A molecular pathology, neurobiology, biochemical, genetic and neuroimaging study of progressive apraxia of speech

Progressive apraxia of speech is a neurodegenerative syndrome affecting spoken communication. Molecular pathology, biochemistry, genetics, and longitudinal imaging were investigated in 32 autopsy-confirmed patients with progressive apraxia of speech who were followed over 10 years. Corticobasal degeneration and progressive supranuclear palsy (4R-tauopathies) were the most common underlying pathologies. Perceptually distinct speech characteristics, combined with age-at-onset, predicted specific 4R-tauopathy; phonetic subtype and younger age predicted corticobasal degeneration, and prosodic subtype and older age predicted progressive supranuclear palsy. Phonetic and prosodic subtypes showed differing relationships within the cortico-striato-pallido-nigro-luysial network. Biochemical analysis revealed no distinct differences in aggregated 4R-tau while tau H1 haplotype frequency (69%) was lower compared to 1000+ autopsy-confirmed 4R-tauopathies. Corticobasal degeneration patients had faster rates of decline, greater cortical degeneration, and shorter illness duration than progressive supranuclear palsy. These findings help define the pathobiology of progressive apraxia of speech and may have consequences for development of 4R-tau targeting treatment.


Supplementary Text
Statistical details for the neuroimaging models In the MRI model, volumes were scaled and centered within region to bring measurements from inherently different sized regions to approximately the same magnitude, improving estimation of the single hierarchical model using these scaled volumes. FDG SUVR and DTI FA already were on a comparable scale across regions; raw values were used as the outcome in these modalities and values from the left and right hemisphere were entered into the models.
Algebraically, these three models can equivalently be expressed as: where indicates individual, indicates timepoint, and indicates region or tract in the case of the DTI model (simply referred to as region hereafter). is the outcome (for a given modality) for person at timepoint for region . The terms and are indicator functions for whether individual was given a diagnosis of PSP or CBD, thereby only including the relevant model terms in the estimation at that data point.
indicates the region-specific intercept (baseline value) in the entire cohort, the region-specific annual change in the entire cohort, the region-specific intercept shift for CBD, the region-specific annual change shift for CBD, the region-specific intercept shift for PSP, and the region-specific annual change shift for PSP.
is the person-specific intercept shift, allowing us to use multiple regions-perscan and multiple scans-per-person in a single modality model.
Of particular interest in these models is comparing and , the diagnosis-specific differences at baseline, and comparing and , the diagnosis-specific differences in annual change. In addition, the model fits for each diagnosis can be compared at any timepoint to assess whether outcome measures are diverging, converging, or are essentially parallel in these diagnostic groups.
The prior distributions of these parameters were auto scaled to find efficient and appropriate distributions (after internally centering the predictors), the default behavior of the rstanarm software. This means the candidate terms were drawn from independent (0, 11.7) and candidate terms were drawn from independent (0, 4.0) distributions in the MRI model, (0, 3.1) and (0, 1.1) respectively for the FDG model, and (0, 2.0) and (0, 0.7) respectively for the DTI model. Group wise estimation was used for diagnosis specific effects, drawing the and parameters simultaneously from, essentially, a bivariate standard normal distribution, allowing for nonzero covariance, in each model. More details can be found in the documentation for the stan_lmer function from the package rstanarm and at this website. Similarly, and were drawn from an identical (but independent) bivariate distribution in each model. The prior distribution for the terms was also, essentially, a standard normal distribution in each model. The overall Sigma, the variability in the outcome measure not described by this model formulation and the analog of the error term in ordinary least squares regression, was drawn from an exponential distribution with rate parameter 1 for the MRI model, 5.3 for the FDG model, and 11 for the DTI model.
Model diagnostics were adequate in all three models. The Monte Carlo standard error of all parameters in all models was approximately zero, the effective sample size of all parameters in all models was in the thousands or tens of thousands except the variance and covariance parameters of the longitudinal group effects in the DTI model, which were above 900. The mixing parameter, Rhat or , was approximately 1 for all parameters in all models. Finally, posterior fits were qualitatively inspected for all regions in all models by comparing model expectations to a scatter plot of the raw data, analogous to the regions selected in Figure 5.The posterior model fits appear to describe these data adequately across regions and modalities when comparing observed and expected outcomes. Results of the hierarchical models were based on 14 Monte Carlo Markov Chains run in parallel, each consisting of 5000 posterior samples.
Statistical details for the clinical trajectory models To model the change over time in four clinical measures, MDS-UPDRS III, MoCA, WAB-AQ, and ASRS, we fit four mixed effects models, one per clinical test, using clinical score as the outcome predicted by diagnosis-specific intercept and time (years) terms, including diagnosisspecific quadratic terms for time to allow for nonlinearity in the change over the disease course. We also included a random intercept per person, to allow for multiple observations, i.e. multiple clinical visits, per person. In the MDS-UPDRS III and ASRS models, we additionally included a person-specific random effect for linear change which was allowed to be correlated with the person-specific intercept. Convergence issues prevented the inclusion of this person-specific rate term in the MoCA and ASRS models. Time was centered at 5 years from onset of disease to improve estimation in the model.

Statistical methods for correlations between neuroimaging and clinical measures
Spearman rank correlations were calculated within PSP and CBD groups to compare and test for an association between measures of language and AOS and regional metabolism measures. The Western Aphasia Battery Aphasia Quotient and Token Test were compared with the left Broca's and left superior temporal gyrus and the Apraxia of Speech Rating Scale version 3 total score was compared to SMA and Precentral regional metabolism.
Statistical methods for neuroimaging versus AOS subtype Consistent with our overall neuroimaging approach, the cortical regions (SMA and precentral/motor cortex) were assessed using FDG-PET SUVR, and the subcortical structures (striatum, globus pallidus, subthalamic nucleus and substantia nigra) were assessed using MRI volumes. Scans closest to death were utilized. In order to account for confounding effects of age in brain volume, MRI volumes were converted to age-and total intracranial volume (TIV)corrected Z scores using a model predicting volume by age and TIV in 36 cognitively normal controls. Specifically, to calculate these Z scored volumes, we used the model fit in controls to predict expected volume for each case based on age at scan and TIV. Then we subtracted the expected volume from the observed volume and divided this difference by the standard deviation of the residuals from the original model fit. We then performed non-parametric Wilcoxon Rank Sum tests using regional imaging measures to compare regional metabolism and age-and TIVadjusted volumes between phonetic and prosodic subtypes of AOS. The Token Test and Western Aphasia Battery Aphasia Quotient (WAB-AQ) were included as measures of language ability and were related to the language areas of Broca's area and left superior temporal gyrus. The Apraxia of Speech Rating Scale (ASRS) was included as a measure of apraxia of speech severity and was assessed in relation to the supplementary motor area (SMA) and precentral cortex; two cortical regions that are commonly abnormal in progressive apraxia of speech. The last available visit was used for all patients for each test. Spearman correlations were assessed separately for progressive apraxia of speech patients with corticobasal degeneration (CBD) and progressive supranuclear palsy (PSP), with CBD cases, trend-lines and p-values shown in orange, and PSP shown in blue. Spearman rho correlations were calculated using a two-sided test via an asymptotic t distribution approximation. Source data are provided as a Source Data file.
Supplementary Fig. 6. Neuroimaging comparisons across phonetic and prosodic apraxia of speech (AOS) targeting the corticostriatal and pallidonigraluysian networks. Cortical regions were assessed using FDG-PET standard uptake value ratios (SUVRs) and subcortical structures were assessed using age-and total intracranial volume-corrected Z scores of MRI volume. The last available scan was used for all patients. The corticobasal degeneration (CBD) and progressive supranuclear palsy (PSP) patients are shown in orange and blue respectively. SMA = supplementary motor area. N=15 phonetic cases and N=9 prosodic cases. Boxes represent lower quartile, median and upper quartile, with whiskers extending to the farthest point at most 1.5*inter-quartile range from each quartile. P values are from two sample Wilcoxon rank-sum test. One CBD phonetic case with z-scored volume of 13.2 in the pallidum was excluded from the pallidum plot to maintain readability of the y-axis. Source data are provided as a Source Data file.