Abstract
Mixed pathologies are common in neurodegenerative disease; however, antemortem imaging rarely captures copathologic effects on brain atrophy due to a lack of validated biomarkers for non-Alzheimer’s pathologies. We leveraged a dataset comprising antemortem MRI and postmortem histopathology to assess polypathologic associations with atrophy in a clinically heterogeneous sample of 125 human dementia patients (41 female, 84 male) with T1-weighted MRI ≤ 5 years before death and postmortem ordinal ratings of amyloid-$$mathtex$${\bi \beta }$$mathtex$$, tau, TDP-43, and $$mathtex$${\bi \alpha }$$mathtex$$-synuclein. Regional volumes were related to pathology using linear mixed-effects models; approximately 25% of data were held out for testing. We contrasted a polypathologic model comprising independent factors for each proteinopathy with two alternatives: a model that attributed atrophy entirely to the protein(s) associated with the patient’s primary diagnosis and a protein-agnostic model based on the sum of ordinal scores for all pathology types. Model fits were evaluated using log-likelihood and correlations between observed and fitted volume scores. Additionally, we performed exploratory analyses relating atrophy to gliosis, neuronal loss, and angiopathy. The polypathologic model provided superior fits in the training and testing datasets. Tau, TDP-43, and $$mathtex$${\bi \alpha }$$mathtex$$-synuclein burden were inversely associated with regional volumes, but amyloid-$$mathtex$${\bi \beta }$$mathtex$$ was not. Gliosis and neuronal loss explained residual variance in and mediated the effects of tau, TDP-43, and $$mathtex$${\bi \alpha }$$mathtex$$-synuclein on atrophy. Regional brain atrophy reflects not only the primary molecular pathology but also co-occurring proteinopathies; inflammatory immune responses may independently contribute to degeneration. Our findings underscore the importance of antemortem biomarkers for detecting mixed pathology.
Significance Statement
Gross brain atrophy is an essential marker of neurodegenerative disease, relied on both in quantitative research and clinical assessment. Recent autopsy studies show that dementia patients typically have more than one form of neuropathology in their brains, but atrophy is typically ascribed to the presumed pathology associated with patients’ primary diagnosis, without consideration of copathologies. The present study related postmortem histopathology to antemortem MRI collected close to death in a clinically heterogeneous sample of dementia patients. We found that polypathology was frequent at the regional level, particularly in limbic cortex. Moreover, we report a significant mediating effect for gliosis on the relationship between protein accumulation and atrophy, suggesting a role for inflammatory immune responses in driving neurodegeneration.
Introduction
Clinical and research guidelines specify brain-imaging results as supportive features in diagnosis of neurodegenerative diseases (McKhann et al., 2011; Höglinger et al., 2017); and atrophy observed on MRI or CT is one of the most common imaging markers employed in clinical evaluation and clinical trials. Understanding the relationship between neurodegenerative pathologies and atrophy is thus important for accurate clinical diagnosis and development of therapeutic strategies. However, patients often have multiple co-occurring pathologies at autopsy (Robinson et al., 2018; Karanth et al., 2020; Nelson et al., 2022), and a monopathologic perspective may overlook comorbid factors that contribute to patients’ degeneration and cognitive decline. For example, recent work shows that transactive response DNA-binding protein of 43 kDA (TDP-43) pathology, manifesting as limbic-predominant age-associated encephalopathy (LATE), frequently co-occurs with Alzheimer’s disease (AD) neuropathologic change in population-based cohorts (Nelson et al., 2022). Both pathologies are reported in patients with amnestic symptoms and have overlapping anatomic distributions (particularly medial temporal lobe limbic structures) (Nelson et al., 2022). Several recent studies have highlighted associations between TDP-43 pathology and medial temporal lobe atrophy (Josephs et al., 2017; Wisse et al., 2021). Likewise, $$mathtex$$\alpha$$mathtex$$-synuclein commonly co-occurs with AD pathology and other forms of TDP-43 (James et al., 2016; Spotorno et al., 2020), creating uncertainty about which pathologies may drive neurodegeneration. Even in “pure” AD—a dual proteinopathy—debate has persisted about the relative contributions of amyloid-$$mathtex$$\beta$$mathtex$$ and tau to neurodegeneration, although recent empirical results support tau as the most direct driver of atrophy in AD (Iaccarino et al., 2018), while amyloid-$$mathtex$$\beta$$mathtex$$ appears to have an indirect association via promotion of tau propagation (Busche and Hyman, 2020; Lee et al., 2022). Finally, the effect of pathologic proteins on brain structure is likely mediated by sequelae including neuron loss (Schuff et al., 1997) and inflammatory processes that exacerbate degeneration. In reactive gliosis, astrocytes can undergo both hypertrophy and atrophy (Garwood et al., 2017); and astrocytes, oligodendrocytes, and microglia may further the propagation of tau (Maphis et al., 2015; Amro et al., 2021; Ferrari-Souza et al., 2022). Vascular pathologies such as cerebral amyloid angiopathy (CAA), which is associated with impaired clearance of amyloid-$$mathtex$$\beta$$mathtex$$ (Greenberg et al., 2020), may also contribute to brain atrophy (Smith, 2018).
There is thus a clear need to better understand pathologic basis of brain atrophy. However, antemortem investigations of polypathology are limited by a lack of validated biomarkers: while amyloid-$$mathtex$$\beta$$mathtex$$ and tau burden can be visualized from positron emission tomography (PET) or estimated globally from biofluid assays, development and validation of markers for TDP-43 and $$mathtex$$\alpha$$mathtex$$-synuclein is ongoing (Fairfoul et al., 2016; Brendel et al., 2020; Alzghool et al., 2022). We interrogated associations between regional polypathologic burden and atrophy in a sample of 125 neurodegenerative disease patients with antemortem structural MRI and postmortem ordinal ratings of amyloid-$$mathtex$$\beta$$mathtex$$, tau, TDP-43, and $$mathtex$$\alpha$$mathtex$$-synuclein from the brain bank of the University of Pennsylvania’s Center for Neurodegenerative Disease Research (CNDR). We sought to estimate the relative contributions of these pathologies to regional atrophy and demonstrate the importance of disease models that account for copathologies. We hypothesized that polypathology would be associated with more severe neurodegeneration, and that modeling independent effects of these proteinopathies would better explain observed brain atrophy than monopathologic models. In secondary analyses, we tested the hypothesis that non-proteinopathic disease markers—including gliosis, neuronal loss, and angiopathy—would mediate the effect of neurodegenerative proteinopathies and improve modeling of atrophy.
Materials and Methods
Participants
We retrospectively selected participants with one or more autopsy-confirmed neurodegenerative pathologies from the University of Pennsylvania’s Integrated Neurodegenerative Disease Database (Fig. 1). Patients were clinically diagnosed during life according to published criteria (McKhann et al., 2011; Rascovsky et al., 2011; Armstrong et al., 2013; Postuma et al., 2015; Crutch et al., 2017; Höglinger et al., 2017) by board-certified neurologists from the Penn Frontotemporal Degeneration Center, Alzheimer’s Disease Research Center, Movement Disorder Center, or Comprehensive Amyotrophic Lateral Sclerosis Center and provided informed consent to research as required by the University of Pennsylvania Institutional Review Board. All patients had antemortem T1-weighted MRI and a primary postmortem diagnosis of one of four proteinopathies, including AD, TDP-43, frontotemporal lobar degeneration (FTLD) due to tau, or $$mathtex$$\alpha$$mathtex$$-synuclein. AD was defined by intermediate or high levels of AD neuropathologic change (Montine et al., 2012); $$mathtex$$\alpha$$mathtex$$-synucleinopathies included Parkinson’s disease and Lewy body disease; TDP-43 included cases with LATE, FTLD with TDP-43 inclusions, and amyotrophic lateral sclerosis (ALS) with frontotemporal dementia (FTD); and FTLD-tauopathies included progressive supranuclear palsy, corticobasal degeneration, Pick’s disease, argyrophilic grain disease, and FTD with parkinsonism linked to chromosome 17 (Table 1). We identified 10 regions of interest with the most frequent coverage across all cases in the CNDR brain bank; participants were excluded from the current sample if they lacked ordinal pathology ratings for some or all regions. Because the sampled regions did not include primary motor cortex, participants who presented with ALS without cognitive impairment were excluded. Participants were also excluded if their autopsy showed no/low levels of AD pathologic change with no co-occurring pathologies. To ensure high-quality MRI data acquired close in time to autopsy, we further restricted the sample to those with MRI collected since 2004 and within 5 years of death. This 5 year cutoff was motivated by the logic of the study design, which assumes that postmortem ratings of proteinopathies reflect their approximate relative burden and contribution to neurodegeneration at the time of MRI. Five years was chosen as a compromise between restricting the antemortem interval to ensure that assumption was met versus maximizing sample size. We felt this value was reasonable, given evidence that many age-related neurodegenerative pathologies develop years before symptom onset (Buchhave et al., 2012; Jack et al., 2013; Chen et al., 2019; Chen and Kantarci, 2020; Hansson, 2021). Moreover, atrophy rates may decrease in advanced disease, as pathologic burden plateaus and relatively little tissue remains (Sabuncu et al., 2011; Insel et al., 2015), suggesting that atrophic effects may be mostly realized months or years before death. Figure 2 illustrates co-occurrence of pathologic diagnoses, ordered by frequency. Copathologies included diagnoses not specifically selected for or excluded, such as secondary/tertiary cerebrovascular disease. Analyses of variance across pathology types (AD, TDP-43, FTLD-tau, and $$mathtex$$\alpha$$mathtex$$-synuclein) showed no differences in education (p = 0.94) or age at onset (p = 0.19). Age at death had a marginally significant difference across groups [F(3, 121) = 2.38, p = 0.073], reflecting heterogeneity in rates of disease progression (Table 1). The ratio of male to female participants also differed by diagnosis ($$mathtex$$\chi ^2$$mathtex$$ = 9.85, p = 0.019), likely reflecting the male bias among patients with Lewy body spectrum disorders.
Flowchart of participant selection steps. Numbers indicate the count of cases retained at each step. AD, Alzheimer’s disease; ALS, amyotrophic lateral sclerosis; LATE, limbic-predominant age-associated TDP-43 encephalopathy; FTLD, frontotemporal lobar degeneration; TDP, transactive response DNA-binding protein of 43 kDa (TDP-43).
Co-occurrence of neuropathologic diagnoses in MRI sample. Top: histogram shows the number of cases across the training and testing datasets with the combination of pathologies indicated in the bottom portion of the figure. Bottom: Horizontally oriented histogram at the left shows the frequency of each individual neuropathologic diagnosis. The black dots and connecting lines show each unique combination of diagnoses that occurs in the dataset. TauNOS, tauopathy, not otherwise specified; FTDP17, frontotemporal dementia with parkinsonism linked to chromosome 17; AGD, argyrophilic grain disease; CAA, cerebral amyloid angiopathy; ALS, amyotrophic lateral sclerosis; PiD, Pick’s disease; CBD, corticobasal degeneration; HS, hippocampal sclerosis; CVD, cerebrovascular disease; PSP, progressive supranuclear palsy; PART, primary age-related tauopathy; LATE, limbic-predominant age-associated TDP-43 encephalopathy; FTLD-TDP, frontotemporal lobar degeneration due to TDP-43; LBD, Lewy body disease; AD, Alzheimer’s disease.
Values for age and education represent group medians [interquartile ranges]
Autopsy and histopathologic rating methods
At autopsy, a single hemisphere was selected at random for histopathology. Tissue was fixed overnight in neutral buffered formalin or ethanol and embedded in paraffin for preparation of sections. Regional tissue sampling was performed as previously described (Toledo et al., 2014), and immunohistochemistry was performed for tau (mouse antibody PHF1), amyloid-$$mathtex$$\beta$$mathtex$$ (NAB228), $$mathtex$$\alpha$$mathtex$$-synuclein (Syn303), and TDP-43 (rat antibody 1D3) (Robinson et al., 2021). Regional tau measures did not differentiate AD- and FTLD-associated tau; while the current study thus examines general effects of tauopathy, it cannot estimate independent effects of 3-repeat, 4-repeat, and mixed 3/4-repeat tau. For cases with FTLD-tau pathology, global Braak staging was performed based on a GT-38 stain for AD-specific paired helical filament tau (Gibbons et al., 2019). Pathologic diagnoses were rendered according to published criteria (Cairns et al., 2007; Hyman et al., 2012; McKeith et al., 2017; Nelson et al., 2019) by trained neuropathologists in the Center for Neurodegenerative Disease Research. Hematoxylin and eosin histology was performed for the regional semi-quantitative assessment of neuronal loss and gliosis. Loss of hematoxylin was the primary determiner of neuronal loss. Gliosis was determined by an increase in both hematoxylin and eosin-positive (primarily astrocytic) glia. Expert raters assigned ordinal ratings in regions of interest for each pathology on a 0–3 scale (0 = none; 1 = low; 2 = intermediate; 3 = severe). Values of “rare” or “sparse” were recoded as zeros. We restricted analysis to pathologic regions of interest that had complete ordinal ratings for amyloid-$$mathtex$$\beta$$mathtex$$, tau, $$mathtex$$\alpha$$mathtex$$-synuclein, and TDP-43 in more than 50% of cases in the CNDR database, including superior/middle temporal and angular gyri; middle frontal, cingulate, entorhinal cortices; and amygdala, hippocampal CA1/subiculum, caudate/putamen, globus pallidus, and thalamus/subthalamus. The substantia nigra met this criterion; however, it was excluded from analysis because it could not be reliably segmented on available T1-weighted MRI. In the secondary analyses, we examined the effects of regional angiopathy, gliosis, and neuronal loss (rated on the same 0–3 scale); values for these exploratory variables were missing in 71 out of 1,216 observations (6.2%) over 26 participants.
Neuroimaging methods
T1-weighted MRI was acquired on a Siemens 3-Tesla scanner outfitted as a TIM Trio (n = 70) and subsequently as a Prisma Fit (n = 26); on an alternate TIM Trio (n = 8); and on Siemens and GE 1.5-Tesla scanners (n = 18 and n = 3, respectively). Bias correction, brain extraction, and tissue segmentation were performed as described (Phillips et al., 2019), and intracranial volume was estimated using a multi-atlas approach (Xie et al., 2019). Gray matter volumes were quantified for cortical regions in the 250-label parcellation of Cammoun et al. (2012) and subcortical regions in the BrainColor atlas (Fortin et al., 2018). Imaging regions representing postmortem sampling (Fig. 3) were selected based on consensus between neuropathologists and imaging experts (Spotorno et al., 2020). Volumes for superior and middle temporal gyri were summed to represent the superior/middle temporal autopsy region; similarly, caudate and putamen volumes were summed to represent the corresponding autopsy region. Whole hippocampal volume represented the CA1/subiculum region, as available T1-weighted MRI did not permit segmentation of hippocampal subfields, and whole thalamus volume represented the postmortem thalamus/subthalamus label. These volumetric measures were converted to standardized scores adjusting for age, sex, and intracranial volume (i.e., W-scores) based on regression models computed for each region in an independent sample of 289 T1-weighted scans from 147 cognitively normal adults (median age = 66 years; interquartile range = 59–71 years; 45.6% female). Negative standardized volume scores thus indicate greater-than-expected atrophy relative to typical aging; positive scores indicate greater-than-expected volumes.
Top: overview of the study analysis. Equations describe linear mixed-effects (LME) regression models relating pathology scores to standardized regional volume scores. Bottom: regions included in joint imaging/pathology analysis. GM, gray matter; β0, model intercept; βi, random intercept per participant; ε, model error.
Experimental design and statistical analyses
The dataset was divided into training and testing splits in a 75:25% ratio (Table 1) using random sampling stratified for sex and primary pathology type (AD, FTLD-tau, TDP-43, or $$mathtex$$\alpha$$mathtex$$-synuclein). Two-sample t tests and chi-square tests were used to assess potential differences between the train and test splits for each pathology type on age of onset, age at death, sex ratios, race, education, and levels of AD neuropathologic change. Outlier correction was performed using W-scores at the 1st and 99th percentiles in the training dataset to replace scores above and below those values, respectively. By this procedure, 10 out of 918 observations in the training data and 5 out of 328 observations in the test dataset that were below the 1st percentile were replaced with a value of −5.74; 10 out of 918 training data points and 2 out of 328 test data points above the 99th percentile were replaced with a value of 2.81.
To examine polypathologic associations with regional atrophy W-scores in the training dataset, we used a linear mixed-effects model including ordinal scores for the four proteinopathies of interest (amyloid-$$mathtex$$\beta$$mathtex$$, tau, TDP-43, and $$mathtex$$\alpha$$mathtex$$-synuclein); additional fixed effects for disease duration at MRI (defined as the interval between symptom onset reported by patient or caregivers and the scan date), number of years from MRI to death, and MRI scanner; and a random intercept per participant (Fig. 3). Age of onset was missing for one participant with a primary neuropathologic diagnosis of AD and a secondary diagnosis of amygdala-predominant Lewy body disease. For this participant, disease duration at MRI was imputed as the mean duration of the remaining participants (5.41 years). Education level for one individual with a primary pathologic diagnosis of corticobasal degeneration was imputed as the median of the remaining participants. We compared observed F statistics using marginal sums of squares for each proteinopathy to null distributions obtained by randomly permuting each pathology score 10,000 times; significance was assessed at $$mathtex$$\alpha \,$$mathtex$$ = 0.05. This approach allowed us to determine the information about gray matter atrophy conveyed by each proteinopathy, relative to null models of the same complexity based on pathology score distributions with identical statistical properties as the observed data.
To further assess the effect of modeling copathologies, we compared the polypathologic model with one that included only the protein (s) corresponding to a patient’s primary neuropathologic diagnosis; proteins corresponding to copathologies were set to zero. Thus, for patients with a primary FTLD-tauopathy, the scores for amyloid-$$mathtex$$\beta$$mathtex$$, TDP-43, and $$mathtex$$\alpha$$mathtex$$-synuclein were set to zero; for primary TDP-43 cases, all pathologies except TDP-43 were set to zero; and for cases with primary Lewy body disease, all but $$mathtex$$\alpha$$mathtex$$-synuclein were set to zero. In patients with a primary neuropathologic diagnosis of AD, we retained the observed scores for amyloid-$$mathtex$$\beta$$mathtex$$ and tau while zeroing out TDP-43 and $$mathtex$$\alpha$$mathtex$$-synuclein. This model maintained the same degrees of freedom as the polypathologic model but assumed that copathologies would have negligible contribution to observed atrophy. We hypothesized that the polypathologic model would better fit the observed W-scores than this primary diagnosis model, supporting the importance of modeling copathologic effects. We additionally tested a “summed burden” model, which replaced ordinal scores for the four pathologies of interest with their sum, thus assessing overall pathologic burden in a protein-agnostic manner; all other terms were as in the polypathologic model. Model performance was assessed by comparing log-likelihood values for the polypathologic, primary diagnosis, and summed burden models. To assess replicability and disease-specificity of effects, we computed Pearson’s correlations between observed regional volumes and fixed-effects predictions (i.e., not including subject-specific random intercepts) for each primary pathology group in both training and testing splits.
In secondary analyses collapsing across training and testing splits, we assessed whether non-proteinopathic measures including regional angiopathy, gliosis, and neuronal loss were associated with residuals from the polypathologic model and could account for atrophy not explained by pathologic burden itself. Finally, we used the lme4 (version 1.1-26; https://cran.r-project.org/web/packages/lme4) and mediation (version 4.5.0; https://cran.r-project.org/web/packages/mediation) R packages to investigate whether these supplementary measures mediated the association between each pathology of interest and standardized volume scores, with 10,000 bootstrap iterations per model. Mediation models again included covariates of disease duration, MRI-death interval, and MRI scanner.
Results
Co-occurrence of pathologic diagnoses
The most common neuropathologic diagnosis among this sample was AD: 68 out of 125 cases (54%) had intermediate–high levels of AD neuropathologic change, and 29 had low levels. AD pathology commonly co-occurred with Lewy body disease and TDP-43 (both FTLD-TDP and LATE) and less commonly with FTLD-tauopathies such as progressive supranuclear palsy and corticobasal degeneration (Fig. 2). Polypathology was common: only 22 out of 125 cases (17.6%) had a single diagnosis, while 66 (52.8%) had two and 37 (29.6%) had three or more diagnoses. Frequency of copathologies varied by age, consistent with analyses of the CNDR’s full brain bank (Robinson et al., 2018). The median age at death was 71 years; in participants below this median, 44 out of 60 (73.3%) had more than one neuropathologic diagnosis, while among those aged 71 years or older at death, 59 out of 65 (90.8%) had multiple diagnoses [$$mathtex$$\chi ^2$$mathtex$$(1) = 6.54, p = 0.020]. Individuals with a single pathologic diagnosis included 11 of 54 patients (20%) with a primary diagnosis of AD, 3 of 25 cases (12%) with primary FTLD due to TDP-43, 6 of 7 (86%) of Pick’s disease cases, 1 of 12 (8%) progressive supranuclear palsy cases, and the single case of FTD with parkinsonism linked to chromosome 17. Between train and test splits, none of the four primary pathology groups exhibited differences in age at onset (all t < 0.81, p > 0.44), age at death (all t < 0.90, p > 0.39), male/female distribution (all Χ2 < 2.56, p > 0.29), race (all Χ2 < 0.69, p = 1), or distribution of levels of AD neuropathologic change (from none to high; all Χ2 < 4.52, p > 0.22).
Regional polypathology
Across regions, amyloid-$$mathtex$$\beta$$mathtex$$ and tau were the most common pathologies (Fig. 4, top). The highest burdens of $$mathtex$$\alpha$$mathtex$$-synuclein and TDP-43 were observed in limbic structures including the amygdala, entorhinal cortex, hippocampus, and anterior cingulate cortex. When we ranked regions by the average number of proteins with ordinal scores >1 in each region (distinguishing low vs clinically significant levels of pathology), the highest rates of copathology were observed in limbic structures, followed by neocortical regions, then the hippocampus and subcortical structures (Fig. 4, bottom). This ranking partially reflected the co-occurrence of amyloid-$$mathtex$$\beta$$mathtex$$ and tau in patients with intermediate–high AD pathologic change; when we ranked co-occurrence of tau, TDP-43, and $$mathtex$$\alpha$$mathtex$$-synuclein alone, hippocampus had the fourth-highest copathology rank, between anterior cingulate and superior middle/temporal cortex; and angular gyrus ranked seventh, between middle frontal cortex and thalamus. The relative ranks of other regions remained the same.
Top: frequency of ordinal scores for amyloid-$$mathtex$$\beta$$mathtex$$, tau, TDP-43, and $$mathtex$$\alpha$$mathtex$$-synuclein in the training dataset. In each plot, bars represent (from left to right) amyloid-$$mathtex$$\beta$$mathtex$$(red), tau (green), TDP-43 (turqoise), and $$mathtex$$\alpha$$mathtex$$-synuclein (purple). Increasing opacity represents increasing ordinal values (0–3). Bottom: bars indicate the average number of pathology scores >1 in each region, averaged across participants. Error bars display the standard error of the mean.
Polypathologic associations with gray matter atrophy
Permutation analyses on training data indicated that gray matter volume was inversely associated with tau, TDP-43, and $$mathtex$$\alpha$$mathtex$$-synuclein, although not with amyloid-$$mathtex$$\beta$$mathtex$$ (Table 2A). The polypathologic model outperformed the primary diagnosis model (log-likelihood values of −1,564.04 and −1,577.34, respectively), which produced broadly similar results but indicated a significant effect of amyloid-$$mathtex$$\beta$$mathtex$$ and a marginally significant effect of $$mathtex$$\alpha$$mathtex$$-synuclein (Table 2B). The summed burden model, which predicted atrophy from the sum of ordinal scores for the four pathologies of interest, had the lowest performance (log-likelihood of −1,579.91). In this model, a higher total pathologic burden (irrespective of the specific pathologies involved) was inversely associated with regional volumes (Table 2C).
Coefficients of interest for polypathologic, primary diagnosis, and summed burden models
In the training dataset, fitted values for the polypathologic model were more highly correlated with observed W-scores than fits from the summed burden and primary diagnosis models (Fig. 5). In the testing dataset, the polypathologic model again produced the highest correlations in the AD and $$mathtex$$\alpha$$mathtex$$-synuclein subgroups. However, in the FTLD-tau and TDP-43 subgroups, the results in the testing sample were nonsignificant for the summed burden, primary diagnosis, and polypathologic models.
Scatterplots show correlation between observed W-scores and fitted values based from polypathologic, primary pathologic, and summed burden models in the training and testing datasets.
As observed above (Regional polypathology), the rates of polypathology were highest in limbic areas, intermediate in neocortical areas, and lowest in subcortical areas. We thus repeated permutation analyses for the polypathologic model separately in the training data for each of these three region types. Across limbic areas, TDP-43 [F(3, 264) = 5.48, p = 0.0015] and tau [F(3, 264) = 4.95, p = 0.0027] were again associated with gray matter volume, but amyloid-$$mathtex$$\beta$$mathtex$$ and $$mathtex$$\alpha$$mathtex$$-synuclein were not (both p > 0.26). A similar pattern of associations was observed for subcortical areas [TDP-43: F(3, 170) = 5.61, p = 0.0013; tau: F(3, 170) = 3.90, p = 0.01; all others, p > 0.20]. In neocortical areas, ANOVAs indicated a significant effect for TDP-43 only [F(3, 172) = 4.68, p = 0.0036; all others p > 0.16].
Exploratory analysis of alternative predictors of atrophy
In the secondary analyses, we asked whether non-proteinopathic measures of neurodegeneration, including angiopathy, gliosis, and neuronal loss, were associated with the residuals of the polypathologic model recomputed in the full sample. Gliosis and neuronal loss ratings were highly correlated with one another (R = 0.91, p < 0.0001). Correlations of gliosis and neuronal loss with angiopathy were significant but smaller (R = 0.23 and R = 0.19, respectively, both p < 0.0001). Gliosis [F(3, 1084) = 7.38, p = 6.77 × 10−5] and neuronal loss [F(3, 1091) = 9.70, p = 2.55 × 10−6] but not angiopathy [F(3, 1088) = 1.45, p = 0.23] were associated with residual atrophy. Residual atrophy was marginally worse for regions with a gliosis rating of 1 [$$mathtex$$\beta$$mathtex$$ = −0.173, t(1084) = −1.74, p = 0.08], nonsignificant for gliosis = 2 [$$mathtex$$\beta$$mathtex$$ = 0.093, t(1084) = 0.88, p = 0.38], and significantly worse for regions with gliosis = 3 [$$mathtex$$\beta$$mathtex$$ = −0.484, t(1084) = −4.06, p = 5.31 × 10−5], relative to regions with no/minimal gliosis. Similarly, atrophy did not significantly differ between regions with no/minimal neuronal loss and those with neuronal loss = 1 [$$mathtex$$\beta$$mathtex$$ = −0.153, t(1091) = −1.59, p = 0.13] or 2 [$$mathtex$$\beta$$mathtex$$ = 0.166, t(1091) = 1.54, p = 0.12] but was significantly worse for neuronal loss = 3 [$$mathtex$$\beta$$mathtex$$ = −0.525, t(1091) = −4.45, p = 9.35 × 10−6].
We additionally hypothesized that gliosis, neuronal loss, and angiopathy—as possible sequelae of pathologic protein accumulation—would mediate associations between pathology and atrophy. Mediation analysis results are summarized in Table 3. Because amyloid-$$mathtex$$\beta$$mathtex$$ was not significantly associated with atrophy in the polypathologic model, it was omitted from these analyses. Both gliosis and neuronal loss significantly mediated the association of atrophy with tau and TDP-43; in contrast, we observed no mediating effects for the association between $$mathtex$$\alpha$$mathtex$$-synuclein and atrophy. TDP-43 exhibited slightly greater mediation than tau, both in terms of unstandardized mediation effects and proportion of the total effect. The mediating effects of gliosis were slightly larger than those of neuronal loss. Angiopathy had a significant positive mediation effect on the association between atrophy and tau.
Mediation of atrophy–pathology relationships by gliosis, neuronal loss, and angiopathy
Discussion
We investigated polypathologic associations with gray matter atrophy, a useful but nonspecific marker of neurodegenerative disease. This analysis leveraged a dataset comprising antemortem MRI and postmortem ratings of regional tau, amyloid-$$mathtex$$\beta$$mathtex$$, TDP-43, and $$mathtex$$\alpha$$mathtex$$-synuclein. A polypathologic model produced more accurate and replicable estimates (Fig. 5) of atrophy than a model which assumed atrophy could be explained by primary neuropathologic diagnosis, with no contribution from copathologies. This polypathologic model also outperformed a model based on the summed burden of all pathologies. Gliosis and neuronal loss not only mediated the observed associations between pathology and atrophy but accounted for variance not explained by the four proteinopathies.
Copathologies are frequent at global and regional levels
We previously reported that copathologies are frequent among older patients and those with more severe pathology (Robinson et al., 2018). In the current sample, nearly all patients had multiple pathologic diagnoses. Moreover, proteinopathies frequently co-occurred at the regional level, particularly in limbic areas. Recent research has identified the hippocampus as a focus of tau and TDP-43 copathology (Nelson et al., 2019; Wisse et al., 2021); the current results suggest that the amygdala, anterior cingulate, and entorhinal cortex have equal or greater polypathologic burden. High rates of copathology suggest that therapies targeting a single protein could fail to fully address the sources of degeneration.
Effects of copathology on disease severity
Prior research has addressed how polypathology affects neurodegenerative disease trajectory. Josephs et al. (2017) found that TDP-43 copathology was associated with faster hippocampal atrophy among participants with intermediate tau; TDP-43 copathology may also be associated with accelerated neocortical atrophy (Josephs et al., 2020). However, some findings cast doubt on the relevance of polypathologies for predicting disease progression: an autopsy study of progressive supranuclear palsy concluded that copathologies did not have an effect on disease milestones, including age at onset and death (Jecmenica Lukic et al., 2020). In the current analysis, higher regional copathology was associated with more severe degeneration. Additional research is needed to resolve relationships between specific pathologies, their relative and chronological onset, and their consequences for atrophy. Biomarkers for FTLD-tau, TDP-43, and $$mathtex$$\alpha$$mathtex$$-synuclein may bring greater clarity to these associations.
Differential associations of four proteinopathies with gray matter atrophy
While tau, TDP-43, and $$mathtex$$\alpha$$mathtex$$-synuclein were associated with atrophy, amyloid-$$mathtex$$\beta$$mathtex$$ was not. In primary AD cases, accounting for TDP-43 and $$mathtex$$\alpha$$mathtex$$-synuclein copathology produced more accurate volume estimates, suggesting that copathologies contribute to atrophy in AD.
Prior research on the association between $$mathtex$$\alpha$$mathtex$$-synuclein and atrophy has been inconsistent. We previously found no significant atrophy for Lewy body patients versus controls after multiple comparisons correction (Howard et al., 2022). A recent review (Saeed et al., 2020) reported that Parkinson’s disease is often characterized by atrophy in basal ganglia and the frontal lobes, but that other studies have found no significant atrophy. Here we found that $$mathtex$$\alpha$$mathtex$$-synuclein had a modest association with atrophy; but in patients with primary Lewy body disease, including copathologies led to better model fits. Further research is necessary to ensure that associations between $$mathtex$$\alpha$$mathtex$$-synuclein and atrophy do not indicate an indirect or spurious association.
In FTLD-tau and TDP-43 subgroups, the polypathologic model again yielded the most accurate modeling in the training dataset. Neuropathologic diagnoses of FTLD-tau and TDP-43 proteinopathies co-occurred only rarely with each other or with Lewy body disease, but most patients had at least low levels of AD copathology, suggesting that the benefit of the polypathologic model might be in accounting for AD copathology. However, the results in the FTLD-tau and TDP-43 subgroups are tempered by the failure of all three models (summed burden, primary diagnosis, and polypathologic) to replicate in the testing samples. These null results may reflect small sample sizes or clinicopathological heterogeneity in the FTLD-tau and TDP-43 groups.
The lack of association between amyloid-$$mathtex$$\beta$$mathtex$$ and atrophy in the polypathologic model adds an expected but informative data point to the literature. Recent studies have reported null or weak associations between amyloid-$$mathtex$$\beta$$mathtex$$ and atrophy (e.g., Iaccarino et al., 2018). Moreover, longitudinal PET studies indicate that tau but not amyloid-$$mathtex$$\beta$$mathtex$$ can predict future atrophy (La Joie et al., 2020). Although both biofluid (Kaffashian et al., 2015) and PET (Oh et al., 2014) measures of amyloid-$$mathtex$$\beta$$mathtex$$ are associated with atrophy, many studies have not assessed tau, raising the possibility that observed effects reflected collinearity between amyloid-$$mathtex$$\beta$$mathtex$$ and tau. The current study, which adjusts for tau, $$mathtex$$\alpha$$mathtex$$-synuclein, and TDP-43 burden, does not support a direct association with brain atrophy.
We repeated the polypathologic model separately for limbic, subcortical, and neocortical regions. TDP-43 again emerged as the most significant correlate of atrophy; tau was significant for limbic and subcortical areas but not for neocortical areas, contrary to expectations. Null results for $$mathtex$$\alpha$$mathtex$$-synuclein in all the three models may reflect reduced statistical power to detect a weak association with gray matter atrophy. Similarly, the null result for neocortical tau may reflect a reduced and/or greater atrophy severity in individuals with primary TDP-43 proteinopathies relative to primary AD cases. One possibility that merits further investigation is that neocortical atrophy typically ascribed to tau is partially driven by TDP-43 copathology.
Finally, the relative magnitudes of regression coefficients for the four proteinopathies warrant cautious interpretation. Across the ordinal scale, TDP-43 and tau were associated with greater atrophy than $$mathtex$$\alpha$$mathtex$$-synuclein. However, ordinal ratings represent subjective judgments by expert observers, and it is unclear how to define equivalent levels of distinct pathologies. Prior research has reported more rapid degeneration in FTLD, including ubiquitin-reactive cases with likely TDP-43 pathology (Whitwell et al., 2008), than in AD. Retrospective studies of longitudinal atrophy in autopsy-confirmed cases may clarify relative rates of progression for AD, FTLD, and other proteinopathies.
Associations with gliosis and neuron loss
Although the polypathologic model outperformed alternative models, the low correlations observed in Figure 5 suggested an incomplete model. In exploratory analyses, we found that both gliosis and neuronal loss mediated associations of tau and TDP-43 with atrophy. Furthermore, ordinal ratings of gliosis and neuronal loss had significant associations with residuals of the polypathologic model. Angiopathy had a positive mediation effect on the relationship between tau burden and atrophy; this counterintuitive finding may reflect milder atrophy among older patients with both vascular disease and low levels of AD pathologic change.
Taken together, the results suggest that gliosis and neuron loss may not only explain a significant portion of the relationship between atrophy and proteinopathy but account for additional atrophy that is not attributable to tau, TDP-43, or $$mathtex$$\alpha$$mathtex$$-synuclein. Frisoni et al. (2022) noted that combined amyloid-$$mathtex$$\beta$$mathtex$$ and tau positivity in cognitively unimpaired individuals did not predict cognitive decline. These observations, coupled with the current findings, suggest that disease progression is driven not only by proteinopathies themselves but also sequelae such as innate immune responses, which may accelerate disease progression (Schonhoff et al., 2020). Gliosis and neuronal loss ratings were highly correlated, likely reflecting a relationship between pathologic glial responses and cell death. Future research will help determine whether gliosis differs consistently between proteinopathies or reflects inter-individual heterogeneity.
Limitations
The current study is limited by the use of ordinal pathology scoring, which constrains precision and statistical methodology. Digital quantification methods for histopathology (Giannini et al., 2019) may improve precision and allow us to examine interactions between pathologies; in the current study, we investigated simple additive effects due to the explosion of model complexity when examining interactions of ordinal variables. Additionally, regional tau ratings in the current dataset did not distinguish between AD- and FTLD-type tau, although GT-38 staining (Gibbons et al., 2019) was used for Braak staging of FTLD-tau cases. In future studies, isoform-specific tau markers may clarify AD- and FTLD-specific associations with neurodegeneration in mixed-pathology cases. Similarly, available data did not allow us to distinguish TDP-43 subtypes, and regional values reflected a combination of LATE, FTLD-TDP, and ALS pathology. A third limitation involved small sample sizes in the testing split for the FTLD-tau, TDP-43, and $$mathtex$$\alpha$$mathtex$$-synuclein subgroups. This limitation may explain the replication failures in FTLD-tau and TDP-43 subgroups. Additionally, the sample was enriched for individuals with rare neurodegenerative conditions. The rates of copathology may thus differ from the rates in the general population; however, comparable results have been reported in population-based studies (Kovacs et al., 2013; Karanth et al., 2021).
Finally, one parameter of the study design which warrants further consideration is the maximum interval between MRI and autopsy. In individuals with rapid disease progression—such as those with a combined ALS and FTD syndrome (Grossman et al., 2023)—postmortem ratings of copathologies may not reflect their approximate burden at the time of MRI. Future research with in vivo assessments of $$mathtex$$\alpha$$mathtex$$-synuclein and TDP-43 may provide better insight into the temporal evolution of these proteinopathies and their relationship to atrophy.
Despite these limitations, the current study leveraged a unique multimodal dataset to investigate polypathologic correlates of atrophy. We found that brain atrophy was related not only to the TDP-43, tau, and $$mathtex$$\alpha$$mathtex$$-synuclein accumulation but also gliosis and neuron loss. Currently, copathologies are typically inferred from neurodegeneration out of proportion with amyloid-$$mathtex$$\beta$$mathtex$$ and tau (Das et al., 2021). The current findings highlight the importance of spatially specific biomarkers for TDP-43, $$mathtex$$\alpha$$mathtex$$-synuclein, and inflammatory immune responses to fully assess the factors driving individual patients’ brain atrophy.
Data Sharing
Analysis code for this manuscript will be made available on Dr. Phillips’ Github site (https://github.com/jeffrey-phillips). De-identified imaging and pathology data will be made available to researchers with an approved request to the Penn Neurodegenerative Data Sharing Committee (https://www.pennbindlab.com/data-sharing).
Footnotes
This work was supported by grant supports from the National Institute on Aging (R01-AG054519, K01-AG061277, and R01-AG076832) and the Alzheimer’s Association (AARG-22-926144). We thank the patients and caregivers who made this research possible through research participation and brain donation. Additionally, we thank Manuela Neumann and Elisabeth Kremmer for providing the phosphorylation-specific TDP-43 antibody 1D3.
↵† Deceased.
D.I has served on the Scientific Advisory Board of the Lewy Body Dementia Association, which advocates for patients with one of the diseases investigated in this study. D.W. has received financial support from and serves as site principal investigator for interventional clinical trials conducted by Biogen, which co-developed the anti-amyloid therapies aducanumab and lecanemab; he has also received consulting fees from Eli Lilly and Company. None of the other authors have conflicts of interest relevant to the current study.
- Correspondence should be addressed to Jeffrey S. Phillips at jeffrey.phillips{at}pennmedicine.upenn.edu.
