Anatomic & metabolic brain markers of the m.3243A>G mutation: A multi-parametric 7T MRI study.

One of the most common mitochondrial DNA (mtDNA) mutations, the A to G transition at base pair 3243, has been linked to changes in the brain, in addition to commonly observed hearing problems, diabetes and myopathy. However, a detailed quantitative description of m.3243A>G patients' brains has not been provided so far. In this study, ultra-high field MRI at 7T and volume- and surface-based data analyses approaches were used to highlight morphology (i.e. atrophy)-, microstructure (i.e. myelin and iron concentration)- and metabolism (i.e. cerebral blood flow)-related differences between patients (N = 22) and healthy controls (N = 15). The use of quantitative MRI at 7T allowed us to detect subtle changes of biophysical processes in the brain with high accuracy and sensitivity, in addition to typically assessed lesions and atrophy. Furthermore, the effect of m.3243A>G mutation load in blood and urine epithelial cells on these MRI measures was assessed within the patient population and revealed that blood levels were most indicative of the brain's state and disease severity, based on MRI as well as on neuropsychological data. Morphometry MRI data showed a wide-spread reduction of cortical, subcortical and cerebellar gray matter volume, in addition to significantly enlarged ventricles. Moreover, surface-based analyses revealed brain area-specific changes in cortical thickness (e.g. of the auditory cortex), and in T1, T2* and cerebral blood flow as a function of mutation load, which can be linked to typically m.3243A>G-related clinical symptoms (e.g. hearing impairment). In addition, several regions linked to attentional control (e.g. middle frontal gyrus), the sensorimotor network (e.g. banks of central sulcus) and the default mode network (e.g. precuneus) were characterized by alterations in cortical thickness, T1, T2* and/or cerebral blood flow, which has not been described in previous MRI studies. Finally, several hypotheses, based either on vascular, metabolic or astroglial implications of the m.3243A>G mutation, are discussed that potentially explain the underlying pathobiology. To conclude, this is the first 7T and also the largest MRI study on this patient population that provides macroscopic brain correlates of the m.3243A>G mutation indicating potential MRI biomarkers of mitochondrial diseases and might guide future (longitudinal) studies to extensively track neuropathological and clinical changes.

Introduction

The brain contributes to about 20% of the human body's energy consumption (Magistretti and Allaman, 2015; Sokoloff et al., 1977). Mitochondria serve as the biological units that utilize oxygen and glucose through oxidative phosphorylation to produce ATP from ADP (adenosine diphosphate). Approximately 4.7 billion ATP molecules per second are used by a single cortical neuron in a resting human brain, illustrating the extremely high energy demand of the brain (Zhu et al., 2012). As a result, a reduced energy production due to mitochondrial dysfunctioning may affect neuronal and glial integrity, leading to changes in the brain's structure and functioning (Keogh and Chinnery, 2015). For example, using MRI, qualitative cortical, white matter, brainstem and cerebellar morphological changes were observed in patients with respiratory energy chain defects (Saneto et al., 2008). In addition, changes are observed related to the homeostasis of biochemical compounds (such as iron, myelin and calcium), as particularly highlighted by deep gray matter MRI signal intensity changes (Dinopoulos et al., 2005).

Mitochondrial dysfunctioning can be caused by mutations in the mtDNA. Germline mutations can either occur de novo or be inherited from the mother. These mutations can be either homoplasmic, i.e. all mtDNA copies mutated, or heteroplasmic, i.e. a mixture of mutated and wild-type mtDNA. For heteroplasmic mutations, the mutation load may differ among different tissue types. As mitochondria are present in most eukaryotic cell types, mitochondrial mutations lead to a wide-range of symptoms, affecting multiple systems in the body, especially those with a high energy requirement, including the central nervous system (Lightowlers et al., 2015).

One of the most common mtDNA mutations is the m.3243A>G mutation within the mtDNA-encoded tRNA leucine 1 (MTTL1) gene (Goto et al., 1990), affecting proper incorporation of the amino acid leucine in mtDNA encoded proteins. The m.3243A>G mutation causes a broad variety of clinical manifestations, of which the MELAS or MIDD syndromes are the most common (i.e. 10% and 38% of m.3243A>G carriers, respectively). These patients are characterized by a wide range of phenotypic expressions, but most often suffer from sensorineural (due to dysfunctioning cochlear hair cells or nerve) hearing loss (51%), diabetes (42%) and proximal myopathy (27%) (Nesbitt et al., 2013). m.3243A>G specific brain symptoms are (transient) SLEs, visible as T2 MRI signal hyperintensities and characterized by an increased CBF, as measured using ASL. Increased CBF is detected both before (Ikawa et al., 2013) and after onset of SLEs (Li et al., 2017), possibly to counterbalance observed neuronal hyperexcitability and pathology (Iizuka et al., 2002). Since SLEs are an important feature of the brain pathology in m.3243A>G patients, several studies have attempted to characterize the pathogenesis, as well as the temporal and spatial progression of these lesions. For example, these studies have shown locally increased lactate levels (Tsujikawa et al., 2010), but also a reduced apparent diffusion coefficient and cortical laminar necrosis (Iizuka et al., 2003). In addition, increased baseline CBF was observed in m.3243A>G patients without SLEs (Rodan et al., 2015).

However, not all patients fall within the ‘standard’ disease criteria, leading to sub-optimal and delayed diagnosis and prognosis. Several attempts have been made to link m.3243A>G mutation load in different tissue types (e.g. skeletal muscle, blood leukocytes and urine) with disease severity (based on, for example, number of symptoms or severity scores) to predict clinical prognosis (Chinnery et al., 1997; de Laat et al., 2012; Whittaker et al., 2009). Mutation load in UECs was preferred, based on its (1) strong correlation with clinical parameters, (2) stability throughout life and (3) non-invasive assessment. Based on the above criteria, it is evident that the m.3243A>G mutation induces a large diversity of symptoms, including different types of brain changes. While the severity of the phenotype may depend on mutation load, no previous study focussed on the mutation load-brain relationship in m.3243A>G patients. Thus, it is not clear how the mutation load is associated with cortical, subcortical and cerebellar pathologies on microscopic and macroscopic levels and how those relate to the individual's clinical symptoms.

In the present neuroimaging study, we aim to quantify the relationship between the m.3243A>G mutation load percentage and brain phenotype in a large group of patients (N = 22) and healthy controls (N = 15) using UHF 7T MRI. State-of-the-art quantitative brain imaging and analysis approaches are utilized that enable a direct comparison between groups and within patients, with a focus on the brain's morphological (e.g. brain volume and cortical thickness), microstructural-related (T1 and T2* relaxometry) and metabolism-related (CBF) properties. More specifically, we investigate (1) the relationship between (sub-)cortical and cerebellar atrophy with the m.3243A>G genotype (i.e. no mutation vs. low to high mutation load) and (2) the potential differences in myelin- and iron-sensitive T1, T2*, as well as CBF, colocalising with the morphological changes in patients.

The relative gain in SNR at 7T compared to clinical 1.5 T and 3 T depends on the MRI contrast used (Norris, 2003; Pohmann et al., 2016; Vaughan et al., 2001). For CBF using ASL the gain is approximately linear for high-resolution studies (e.g. below 2 mm isotropic). For example, in the study of Ivanov et al. (2017a), 7T outperformed 3 T by a factor of ~1.75 at 1.5 mm in-plane resolution for CBF data. For susceptibility contrasts, such as T2* and functional MRI, the gain can be more than linear. Uludag and Blinder (2017)have theoretically estimated that the SNR of fMRI at high spatial resolution scales approximately quadratically with magnetic field strength. For T1 contrast, however, a signal gain is only expected for very high resolution, higher than typically utilized in in vivo human studies.

Several studies already provided clear evidence of the enhanced detection of disease-specific morphological changes in the brains of, for example, multiple sclerosis and Alzheimer's disease patients using 7T compared to 3 T (Kollia et al., 2009; Nakada et al., 2008). Studies, comparing lower to UHF-MRI for a wide range of clinical applications, have been discussed recently by Trattnig et al. (2016). Here, the improved SNR for 7T enriches the images with spatial details not visible at lower field strengths due to the possibility to acquire data with a higher resolution and CNR, but without increasing the scan time (Duyn et al., 2007). As a result, 7T imaging exhibits increased diagnostic power due to enhanced sensitivity to detect fine-scale structural (Keuken et al., 2017; Stuber et al., 2014), perfusion (Gardener et al., 2009; Ivanov et al., 2017a) and functional (Pfeuffer et al., 2002; Uludag and Blinder, 2017; Yacoub et al., 2008) properties and changes in the brain, see also recent reviews by van der Zwaag et al. (2016) and Ugurbil (2017). As such, the current results contribute to a better understanding of the disease pathology and classification of m.3243A>G patients, in particular, and mitochondrial disease patients, in general.

Materials & methods

Participants

Twenty-two m.3243A>G patients and fifteen healthy age-, gender- and education-matched controls were included in this study after providing written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the ethics review board of the MUMC+ in Maastricht, The Netherlands. We included patients only when they were ADL-independent, as measured using the Barthel disability index and had a score < 30 on the NMDAS (Schaefer et al., 2006). We used three (current function, system specific involvement, and current clinical assessment) out of the four NMDAS sections, which resulted in a spectrum with less severe phenotypes to demonstrate the usefulness of quantitative MRI at 7T for detecting brain changes beyond lesions and atrophy, as it is typically assessed in imaging studies. For this reason, patients in the acute phase and/or with a history of SLEs were excluded. For severe phenotypes (i.e. Barthel index < 15 and/or NMDAS > 30), the brains of m.3243A>G patients strongly deviate from those of age-matched controls, already detectable with standard MRI methods at 1.5 and 3 T. Patients were only eligible for inclusion after evaluation of the clinically-relevant symptoms by an experienced clinician (I.F.M.d.C, see Fig. 1c for an overview of the symptoms). Patients with metallic implants (e.g. cochlear or stents) were excluded, as these would interfere with the magnetic field of the MRI scanner. Assessed parameters were: (1) m.3243A>G mutation loads (%) in blood and UECs following the procedure described in (Sallevelt et al., 2013), (2) fasting blood glucose (mmol/L) and HbA1c (% and IFCC) levels measured using standard procedures and (3) cognitive performance scores. Several cognitive domains were assessed using the visual 15-WLT to test memory, recall and recognition (Van der Elst et al., 2005), the Stroop colour-word test to test attention (Van der Elst et al., 2006b) and the LDST to test information processing speed (Van der Elst et al., 2006a). Raw test scores were Z-scored based on the average control scores for each cognitive task (see Table 1).

Fig. 1

Correlation analysis between clinical and genetic data. Scatter plots show the correlation between m.3243A>G mutation load (%) measured in UECs and blood (A) and the age- and gender-adjusted correlation between mutation load in blood and NMDAS score (B). Filled and empty dots represent Plow and Phigh patients, respectively. Solid lines represent the best fit ±95% CIs. Boldface p-values indicate a significant correlation. Spiderplot visualizes the prevalence of each symptom within both the Plow (solid line) and Phigh (dashed line) m.3243A>G patient groups (C). Diabetes was true in case of known history from physician and HbA1c (IFCC) level higher than 45. Low weight was true in case of a BMI lower than 18.50.

Table 1

Main characteristics of healthy controls and m.3243A>G patients. Values represent mean (±S.D.) if not stated otherwise. Significant differences between groups are indicated in bold-face.

	Controls (n = 15)	m.3243A>G patients (n = 22)	p-Value
Demographics
Age, yr	38.40 (14.24)	41.23 (10.29)	0.487
Sex, % women	73.3	81.8	0.538
BMI, kg/m2	24.43 (4.24)	23.04 (3.59)	0.289
Education, scalea	5.20 (1.21)	5.09 (0.92)	0.838
Glucose status
Fasting glucose, mmol/L	4.99 (0.45)	6.56 (1.70)	0.001b
HbA1c, %	5.02 (0.64)	6.55 (1.24)	<0.001b
HbA1c, IFCC	31.20 (7.21)	45.36 (12.40)	<0.001b
Clinical-relevant symptoms
Mutation load
UECs, %	0	53.14 (26.09)	–
Blood, %	0	20.23 (11.40)	–
Barthel index	–	19.82 (0.83)	–
NMDAS	–	8.50 (4–13)	–
Number of symptoms	0	3.64 (2.46)	–
See Fig. 1C for more details
Cognitive performance
MMSE	29.13 (1.30)	28.27 (2.47)	0.226
LDST, z-score	0 (1.0)	−1.08 (2.18)	0.083
Stroop, z-score
Words only	0 (1.0)	0.62 (1.35)	0.054
Colours only	0 (1.0)	0.95 (1.63)	0.081
Words and colours	0 (1.0)	1.40 (2.89)	0.127
15-WLT, z-score
Total	0 (1.0)	−0.38 (1.04)	0.282
Recall	0 (1.0)	0.01 (0.99)	0.973
Recognition	0 (1.0)	−0.84 (3.01)	0.310

Abbreviations: BMI = body mass index; UEC = urinary epithelial cells; NMDAS = Newcastle Mitochondrial Disease Adult Scale; MMSE = mini-mental state examination; LDST = letter-digit substitution task; 15-WLT = 15-words learning task.

Educational scale ranges from 1 (no education) to 8 (university).

ANOVA, corrected for age, gender and BMI.

Data acquisition

MRI data were acquired using a whole-body 7T magnet (Siemens Medical Systems, Erlangen, Germany) and a 32-channel phased-array head coil (Nova Medical, Wilmington, MA, USA). High resolution (0.7 mm isotropic nominal voxel size) whole-brain quantitative T1 and T2* images were obtained with MP2RAGE (Marques et al., 2010) and ME-GRE sequences, respectively. The SA2RAGE (Eggenschwiler et al., 2012) sequence was used to map B1+ across the brain. In addition to the anatomical scans, baseline partial-brain (excluding cerebellum and lower temporal lobes) perfusion was quantified using FAIR QUIPSS II ASL (Ivanov et al., 2017b). The eye centers were taken as reference for the magnet isocenter position to improve the ASL labelling efficiency. See Supplementary Table 1 for the relevant sequence parameters. In addition, a standard dual-echo field map was acquired to correct the ASL data for EPI readout-related geometrical distortions. Dielectric pads containing a 25% suspension of barium titanate in deuterated water were placed proximal to the temporal lobe areas to locally increase the transmit B1+ field and to improve its homogeneity across the brain (Teeuwisse et al., 2012).

Surface-based analyses

Images were pre-processed to improve subsequent automatic volume- and surface-based processing. An extensive description of the pipeline can be found in the Supplementary Methods 1. The pre-processing included brain masking, post-hoc T1 correction and computation of T2* and ASL maps. Cortical reconstruction and submillimeter volumetric segmentation was then performed with the FreeSurfer (v6.0, http://surfer.nmr.mgh.harvard.edu/) image analysis suite using the pre-processed MP2RAGE UNI images as input (Dale et al., 1999). Manual corrections of the tissue classifications were performed when necessary. Boundary-based registration (i.e. ‘bbregister’) was used to co-register the T2* and CBF maps to the MP2RAGE data with a 6 DOF transformation and spline interpolation. In addition, the fieldmap was used to correct for EPI readout geometrical distortions and improve the co-registration of the CBF map, particularly near the sinuses. Co-registered CBF maps were then corrected for partial volume effects by dividing it with a GM probability map obtained using SPM12 (http://www.fil.ion.ucl.ac.uk/spm). For each subject, all modalities were projected onto the surface using FreeSurfer's mri_vol2surf function by averaging between 20 and 80% of the cortical thickness (with steps of 0.05%) to reduce potential partial voluming with WM and CSF. In addition, WM surface maps were computed by averaging between −0.5 mm and −2 mm distance (with steps of 0.05 mm) from the WM-GM boundary. All surface maps, including surface-based morphology metrics generated by FreeSurfer (e.g. cortical volume and thickness), were coregistered to the ‘fsaverage’ subject using sphere-based alignment (Fischl et al., 1999) and smoothed with FWHM = 10 mm for further statistical analyses. Final surface maps were visualized using the Connectome Workbench v1.2.3 viewer (Washington University School of Medicine, Saint Louis, MO, USA) after conversion of the inflated surfaces and overlays to a compatible format. Non-cortical tissue in between the hemispheres was masked using FreeSurfer's parcellation scheme to avoid inappropriate scaling of the surface maps.

Volume-based analyses

In addition to the surface-based data, volumetric data were assessed for subcortical structures and cerebellum. For the CN, GP and Pu, the automatic subcortical parcellation by FreeSurfer (Fischl et al., 2002) was manually corrected by taking into account the microstructural information (i.e. the values) from both the quantitative T1 and T2* maps using ITK-SNAP v3.6.0 (Yushkevich et al., 2006). In addition, RN, SN and DN were semi-automatically delineated, navigated by a threshold-based approach implemented in ITK-SNAP. The cerebellar segmentation tool (CERES) was utilized to accurately segment the cerebellum into GM and WM (Romero et al., 2017). Final subcortical and cerebellar labels were used to export volume (mm3) and average T1, T2* and CBF values for each of the structures to MATLAB and/or SPSS for further statistical analyses.

Statistical analyses

Descriptive participants' characteristics are reported as mean ± standard deviation (S.D.), while the median and IQR (25th and 75th percentiles) are shown for the NMDAS. Between-group (controls vs. m.3243A>G) demographic characteristics were tested by use of independent 2-sample t-tests (continuous data) or Pearson χ2-tests (categorical data) with SPSS (Statistical Package for Social Sciences, version 23, IBM Corp., USA) and α = 0.05.

For the volumetric data, a MANOVA test with Bonferroni correction was conducted to compare cGM, WM, subcortical GM (global or per structure) and cerebellar GM and WM volumes, T1, T2* and CBF between groups. Each patient was assigned to one of two groups based on mutation load. However, due to differences in mutation load across tissue types (and, in blood, also with age, see for example de Laat et al. (2012)), it is difficult to predefine the pathogenic threshold level (Chomyn et al., 1992) for brain tissue based on the blood and/or UEC data. Therefore, the median was used as classifier within the entire patient population. This was based on pragmatic reasons (i.e. to have equal sub-group sizes with comparable statistical power) in order to perform group-wise analyses (besides correlation). In addition, linear regression analysis was used to explore the relationship between mutation load and the structures volumes, T1, T2* and CBF, and was therefore restricted to m.3243A>G patients. All tests were corrected for age, gender and eTIV.

For the cortical surface data, a two-step analysis approach was used, due to the high clinical variability between the m.3243A>G patients, for detection of (1) within- and (2) between-group brain differences. First, a whole-brain vertex-wise linear regression analysis was performed to detect clusters of vertices, for which cortical thickness (as measure of cortical atrophy), T1, T2* and CBF are significantly negatively or positively correlated with mutation load (in blood) in m.3243A>G patients. Here, age and gender were used as nuisance variables and corrected for multiple comparisons (i.e. number of vertices in left and right hemispheres) using Monte Carlo Null-Z simulation with a voxel-wise threshold of 2.0 (cluster-wise p-value of <0.01). A more lenient cluster-wise p-value of <0.05 was used for T1, T2* and CBF in a separate, additional analysis to also highlight subtle, but still significant, changes as a function of mutation load. This was motivated by the hypothesis that T1, T2* and CBF differences do not per se have to concur with, but can also pre- and/or proceed cortical thickness alterations. Second, significant clusters (per parameter) were then used to compare between groups using an ANOVA test, corrected for age and gender.

Results

The healthy control subjects (N = 15) were matched to the m.3243A>G patients (N = 22) on the basis of age (Pearson χ2 test, p = 0.487), gender (p = 0.538) and educational level (p = 0.838), see Table 1. Significant higher fasting glucose (ANOVA, F1,32 = 11.14, p = 0.002) and HbA1c (F1,32 = 12.75, p = 0.001) levels were detected in m.3243A>G patients compared to controls, after accounting for age, gender and BMI (body mass index). The Barthel index was, except for one patient with a score of 16 (i.e. moderate to good self-supporting), maximal. Scores for the NMDAS ranged between 0 and 26 with a median score of 8.5 (IQR = 4–13).

m.3243A>G clinical characteristics

The mutation load in blood was significantly correlated with the mutation load in UECs (Pearson correlation, r18 = 0.757, p < 0.001, see Fig. 1A) and was on average 2.95 (±1.40) times lower than in UECs. The median m.3243A>G percentage in blood of 23% was used to subdivide patients in Plow (<23%, filled dots in Fig. 1A) or Phigh (≥23%, empty dots). Correlation analyses revealed that both the mutation loads in blood (r18 = 0.473, p = 0.035, see Fig. 1B) and UECs (r18 = 0.550, p = 0.012, not shown) are positively correlated (after correcting for age and gender) with the NMDAS score. Increasing mutation load in blood, not UECs, was correlated with worse cognitive test performances for the LDST (Pearson correlation, r18 = −0.745, p = 0.005) and Stroop tasks (r18 = 0.599, p = 0.025, not shown), corrected for age, gender and education.

Hearing loss was the most prevalent observed symptom within the m.3243A>G patient group (Plow vs. Phigh: 50 vs. 75%), as can be observed in Fig. 1C. This was followed by diabetes mellitus (60 vs. 58.3%), exercise intolerance (30 vs. 75%) and tiredness (40 vs. 66.7%). Migraine (30 vs. 41.7%) and muscle cramps (10 vs. 41.7%) are moderately observed, while cardiomyopathy (10 vs. 25%), underweight (0 vs. 33.3%), cognitive decline (10 vs. 16.7%), epilepsy (10 vs. 8.3%), swallowing problems and SLEs (both 0 vs. 8.3%) are only observed in few patients. There was a significantly higher prevalence of exercise intolerance (Pearson χ2 test, p = 0.035) in the Phigh group compared to the Plow group.

Global brain morphology

Example quantitative T1, T2* and CBF maps for a control subject and an m.3243A>G patient of comparable age (66 vs. 67 yrs. old) are depicted in Fig. 2 across several axial slices. Note that, because we acquired quantitative MRI parameters, we can choose the same absolute scale for all subjects allowing us to examine quantitative MRI differences in addition to local contrast differences, typically indicative of lesions and atrophy. Clear signs of WMLs (see dashed blue box) were visible in both the patient's T1 and T2* maps and were characterized by increased T1 and T2* relaxation times. In addition, a larger sulcal cerebrospinal fluid volume (i.e. increased space between neighboring gyri) was visible for the patient, suggesting gray and/or white matter volume loss. In addition, with respect to the cerebellum, larger inter-folial spaces were visible for several patients (not shown). Possible differences regarding (sub)cortical perfusion (bottom row) were more difficult to detect visually and needed quantitative analysis. Therefore, we examined the normalized T1, T2* and CBF distributions for both cGM and WM tissue. Both cGM and WM T1 showed a significant increase as a function of mutation load (see Supplementary Fig. 1).

Fig. 2

Example MRI data. T1 (msec, A), T2* (msec, B) and CBF (ml/100 g/min, overlaid onto the T1 map, C) maps are shown for a single control subject (left, age = 66 yr, M) and m.3243A>G patient (right, age = 67 yr, F). White matter lesions are highlighted by the dashed blue box and z-coordinates are shown (MNI-space) as reference.

The total volume (in mm3, including left and right hemispheres, corrected for eTIV) was computed for each structure (cGM, ventricles, subcortical GM and cerebellar GM, see Figs. 3A-D, respectively) to characterize the morphological differences between groups. WM and cerebellar WM volumes are shown in Supplementary Fig. 2A and B, respectively. Significantly different brain volumes were detected across groups (MANOVA, Wilks' Lambda = 0.438, F12,54 = 2.298, p = 0.019). Here, the cGM as well as the WM volume were significantly different between groups (F2,32 = 7.783, p = 0.002 and F2,32 = 5.090, p = 0.012, respectively). Similarly, group differences were observed for the ventricles (F2,32 = 4.101, p = 0.026) and subcortical (F2,32 = 5.670, p = 0.008) and cerebellar GM (F2,32 = 8.506, p = 0.001) volumes, but not for the cerebellar WM volume. Significant differences were only found when comparing the control subjects with the Phigh group (p = 0.002 for cGM, p = 0.010 for WM, p = 0.024 for the ventricles, p = .035 for subcortical GM, p = 0.001 for cerebellar GM) but not with the Plow group, except for the subcortical GM (p = 0.021). Significant correlations were observed between mutation load and each structure's volume (Pearson correlation, r18 = −0.282, p = 0.005 for cGM, r18 = −0.228, p = 0.024 for WM, r18 = −0.604, p = 0.008 for the ventricles and r18 = −0.074, p = 0.008 for cerebellar GM) across patients, except for the subcortical GM and cerebellar WM. Note that we did not correct for fasting glucose status, as this did not significantly affect the structures' volume within the patient group. Finally, no differences were observed across groups with regards to eTIV.

Fig. 3

Between and within (i.e. mutation load effect) group comparison of the brain structure's volumes. Bar plots display unadjusted mean (±S.D.) volumes for total gray matter (A), ventricles (B), subcortical (C) and cerebellar (D) gray matter volumes for controls (black) and Plow (filled red) and Phigh (dashed red bars) mutation load in blood. Boldface p-values indicate a significant difference between groups after correction for age, gender and multiple comparisons. Scatterplots show the correlation between volume and mutation load in blood for m.3243A>G patients (Plow: filled and Phigh: empty red dots), adjusted for age and gender. Mean volume for each control subject (black dots) are displayed for reference and red solid lines represent the best fit ±95% CIs. Boldface p-values indicate a significant correlation between mutation load in blood and volume after correcting for age and gender. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fine-scale cortical morphology, microstructure and perfusion

Several cortical ROI clusters of vertices showed a significant (cluster-wise p-value of <0.01) negative correlation between cortical thickness and mutation load in blood, see Fig. 4A (delineated by yellow lines). In general, the clusters encompassed regions predominantly within (respective to the FreeSurfer's annotation) the middle frontal gyri, parietal (including precuneus) and superior temporal lobes. Strikingly, no significantly correlated clusters remained after multiple comparison correction when repeating the same analyses using mutation load in UECs. Similar analyses were performed using T1, T2* or CBF instead of cortical thickness (see Fig. 4B–D, respectively). T1 and T2* increased as a function of mutation load (cluster-wise p-value of <0.05) in pre- and post-central sulci, as well as in the occipital lobes, precuneus and insula for T2*. Increasing CBF was found in the superior frontal sulcus. To visualize the overlap of the findings using the different MRI modalities, significant clusters (vertex-value = 1) were summed and color-coded, ranging between 1 (green)-4 (purple), see Fig. 4E. Most overlap was observed near the central sulci, frontal gyri, occipital lobes and precuneus.

Fig. 4

Surface-based MRI data analysis. Vertex-wise linear regression analyses between mutation load in blood and cortical thickness (A), T1 (B), T2* (C) and CBF (D) within the m.3243A>G patients. Statistical z-maps, overlaid onto the left and right inflated (unfolded) surface reconstructions, show the vertices where cortical thickness (negatively), T1, T2* or CBF (positively) correlated with mutation load (after correction for age and gender). Significant clusters that remained after cluster-wise multiple comparison correction are delineated by yellow borders. Clusters were summarized across parameters by color-coding (1, green, to 4, purple) vertices based on the number of significant clusters (E). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Z-scores were computed per cluster for each parameter and subsequently averaged to compare across groups. Group-wise averages revealed a significant lower cortical thickness in the m.3243A>G patients (including both Plow and Phigh patients, ANOVA, F1,33 = 5.646, p = 0.023) compared to controls (see Fig. 5A). No significant differences were observed across groups for the T1, T2* and CBF data (see Fig. 5B–D), but compared to the controls, lower T1 was observed for Plow, while a slightly higher T1 is observed for Phigh. A similar, but less apparent, pattern was observed for the T2* and CBF data.

Fig. 5

Between group comparison of significant clusters. Dot plots display unadjusted mean (±S.D) cGM thickness (A), T1 (B), T2* (C) or CBF values (D), across corresponding significant clusters in Fig. 4, for controls (black), Plow (filled red) and Phigh (empty red dots). Boldface p-values indicate a significant difference between groups after correction for age, gender and multiple comparisons. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Analogous to cGM, subcortical (i.e. alongside the cGM clusters) WM T1 and T2* were compared across groups (see Supplementary Fig. 2A). WM T1 and T2* were not significantly different across groups. However, WM T2* increased with mutation load in blood within the m.3243A>G patients (Pearson correlation, r18 = 0.454, p = 0.044). Similarly, a trend was observed for WM T1 (r18 = 0.421, p = 0.065).

Subcortical and cerebellar morphology, microstructure and perfusion

Reduced subcortical GM volume was observed for both the Plow and Phigh groups compared to the controls, but no effect of mutation load (see Fig. 3D). Group-averaged volumes for each subcortical structure (see Fig. 6A) individually are plotted (see Fig. 6B) and highlights significantly reduced volume in the m.3243A>G patients compared to controls (MANOVA, Wilks' Lambda = 0.616, p = 0.025) for the Pu (F1,33 = 8.959, p = 0.005) and GP (F1,33 = 4.621, p = 0.039). No clear microstructural changes, i.e. between groups or as function of mutation load, in the subcortical structures were present. However, several patients showed signs of spongiotic lesions in the Pu and GP that were characterized by high T1 and T2*, locally (see for example Fig. 6A for one subject). CBF increased with mutation load (Pearson correlation, r18 = 0.489, p = 0.029).

Fig. 6

Between and within group comparison of subcortical structures. ROIs (1 = RN, 2 = SN, 3 = CN, 4 = Pu, 5 = GP and 6 = DN) are displayed schematically in A. Mean subcortical gray matter volume (B) and T1, T2* and CBF (C) are compared across groups (controls: filled black, Plow: filled red and Phigh: dashed red bars) per ROI. Scatterplots display the correlation between mean (z-scored relative to controls and averaged across ROIs) T1, T2* and CBF vs. mutation load in blood (D). Mean values for controls (black dots) are displayed for reference and red solid lines represent the best fit ±95% CIs. Boldface p-values indicate a significant correlation between mutation load in blood and cortical thickness, T1, T2* or CBF after correction for age and gender. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

More pronounced effects were observed for cerebellar GM (see Fig. 7). Here, T1 was significantly higher for m.3243A>G (ANOVA, F1,33 = 12.947, p = 0.001) and especially for the Phigh group compared to controls (F2,32 = 9.915, p < 0.001, see Fig. 7A, top row). Moreover, T1 increases as a function of mutation load (r18 = 0.485, p = 0.030, see Fig. 7B). In contrast, T2* was comparable across groups and did not correlate with mutation load. Similar but fewer significant findings were shown for cerebellar WM (see Supp. Fig. 2B).

Fig. 7

Between and within group comparison of cerebellar GM. Dot plots in A display unadjusted mean (± S.D.) cerebellar GM T1 (top) or T2* (bottom) values for controls (black), Plow (filled red) and Phigh (empty red dots). Boldface p-values indicate a significant difference between groups after correction for age, gender and multiple comparisons. Scatterplots in B show the correlation between T1 or T2* and mutation load in blood for m.3243A>G patients (Plow: filled and Phigh: empty red dots), adjusted for age and gender. Mean T1 or T2* for each control subject (black dots) are displayed for reference and red solid lines represent the best fit ±95% CIs. Boldface p-values indicate a significant correlation between mutation load in blood and cerebellar T1 or T2* after correction for age and gender. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Discussion

Mitochondrial dysfunctioning, either induced by acquired damage throughout life (such as by nucleases and reactive oxygen species) and/or by de novo or maternally inherited germline variations of the mtDNA, has been consistently linked to neurodegeneration (Keogh and Chinnery, 2015). While several mtDNA mutations are known to affect the brain (Reeve et al., 2008), the current study focussed on the effects of the m.3243A>G genotype (i.e. mutation load) on the brain tissue using multiple parametric maps acquired with 7T MRI and compared these data to values obtained on healthy controls.

Mutation load and clinical characteristics

In the current study, we assessed mutation loads in both UECs and blood. Despite the mutation load in blood (leukocytes) being significantly lower compared to that in UECs, albeit comparable to that observed in earlier studies (Fayssoil et al., 2017; Shanske et al., 2004; Whittaker et al., 2009), both measures strongly correlated (de Laat et al., 2012; Frederiksen et al., 2006). It is known that the mutation load in blood decreases over lifetime with a constant rate of 1.4% per year (Rahman et al., 2001), while the mutation load in UECs is considered to be more consistent with age. Here, the loss of the m.3243A>G mutation in blood presumably originates from a selection against pathogenic mtDNA mutations in a stem cell population (Rajasimha et al., 2008).

Hearing loss was the most prevalent symptom within the current study population, followed by exercise intolerance and diabetes mellitus. This matches with the MELAS and MIDD-like phenotypes associated with the m.3243A>G mutation and is in line with previous findings, reporting comparable numbers across their study cohorts (de Laat et al., 2012; Fayssoil et al., 2017; Nesbitt et al., 2013). In addition, we found a significant correlation between mutation load (in blood and UECs) and NMDAS score, as also shown in previous studies (de Laat et al., 2012; Whittaker et al., 2009). The observed differences between the Plow and Phigh groups in terms of symptoms prevalence, especially in the myopathy-related symptoms (e.g. exercise intolerance), can partly be explained by the “mitochondrial threshold effect” (Rossignol et al., 2003). The significant correlation between the proportion of apoptotic muscle fibers and the proportion of mutant mtDNA exemplifies this hypothesis (Aure et al., 2006). It is known that the mitochondrial copy number can compensate and can be increased by physical exercise, implying the mutation load is not the only factor involved. Like muscle cells, neurons are post-mitotic and therefore are likely to behave similarly. Comparable mutation loads were detected across several brain regions to those in skeletal muscle. However, it is not known how the mutation load in neurons develops over lifetime (Betts et al., 2006). We have found that the mutation load in blood is a better marker for cognitive performances and brain functioning in m.3243A>G patients than the mutation load in UECs, even though both mutation loads are correlated to each other. The reason for this discrepancy is unknown, but may be specific to the age-range of the current patient population.

Brain morphological markers of the m.3243A>G mutation

The current results present strong evidence of a mutation load effect within the patient population on global brain volume, in addition to widespread lower brain volume in the patients compared to controls. While patients with a low mutation load are characterized by normal-appearing brains, stronger brain atrophy is observed in patients with higher mutation load. This was especially apparent in the cortical and cerebellar GM, but not for the subcortical structures, which is in accordance with biochemical analyses of post-mortem brains (Sparaco et al., 2003). Strongest local cGM atrophy and reduced cortical thickness as a function of m.3243A>G mutation load was observed in regions linked to: (i) auditory processing (superior temporal cortices), (ii) attentional control (middle frontal and parietal gyri and anterior cingulate cortices), (iii) sensorimotor network (central sulci and neighboring gyri) and (iv) default mode network (precuneus). These latter two structures were also observed by a recent study that used voxel-based morphometry to characterize morphological differences between m.3243A>G patients and healthy controls (Tsujikawa et al., 2016). Notably, significant local correlations, after correction for multiple comparisons, between cortical thickness and mutation load were only found for values from the blood.

The neuropathological effects of the mutation load in some brain areas, but not in others, signify the differential abilities of neurons to survive respiratory chain deficiencies (Betts et al., 2006). More work has to be performed, particularly using invasive cellular methods, to establish the link between mitochondrial dysfunction and macroscopic brain changes, as detected with non-invasive MRI. Nevertheless, the observed brain patterns can be readily related to the different clinical symptoms often observed in m.3243A>G patients, such as the currently high prevalence of hearing impairment. Since auditory brainstem responses are often preserved, it is likely of cochlear origin (Kullar et al., 2016). In addition, the severity of the hearing loss correlates with the mutation load, being worse in patients with a higher mutation load (Chinnery et al., 2000). Therefore, decreasing cortical thickness and atrophy in the regions linked to auditory processing (auditory cortex, planum temporale and superior temporal sulcus and gyrus) with increasing m.3243A>G mutation load, are most likely caused by impaired cochlear functioning (Eckert et al., 2012; Formisano et al., 2008; Syka, 2002).

Even though the cortical thinning in the auditory regions may predominantly occur after cochlear damage, the morphological differences in the other affected brain regions may be induced through direct interferences of the mutation on neuronal integrity through, for example, neuronal hyperexcitability (Iizuka et al., 2003; Iizuka et al., 2002). Impaired mitochondrial protein (i.e. OXPHOS subunits) synthesis due to a defective tRNALeu(UUR) has been shown in the cerebral and cerebellar cortex of post-mortem m.3243A>G brains (Sparaco et al., 2003). Here, changes were mostly found in the intermediate layers of cGM regions, with a higher degree of cellular dependence on oxidative metabolism, and in the cerebellar Purkinje cells, respectively. As such, neuronal changes may occur especially in brain regions characterized by a high (baseline) energy demand and/or functional connectivity or cerebellum (i.e. metabolic hypothesis), depending on the mutation load similar to that observed in skeletal muscle cells (Aure et al., 2006). These high energy consuming brain areas include regions related to the default mode, dorsal attention and sensorimotor networks (Tomasi et al., 2013), in agreement with our findings of atrophy in these regions, such as in the precuneus, middle frontal gyrus and central sulcus (somatosensory and motor cortices), respectively. The cortical thinning observed in regions associated with attentional control/executive functioning, such as the middle frontal and parietal gyri and anterior cingulate cortices (Milham et al., 2002), are consistent with behavioral symptoms, e.g. patients with a higher mutation load performed worse on the Stroop task. In particular, OXPHOS deficits-susceptible inhibitory interneurons may play a prominent role in the cognitive dysfunctioning observed in mitochondrial disease patients (Finsterer, 2012; Lax et al., 2016).

Comparable to the cGM atrophy, cerebellar GM atrophy was most pronounced in the high mutation load group compared to patients with a lower mutation load, quantified by a volume loss or illustrated by larger inter-folial spaces. Previously, studies have shown that the cerebellum is highly vulnerable to mitochondrial dysfunctioning (Scaglia et al., 2005) due to respiratory chain deficiencies of complex I and IV, leading to cerebellar atrophy by Purkinje and granule cell loss (Lax et al., 2012; Sparaco et al., 2003). This is consistent with the fact that cerebellar ataxia is commonly described in mtDNA patients (de Laat et al., 2012; Lax et al., 2012). As for cerebellar GM, smaller subcortical volume was observed for the patients compared to the healthy controls, in particular in the Pu and GP (i.e. lentiform nucleus), structures especially involved in motor execution (Middleton and Strick, 2000). The volume loss correlates with the described neuronal vacuolization (i.e. spongiosis) in both these structures for several patients, potentially resulting from abnormalities within the vascular smooth muscle and endothelial cells (i.e. vascular hypothesis, Betts et al. (2006)) or from astroglial pathogenesis (Bugiani et al., 2017; van der Knaap et al., 2012). However, in contrast to the cortical and cerebellar GM, the mutation load did not affect the degree of subcortical atrophy. This may implicate that the causal mechanism is different compared to that underlying cortical atrophy (e.g. vascular vs. metabolic hypotheses). Generally, by taking together the cortical atrophy in sensorimotor regions, subcortical (i.e. lentiform nucleus) and cerebellar GM volume losses, the current results suggest a global brain-wide impaired motor functioning (Seidler et al., 2015). These changes are excellent brain correlates to some of the symptoms often observed in m.3243A>G and mitochondrial disease patients, such as cerebellar (and/or sensory) ataxia and gait instability (de Laat et al., 2012).

Cortical microstructure- and metabolic-related markers of the m.3243A>G mutation

Besides macroscopic morphological differences, T1, T2* and CBF significantly varied (i.e. increased) as a function of mutation load in several cGM regions. Differences in quantitative T1 and T2* across the brain are predominantly attributed to varying myelin and iron content between brain structures and regions (Stuber et al., 2014). In line with earlier work that showed delayed or decreased myelination in patients affected by respiratory chain defects (Dinopoulos et al., 2005; Lax et al., 2016; Sofou et al., 2013), positive correlations between T1 and T2* values with mutation load within the patient population were observed, in particular in the primary- and heavily-myelinated regions of the cGM, including the somatosensory- and motor cortices (Haast et al., 2016), but also in cerebellar GM. In general, increased T1 and T2* implicate lower levels of myelin (Stuber et al., 2014). Decreased myelin synthesis, due to the defective cellular energy production during the critical time of myelin development in early life, may have resulted in the increasing T1 and T2* with mutation load. Furthermore, we observed higher CBF in the subcortical structures as well as in the superior frontal gyrus, for patients with a higher mutation load. These effects could be induced (i) by the altered vascular functionality (e.g. related to the subcortical spongiosis and vascular abnormalities), (ii) by a compensatory mechanism for the impaired neuronal energy production or (iii) as an attempt to reduce the increased brain lactate (or other potentially harmful compounds) levels, as argued by Rodan et al. (2015).

However, when the patient populations are compared against controls, surprisingly, the metabolic-, but also microstructural-related characteristics do not follow the expected linear change as a function of mutation load, but a U-shaped dependency across groups. That is, MRI values over the entire cortex of Phigh patients and healthy subjects are similarly different than those of the Plow patients. Two hypotheses could explain this observed pattern: i) the dip in cortical CBF, T1 and T2* values for the Plow patients may signify that certain compensatory mechanisms are initiated only in the Phigh patients (i.e. mitochondrial threshold effect). That is, for the Plow group, there is only a mild loss in e.g. myelination, which does not evoke/necessitate activation of local molecular pathways to compensate for this loss. However, for the Phigh group, compensatory mechanisms are necessary to maintain functioning of the respective brain area. In fact, delayed myelination in the Phigh group, as described above, may have led to global myelination levels similar to control ranges, as suggested by the correspondence between the range of T1 and T2* values observed in the group-wise comparison between controls and patients, confirming earlier claims (Dinopoulos et al., 2005; Sofou et al., 2013). Alternatively, ii) T1, T2* and brain metabolic changes have many biochemical and microscopic contributions, which do not have to follow the same dependency on mutation load. For example, T1 increases as a function of both myelin and iron. That is, a decrease in myelin has an opposing effect on the resulting T1 values than an increase in iron content, and the latter change is a common observation in many neurodegenerative diseases. Thus, it is reasonable to assume that myelination decreases and iron content increases as a function of mutation load but they have different dependencies on the m.3243A>G mutation load, leading to the U-shape function observed pattern for T1. In contrast to the observed cortical pattern, the increased cerebellar GM T1 observed in patients compared to controls, in addition to the unchanged T2*, may implicate predominantly myelin loss secondary to neuronal and axonal loss (Lax et al., 2012; Virtanen et al., 2011).

Interestingly, T1 and T2* differences in the subcortical structures were absent, which is indicative for subcortical biochemical changes that may be more specific for certain mtDNA mutations (Barragan-Campos et al., 2005; Bindu et al., 2015). This is in sharp contrast to previously observed subcortical T1w and T2w changes (Tschampa et al., 2013). However, it is important to note that most of the available MRI studies on mitochondrial diseases, if not all, employed data acquired using conventional clinical, non-quantitative MRI approaches, such as FLAIR. These do not provide quantitative markers of the underlying biochemical composition of the tissue, as the MRI signal intensity are additionally influenced by MRI sequence parameters and the hardware setup (such as transmit and receive imaging coils) (Bock et al., 2013; Lorio et al., 2016). Moreover, clinical scans typically have anisotropic voxel sizes, increasing partial voluming effects and mixing signals originating from WM and GM or spongiotic tissue and subcortical GM. Therefore, quantitative interpretation of data acquired for diagnostic purposes is challenging, as interpretations of such data contributions are confounded by many other MRI acquisition and physiological parameters (e.g. transmit and receive RF fields, proton density, relaxation rates, macromolecule concentrations (Haast et al., 2016)). In summary, these results exemplify the benefits of a multi-parametric MRI protocol, as they allow discerning the biochemical causes of brain changes in disease and reliable intra- and inter-subject comparisons.

Conclusions

In conclusion, the current work utilized high-resolution MRI data acquired using 7T MRI to quantify brain changes in a representative population of m.3243A>G patients. The data revealed neuroradiological (i.e. morphology, microstructure- and metabolism-related) changes in patients with varying levels of mutation load, including those with a mutation load lower than previously observed threshold levels, compared to healthy controls (see Supplementary Table 2 for an overview). Interestingly, mutation load in blood compared to UECs provides a better estimate of the brain changes. The observed changes are widespread, but mainly affect GM, and are consistent with the different clinical symptoms. However, the underlying biological mechanisms and timing may vary across brain structures/regions. Our results, thus, indicate the value of quantitative MRI at 7T to determine pathological effects of the m.3243A>G mutation on brain structure and function for various mutation loads. Future longitudinal quantitative MRI but also in vitro studies are necessary to track the progression of the detected MRI-based biomarkers to improve the understanding of the disease pathogenesis and clinical symptoms development.