (b) A comparison of connections surviving a 95% confidence interval threshold, estimated using cohort resampling, for the APOE-ɛ4 versus young controls experiment (top row) and the AD versus elderly controls experiment (bottom row). Here we display the thresholded connections on both the circle plots (left column) and three views of a template brain (columns 2–4). (c) A direct comparison of effect sizes. In the left column, unthresholded t-statistics are displayed for the two experiments, with the opacity of the lines reflecting the magnitude of the t-statistic. In the middle, the two sets of t-statistics are multiplied together for display. In this panel, strong lines represent connections with high t-statistics for both the APOE-ɛ4 experiment and the AD experiment, with blue colours representing connections which show opposite effect directions for example hyperconnectivity in APOE-ɛ4 and hypoconnectivity in AD. In the right panel, the scatter plot shows the effect-size (i.e. unpaired t-statistic, thresholded at p<0.05 uncorrected) for the AD versus elderly control experiment, plotted against the effect-size for the APOE-ɛ4 carriers versus non-carriers experiment. A clear negative relationship is observed. Data and scripts needed to generate this figure can be found at http://dx.doi.org/10.17605/OSF.IO/E4CJX.

https://doi.org/10.7554/eLife.36011.007

Machine-learning results

Table 3 shows the cross-validated results of the support-vector-machine (SVM) classification tests. Using pooled classification, based on valid edges in all frequency ranges, the classifier was, on average, able to distinguish AD patients from elderly controls with reasonable specificity and sensitivity and a significant AUC of 77.6% (p<0.02). Performance classifying APOE-ɛ4 carriers from non-carriers was also significantly above the chance level of 50%, with an AUC of 63.5% (p<0.005). Interestingly, when we trained a classifier to distinguish APOE-ɛ4 carriers, using all data from the young cohort, and then predict AD cases from controls, performance was significantly below chance, with an AUC of 31.8%. If we assume that the classifier is predicting reverse labelling in the disease group, this was equivalent to an AUC of 68.2% that is above chance, but only a trend to significance (p<0.14). However, this apparent reversed labelling reinforces the principal findings of this paper that APOE-ɛ4 carrier status in young controls is principally revealed by hyperconnectivity, whilst Alzheimer’s disease leads to hypoconnectivity in a larger, but overlapping, set of connections.

Table 3
Leave-one-out Cross-validated performance of within experiment (top two rows) and cross-experiment (bottom row) SVM classification tests.
https://doi.org/10.7554/eLife.36011.008
ProblemPerformance using all valid features
Classify AD versus elderly controls eyes openSpecificity/Sensitivity = 76.7 % / 65.3%
AUC = 77.6%, p<0.019
Classify E4 versus nonE4
eyes open
Specificity/Sensitivity = 56.8 % / 58.6%
AUC = 63.5%, p<0.005
Train on E4 versus nonE4 then classify AD versus elderly controls eyes openSpecificity/Sensitivity = 36.4 % / 28.6%
AUC = 31.8% (68.2% using reversed labels)
p<0.14

Discussion

In a large sample of young healthy APOE-ɛ4 carriers, and using a conservative approach in which we restrict our analysis to highly reproducible (top 20%) connections, we found consistent increases in resting-state oscillatory activity and connectivity in a right-dominant network of parieto-occipital, parieto-temporal and parieto-frontal connections. The largest and most consistent cluster of effects arose in the beta-band, representing rhythmic synchronized activity thought to reflect global input at the synapse (Schnitzler and Gross, 2005) and in our analysis of Combined connectivity maps. These findings occurred in absence of cognitive or structural brain differences. We also showed that those connections showing the largest average increases in connectivity tended to overlap with those connections that showed decreased parieto-temporal connections in AD patients, and was most profound in connections to/from right-hemisphere parietal areas. Our results therefore suggest that the right parietal cortex is a particularly important area in the risk for, and development of, AD. Although our study can only show an implied association between early hyperconnectivity and later hypoconnectivity, we could speculate that an initial hyperconnectivity/hyperactivity centred around (right) parietal cortex has a cascade of effects that may lead to eventual disconnection in AD. Although initially a right-dominant effect, this may spread over time to more widespread effects across the cortex and hemispheres. A larger study, using the methods we present here, over a wider age-range, may shed further light on this.

Interestingly, when analysing activity differences in oscillatory envelopes, rather than connectivity, the only significant differences found were APOE-ɛ4 related hyperactivity in the gamma bands, both 40–60 Hz and 60–140 Hz. These gamma activity increases overlapped with those right-hemisphere brain regions showing hyperconnectivity in the lower bands (alpha/beta). This result is consistent with theories suggesting that gamma band oscillations mostly reflecting local, rather than global processing, whilst alpha/beta underpin longer-range interactions (Donner and Siegel, 2011). Our results are also consistent with a recent computational model in which it was shown that a realistically-coupled set of local gamma oscillators, at multiple simulated brain regions, led to long-range, slowly varying, coupled amplitude-envelopes, but in the alpha/beta range (Cabral et al., 2014). Finally, it has been suggested that higher-gamma band activity, such as our 60–140 Hz band, may be correlated with neuronal spiking rates (Ray and Maunsell, 2011). If true then the higher gamma-band activity we see in APOE-ɛ4 carriers may be a reflection of the neuronal hyperactivity, associated with increased amyloid-β, seen in rodent AD models (Stargardt et al., 2015).

The highly specific spatial patterns of increased resting-state oscillatory connectivity in young APOE-ɛ4 carriers, identifying the right parietal cortex as the most affected area, including inter-hemispheric connections and connections to the precuneus, complement and extend current neurophysiology findings. MEG studies using Graph Theory network measures have also identified a loss of connectivity in AD, with the parietal cortex as the most affected hub (de Haan et al., 2012b; Tijms et al., 2013). Effects were found specifically in the beta band – thought to play an important role in long-range functional connectivity (Schnitzler and Gross, 2005). Furthermore, previous EEG studies on APOE-ɛ4 have identified differences in power or connectivity in the alpha band, but these have only been investigated in older participants (Babiloni et al., 2006; Canuet et al., 2012; de Waal et al., 2013; Jelic et al., 1997; Kramer et al., 2008; Ponomareva et al., 2008). Using the higher spatial resolution of MEG, Cuesta et al. (Cuesta et al., 2015) studied source-localised connectivity in elderly participants with and without MCI, and found that APOE-ɛ4 carriers showed differences only in a much lower frequency band than alpha (delta, 1–4 Hz), focused around a frontal brain hub. Cuesta et al. measured phase-based connectivity, using the phase-locking value, a very different measure than the amplitude correlations we used, and more prone to source leakage artefacts (Colclough et al., 2016). Interestingly, in our previous study we found clear changes in oscillatory amplitude, but not phase correlations, in the brains of AD patients (Koelewijn et al., 2017), though we did not investigate MCI patients. For these reasons, and to reduce multiple comparisons, we did not investigate phase-based connectivity here.

To our knowledge, only one small MEG study (N = 8) has previously investigated young APOE-ɛ4 carriers, and found increases in mainly medial-frontal theta-band (3.5–7 Hz) oscillatory power during a working memory task (Filbey et al., 2006). This study did not investigate connectivity, nor analyse the alpha or beta band. Of note is that many previous neuroimaging studies used very low numbers of participants, with the resulting APOE-ɛ4 carrier groups being typically below N = 20, with some notable exceptions in elderly participants (Babiloni et al., 2006; de Waal et al., 2013; Sheline et al., 2010). Our more highly-powered MEG study in young participants therefore provides valuable, reliable, and unique insights into the early effects of APOE-ɛ4 on oscillatory brain function.

Functional MRI (fMRI) resting-state studies typically show that the blood-oxygen-level-dependent (BOLD) response within the default-mode network (DMN) is reduced in healthy ageing (Damoiseaux et al., 2008). This reduction is accelerated in AD (Buckner et al., 2008) and is also modified in elderly APOE-ɛ4 carriers compared to non-carriers (Sheline et al., 2010). Filippini et al. (Filippini et al., 2009a) previously investigated DMN BOLD activity in a small sample of young APOE-ɛ4 carriers versus non-carriers (N = 18 each) using an independent component analysis (ICA) technique. They found increased co-activation of areas within the DMN for ɛ4 carriers, which were focused on hippocampal and frontal, but posterior regions were not significantly different. Furthermore, they also found an increase in the sensorimotor network. Despite the different spatial focus of effects, these fMRI findings are similar to ours in that they reveal increased brain connectivity in young APOE-ɛ4 carriers. In a more recent resting state-fMRI study of older, but healthy, APOE-ɛ4 carriers it was found that, as well as replicating Fillipini et al.’s finding of increased hippocampal DMN connectivity, significant increases in network activity were also found for APOE-ɛ4 carriers in posterior regions of the DMN, including the precuneus, and lateral parietal regions as well as frontal cortex (Westlye et al., 2011). Importantly, these frontal and posterior-lateral regions were only significantly increased in the right hemisphere (their Figure 1D and Table 3), showing good concordance with our findings of oscillatory hyperconnectivity in similar right-hemisphere regions. Finally, we note that some recent task-based fMRI studies have also shown APOE-ɛ4 carrier hyperactivity in posterior parts of the DMN (Hodgetts et al., 2019; Shine et al., 2015). Interestingly, the authors of these studies speculate that hyperactivity/hyperconnectivity of the DMN in these young APOE-ɛ4 carriers may reflect reduced network efficiency/flexibility.

Note that in our cohort, and other previous studies, there is little or no evidence of cortical volume changes, including hypothesised hippocampal volume decreases (Filippini et al., 2009a; Heise et al., 2011; O'Dwyer et al., 2012; Wishart et al., 2006), suggesting that changes in both haemodynamic and oscillatory brain function can be detected before changes in brain structure in the pathway of increased risk for AD (Figure 5, lower panel, (Hampel et al., 2011)).

Schematic representation of a speculated mechanism (top tow) and order of events (bottom row) in brain structure and function when a healthy young individual develops Alzheimer’s disease.

Our results are consistent with theories that APOE-ɛ4 status may lead to early neuronal hyperactivity/hyperconnectivity, either directly via modification of the excitation/inhibition balance (Nuriel et al., 2017) or linked with amyloid deposition (Stargardt et al., 2015). Most biomarkers of disease show a progressive incline or decline, whereas our results suggest that functional connectivity shows a profile of an early increase, before structural, cognitive and neurobiological markers are evident. Eventually a more profound connectivity decrease is seen after a clinical diagnosis of Alzheimer’s disease has been established. Inspired by Hampel et al. (2011), Figure 4.

https://doi.org/10.7554/eLife.36011.009

The results presented here are, of course, only correlational in nature but we can speculate, in the presence of prior in-vitro and in-vivo animal studies, what underlying neurophysiological changes underpin our hyperactivity and hyperconnectivity findings. It is known that increased amyloid-β is somehow related to increased neural activity, although the sequence of cause and effect is not known (Stargardt et al., 2015). For example, it was shown in a rodent disease model that during Braak pre-symptomatic stages 0, I and II, amyloid-β starts to accumulate within neurons, whilst at the same time synaptic activity rises within the neurons (i.e. hyperactivity). As the disease progresses to Braak stage VI, and amyloid-β becomes significant enough to cause plaques, synaptic activity then falls (i.e. hypoactivity). Interestingly, it appeared that the rise in amyloid-β appeared to be linked to a reduction of activity in the Insulin-Degrading Enzyme (IDE).

However, what it is not clear is how APOE-ɛ4 carrier status might lead to an increased risk of the hyperactivity <>amyloid-β relationship described above in young or even middle-aged people. Experimental studies have demonstrated that carrying the APOE-ɛ4 allele leads to a wide range of effects in the brain, including impaired myelination, mitochondrial dysfunction, reduced cholinergic function and vascular pathology (Reinvang et al., 2013). Healthy elderly APOE-ɛ4 carriers also show increased uptake of amyloid-binding PET ligands (PiB) (Morris et al., 2010; Reiman et al., 2009).

Already in middle age, post-mortem studies of APOE-ɛ4 carriers have shown an increase in senile amyloid-β plaques (Kok et al., 2009), with some evidence of this even at ages as young as 30. It is therefore reasonable for us to speculate that in our young cohort, although amyloid-β plaques may not have yet developed, amyloid-β may be accumulating and, as in the rodent model, be associated with neural hyperactivity (Figure 5, upper panel). The mechanism for how APOE-ɛ4 carrier status leads to increased amyloid-β deposition is complex and may involve multiple mechanisms, including disruptions of amyloid-β clearance systems dependent on receptor-mediated, cerebrovascular or proteolytic degradation (Kanekiyo et al., 2014; Zhao et al., 2018). For example, one study has shown that APOE-ɛ4 leads to reduced levels of insulin-degrading enzyme (IDE), one of the most important enzymes for amyloid-β clearance (Du et al., 2009) and, as described above, is reduced during pre-symptomatic stages in rodent disease models which also show synaptic hyperactivity (Stargardt et al., 2015).

It has also been shown in a recent study, however, that neural hyperactivity can also occur in APOE-ɛ4 mouse models in the absence of overt amyloid or tau pathology (Nuriel et al., 2017) via a mechanism in which APOE-ɛ4 leads to an overall decrease in inhibition caused by a reduced responsiveness of excitatory neurons to GABAergic input. The authors therefore suggest that it is this APOE-ɛ4 related shift in the excitation/inhibition balance that leads to hyperactivity that later results in increased amyloid deposition, and hence increased risk of AD.

It is important to note, though, that increased amyloid-β is not sufficient to cause AD (Herrup, 2015; Sheline et al., 2010) and the presence of APOE-ɛ4 does not guarantee progression to AD. It remains to be investigated what role other important factors for AD play in this cascade of events, such as the tau protein (Ossenkoppele et al., 2016). Interestingly, PET studies of AD have also demonstrated AD-related hypometabolism in bilateral parieto-temporal regions that are very similar to the regions we report here (Kato et al., 2016).

Our findings therefore support theories based on rodent models (Stargardt et al., 2015), human computational models (de Haan et al., 2012b) and MCI data (Maestú et al., 2015) that propose that hyperactivity, ultimately related to hyperconnectivity, may be at the heart of brain dysfunction that could eventually lead to AD, and provides compelling evidence that this may already happen in early adulthood. Areas predisposed to hyperconnectivity are thought to be prone to yield increased accumulation of amyloid-β (Myers et al., 2014), supported by longitudinal PET and fMRI studies which found that hippocampal hyperactivation in elderly individuals leads to widespread increased deposition of amyloid-β as well as hippocampal signal loss and memory decline, although these studies found no effect of APOE-ɛ4 (O'Brien et al., 2010) or there was insufficient power to test this (Leal et al., 2017). The present study suggests that the parietal cortex is an important common early focus for hyperconnectivity, possibly as part of the posterior DMN. Importantly, this hyperconnectivity is present during the task-free resting-state, both avoiding and highlighting the potential confound of baseline-level differences in task-based group studies.

One important methodological note, for future work, is that our exploratory analysis, in which we combined the connectivity maps from different frequency bands using a vector sum approach, revealed both more robust connectivity patterns across participants, with increased number of valid edge connections, and increased sensitivity to both APOE-ɛ4 carrier effects and Alzheimer disease-related changes in connectivity. This simple combination of effects may therefore be worth adding to future studies in order to, potentially, provide increased sensitivity to disease-related modulations.

In addition to classical statistics we used machine-learning classification approaches to investigate how much discriminative information is present within these static oscillatory connectivity matrices. Here, the support-vector-machine (SVM) was trained used a pooled set of features taken from all valid connectivity edges from all frequency bands and the combined map. The SVM will, as part of its decision boundary optimisation, choose the best weighting of all these features. When tested using leave-one-out cross validation, the classifier was able to decode APOE-ɛ4 carriers from controls and AD patients from elderly controls, with significantly better than chance performance. In our small AD cohort, prediction performance appeared close to 70%, however, given the low number of participants in this study, this needs to be evaluated using a larger independent sample. Although our results here must be considered preliminary, they are encouraging – particularly as the AD patients in our study were relatively early in their diagnosis.

In the current study, we employed MEG to show patterns of oscillatory hyperactivity and hyperconnectivity in neural activity in young-adult APOE-ɛ4 carriers, identifying the right parietal cortex as an important centre in these hyperconnectivity patterns. One particularly attractive feature of MEG is that it provides non-invasive high-temporal and spatial-resolution measures of neurophysiology that allow for comparison with cellular hyperactivity data in a translational context (Busche et al., 2008), providing potential biomarkers for new therapeutic and preventative studies in the pathological processes leading to AD. Finally, an important question is why and at what stage hyperconnectivity ‘collapses’ to convert to the profound disconnection and symptomology of AD, and what further factors determine whether ɛ4 carriers indeed go on to develop AD. Future studies focusing on the often-neglected middle-age group, between the very young and the elderly, as well as longitudinal studies will be invaluable to unravel the time course of these events.

Limitations of the study

There are several methodological limitations of this study that need to be acknowledged. Firstly, all of the analyses we present here are associations or correlations than can only imply a link between APOE-ε4 status, hyperconnectivity and future risk of developing Alzheimer’s disease – no causal relationships can be proved in this type of study. Similarly, the partial spatial overlap between the hyperconnectivity we observe in young APOE-ε4 carriers, compared to their control group, and hypoconnectivity shown in elderly AD patients, compared to their control group, is interesting but again does not demonstrate a causal link or a direct trajectory. Longitudinal studies that use our methods in individuals as they age, perhaps in older people with and without Mild Cognitive Impairment (MCI), to see if conversion to AD can be predicted would be an exciting extension of our work.

Although our young APOE-ε4 sample size is larger (51 versus 108) than much of the previous work in this field, it is still relatively small and our findings need replicating in an independent cohort. Such cohorts are starting to become available as online resources (e.g. CamCAN, http://www.cam-can.org/index.php?content=dataset) and it would be relatively straightforward to replicate our specific analysis pipeline, including our edge-thresholding procedure, to confirm the findings we report here. Second, as this is an exploratory study, we did not perform an exhaustive test of all connectivity metrics we could have employed, such as phase-based measures and dynamic connectivity analyses. We took the simplest possible measure – static oscillatory amplitude-amplitude envelope correlation as this has previously demonstrated the best within-participant reproducibility (Colclough et al., 2016) but this does not guarantee that such robust and repeatable measures are the most sensitive to cohort differences. It is entirely possible that other measures, particularly time-resolved dynamic connectivity analyses (Vidaurre et al., 2016), could be more sensitive and informative than the static amplitude-amplitude measure we used here.

Similarly, we took a conscious decision to only analyse connections that were assessed to be the strongest and thus, the most repeatable, across participants in our cohorts. In a way, this is a form of quality-control that rejects features (here connectivity edges) that are of low SNR and hence are poorly estimated in individuals and we note that such a procedure is common in the application of graph-theory methods for the analysis of network properties. It is entirely possible that we have chosen a procedure, and in particular a threshold (mean rank of 80%), that is sub-optimal for revealing cohort differences. This is particularly relevant for MEG in which SNR is region-dependent and so we may have biased our analysis here to be more sensitive to superficial cortical regions (such as the parietal lobes) whilst suppressing an analysis of poor SNR regions, such as deeper regions in the brain. Future work could tune this parameter (as well as varying other parameters in our analysis pipeline) to try and maximise detection and discriminability of cohort effects across the brain. However, it is worth noting that MEG source-localisation algorithms, including the beamformer we use here, have non-uniform spatial resolution that means that regions with poor SNR, including deeper regions, are poorly resolved and hence activity is spread over wider regions. By restricting our statistical analyses to regions with higher SNR, we are protecting ourselves, somewhat, from localisation errors that could result from this spatial spreading.

In common with previous studies, we used a procedure in which, after trial-rejection and beamformer source localisation, trials were concatenated to form a continuous virtual-sensor timeseries for subsequent analysis. One potential problem with this is that it introduces discontinuities in the timeseries that could lead to sub-optimal estimations of connectivity. The precise impact of this is difficult to predict as it depends on the exact temporal distribution of rejected trials. However, it is likely to have the biggest impact on the very lowest frequencies, such as Delta, which only has relatively few oscillatory cycles in each trial. In our experiment there was no significant difference in the number of rejected trials between the groups (e.g. APOE-ɛ4 carriers and their controls) but this concatenation procedure may have reduced our sensitivity to effects in the Delta and Theta ranges, which showed relatively few robust connections compared to Alpha and Beta.

We chose to use a common anatomical parcellation, the AAL (Tzourio-Mazoyer et al., 2002) atlas, which ourselves and other groups (Hillebrand et al., 2016; Hunt et al., 2016; Koelewijn et al., 2017; Pang et al., 2016; Routley et al., 2017) have used in multiple MEG connectivity analysis papers, including our recent paper on oscillatory dysconnectivity in AD (Koelewijn et al., 2017). We therefore chose, a-priori, a commonly used atlas for our spatial reduction strategy. However, this atlas has 90 regions of largely varying size and shape and has been defined without regard to the sensitivity profile of our MEG system to different regions of the brain. We are therefore summarising activity, using a single virtual sensor, in what are large extended regions of widely varying SNR. This will bias our results to detecting connectivity in certain regions and may well be sub-optimal for detecting interactions between regions in which a finer, or coarser, parcellation may well have been beneficial. The principled way forward here might be to derive parcellations that use information regarding the MEG lead-field and the resulting resolution kernels in source-space (Farahibozorg et al., 2018).

A similar criticism might also be levelled at our source-leakage correction procedure in which we have used multivariate orthogonalisation (Colclough et al., 2015) simultaneously across all 90 timecourses to minimise all cross parcel correlations in the raw timeseries. This procedure may well be most successful in conditioning, and hence detecting, high-SNR signals which will tend to be more superficial signals with higher oscillatory power, such as the lateral parietal regions we describe in this paper. This differential sensitivity, which stems fundamentally from the relative lack of depth-sensitivity in MEG recordings but may be worsened by these analysis procedure choices, may well explain why our findings do not fully match those found in previous fMRI studies (Filippini et al., 2009a; Westlye et al., 2011), at least for deep and medial regions.

Materials and methods

In this paper we present the results of analysing two MEG resting-state experiments: 1) A study of young healthy participants who have been genotyped for APOE-ɛ4 and 2) A study of Alzheimer’s disease (AD) patients compared to a matched cohort of elderly controls.

Participants in experiment 1: young APOE-ε4 cohort

Request a detailed protocol

The APOE analysis was performed on resting-state MEG data acquired at CUBRIC as part of the ‘100 Brains’ and ‘UK MEG Partnership’ projects, which together yielded a total of 183 usable datasets. Inclusion criteria were aged 18–65, completed or be undertaking a degree, normal or corrected-to-normal vision, and no history of neurological or neuropsychiatric disorder. Of the 183 participants, 51 participants were identified as heterozygote or homozygote APOE-ɛ4 carriers, and 108 as not carrying an ɛ4 allele (demographic details in Table 1). Genotyping was unsuccessful in the remaining 24 participants, who were excluded from analysis. All participants gave informed consent according to the Declaration of Helsinki and all procedures were approved by the Ethics Committee of the School of Psychology, Cardiff University (EC.12.01.10.3071).

APOE genotyping in Experiment 1

Request a detailed protocol

Genomic DNA was obtained from saliva using Oragene OG-500 saliva kits (DNA Genotek, Inc., Ontario, Canada). Genotyping was performed using custom Illumina HumanCoreExome-24 BeadChip genotyping arrays, which contain approximately 500,000 common genetic variants (Illumina, Inc., San Diego, CA). Quality control and imputation were implemented in plink 1.9 (Chang et al., 2015). Individuals were excluded for any of the following reasons: 1) ambiguous sex (where samples with undetermined X chromosome heterozygosity were excluded); 2) cryptic relatedness up to third-degree relatives as ascertained using identity by descent; 3) genotyping completeness less than 98%; 4) non-European ethnicity admixture which was determined via population stratification, where samples that clustered outside the CEU HapMap3 population using principal component analysis were excluded; and 5) outliers from an autosomal heterozygosity filter. Single nucleotide polymorphisms (SNPs) were excluded where the minor allele frequency was less than 1%, if the call rate was less than 98%, or if the χ2 test for Hardy-Weinberg equilibrium had a p value less than 1e-6. A total of 233,054 genotyped SNPs remained after quality control. Autosomal chromosomes were imputed using the reference panel HRCv1.1 (hrc.r1.1.2016) using a mixed population panel (McCarthy et al., 2016). Phasing was completed using Eagle v2.3 (Loh et al., 2016) and imputation was performed using Mimimac3 (Das et al., 2016). Imputed data was converted to best-guess genotypes (Chang et al., 2015) with multi-allelic sites removed. SNP filters for HWE (1e-6) and minor allele frequency (1%) were re-applied. A total of 7,545,595 SNPs were successfully imputed. Variants rs7412 and rs429358 were used to estimate APOE genotype for all 183 individuals (Corneveaux et al., 2010).

Experiment 1 young APOE-ε4 cohort: data acquisition and analysis

Data acquisition

Request a detailed protocol

Five-minute whole-head MEG recordings were acquired at a 1200 Hz sample rate using a 275-channel CTF radial gradiometer system. An additional 29 reference channels were recorded for noise cancellation purposes and the primary sensors were analysed as synthetic third-order gradiometers (Vrba and Robinson, 2001). Subjects were seated upright in the magnetically shielded room with their head supported with a chin rest to minimise movement. Participants were asked to rest and fixate their eyes on a central red fixation point, presented on either a CRT monitor or LCD projector. Horizontal and vertical electro-oculograms (EOG) were recorded to monitor eye blinks and eye movements.

Participants also underwent a magnetic resonance imaging (MRI) session in which a T1-weighted 1 mm anatomical scan was acquired, using an inversion recovery spoiled gradient echo acquisition (3T, General Electric).

Participants further performed the MATRICS Consensus Cognitive Battery (MCCB) (Kern et al., 2008; Nuechterlein et al., 2008). Group performance was compared using two-sample t-tests, excluding missing or incomplete data (leaving per task N = 47–48 for ɛ4 carriers and N = 98–100 for non-carriers, see Table 1).

MEG data analysis

Request a detailed protocol

MEG generates multi-dimensional data, which can be analysed in a large variety of ways. We sought to conduct an analysis investigating only consistent functional connectivity across the brain, while reducing noise. We focused on amplitude-amplitude connectivity of beamformer-derived oscillatory source signals, one of the most robust and repeatable types of MEG connectivity measures (Colclough et al., 2016), across six frequency bands and 90 AAL atlas brain areas. The analysis pipeline also allowed a voxelwise assessment of resting-state activity within each frequency band and is displayed in Figure 6.

Schematic overview of the analysis pipeline.

Note that in the labelled circular plot (bottom right) the nodes are in approximate anatomical locations as if the brain was viewed from above (left is left/right is right) and are coloured according to their originating lobes (Blue: frontal, Gold: Insula/Anterior-Posterior cingulate, Purple: Medial temporal, Green: Temporal: Teal: Parietal and Sensorimotor, Maroon: occipital, Lilac: Precuneus. When statistical effects are plotted, red represents cohort increases and blue cohort decreases.

https://doi.org/10.7554/eLife.36011.010

All datasets were down-sampled to 600 Hz, and filtered with a 1 Hz high-pass and a 150 Hz low-pass filter. The datasets were then segmented into 2 s epochs. The data were visually inspected and epochs with major artefacts such as head movements, large muscle contractions, or ocular artefacts were excluded from subsequent analysis.

To achieve MRI/MEG co-registration, fiduciary markers were placed at fixed distances from three anatomical landmarks identifiable in the subject’s anatomical MR scan, and their locations were manually marked in the MR image. Head localization was performed at the start and end of each recording.

The MEG sensor data were source-localised using FieldTrip (RRID:SCR_004849) version 20161011 (Oostenveld et al., 2011), with an LCMV beamformer on a 6 mm grid, using a single-shell forward model (Nolte, 2003), where the covariance matrix was constructed per each of the following frequency bands: 1–4, 3–8, 8–13, 13–30, 40–60, and 60–140 Hz. For each band, the beamformer weights were normalized using a vector norm (Hillebrand et al., 2012), data were normalized to the MNI template, and reduced to 90 nodes following the Automatic Anatomical Labelling (AAL) atlas (Tzourio-Mazoyer et al., 2002). The selection of one node per AAL region was achieved by performing a frequency analysis on all virtual channels within the AAL region, and selecting the virtual channel with the greatest temporal standard-deviation within the region. Epochs were concatenated to generate a continuous virtual-sensor timecourse for each voxel and then band-passed into the relevant frequency-band of interest (Delta: 1–4 Hz, Theta: 3–8 Hz, Alpha: 8–13 Hz, Beta-13–30 Hz, Low Gamma: 40–60 Hz and High-Gamma: 60–140 Hz).

The resulting 90-node time series were orthogonalized to avoid spurious correlations using a multivariate regression approach known as symmetric orthogonalization (Colclough et al., 2015). The data then underwent a Hilbert transform to obtain the oscillatory amplitude envelope, and were subsequently despiked to remove artefactual temporal transients using a median filter, downsampled to 1 Hz and trimmed to avoid edge effects (removing the first 2 and last three samples). Amplitude correlations were then calculated by correlating the 90 downsampled Hilbert envelopes to each other, and converted to variance-normalized Z-scores by applying a Fisher transform. In addition to analysing effects within each frequency band separately, we also synthesised a combined measure of connectivity by calculating the vector-sum of all connectivity matrices, using the following formula for each element (i,j) in the square connectivity matrix:

Combinedij=sqrt(Deltaij2+Thetaij2+Alphaij2+Betaij2+LowGammaij2+HighGammaij2)

Assessment of activity differences

Request a detailed protocol

In order to assess possible cohort differences in oscillatory activity, rather than connectivity, we also calculated a measure of temporal variance at each beamformer-reconstructed voxel in the brain, for each separate frequency band, via an assessment of the amplitude envelopes. Due to large variations in signal sensitivity throughout the brain, beamformer reconstructions of virtual-sensors result in signals in which the mean and temporal standard-deviation of the amplitude envelopes are highly correlated and show large variations across the source space. For that reason, we choose to use a normalized measure, the coefficient of variation, which is the simply the standard-deviation of the amplitude envelope divided by the mean of the amplitude envelope. Using this measure, voxels which show a high fractional temporal variability, compared to the mean, will have a high value, whilst those voxels that are of more stable oscillatory power will have a low value. We note that the measure is similar, by analogy, to common measures employed in BOLD fMRI, such as percentage-change from baseline. The statistical significance of group-effects was assessed using univariate voxelwise unpaired t-tests of this activity measure and then correcting for multiple comparisons using randomisation testing using FSL randomise (Nichols and Holmes, 2003; Singh et al., 2003; Winkler et al., 2014). The latter procedure was performed using threshold-free cluster enhancement (TFCE) and 5000 iterations, without variance smoothing. This procedure tests positive and negative cohort contrasts separately so a null hypothesis corrected p-value of 0.025 was chosen, so as to implement a corrected two-tailed test at p<0.05. Separately for each frequency band, thresholded t-statistics are displayed on a template brain surface mesh, extracted using FreeSurfer.

Statistical analysis of group differences in connectivity

Request a detailed protocol

First, each participant’s connectivity matrix was z-scored to have zero mean and unit variance across all connections, in order to correct for possible global differences, due to variability in data quality, either across participants or cohorts (Siems et al., 2016). All subsequent analyses are therefore assessments, at each connection edge, of the relative strength of connectivity, compared to the mean, in each participant. This procedure has the advantage of ensuring that any systematic differences, either at the participant or cohort-level, are controlled for.

To avoid analysing noise, we then independently assessed which functional connections were consistently present across all 159 datasets, for each frequency band and in the Combined maps. This was done by calculating the rank of every connection in each participant’s connectivity map and then averaging this map rank across all participants. The resulting average rank-map was thresholded at 80%, such that we only consider as ‘valid’ these connectivity edges that are consistently in the top 20%. Importantly, to ensure that large cohort differences are not missed, we performed this assessment of valid connections separately for each cohort, and labelled connections as valid if they passed our rank-mask threshold in either cohort. The remaining connections were discarded.

We statistically tested the difference between groups using three different statistical approaches. In the first, we used unpaired t-tests of the corrected Z-scores, at each of the ‘valid’ connectivity edges, and looked for significant edges at both p<0.05 (uncorrected) and at p<0.05 (corrected using a 5000-permutation test with omnibus thresholding).

Secondly, we tested for significant differences in connected network cluster size using methods similar to those described in the NBS toolbox (Zalesky et al., 2010). Here, connection differences are first thresholded at p<0.05 (uncorrected) before Matlab’s (RRID:SCR_001622) graph function is used to convert this difference map to a binarized graph. Matlab’s conncomp function is then used to find the largest connected cluster within this graph. The corrected significance of this largest cluster is assessed using randomisation testing with 5000 iterations and thresholded at p<0.05 (corrected). This network-based statistic has previously been shown to be more sensitive to cohort effects than the simple mass-univariate testing we use in our first approach (Zalesky et al., 2010). As previously described, we test positive and negative cluster sizes separately.

Thirdly, we tested the likely generalisation of our findings by assessing the confidence interval on cohort-related increases and decreases in connectivity. This was done by randomly sampling half the APOE-ɛ4 carriers (i.e. N = 26) and half the non- APOE-ɛ4 carriers (i.e. N = 54) and for each of the valid edges in the map we tabulate whether the mean edge connectivity is increased or decreased in the APOE-ɛ4 group. This sampling strategy is repeated 5000 times and those edges that do not have a consistent sign – either increases or decreases – on more than 95% of samples are discarded. The end result is a set of connections that are robustly increased, or decreased, in the APOE-ɛ4 carriers compared to non-carriers, irrespective of the specific individuals entered into the test.

This is an exploratory study that used previously collected data to search for differences related to APOE-ɛ4 carrier status. To our knowledge there has not been another published MEG study in young volunteers but we can estimate statistical power based on other studies. Power was estimated using the pwr function in R (Champely et al., 2018). The area under the curve (AUC) for APOE-ɛ4 status, calculated in a recent Alzheimer’s disease case-control study (Escott-Price et al., 2015) was transformed to Cohen’s d (Ruscio, 2008), yielding a moderate effect size (d = 0.653). For this effect size, for our sample of 51 carriers and 108 non-carriers, we have approximately 87% power to detect an effect of APOE-ɛ4 status on our measures.

All network plots and brain connectivity diagrams were generated using custom in-house scripts in Matlab (Shaw, 2019).

Supervised Machine-Learning analysis

Request a detailed protocol

An alternative way of investigating whether the static oscillatory connectivity matrices contain information that distinguishes cases, compared to controls, is to attempt to train a classifier that can distinguish between the two groups. Here we used a support vector machine (SVM) (Boser et al., 1992) analysis in which we took all valid edges, pooled across all estimated maps for each person (Delta, Theta, Alpha, Beta, Low Gamma, High Gamma, Combined) and used these as features for the SVM. Note that, in order to prevent biased performance, features were selected by taking the top 20% of edges, estimated over the whole cohort so that features were not selected separately from each group. In machine-learning it is important to use cross-validation to prevent over-fitting (as we have more features than participants). Ideally this would be done by training on one dataset and testing on another, but here we only have one dataset. We therefore used a Leave-One-Out (LOO) sampling approach in which we randomly select one case (i.e. an APOE-ɛ4 carrier) and one control (i.e. a non-carrier) as our test sample and hence train the classifier using all other participants. We then test how well this classifier performs on the two test participants we held back. This procedure is repeated 5000 times to build up an estimate of how well the classifier can successfully categorise each participant. We then assess the accuracy of the classifier by pooling these out-of-training estimates to obtain mean measures of Sensitivity (i.e. the true-positive rate), Specificity (true-negative rate) and the area under the receiver operating characteristic curve (AUC of the ROC). For this binary classification the chance value for all of three measures is 0.5. One important note is that, for the APOE-ɛ4 experiment, we have a significantly unbalanced classification problem in that there are twice as many non-carriers as carriers. This is important because the classifier could achieve a performance of 66% by simply labelling all participants as non-carriers. We guard against this by taking a conservative approach of balancing the classification, at each iteration, by only using a random sample of 50 non-carriers and 50-carriers to train the classifier and, in addition, reporting the three evaluation metrics we describe above. Finally, the performance of all classification analyses is ultimately assessed using the AUC. To assess the statistical significance compared to chance, we use random permutation (50,000 iterations) of the true labels to build an empirically estimated null distribution of the AUC and then tested the achieved AUC against this.

Cortical volume analysis

Request a detailed protocol

To investigate whether brain structure was affected in APOE-ɛ4 carriers, we performed cortical segmentation using FreeSurfer v6.0 (RRID:SCR_001847, surfer.nmr.mgh.harvard.edu). We calculated hippocampal volumes per hemisphere and total intracranial volume (ICV). Relative hippocampal volumes were calculated by dividing hippocampal volumes by ICV. Following quality control, performed using a protocol from ENIGMA (http://enigma.ini.usc.edu/) (Stein et al., 2012), volumes were excluded for 1 ɛ4 carrier and two non-carriers. Using the same data and analysis procedure, cortical thickness, volume and area were also assessed within each of the subdivisions of the Desikan-Killiany atlas (Desikan et al., 2006) included within FreeSurfer.

Alzheimer’s disease and elderly control cohort: MEG acquisition and analysis

Request a detailed protocol

Recruitment, acquisition and pre-processing of the AD patients (N = 14, 5F/9M, mean age 77.7 years, Mini-Mental State Examination (MMSE) score 19–27, all fulfilled criteria for probable AD [McKhann et al., 1984]) and age-matched controls (N = 11, 7F/4M, mean age 74.4 years, MMSE score 28–30) has been described in detail previously (Koelewijn et al., 2017). In summary though, for inclusion as probable AD, criteria were that dementia had been established by clinical and cognitive assessment (ACE-R), that cognitive impairments were progressive and included episodic memory deficit and deficit in at least one other area of cognition, and there was no evidence of other factors that could explain the dementia syndrome such as cerebrovascular disease identified from a previous CT scan. The study was approved by the South-East Wales Research Ethics Committee (10/WSE04/35), and all participants provided informed consent. The procedure was identical to the APOE analysis, except that participants were asked to rest while opening and closing their eyes following an auditory verbal cue (24 × 15 s alternations in total). Each dataset was split into one eyes-open and one eyes-closed dataset, following which the statistical analysis of AD patients, compared to the matched elderly control sample, proceeded in exactly the same way as that described above for the APOE-ɛ4 study in young controls. Here, we report comparison results only for the eyes-open resting data. This was done in order to compare the pattern of effects revealed by comparing young APOE-ɛ4 carriers to non-carriers with those obtained by comparing AD participants to an age-matched sample of elderly controls in the same resting-state paradigm. Note that we have previously reported results from this AD-control comparison for eyes-open resting state using a different analysis pipeline (Koelewijn et al., 2017) but, in that dataset, 14/16 AD patients and 11/21 controls viewed a visual grating instead of a simple blank screen. In this analysis, we decided to only include those participants who saw a visual grating during the eyes-open period, in order to make sure that patient-related effects were not due to presence, or not, of the visual grating as this was not matched across the cohorts. This means we have a reduced subset of those analysed previously that is N = 14 AD patients and N = 11 elderly controls who all performed an identical version of the resting-state paradigm.

References

  1. 1
  2. 2
  3. 3
  4. 4
    A training algorithm for optimal margin classifiers
    1. BE Boser
    2. IM Guyon
    3. VN Vapnik
    (1992)
    In Proceedings of the Fifth Annual Workshop on Computational Learning Theory, 10.1145/130385.130401.
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
  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
  39. 39
  40. 40
  41. 41
  42. 42
  43. 43
  44. 44
  45. 45
  46. 46
  47. 47
  48. 48
  49. 49
  50. 50
  51. 51
  52. 52
    Nonparametric permutation tests for functional neuroimaging
    1. T Nichols
    2. A Holmes
    (2003)
    Human Brain Function 15:1–25.
  53. 53
  54. 54
  55. 55
  56. 56
  57. 57
  58. 58
  59. 59
  60. 60
  61. 61
  62. 62
  63. 63
  64. 64
  65. 65
  66. 66
  67. 67
  68. 68
  69. 69
    SourceMesh Features. Github, version 08aa50a
    1. A Shaw
    (2019)
    SourceMesh Features. Github.
  70. 70
  71. 71
  72. 72
  73. 73
  74. 74
  75. 75
    Identification of common variants associated with human hippocampal and intracranial volumes
    1. JL Stein
    2. SE Medland
    3. AA Vasquez
    4. DP Hibar
    5. RE Senstad
    6. AM Winkler
    7. R Toro
    8. K Appel
    9. R Bartecek
    10. Ø Bergmann
    11. M Bernard
    12. AA Brown
    13. DM Cannon
    14. MM Chakravarty
    15. A Christoforou
    16. M Domin
    17. O Grimm
    18. M Hollinshead
    19. AJ Holmes
    20. G Homuth
    21. JJ Hottenga
    22. C Langan
    23. LM Lopez
    24. NK Hansell
    25. KS Hwang
    26. S Kim
    27. G Laje
    28. PH Lee
    29. X Liu
    30. E Loth
    31. A Lourdusamy
    32. M Mattingsdal
    33. S Mohnke
    34. SM Maniega
    35. K Nho
    36. AC Nugent
    37. C O'Brien
    38. M Papmeyer
    39. B Pütz
    40. A Ramasamy
    41. J Rasmussen
    42. M Rijpkema
    43. SL Risacher
    44. JC Roddey
    45. EJ Rose
    46. M Ryten
    47. L Shen
    48. E Sprooten
    49. E Strengman
    50. A Teumer
    51. D Trabzuni
    52. J Turner
    53. K van Eijk
    54. TG van Erp
    55. MJ van Tol
    56. K Wittfeld
    57. C Wolf
    58. S Woudstra
    59. A Aleman
    60. S Alhusaini
    61. L Almasy
    62. EB Binder
    63. DG Brohawn
    64. RM Cantor
    65. MA Carless
    66. A Corvin
    67. M Czisch
    68. JE Curran
    69. G Davies
    70. MA de Almeida
    71. N Delanty
    72. C Depondt
    73. R Duggirala
    74. TD Dyer
    75. S Erk
    76. J Fagerness
    77. PT Fox
    78. NB Freimer
    79. M Gill
    80. HH Göring
    81. DJ Hagler
    82. D Hoehn
    83. F Holsboer
    84. M Hoogman
    85. N Hosten
    86. N Jahanshad
    87. MP Johnson
    88. D Kasperaviciute
    89. JW Kent
    90. P Kochunov
    91. JL Lancaster
    92. SM Lawrie
    93. DC Liewald
    94. R Mandl
    95. M Matarin
    96. M Mattheisen
    97. E Meisenzahl
    98. I Melle
    99. EK Moses
    100. TW Mühleisen
    101. M Nauck
    102. MM Nöthen
    103. RL Olvera
    104. M Pandolfo
    105. GB Pike
    106. R Puls
    107. I Reinvang
    108. ME Rentería
    109. M Rietschel
    110. JL Roffman
    111. NA Royle
    112. D Rujescu
    113. J Savitz
    114. HG Schnack
    115. K Schnell
    116. N Seiferth
    117. C Smith
    118. VM Steen
    119. MC Valdés Hernández
    120. M Van den Heuvel
    121. NJ van der Wee
    122. NE Van Haren
    123. JA Veltman
    124. H Völzke
    125. R Walker
    126. LT Westlye
    127. CD Whelan
    128. I Agartz
    129. DI Boomsma
    130. GL Cavalleri
    131. AM Dale
    132. S Djurovic
    133. WC Drevets
    134. P Hagoort
    135. J Hall
    136. A Heinz
    137. CR Jack
    138. TM Foroud
    139. S Le Hellard
    140. F Macciardi
    141. GW Montgomery
    142. JB Poline
    143. DJ Porteous
    144. SM Sisodiya
    145. JM Starr
    146. J Sussmann
    147. AW Toga
    148. DJ Veltman
    149. H Walter
    150. MW Weiner
    151. JC Bis
    152. MA Ikram
    153. AV Smith
    154. V Gudnason
    155. C Tzourio
    156. MW Vernooij
    157. LJ Launer
    158. C DeCarli
    159. S Seshadri
    160. OA Andreassen
    161. LG Apostolova
    162. ME Bastin
    163. J Blangero
    164. HG Brunner
    165. RL Buckner
    166. S Cichon
    167. G Coppola
    168. GI de Zubicaray
    169. IJ Deary
    170. G Donohoe
    171. EJ de Geus
    172. T Espeseth
    173. G Fernández
    174. DC Glahn
    175. HJ Grabe
    176. J Hardy
    177. HE Hulshoff Pol
    178. M Jenkinson
    179. RS Kahn
    180. C McDonald
    181. AM McIntosh
    182. FJ McMahon
    183. KL McMahon
    184. A Meyer-Lindenberg
    185. DW Morris
    186. B Müller-Myhsok
    187. TE Nichols
    188. RA Ophoff
    189. T Paus
    190. Z Pausova
    191. BW Penninx
    192. SG Potkin
    193. PG Sämann
    194. AJ Saykin
    195. G Schumann
    196. JW Smoller
    197. JM Wardlaw
    198. ME Weale
    199. NG Martin
    200. B Franke
    201. MJ Wright
    202. PM Thompson
    203. Alzheimer's Disease Neuroimaging Initiative
    204. EPIGEN Consortium
    205. IMAGEN Consortium
    206. Saguenay Youth Study Group
    207. Cohorts for Heart and Aging Research in Genomic Epidemiology Consortium
    208. Enhancing Neuro Imaging Genetics through Meta-Analysis Consortium
    (2012)
    Nature Genetics 44:552–561.
    https://doi.org/10.1038/ng.2250
  76. 76
  77. 77
  78. 78
  79. 79
  80. 80
  81. 81
  82. 82
  83. 83
  84. 84
  85. 85
  86. 86
  87. 87

Decision letter

  1. Huda Y Zoghbi
    Senior and Reviewing Editor; Texas Children's Hospital, United States
  2. Hanna Renvall
    Reviewer; Aalto University, Finland
  3. Allison Nugent
    Reviewer
  4. Annika Hulten
    Reviewer; Aalto University, Finland

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 "Disrupted parietal oscillatory connectivity links young APOE-ɛ4 carriers to Alzheimer's disease" for consideration by eLife. Your article has been reviewed by David Van Essen as the Senior Editor, a Reviewing Editor, and three reviewers. The following individuals involved in review of your submission have agreed to reveal their identity: Hanna Renvall (Reviewer #1); Allison Nugent (Reviewer #2); Annika Hulten (Reviewer #3).

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

Summary:

The study by Koelewijn et al., involves a data set of 183 subjects in an attempt to link APOE-ɛ4 carriers to preclinical changes neural activity, in relation to their increased genetic risk of developing AD. The authors investigated resting-state oscillatory network connectivity across different frequency bands. Key findings: APOE-ɛ4 carries have increased alpha- and beta-band connectivity in bilateral occipital areas and parietal cortices, as compared to non-carriers. The connectivity changes were partially overlapping with decreased connectivity between right hemisphere parietal and temporal areas in a separate population of AD patient. Based on these findings the authors suggest that AD can be characterized by pre-onset hyperconnectivity that, once the disease manifests, turns into hypoconnectivity particularly in the parietal cortex.

In general, the referees were positive and found the study exciting. However, in particular in regard to methodology there were several concerns that require clarifications and justification. Also, it was a general feeling that the Discussion section requires improvement: in its current form the paper lacks deeper discussion and theoretical background needed to interpret the present findings with respect to the majority of the existing literature in AD, as well as on the possible clinical value of the results.

One referee asks for further analysis (see point 14): In the description of the AD subjects, the manuscript states that there was an age-matched cohort. However, it appears they were not compared to the AD subjects. It is proposed to use these as a comparison group rather than the young E4 carriers. The authors are encouraged to follow this advice or provide a strong rationale for why this was not done.

Essential revisions:

Methodology

1) It is unclear why the authors decided to use only eyes open condition here (cf. their earlier study on AD patients with both eyes open and closed data), and not e.g. the difference between eyes closed (with prominent alpha activity) and eyes open conditions that might be a more reliable functional measure? Furthermore, it appears that the differences between AD patients and controls were even more prominent in the eyes closed condition? Please motivate.

2) The section on the Gaussian Mixture Model (GMM) is inadequately described. The authors state that a GMM was applied to all 8100 connection to separate signal from noise. Within each frequency band, however, aren't there only 4005 unique connections, since correlations cannot discern directionality? How many subpopulations were included in the mixture model? Is there any published validation that shows that this method effectively separates true correlations from noise? Was this done for each group separately?

3) Is the discarding of connections +/- two scaled median SD's after the connections were removed by the GMM method? So, each subject now has a different number of valid connections?

4. The authors state that the valid connection masks were added to form a single mask – was this a union or an intersection? It says that it included all valid connections present in either group, but does this include connections that were not valid in all subjects?

5) How were the randomization test (akin to a two-sample t-test) and the uncorrected two sample t-test different, other than that one is parametric and the other non-parametric? The authors seem to indicate that the interpretation is different. (i.e. the two-sample t-test can test whether differences were global across the brain).

6) Please clarify how the power was calculated, what parameters were used to connect the effect size for disease prediction accuracy (the AUC from Escott-Price) to the MEG measures?

7) It is stated that the GMM selected valid datasets as well as valid connections, although the GMM methods do not mention how datasets were defined as valid or invalid.

8) One referee mentions that a main concern is the choice of parcellation scheme used in the connectivity analysis. While it is fair to say that there exists no one correct atlas that always should be used, the AAL is definitely non-optimal for a MEG connectivity analysis as the parcels are of very different size and shape. By selecting only one node from each parcel, the nodes will be unevenly distributed across the cortex, which in itself is somewhat problematic. However, a bigger problem is that large elongate parcels (such as those in the temporal lobe) may contain two (or more) active areas but will be represented by only one node (capturing the strongest source). The worst-case scenario is if a strong active area is situated on the border between two or more large parcels. In this case the nodes from each of these parcels will reflect the same underlying neural activity, whereas a clear but somewhat smaller activation in the other end of one of the parcels would go undetected. Assuming that the multivariate leaking correction works as intended, two nodes capturing the same activity is less of a problem (though it may lead to spatially incorrect inferences) but ignoring an active area may alter the network dynamics radically. The most straight forward way around this is to select a parcellation scheme with optimally sized parcels (not too small, not too big) that are of roughly equal size. To my knowledge no existing atlas fulfills these criteria fully, which is probably reflected in the fact that most connectivity papers use modified versions of existing parcellations. If the number of parcels becomes too big, there is the option to select only those that cover the cortical regions associated with the DMN in the literature. Alternatively, you may wish to try a more sophisticated approach and apply a data driven way of finding all active regions based on some form of clustering algorithm or ICA, and the define parcels/nodes based on these. Please comment on the AAL approach in relation to these alternatives.

9) The use of a multivariate orthogonalization approach is an elegant way to address the problem of spurious (aka ghost) interactions. However, like all methods it has its limitations, and it would be good to acknowledge these in the Discussion section. For example, the SNR is not constant across all cortical regions in MEG and areas which will the source estimate specificity and by extension the rotation done in the correction. Therefore, connectivity between areas with high source power (like the occipital and parietal regions in the alpha band) may be more easily detected than those with low source power (like the frontal cortex). This may, in part be the reason for the different results between the previous fMRI study by Filippini et al.

Results/interpretation

10) The major finding suggest hyperactivity in the right parietal connectivity, but after the statistical analysis the result appears quite modest and very local. How can one be sure that it is not the result of signal leakage?

11) The location for the statistically significant hyperactivity is not anatomically the most evident one, taken that the most prominent structural changes appear in the hippocampus and more generally in the medial temporal lobes. The authors have analyzed the hippocampal volumes and total intracranial volumes in all subjects, but what about the parietal volumes? Is their e.g. any hemispheric difference in the parietal lobe volumes in the APOE-ɛ4 carriers that would correlate with the present finding?

12) In Table 2, the only overlapping, affected areas between non-symptomatic APOE-ɛ4 carriers and AD patients were the right parietal areas, whereas 6/10 nodes with greatest group difference in APOE-ɛ4 carriers and controls were not different between AD patients and age-matched controls. How do the authors explain this disappearing difference along with the progressing disease?

13) In Figure 2, from panels D and H, it appears that all connections were greater in the E4 carriers compared to non-carriers. In panel E, it is clear that most of the valid connections are above the x=y line, indicating that the mean Z is higher for the carriers than non-carriers. However, in panel A, it appears that most of the valid connections are below the x=y line, which would indicate that the mean Z is larger in the non-carriers. Please clarify. Also, in panes F and G, it appears that there are valid connections to frontal areas, however these are not apparent on the glass brains.

14) (also see general comment for this point). The description of participants is confusing, given that a re-analysis of data in AD was performed, but these subjects are not well described. Please clarify. There also seems to be a comparison between young E4 carriers and AD patients: I question the utility of this given that age is a confound. In the description of the AD subjects the manuscript states that there was an age-matched cohort, but I do not see where they were compared to AD subjects. This would be a much better comparison group than young E4 carriers.

15) Table 2 shows nodes most "affected" within each group separately – what is being compared? Eyes open vs. eyes closed? Based upon Figure 4, is it right that the comparisons are young APOE-E4 vs. AD, and AD vs. matched controls? Why are the results for E4 carriers vs. non-carriers in Figure 4 different from the results presented earlier?

Discussion

16) In its current form the paper lacks deeper discussion and theoretical background needed to interpret the present findings with respect to the majority of the existing literature in AD, as well as on the possible clinical value of the results.

17) The AD patients who are APOE-ɛ4 carriers are known to develop the clinical AD at an earlier stage than non-carriers, but in the end with very similar structural and clinical changes to the non-carriers. Why this happens, and what is the role of APOE-ɛ4 allele in the course of the disease remains unclear, but it has nevertheless been studied extensively. The manuscript should provide some light on what is known about the cellular-level functions of APOE-ɛ4, and especially on how these functions might be related to the present neuroimaging results. I do see that it is difficult to draw conclusions between cellular level findings and non-invasive neuroimaging results, but at least some hypothesis/discussion would be needed here.

18) Taken that the ultimate goal is to find individual measures on the connectivity changes in the prodromal patients, the decision to concentrate only on the connections that were consistently present in all datasets would benefit some comment/discussion on the selection's potential drawbacks.

19) PET often shows hypometabolic changes already in the stage of mild cognitive impairment in AD patients – is there literature on PET results in young healthy APOE-ɛ4 carriers? If such exists, it should be discussed with respect to the present results.

20) Taken that MEG may not be the optimal method to study deep brain structures or frontal cortex, some comment on the possible limitations of MEG in this respect would be good to add here.

21) When searching for clinical biomarkers, the consistency of the measures over multiple measurements should be demonstrated. Any such information available to be added here?

22) In discussing both their own findings and the previous literature, the authors should avoid inferring causation from correlation. This line of reasoning occurs both in the introduction and discussion. For example, in the Introduction the authors state that the previous findings "suggest that the presence of an APOE-ɛ4 allele alters oscillatory brain function", when in fact the studies that are referred to have only found an association. The fact that the presence APOE-ɛ4 does not always lead to AD (with accompanying changes in oscillatory brain activity) suggests that other genetic or environmental factors (such as e.g. cholesterol levels) are at play as well. Moreover, it is possible (if not probable) that the APOE-ɛ4 is part of a more multifaceted genetic basis of AD. Assigning the causation to one allele would in this case be misleading.

23) A referee was unable to follow the logic in the Discussion section where the authors compare the present findings to those reported by Filippini et al. The author claim that the previous fMRI study corroborates the present findings even though completely different areas are highlighted in each of the respective studies. I am also uncertain what the word "inconsistency" is referring to in the sentence "In contrast, evidence of the hippocampal volume decrease in young carriers is inconsistent […]". You referring to previous findings being inconsistent or that the present study adds to the inconsistency? I suspect that the authors mean that given the inconsistent findings of reduced hippocampal measures, hemodynamic and/or oscillatory network effect may be better markers of an increased AD risk. Still given that there are only 2 network studies even this seems a bit of an overstatement. In either case, please clarify this paragraph.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Oscillatory hyperactivity and hyperconnectivity in young APOE-ɛ4 carriers and hypoconnectivity in Alzheimer's disease" for further consideration at eLife. Your revised article has been favorably evaluated by Huda Zoghbi as the Senior Editor, a Reviewing Editor, and three reviewers.

The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance, as outlined below:

The reviews below provide details about the limitations that need to be addressed. In summary, all reviewers agree that revision of the text (especially the claims) and figures is still needed. Please revise the text putting special emphasis on the limitation/claim parts and on the details of the methodology. While we agreed that no additional analyses need to be done, the epoching of the time series should be mentioned, perhaps in limitations.

I hope that the comments below will help you revised the manuscript.

Reviewer #1:

The authors have evidently gone through a big effort in re-analyzing the data and in replying to the reviewer comments on the earlier manuscript. I think that the machine learning approach gives a significant contribution to the analyses.

As to the present manuscript, I would still be careful in using words like "prevention" from which the present data is still quite far. Providing possible predictive information on AD in very early stage is already an important step.

In general, the language is not always best possible and could still be clarified and polished. Also, the amount of given details on the data analysis continues to vary, which would make it difficult to replicate part of the applied analyses: e.g. the analysis based on Graph Theory (subsection “Statistical Analysis of Group Differences in Connectivity”) mainly refer to existing toolboxes, while the SVM approach is rather extensively explained.

Reviewer #2:

The revised paper is significantly improved in terms of clarity of the methods and results. While many of my concerns have been addressed, the number of revisions performed by the authors have prompted a few additional comments:

I have a concern about the division of the time series into 2 second epochs, removal of segments with artifacts, followed by concatenation. While the number of retained epochs were consistent across groups, what about the number of discontinuities introduced into the time series? For example, removing two adjacent epochs would result in one discontinuity, while removing two separate epochs would result in two discontinuities. What is the impact of this on the connectivity? This may be particularly problematic for the lowest frequency band; there will only be 2-8 cycles of the 1-4 Hz band. I am not suggesting that the analyses be re-done, although some commentary on this point may be warranted.

Why was a visual grating used rather than a more "pure" resting state study?

In Figure 1 – I'm not sure how relevant row 2 is, given that I would expect many connections to survive a p<0.05 uncorrected threshold by chance alone. This is especially true since row 2 is quite similar to row 4. Also, in this figure, given the narrow width of the lines, I can't visually detect any differences in opacity. Perhaps line width would be a better indicator of connection strength?

In Figure 2 - I'm not entirely clear on the utility of the lower panel. Although there is some overlap with the areas showing higher gamma activity and areas showing greater connectivity, it is not entirely convincing that these two things are related in any way, given that more regions that don't overlap than regions that do.

Figure 3 – same comment as for Figure 1.

Figure 4 – it may be useful to show which specific connections do overlap, perhaps a conjunction of the graphs shown in the bottom panel of 4A. Is there a negative correlation between the strengths of the connections that are in this conjunction? How many edges actually overlap?

I am concerned that the results are somewhat overstated in the conclusions. There was at best partial overlap between the oscillatory power and connectivity results. Likewise, there was only partial overlap between the connections found in the young sample and the older sample.

Some comment on the regions showing the greatest abnormality may be helpful to put the results in context.

Reviewer #3:

This is a follow-up review on the study by Koelewijn et al. that looks at changes in the oscillatory connectivity in resting state MEG in APOE-ɛ4 carriers and matched controls. The authors have, in my opinion replied to the reviewers’ comments and suggestions in a satisfactory manner, and I am ready to suggest that the manuscript be accepted for publication in eLife.

Before signing off on this review, I would nonetheless like to express that I am somewhat disappointed that the authors, contrary to my suggestion, chose not to change the parcellation template used in the connectivity analysis. I am sympathetic to the authors' argument that the AAL is widely used, but this does mean that it's a good one. Most likely the reason, that is widely used is merely a consequence of the methodological dominance of fMRI where the template is much less problematic. As the authors themselves note in their Discussion, it remains a fact that the AAL is not well-suited for connectivity analysis of MEG data. Bad practices in science do not change unless someone leads the way. Hence, I encourage the authors to at least in their future studies become trail makers that help change suboptimal practices, and thus advancing also the validity of MEG.

https://doi.org/10.7554/eLife.36011.015

Author response

Summary:

The study by Koelewijn et al., involves a data set of 183 subjects in an attempt to link APOE-ɛ4 carriers to preclinical changes neural activity, in relation to their increased genetic risk of developing AD. The authors investigated resting-state oscillatory network connectivity across different frequency bands. Key findings: APOE-ɛ4 carries have increased α- and β-band connectivity in bilateral occipital areas and parietal cortices, as compared to non-carriers. The connectivity changes were partially overlapping with decreased connectivity between right hemisphere parietal and temporal areas in a separate population of AD patient. Based on these findings the authors suggest that AD can be characterized by pre-onset hyperconnectivity that, once the disease manifests, turns into hypoconnectivity particularly in the parietal cortex.

In general, the referees were positive and found the study exciting. However, in particular in regard to methodology there were several concerns that require clarifications and justification. Also, it was a general feeling that the Discussion section requires improvement: in its current form the paper lacks deeper discussion and theoretical background needed to interpret the present findings with respect to the majority of the existing literature in AD, as well as on the possible clinical value of the results.

One referee asks for further analysis (see point 14): In the description of the AD subjects, the manuscript states that there was an age-matched cohort. However, it appears they were not compared to the AD subjects. It is proposed to use these as a comparison group rather than the young E4 carriers. The authors are encouraged to follow this advice or provide a strong rationale for why this was not done.

There seems to have been a misunderstanding here, and we apologise for not explaining this properly. We did two experimental comparisons. (1) Young APOE-ɛ4 carriers were compared to Young non-carriers and (2) Elderly AD participants were compared to an age-matched control group. We do not compare the elderly group to the young group as this would be dominated by age-related changes. We do, however, directly compare the statistical effects calculated separately for the two experimental comparisons (see Figure 5).

Essential revisions:

Methodology

1) It is unclear why the authors decided to use only eyes open condition here (cf. their earlier study on AD patients with both eyes open and closed data), and not e.g. the difference between eyes closed (with prominent α activity) and eyes open conditions that might be a more reliable functional measure? Furthermore, it appears that the differences between AD patients and controls were even more prominent in the eyes closed condition? Please motivate.

The APOE study is an opportunistic one in which we have post-hoc analysed data from an internal database at Cardiff called “the 100 Brains study”. In this study, the resting-state recording was one of many recordings performed in the MEG, in addition to many further tests and neuroimaging modalities. It was a practical decision at the time of designing this study to only collect eyes-open data to reduce the total time required. As a result, we only have eyes-open data available for this analysis. Furthermore, we actually found a very similar pattern in eyes-open and closed resting-state recordings in our previous paper (Koelewijn et al., 2017), which has also been reported by others (Hillebrand et al., 2016). For this reason, we do not think that results would have been much different had we presented eyes-closed data, or a difference between eyes-open and eyes-closed. Unfortunately, we do not have the data to perform this analysis and as such we restrict our findings to eyes-open resting-state for both experiments. To avoid confusion, the eyes-closed analysis (which was only available for the AD versus elderly control experiment) has been removed from the manuscript.

2) The section on the Gaussian Mixture Model (GMM) is inadequately described. The authors state that a GMM was applied to all 8100 connection to separate signal from noise. Within each frequency band, however, aren't there only 4005 unique connections, since correlations cannot discern directionality? How many subpopulations were included in the mixture model? Is there any published validation that shows that this method effectively separates true correlations from noise? Was this done for each group separately?

We have decided not to use this relatively complex (and new) approach but instead use a simple rank-thresholding procedure. Namely, to take those connections that have a Mean Rank across the cohort of >=80% i.e. the strongest 20% connections. For Alpha and Beta this identified an extremely similar pattern of connections as the GMM approach, but also now reveals high-rank correlations in Delta, Theta and High Gamma.

3) Is the discarding of connections +/- two scaled median SD's after the connections were removed by the GMM method? So, each subject now has a different number of valid connections?

Actually, we were completely removing participants from the analysis, if they had too many or two few valid connections. As we describe above, we have now removed this procedure.

4. The authors state that the valid connection masks were added to form a single mask – was this a union or an intersection? It says that it included all valid connections present in either group, but does this include connections that were not valid in all subjects?

It was a union i.e. a logical ‘OR’ approach. In this, connections identified as ‘valid’ in either group were included to avoid missing out on potential group differences. Otherwise, if a connection was very greatly increased/decreased in one cohort compared to another, the connection would potentially be discarded, and a large group difference missed. Note that connections are accepted/discarded at the group level by assessing their average mean rank across the cohort – not at the single-subject level.

5) How were the randomization test (akin to a two-sample t-test) and the uncorrected two sample t-test different, other than that one is parametric and the other non-parametric? The authors seem to indicate that the interpretation is different. (i.e. the two-sample t-test can test whether differences were global across the brain).

The uncorrected two-sample t test, we believe, is a useful first look at the distribution of connection strength differences across the brain. The randomisation test, in which we use omnibus testing to derive the null distribution for the largest difference in the brain, estimates whether each of those connections would survive a multiple comparison correction i.e. it is simply used to assess corrected p-values. We use this method instead of Bonferroni correction across all 4010 unique non-diagonal correlation pairs because it is less conservative in the presence of known non-independence of the measures see: (Nichols and Holmes, 2003) (Singh, Barnes and Hillebrand, 2003)

6) Please clarify how the power was calculated, what parameters were used to connect the effect size for disease prediction accuracy (the AUC from Escott-Price) to the MEG measures?

This was not used in the MEG measures, only to assess the potential statistical power in the genetics study.

7) It is stated that the GMM selected valid datasets as well as valid connections, although the GMM methods do not mention how datasets were defined as valid or invalid.

Please see our response to Point 3, above.

8) One referee mentions that a main concern is the choice of parcellation scheme used in the connectivity analysis. While it is fair to say that there exists no one correct atlas that always should be used, the AAL is definitely non-optimal for a MEG connectivity analysis as the parcels are of very different size and shape. By selecting only one node from each parcel, the nodes will be unevenly distributed across the cortex, which in itself is somewhat problematic. However, a bigger problem is that large elongate parcels (such as those in the temporal lobe) may contain two (or more) active areas but will be represented by only one node (capturing the strongest source). The worst-case scenario is if a strong active area is situated on the border between two or more large parcels. In this case the nodes from each of these parcels will reflect the same underlying neural activity, whereas a clear but somewhat smaller activation in the other end of one of the parcels would go undetected. Assuming that the multivariate leaking correction works as intended, two nodes capturing the same activity is less of a problem (though it may lead to spatially incorrect inferences) but ignoring an active area may alter the network dynamics radically. The most straight forward way around this is to select a parcellation scheme with optimally sized parcels (not too small, not too big) that are of roughly equal size. To my knowledge no existing atlas fulfills these criteria fully, which is probably reflected in the fact that most connectivity papers use modified versions of existing parcellations. If the number of parcels becomes too big, there is the option to select only those that cover the cortical regions associated with the DMN in the literature. Alternatively, you may wish to try a more sophisticated approach and apply a data driven way of finding all active regions based on some form of clustering algorithm or ICA, and the define parcels/nodes based on these. Please comment on the AAL approach in relation to these alternatives.

This is a very good set of points and we thank the reviewer for these. The AAL is an increasingly used, almost default, atlas in this area of research but, of course, we have to acknowledge these potential problems. With that in mind, we have added a new subsection “Limitations of the study”. The reviewer will see we have added a section on this specific issue to acknowledge it and to point to a potential way forward.

9) The use of a multivariate orthogonalization approach is an elegant way to address the problem of spurious (aka ghost) interactions. However, like all methods it has its limitations, and it would be good to acknowledge these in the discussion. For example, the SNR is not constant across all cortical regions in MEG and areas which will the source estimate specificity and by extension the rotation done in the correction. Therefore, connectivity between areas with high source power (like the occipital and parietal regions in the α band) may be more easily detected than those with low source power (like the frontal cortex). This may, in part be the reason for the different results between the previous fMRI study by Filippini et al.

Again, this is a very important point and so we have added another paragraph describing this, to the end of the subsection “Limitations of the study”.

Results/interpretation

10) The major finding suggest hyperactivity in the right parietal connectivity, but after the statistical analysis the result appears quite modest and very local. How can one be sure that it is not the result of signal leakage?

In the prior draft of the paper, the reviewer is quite right that after statistical correction at individual edges only one edge remained in the Beta band. In fact, when we revised our analysis procedure (see start of this document) this connection dropped below our significance criteria, whereas a single connection in the theta band now becomes significant. Put simply, with the effect sizes we have, and a conservative correction for multiple comparisons, we do not seem to see extensive connectivity differences at the level of individual edges. However, we have augmented our reporting by also assessing connected cluster size and now report this and we also use cohort-resampling to see how reproducible effects are across sub-samples of our participants. These do seem to suggest robust patterns of statistical effects – just not at the individual edge level.

11) The location for the statistically significant hyperactivity is not anatomically the most evident one, taken that the most prominent structural changes appear in the hippocampus and more generally in the medial temporal lobes. The authors have analyzed the hippocampal volumes and total intracranial volumes in all subjects, but what about the parietal volumes? Is their e.g. any hemispheric difference in the parietal lobe volumes in the APOE-ɛ4 carriers that would correlate with the present finding?

This is an excellent suggestion and we have now added an analysis of regional thickness, volume and area to the manuscript, using the standard Desikan-Killiany gyral atlas within Freesurfer. No significant differences in any parameter were found in any of the regions, after correction for multiple region comparisons.

12) In Table 2, the only overlapping, affected areas between non-symptomatic APOE-ɛ4 carriers and AD patients were the right parietal areas, whereas 6/10 nodes with greatest group difference in APOE-ɛ4 carriers and controls were not different between AD patients and age-matched controls. How do the authors explain this disappearing difference along with the progressing disease?

In the revised manuscript we make it clear that some of the effects are overlapping, but there is not a clear one to one correspondence. We have decided to illustrate this visually with multiple visualisations in the new Figure 5 and have removed the table describing individual connections. We believe that the scatter plot shown in new Figure 5 shows that there is a general inverse relationship between connectivity differences observed in Young APOE-ɛ4 carriers (compared to young controls) and AD patients (compared to their age-matched cohort). The plots of the networks themselves, though reveal that the effects are spatially overlapping but not perfectly registered. This is now explained in the text, figures and this paragraph:

“Our results therefore suggest that the right parietal cortex is a particularly important area in the risk for, and development of, AD. Although our study can only show an implied association between early hyperconnectivity and later hypoconnectivity, we could speculate that an initial hyperconnectivity/hyperactivity centred around (right) parietal cortex has a cascade of effects that may lead to eventual disconnection in AD. Although initially a right-dominant effect, this may spread over time to more widespread effects across the cortex and hemispheres. A larger study, using the methods we present here, over a wider age-range, may shed further light on this.”

13) In Figure 2, from panes D and H, it appears that all connections were greater in the E4 carriers compared to non-carriers. In panel E, it is clear that most of the valid connections are above the x=y line, indicating that the mean Z is higher for the carriers than non-carriers. However, in panel A, it appears that most of the valid connections are below the x=y line, which would indicate that the mean Z is larger in the non-carriers. Please clarify. Also, in panes F and G, it appears that there are valid connections to frontal areas, however these are not apparent on the glass brains.

Apologies – the reviewer was correct. We made an error in the cohort resampling procedure so that connections surviving our confidence-interval criteria were always shown in red – even if they were robust decreases. For the Alpha band, the posterior connections should have been blue, reflecting decreases in the APOE-ɛ4 group. This has been fixed in the new analyses.

Also note that the two frontal connections displayed in panels F and G were the right 7rolandic operculum and supplementary motor area. These two areas are positioned quite posterior in the frontal cortex, and therefore appear close together to the various parieto-temporal areas displayed in the glass brain.

14) (also see general comment for this point). The description of participants is confusing, given that a re-analysis of data in AD was performed, but these subjects are not well described. Please clarify. There also seems to be a comparison between young E4 carriers and AD patients: I question the utility of this given that age is a confound. In the description of the AD subjects the manuscript states that there was an age-matched cohort, but I do not see where they were compared to AD subjects. This would be a much better comparison group than young E4 carriers.

Sorry again for the misunderstanding – Please see our opening comments, but in short, we never directly compared young APOE-ɛ4 carriers to AD patients. Statistical comparisons were always to an age-matched control group in both cases. What we do perform (See new Figure 5) is a comparison of both the spatial distribution and magnitude of (APOE-ɛ4-young controls) and (AD-elderly control) effects.

15) Table 2 shows nodes most "affected" within each group separately – what is being compared? Eyes open vs. eyes closed? Based upon Figure 4, is it right that the comparisons are young APOE-E4 vs. AD, and AD vs. matched controls?

Please see our previous comments and we have tried hard in the redrafted manuscript to make this clearer.

Why are the results for E4 carriers vs. non-carriers in Figure 4 different from the results presented earlier?

In the redrafted manuscript it should be clearer that these are now the same.

Discussion

16) In its current form the paper lacks deeper discussion and theoretical background needed to interpret the present findings with respect to the majority of the existing literature in AD, as well as on the possible clinical value of the results.

This is a very good suggestion and we have now extended the Introduction, Discussion section (and redrafted Figure 6) to try and address this.

17) The AD patients who are APOE-ɛ4 carriers are known to develop the clinical AD at an earlier stage than non-carriers, but in the end with very similar structural and clinical changes to the non-carriers. Why this happens, and what is the role of APOE-ɛ4 allele in the course of the disease remains unclear, but it has nevertheless been studied extensively. The manuscript should provide some light on what is known about the cellular-level functions of APOE-ɛ4, and especially on how these functions might be related to the present neuroimaging results. I do see that it is difficult to draw conclusions between cellular level findings and non-invasive neuroimaging results, but at least some hypothesis/discussion would be needed here.

Please see the new extended Introduction and the Discussion section. As the reviewer probably can understand this is a very complex area, so we have tried to synthesise major points and point readers to major review articles.

18) Taken that the ultimate goal is to find individual measures on the connectivity changes in the prodromal patients, the decision to concentrate only on the connections that were consistently present in all datasets would benefit some comment/discussion on the selection's potential drawbacks.

Please see the new subsection “Limitations of the study” in which we describe the rationale for this decision and why it may lead to a relative insensitivity to some cohort effects. In general, as this is an exploratory study describing a new finding, it is probably best to be on the conservative side.

19) PET often shows hypometabolic changes already in the stage of mild cognitive impairment in AD patients – is there literature on PET results in young healthy APOE-ɛ4 carriers? If such exists, it should be discussed with respect to the present results.

As far as we have been able to find, no such PET studies have been performed in young healthy APOE-ɛ4 carriers.

20) Taken that MEG may not be the optimal method to study deep brain structures or frontal cortex, some comment on the possible limitations of MEG in this respect would be good to add here.

Please see the new subsection “Limitations of the study” in which we make this clear.

21) When searching for clinical biomarkers, the consistency of the measures over multiple measurements should be demonstrated. Any such information available to be added here?

We are only aware of one published paper showing the repeatability of our connectivity measures (Colclough et al., 2016) and this is cited now in the Materials and methods section:

“We focused on amplitude-amplitude connectivity of beamformer-derived oscillatory source signals, one of the most robust and repeatable types of MEG connectivity measures (Colclough et al., 2016)”.

And we also discuss/cite this in the new subsection “Limitations of the study” – It is our prime motivation for only using amplitude-amplitude connectivity measures.

22) In discussing both their own findings and the previous literature, the authors should avoid inferring causation from correlation. This line of reasoning occurs both in the introduction and discussion. For example, in the Introduction the authors state that the previous findings "suggest that the presence of an APOE-ɛ4 allele alters oscillatory brain function", when in fact the studies that are referred to have only found an association. The fact that the presence APOE-ɛ4 does not always lead to AD (with accompanying changes in oscillatory brain activity) suggests that other genetic or environmental factors (such as e.g. cholesterol levels) are at play as well. Moreover, it is possible (if not probable) that the APOE-ɛ4 is part of a more multifaceted genetic basis of AD. Assigning the causation to one allele would in this case be misleading.

Yes – this is, again, a very good point. We have been through the manuscript and toned down our implication of a direct link, including the sentence mentioned, which now reads:

“These findings suggest that the presence of an APOE-ɛ4 allele is associated with differences in oscillatory brain function at a stage preceding AD symptomology, but it is not clear how early in the life span this is established”.

We have also added a section to the subsection “Limitations of the study” making it explicit that this type of study only reveals associations, not causal relationships.

23) A referee was unable to follow the logic in the Discussion section where the authors compare the present findings to those reported by Filippini et al. The author claim that the previous fMRI study corroborates the present findings even though completely different areas are highlighted in each of the respective studies. I am also uncertain what the word "inconsistency" is referring to in the sentence "In contrast, evidence of the hippocampal volume decrease in young carriers is inconsistent […]". You referring to previous findings being inconsistent or that the present study adds to the inconsistency? I suspect that the authors mean that given the inconsistent findings of reduced hippocampal measures, hemodynamic and/or oscillatory network effect may be better markers of an increased AD risk. Still given that there are only 2 network studies even this seems a bit of an overstatement. In either case, please clarify this paragraph.

We thank the reviewer for pointing this out. We have redrafted this entire section (Discussion section), but also added a further study from Westlye et al., 2011. Although in an older group than ours (and Fillipini’s) Westlye also found similar increased activation in the hippocampal parts of the DMN for E4 carriers, but importantly for us, also found increased DMN synchronisation in the posterior and medial parts of the DMN. Interestingly for the lateral cortical areas, including parietal, they also found an exclusively right hemisphere increase in DMN synchronisation for APOE-4 carriers, matching quite nicely with our findings.

[Editors' note: further revisions were requested prior to acceptance, as described below.]

The reviews below provide details about the limitations that need to be addressed. In summary, all reviewers agree that revision of the text (especially the claims) and figures is still needed. Please revise the text putting special emphasis on the limitation/claim parts and on the details of the methodology. While we agreed that no additional analyses need to be done, the epoching of the time series should be mentioned, perhaps in limitations.

I hope that the comments below will help you revised the manuscript.

We have carefully reviewed the text and softened some of the statements. A paragraph detailing the potential problem with epoch concatenation has been added to the Limitations section of the Discussion. We have also added sub-figures to Figure 4 and a new Table 2 showing a conjunction analysis (requested by reviewer 2).

There are two other changes that we have made:

1) As discussed below in response to reviewer 1, we reviewed our cross-validation strategy for the SVM classification and realised that it needed amendment. Put simply, we were choosing the strongest 20% of edges as classification features for the SVM separately for the two groups being tested and then pooling all these edges for classification. We assumed that this was un-biased as we were not choosing features that maximised discrimination but, of course, in a real-world situation, this could not be done. Furthermore, to our surprise, using further randomisation testing confirmed that this did result in >50% performance for completely randomly-labelled data in many cases. So, we changed the procedure to select the top 20% of features using the whole cohort. This resolved the bias, but results in reduced performance, albeit still significantly greater than chance. We have therefore changed the results in Table 3 (previously Table 2) to reflect and moderated our discussion points regarding this analysis.

2) Unfortunately, when double-checking our results and packaging them up for open release (see below), we realised that the code used to generate Figure 3 (the AD cohort results) and the right column of Figure 4a had the wrong threshold set. When we re-ran the corrected script, the differences were subtle and for the presented data are hard to spot (it mostly is apparent in the count of significant edges shown on each plot i.e. in the last revision, the bottom right circle plot has 19/56 edges whereas now it has 23/56 edges. Note that no other figure, or analysis is affected by this error – it was purely a display problem. However, in addition, we now find 11 edges in the Low Gamma range, so have added the relevant column to Figure 3. The derived results and Matlab scripts needed to generate Figure 1, the lower part of Figure 2, Figure 3 and Figure 4 are now publicly released on an Open Science Framework project (https://osf.io/e4cjx/ or DOI: 10.17605/OSF.IO/E4CJX) so readers can independently verify our plots, or indeed change the statistical procedures or display parameters to further investigate the results. Please see our modified data availability statement in the manuscript as well.

Reviewer #1:

The authors have evidently gone through a big effort in re-analyzing the data and in replying to the reviewer comments on the earlier manuscript. I think that the machine learning approach gives a significant contribution to the analyses.

Many thanks for this. Unfortunately, as can be seen in Table 3, we have re-evaluated our cross-validation procedure (now described in more detail in our Materials and methods section) and believe our original performance measures were inflated. The results are still significantly above chance but are not as high as previously reported. With that in mind we have altered the text describing the ML results to make it much clearer that these are encouraging results but moderate in effect.

As to the present manuscript, I would still be careful in using words like "prevention" from which the present data is still quite far. Providing possible predictive information on AD in very early stage is already an important step.

In general, the language is not always best possible and could still be clarified and polished.

We’ve deleted the final part of the sentence in the Introduction that mentions prevention strategies.

Also, the amount of given details on the data analysis continues to vary, which would make it difficult to replicate part of the applied analyses: e.g. the analysis based on Graph Theory (subsection “Statistical Analysis of Group Differences in Connectivity”) mainly refer to existing toolboxes, while the SVM approach is rather extensively explained.

Actually, we don’t use an existing toolbox here but a very simple procedure in Matlab. We’ve deleted the reference to “Graph-Theory” (as we think this is confusing) and made it clear how the largest connected cluster size is estimated in subsection “Statistical Analysis of Group Differences in Connectivity”. This would allow the reader to replicate this procedure quite easily, including the subsequent statistical testing. We have left in the reference to the NBS toolbox as the Authors were the first to propose this method and also demonstrated its increased sensitivity.

Reviewer #2:

The revised paper is significantly improved in terms of clarity of the methods and results. While many of my concerns have been addressed, the number of revisions performed by the authors have prompted a few additional comments:

I have a concern about the division of the time series into 2 second epochs, removal of segments with artifacts, followed by concatenation. While the number of retained epochs were consistent across groups, what about the number of discontinuities introduced into the time series? For example, removing two adjacent epochs would result in one discontinuity, while removing two separate epochs would result in two discontinuities. What is the impact of this on the connectivity? This may be particularly problematic for the lowest frequency band; there will only be 2-8 cycles of the 1-4 Hz band. I am not suggesting that the analyses be re-done, although some commentary on this point may be warranted.

Thank you. This is a very import comment and entirely justified in requiring a comment. We have added a section on this to the subsection “Limitations of the study”.

Why was a visual grating used rather than a more "pure" resting state study?

Unfortunately, this was simply the design of the original study. As we mention in the paper, we are re-analysing the data from our first paper (Koelewijn et al.).

In Figure 1 (Figure 3 – same comment as for Figure 1), I'm not sure how relevant row 2 is, given that I would expect many connections to survive a p<0.05 uncorrected threshold by chance alone. This is especially true since row 2 is quite similar to row 4.

Row 2 and Row 4 are expressing different statistical analyses. Row 2 is a test for any effect in the difference of means, between cohorts, using a t-test, whereas Row 4 shows the consistency of the sign of the effect at each connection using randomised split-half testing of the cohort. The reviewer is right that for a large sample of normally distributed values these two things should converge, so it is actually quite encouraging that the maps are so similar – but that is not necessarily guaranteed, so we would prefer to show both results. Finally note that for the smaller cohort in the AD experiment there is more divergence between Row 2 and Row 4.

Also, in this figure, given the narrow width of the lines, I can't visually detect any differences in opacity. Perhaps line width would be a better indicator of connection strength?

This is potentially a good idea, but when we tried this the figure become rather confusing as the thick lines prevented viewing of other connections, particularly in the small versions of these circle plots. It is also worth noting that because many of these maps are thresholded, the variability of the effect magnitude is relatively small and hence the opacity does not vary much. For example, we only show the confidence interval values on Row 4 from 0.95 to 1.0.

In Figure 2 – I'm not entirely clear on the utility of the lower panel. Although there is some overlap with the areas showing higher γ activity and areas showing greater connectivity, it is not entirely convincing that these two things are related in any way, given that more regions that don't overlap than regions that do.

We actually quite like the lower panel as it allows the reader to easily identify (i) Which AAL nodes show hyperconnectivity in E4 carriers, (ii) Which AAL nodes show hypoconnectivity in E4 carriers, (iii) Which nodes show hyperactivity in E4 carriers. This is very difficult to do from the other Figures. The reviewer is absolutely right though, that the overlap is partial. However, note that those nodes in the right hemisphere that show hyperconnectivity also show hyperactivity (but not vice-versa). We have therefore moderated the claims in subsection “APOE-ɛ4 carrier versus non-carrier group differences in oscillatory activity Q” i.e.:

“In summary, gamma hyperactivities were found in mostly right-lateralised regions, some of which also appear to show hyperconnectivity in the lower frequency ranges (principally beta and alpha). This overlap is described graphically in the lower part of Figure 2.”

Figure 4 – it may be useful to show which specific connections do overlap, perhaps a conjunction of the graphs shown in the bottom panel of 4A. Is there a negative correlation between the strengths of the connections that are in this conjunction? How many edges actually overlap?

We have now added this plot to the bottom of 4A as requested, added an unthresholded conjunction plot to 4C and created a Table describing the effect sizes for these conjunction edges in a new Table 2. We’ve also added more details in the Figure caption.

I am concerned that the results are somewhat overstated in the conclusions. There was at best partial overlap between the oscillatory power and connectivity results. Likewise, there was only partial overlap between the connections found in the young sample and the older sample.

This is a good point. We have now changed the statements that discuss overlap to make it clear that the overlap is only partial between power and connectivity. Similarly, for the connectivity patterns in the two experiments. We have now added a table that shows that overlap in connections and the overlap in nodes showing connectivity differences.

Some comment on the regions showing the greatest abnormality may be helpful to put the results in context.

These are described in the new Table 2.

Reviewer #3:

This is a follow-up review on the study by Koelewijn et al. that looks at changes in the oscillatory connectivity in resting state MEG in APOE-ɛ4 carriers and matched controls. The authors have, in my opinion replied to the reviewers comments and suggestions in a satisfactory manner, and I am ready to suggest that the manuscript be accepted for publication in eLife. However, there are a few minor language issues that they should fix (see below), to make the text more readable and suitable for a high impact publication.

Before signing off on this review, I would nonetheless like to express that I am somewhat disappointed that the authors, contrary to my suggestion, chose not to change the parcellation template used in the connectivity analysis. I am sympathetic to the authors' argument that the AAL is widely used, but this does mean that it's a good one. Most likely the reason, that is widely used is merely a consequence of the methodological dominance of fMRI where the template is much less problematic. As the authors themselves note in their Discussion, it remains a fact that the AAL is not well-suited for connectivity analysis of MEG data. Bad practices in science do not change unless someone leads the way. Hence, I encourage the authors to at least in their future studies become trail makers that help change suboptimal practises, and thus advancing also the validity of MEG.

Many thanks for this, accepted and taken on board.

https://doi.org/10.7554/eLife.36011.016

Article and author information

Author details

  1. Loes Koelewijn

    Cardiff University Brain Research Imaging Centre, School of Psychology, Cardiff University, Cardiff, United Kingdom
    Contribution
    Formal analysis, Validation, Visualization, Methodology, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7890-171X
  2. Thomas M Lancaster

    1. Cardiff University Brain Research Imaging Centre, School of Psychology, Cardiff University, Cardiff, United Kingdom
    2. Neuroscience and Mental Health Research Institute, Cardiff University, Cardiff, United Kingdom
    3. MRC Centre for Neuropsychiatric Genetics and Genomics, Cardiff University, Cardiff, United Kingdom
    Contribution
    Data curation, Formal analysis, Validation, Investigation, Writing—review and editing
    Competing interests
    No competing interests declared
  3. David Linden

    1. Cardiff University Brain Research Imaging Centre, School of Psychology, Cardiff University, Cardiff, United Kingdom
    2. Neuroscience and Mental Health Research Institute, Cardiff University, Cardiff, United Kingdom
    3. MRC Centre for Neuropsychiatric Genetics and Genomics, Cardiff University, Cardiff, United Kingdom
    Contribution
    Conceptualization, Resources, Data curation, Supervision, Funding acquisition, Project administration, Writing—review and editing
    Competing interests
    No competing interests declared
  4. Diana C Dima

    Cardiff University Brain Research Imaging Centre, School of Psychology, Cardiff University, Cardiff, United Kingdom
    Contribution
    Methodology
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-9612-5574
  5. Bethany C Routley

    Cardiff University Brain Research Imaging Centre, School of Psychology, Cardiff University, Cardiff, United Kingdom
    Contribution
    Data curation, Investigation, Project administration
    Competing interests
    No competing interests declared
  6. Lorenzo Magazzini

    Cardiff University Brain Research Imaging Centre, School of Psychology, Cardiff University, Cardiff, United Kingdom
    Contribution
    Data curation, Investigation, Project administration
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-8934-8374
  7. Kali Barawi

    MRC Centre for Neuropsychiatric Genetics and Genomics, Cardiff University, Cardiff, United Kingdom
    Contribution
    Data curation, Investigation, Project administration
    Competing interests
    No competing interests declared
  8. Lisa Brindley

    Cardiff University Brain Research Imaging Centre, School of Psychology, Cardiff University, Cardiff, United Kingdom
    Contribution
    Data curation, Investigation, Project administration
    Competing interests
    No competing interests declared
  9. Rachael Adams

    Cardiff University Brain Research Imaging Centre, School of Psychology, Cardiff University, Cardiff, United Kingdom
    Contribution
    Investigation, Project administration
    Competing interests
    No competing interests declared
  10. Katherine E Tansey

    Core Bioinformatics and Statistics Team, College of Biomedical and Life Sciences, Cardiff University, Cardiff, United Kingdom
    Contribution
    Resources, Methodology
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-9663-3376
  11. Aline Bompas

    Cardiff University Brain Research Imaging Centre, School of Psychology, Cardiff University, Cardiff, United Kingdom
    Contribution
    Data curation, Investigation, Project administration
    Competing interests
    No competing interests declared
  12. Andrea Tales

    Department of Psychology, College of Human and Health Sciences, Swansea University, Swansea, United Kingdom
    Contribution
    Resources
    Competing interests
    No competing interests declared
  13. Antony Bayer

    School of Medicine, Cardiff University, Cardiff, United Kingdom
    Contribution
    Resources
    Competing interests
    No competing interests declared
  14. Krish Singh

    1. Cardiff University Brain Research Imaging Centre, School of Psychology, Cardiff University, Cardiff, United Kingdom
    2. Neuroscience and Mental Health Research Institute, Cardiff University, Cardiff, United Kingdom
    Contribution
    Conceptualization, Resources, Software, Supervision, Funding acquisition, Validation, Visualization, Methodology, Project administration, Writing—review and editing
    For correspondence
    singhkd@cardiff.ac.uk
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-3094-2475

Funding

BRACE

  • Loes Koelewijn
  • Aline Bompas
  • Andrea Tales
  • Antony Bayer
  • Krish Singh

Medical Research Council (MR/K004360/1)

  • Thomas M Lancaster
  • David Linden
  • Krish Singh

Medical Research Council (MR/K501086/1)

  • Bethany C Routley

Medical Research Council (MR/K005464/1)

  • Bethany C Routley
  • Lorenzo Magazzini
  • Krish Singh

Wellcome (WT105613/Z/14/Z)

  • Katherine E Tansey

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 Mark Woolrich and Giles Colclough for providing the ‘ROInets’ package code for orthogonalisation and variance-normalization of the connectivity Z-scores, and Alexander Shaw for help with atemplate visualisations. We would also like to thank Suresh Muthukumaraswamy for initial versions of the AAL template code.

Ethics

Human subjects: All participants gave informed consent according to the Declaration of Helsinki and all procedures were approved by the Ethics Committee of the School of Psychology, Cardiff University (EC.12.01.10.3071) and the South Wales NHS Research Ethics Committee (10/WSE04/35).

Senior and Reviewing Editor

  1. Huda Y Zoghbi, Texas Children's Hospital, United States

Reviewers

  1. Hanna Renvall, Aalto University, Finland
  2. Allison Nugent
  3. Annika Hulten, Aalto University, Finland

Publication history

  1. Received: February 26, 2018
  2. Accepted: April 17, 2019
  3. Version of Record published: April 30, 2019 (version 1)

Copyright

© 2019, Koelewijn 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

  • 1,333
    Page views
  • 217
    Downloads
  • 0
    Citations

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

Download links

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

Downloads (link to download the article as PDF)

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)