Inclusion strategy
Unlike biomarker studies that emphasize strict control of sample collection protocols and matching among groups and cohorts, we used more lenient inclusion criteria in an attempt to obtain a better representation of the high variability in the patient population and in various clinical settings. Both the discovery and validation cohorts included samples collected either in the field in a population-based study or in university-hospital-based clinics. Patients with PD were characterized by the diagnosing specialists as definite, probable, or possible PD and patients with MSA included both probable and possible diagnosis. In each MSA category, both the cerebellar type (MSA-C) and parkinsonian type (MSA-P) were included. In the discovery cohort, all the samples were serum, whereas in the validation cohort, ~ 40% of the samples were plasma.
A detailed description of the exosome isolation, enrichment, origin validation, assay reproducibility, and limitation of sample storage period is provided as Supplementary Results and Supplementary figs. 1–5, online resources.
Discovery cohort
Before immunoprecipitation of exosomes using CNS biomarkers, we asked whether the total number of exosomes or the total concentration of α-syn in the serum samples differed among the groups. The number of exosomes can be estimated conveniently using a commercial CD81 ELISA kit in which the CD81 signal is converted to exosome concentration. To improve normalization of the data for statistical analysis, here and in all subsequent analyses, the values were log-transformed and are presented as log values in the figures. However, the untransformed values are discussed in the text to facilitate comparison with other studies.
The analysis showed a decrease in exosome concentration from 4.4 × 1010 ± 3.1 × 1010 exosomes per mL in the control group to 3.7 × 1010 ± 3.6 × 1010 and 3.1 × 1010 ± 3.4 × 1010 in the PD and MSA groups, respectively (Supplementary Fig. 6a, online resource). Although these differences were statistically insignificant, because the same trend was found in the validation cohort, the exosome concentration was included in the final multivariate statistical model used to separate the groups (see below). The exosome concentrations did not correlate with disease duration in either group. Measurement of serum α-syn showed insignificant differences among the groups (Supplementary Fig. 6b, online resource).
In both the putative neuronal and oligodendroglial exosomes, α-syn increased in the order control < PD < MSA (Fig. 1a). The α-syn concentrations in the PD group (putative neuronal 107 ± 124 pg/mL, putative oligodendroglial 81 ± 104 pg/mL) were significantly higher than in the control (putative neuronal 58 ± 55 pg/mL, putative oligodendroglial 53 ± 73 pg/mL) and significantly lower than in the MSA group (putative neuronal 191 ± 131 pg/mL, putative oligodendroglial 286 ± 348 pg/mL). Interestingly, these results contradicted the observations of Jiang et al. [32] and Yu et al. [70] who reported lower α-syn concentrations in putative CNS-originating exosomes from patients with MSA compared to those with PD.
To evaluate the degree of overlap among the groups, we used ROC analyses. The separation between the control and PD groups was low in both the putative neuronal (AUC = 0.674, Fig. 1b) and putative oligodendroglial (AUC = 0.628, Fig. 1c) exosomes, in agreement with results reported previously by Shi et al. (AUC = 0.654 in putative neuronal exosomes from 215 control and 267 PD plasma samples) [61]. In contrast, high separation was found between the control and MSA groups, particularly in the putative oligodendroglial exosomes (AUC = 0.924, Fig. 1c). α-Syn concentration in putative oligodendroglial exosomes also provided better separation between the PD and MSA groups (AUC = 0.867, Fig. 1c) than in putative neuronal exosomes (AUC = 0.769, Fig. 1b).
> 90% of the samples in the control and PD groups were from the UCLA PEG study in this cohort, precluding meaningful testing of a source effect. In the MSA group, 21 samples were from UCLA and 9 from NYU. Comparing the putative neuronal and oligodendroglial α-syn concentrations in the samples from these two sources we found in both cases that the concentrations in the NYU samples were significantly higher (Supplementary Table 1, online resource).
The oligodendroglial: neuronal exosomal α-syn ratio improves the separation of PD from MSA
We asked next whether the known preference for deposition of α-syn in neurons in PD versus oligodendrocytes in MSA could further help distinguish between these groups, even though the measurement in our assay was of total, rather than aggregated α-syn. In agreement with the pattern of pathological α-syn deposition in the brain, in the PD group, the average α-syn concentration in the putative neuronal exosomes was higher than in the putative oligodendroglial exosomes, whereas the opposite was true in the MSA group (Fig. 1a). Therefore, we calculated the ratio between the α-syn concentration in the putative oligodendroglial and putative neuronal exosomes (oligo:neuro ratio) for each sample (Fig. 2a). In most cases, the ratio was as expected, < 1 for PD and > 1 for MSA, yielding AUC = 0.916 (Fig. 2b), corresponding to 90.0% sensitivity and 88.2% specificity. Importantly, unlike the total α-syn concentrations that differed significantly between the UCLA and NYU MSA samples, the difference between the average ratio values was insignificant (Supplementary Table 1, online resource), suggesting that the ratio may serve to remove variations among collection sites.
Validation cohort
The samples used in our discovery cohort were collected from patients diagnosed clinically and not validated pathologically. To our knowledge, serum or plasma samples collected from living patients whose brains were analyzed pathologically after death are not available in sufficient numbers in any current biobank, especially considering that the samples cannot be > 5 years in storage. Therefore, we attempted to use samples collected post-mortem with a short post-mortem interval (PMI) from patients whose diagnosis was validated pathologically. Unfortunately, we found that in such samples, erythrocyte α-syn contaminated the signal and did not allow meaningful analysis of α-syn in putative CNS-originating exosomes (Supplementary Results, Supplementary Table 2, and Supplementary figs. 7, 8, online resources).
In the absence of pathologically validated samples, we obtained next a new set of samples to assemble a validation cohort. Most of these samples were obtained again from the UCLA PEG study (control and PD), the UCLA Clinical Neurogenomics Research Center (MSA), and the NYU Dysautonomia Center (PD and MSA). An additional major source of samples for this cohort was a biobank at Columbia University (control and MSA). Comparison of the two cohorts showed that the groups did not differ significantly in their composition in terms of sex, ethnicity, and age (Supplementary Table 3, online resource). A few samples from the UCLA Easton Center were included originally but later eliminated, because their storage time was > 5 years. Because this cohort contained both serum and plasma samples, an adjustment of the raw data was necessary to allow analysis of these samples together (Supplementary Results and Supplementary Fig. 9, online resources).
Similar to the discovery cohort, before immunoprecipitation, there was a trend toward reduced exosome concentration in the serum/plasma samples in the order: control > PD > MSA (Supplementary Fig. 10a, online resource). Interestingly, unlike the discovery cohort, α-syn concentration in the serum/plasma was substantially higher in the MSA group (486 ± 479 pg/mL) than in the control (187 ± 241 pg/mL) or PD (208 ± 183 pg/mL) groups (Supplementary Fig. 10b, online resource). As the majority of the MSA samples in this group were from Columbia University, we asked if differences among the sample sources might have accounted for the higher serum/plasma α-syn levels in this cohort. Comparison among the sources showed that despite the larger number of samples from Columbia University, the variability in these samples was lower than in the samples from UCLA or NYU, yet the differences among the groups were statistically insignificant (Supplementary Fig. 10c, online resource). The increased α-syn concentration in the MSA group could be partially attributed to the differences between the serum and adjusted plasma concentrations in these samples (Supplementary Fig. 9c, online resource). In addition, there was a larger fraction of MSA-P in the validation cohort (33.3%, including 4 samples with a mixed MSA-P/MSA-C diagnosis) compared to the discovery cohort (13.3%). On average, samples from patients diagnosed with MSA-C had lower serum/plasma α-syn concentrations (353 ± 251 pg/mL) than samples from patients with MSA-P or mixed diagnosis (547 ± 669, p= 0.062, Student’s t test). Thus, the larger fraction of the latter in the validation cohort possibly also contributed to the increased concentration of serum/plasma α-syn in this cohort’s MSA group.
Similar to the discovery cohort, the α-syn concentrations increased in the order: control < PD < MSA in both the putative neuronal and oligodendroglial exosomes, though the difference between the control and PD groups was statistically significant only for the latter (Fig. 3a). Higher concentrations of α-syn were observed in the immunoprecipitated exosomes from the MSA group in this cohort, possibly for the same reasons discussed above for serum/plasma α-syn. The average α-syn concentration in the PD group (putative neuronal 110 ± 136 pg/mL, putative oligodendroglial 96 ± 148 pg/mL) was closer to the control group (putative neuronal 96 ± 110 pg/mL, putative oligodendroglial 60 ± 91 pg/mL) and substantially lower than the MSA group (putative neuronal 284 ± 251 pg/mL, putative oligodendroglial 497 ± 360 pg/mL). Accordingly, the separation between the control and PD groups was moderate in both the putative neuronal (AUC = 0.611, Fig. 3b) and putative oligodendroglial (AUC = 0.645, Fig. 3c) exosomes, whereas the separation between the control and MSA groups, particularly in the putative oligodendroglial exosomes was high (AUC = 0.947, Fig. 3c). α-Syn concentration in putative oligodendroglial exosomes provided better separation between the PD and MSA groups (AUC = 0.920, Fig. 3c) than in putative neuronal exosomes (AUC = 0.824, Fig. 3b), mirroring the discovery cohort.
Analysis of the oligo:neuro ratio (Fig. 4a) showed that the number of PD samples for which the ratio was < 1 was lower in this cohort (69.8%) than in the discovery cohort (88.2%), whereas the fraction of MSA samples for which the ratio was > 1 was similar in the validation cohort (90.0%) to the discovery cohort (86.7%). Accordingly, the separation between the PD and MSA groups was somewhat lower in the validation cohort, AUC = 0.871.
Multivariable models for separation between PD and MSA
To further explore the degree to which the measured biomarkers could help improve distinguishing between PD and MSA, we tested several statistical models, including: (1) a multinomial logistic model with LASSO variable selection [16, 64]; (2) a linear discriminant model; (3) a classification tree model; and (4) a K-nearest-neighbor model. In each case, the model was trained on the discovery cohort and was challenged with selecting among the putative neuronal exosomal α-syn, putative oligodendroglial exosomal α-syn, oligo:neuro ratio, and serum/plasma exosome concentration of the combination providing the best separation among the groups in a ROC analysis. Because the first three parameters are interdependent, the models were expected to select two out of the three and fulfilled this expectation in all cases. The formula created by the model then was applied to the validation cohort to test to what extent the discrimination power could be reproduced.
All the models, except the K-nearest-neighbor model, which yielded poor accuracy (data not shown), provided similar results (Supplementary Table 4, online resource), yet the multinomial logistic model yielded the highest discrimination power. The model selected the putative neuronal exosomal α-syn, oligo:neuro ratio, and serum/plasma exosome concentration to create the discriminative formula, which in the discovery cohort separated the control and PD groups with AUC = 0.762, control and MSA with AUC = 0.961, and PD and MSA groups with AUC = 0.928. When the formula created by the model was applied to the validation cohort, it separated the control and PD groups with AUC = 0.610, control and MSA with AUC = 0.962, and PD and MSA groups with AUC = 0.902, corresponding to 89.8% sensitivity and 86.0% specificity.
A limitation of our study’s design was that it did not emphasize inclusion of early stage patients, in which the potential for misdiagnosis is highest. Most patients with PD in our study were 5–8 years from diagnosis, whereas the disease duration in the MSA group was mostly 3–5 years (Supplementary Fig. 11a, online resource). Examination of the biomarkers reported here in patients with early stage disease will be pursued in the future. Nonetheless, to test whether the main biomarker, the oligo:neuro ratio, can be detected in early stage disease or only develops at later stages, we tested whether this biomarker correlated with disease duration. Spearman analysis showed that the oligo:neuro ratio did not correlate with disease duration (Supplementary Fig. 11b–e, online resource), suggesting that it could be a useful biomarker already at the time of diagnosis.
Another potential limitation of the study is that the reference used for biomarker accuracy is the clinical diagnosis, which as discussed in the Introduction, is prone to error. Though validation of the biomarker’s accuracy through a neuropathological examination was not possible in most cases, toward the end of the study, we went back and checked whether any clinical diagnosis was validated after patients passed away or changed by the diagnosing clinicians. The data are presented in Supplementary Table 5 (online resource).
In the PD groups (both cohorts combined), one diagnosis was changed from PD to MSA-P, which was predicted correctly by the model. Two patients’ diagnosis was changed from PD to Parkinson’s disease dementia (PDD), of which one was predicted by the model to be MSA. No diagnosis was pathologically validated in this group. In the MSA group, the model predicted correctly the diagnosis of 11 out of 13 patients whose diagnosis with MSA, or in two cases MSA mixed with AD, was confirmed pathologically. In two cases, both of which were diagnosed clinically with probable MSA-C and later validated pathologically as MSA, the model predicted incorrectly a diagnosis of PD. These numbers are in agreement with the degree of sensitivity and specificity of the model described above.
In contrast to the diagnostic power of the biomarkers we measured, we did not find cross-sectional correlation with disease severity for any of the biomarkers, including measurements of motor deterioration (UPDRS-III, UMSARS, H&Y) or cognitive function (MMSE).