Immortalized murine embryonic fibroblast (MEF) cell lines were prepared from heterozygous knockin mice bearing the DYT-TOR1A mutation (Tor1a^ΔGAG/+ genotype hereafter abbreviated as DYT-TOR1A or DYT) (19) and wildtype (WT) littermate embryos according to standard methodology (Methods). Three independent cell lines for each genotype were used. Genotype and all drug treatment conditions were tested in a blinded experimental design and in parallel by splitting the parental cell line flask into separate flasks for each condition. EVs produced during the 24-h period following media exchange with an EV-depleted media were isolated from the conditioned media by ultracentrifugation (21). Protein was isolated from the resultant EV pellet. DYT-TOR1A did not significantly modify recovery of total protein or amount of the constitutive EV marker, TSG101 (Figures 1B, C). Specific EV enrichment was confirmed by Western blot for TSG101 compared to non-EV markers (calnexin, actin) (Figures 1C, D) (22). Samples were then subjected to unbiased, quantitative LC-MS/MS proteomics analysis. Quantitative proteomic measurements also demonstrated that EV protein abundances of classic EV markers (TSG101 and the tetraspanins CD9, CD81, and CD63) were not modified by genotype (Figure 1E).

We characterized the EV proteome to identify genotype-dependent changes in EV protein abundances between DYT-TOR1A and WT EV samples. Following alignment of peptide signals to unique identifying peptides (UIPs) and removal of proteins with fewer than two detected UIPs, 1974 proteins were detected across all cell lines. Using a Bonferroni-adjusted p-value threshold for multiple hypothesis testing (p < 2.5e-5) (23), no significant genotype effects were identified. We next used this discovery dataset to identify putative DYT biomarkers for testing in follow-on experiments. Using an uncorrected p-value cutoff of less than 0.05, we identified 363 of 1974 proteins with significantly different abundances in DYT-TOR1A versus WT EVs (Figure 1F). This differential subset of 363 is more than 3.5 times larger than would be predicted by chance (e.g., 99 proteins from the total of 1974, based on the expected proportion α = 0.05).

We further noted that among the 363 differential proteins, there was an asymmetric distribution of genotype effects. The DYT-TOR1A effects showed a bias towards decreased abundances, with 320 proteins being significantly less abundant compared to only 43 being more abundant in DYT-TOR1A relative to WT (two-tailed binomial sign test, p < 0.0001). This skewed distribution was also maintained across all EV proteins (1491 less, 483 more; two-tailed binomial sign test, p < 0.0001).

We therefore considered technical reasons that could artifactually cause such a distribution bias, e.g., lower EV yields and/or detection thresholds not being met preferentially in DYT samples. As Figure 1E demonstrates, there were no significant genotype-dependent differences in abundance of EV constituents detected in the LC-MS/MS data. Secondly, when a protein is not detected in a sample, an imputed value is given as described in Methods prior to sample loading normalization. We therefore examined whether the DYT genotype effects came preferentially from proteins with multiple imputed values. Instead, we observed that hits were distributed proportionally across proteins with 0, 1, 2 or 3 imputed values and the vast majority of hits came from proteins with no imputed values (Supplementary Figure S2). These observations rule out LC-MS/MS detection thresholds as a systematic confound. In summary, we have performed proteomic analysis of MEF culture-derived EV preparations and identify 363 candidate proteins for DYT-TOR1A genotype biomarkers.

Recent studies have shown corrective effects of the HIV protease inhibitor ritonavir on cell and brain phenotypes in DYT-TOR1A preclinical models (4). For translation to human clinical trials, it is desirable to have pharmacodynamic biomarkers to aid early dose-finding studies and to assess target engagement (7, 8, 11). To explore the potential for the DYT-TOR1A genotype-associated EV changes that we identified to be used as pharmacodynamic biomarkers of disease state, we exposed DYT-TOR1A MEF cultures to 20 μM ritonavir throughout the 24 h of media conditioning preceding EV isolation. EV protein fractions were analyzed by quantitative LC-MS/MS proteomics performed in the same batch run as all conditions reported in this study.

Of the subset of 363 proteins significantly disrupted by DYT-TOR1A genotype basally, we found that >60% (230/363) had significant changes in abundance following ritonavir treatment at a threshold of p ≤ 0.05. This number of hits is 12 times greater than would be predicted by chance if ritonavir had no true effect on the genotype-dependent hits (18.15 proteins by α = 0.05). We further noticed that when examining the behavior of the 363 putative DYT biomarkers independent of p-value, the overwhelming majority of proteins showed ritonavir effects on protein abundance that were in the corrective direction (344/363) (Figure 2A). The putative DYT biomarker subset of proteins also showed strong and inverse correlations between genotype and ritonavir effects (Pearson r = −0.78, p < 0.0001) (Figure 2B). Noticing the very large number of proteins modified by ritonavir, we further examined the relationship between DYT genotype disruptions and DYT+RTV effects across the entire proteome and found that the strong inverse correlation was maintained (n = 1974, Pearson r = −0.74, p < 0.0001) (Figure 2B). Proteome-wide, ritonavir significantly modified 29% of the DYT EV proteome (uncorrected p ≤ 0.05, 582/1974) in a direction that was opposite to the genotype effect, with an asymmetric distribution toward increasing abundances for both significant and non-significant abundance changes (two-tailed binomial sign test: 508/582, with log₂ fold change >0, p < 0.0001; 1372/1974 with log₂ fold change >0, p < 0.0001). Lastly, we deployed hierarchical clustering to evaluate ritonavir’s effects on the putative DYT-TOR1A genotype biomarker proteins (n = 363). This analysis showed that ritonavir-treated DYT-TOR1A EV samples clustered more closely with WT than DYT-TOR1A samples (Figure 2C).

In summary, DYT-TOR1A genotype disruptions of EV protein composition show potential as pharmacodynamic markers of disease state. Ritonavir treatment acutely modified a substantial fraction of DYT-TOR1A genotype-dependent protein disruptions (95%) and caused dendrogram clustering of the EV proteome to become more closely related to WT samples than the DYT-TOR1A genotype.

DYT-TOR1A and other dystonias show dysfunction in a biochemical pathway, the integrated stress response (ISR), that has wide-reaching effects on the proteome because it regulates global protein synthesis (20). This prompted us to ask how the broad EV compositional differences we observed in the previous 2 experiments were related to ISR pathway effects.

We used ISR tool compounds to modify ISR activity. Our prior studies established the corrective directionality of the eIF2α phosphatase inhibitor salubrinal in DYT-TOR1A cell and mouse model phenotypes and sufficiency of the ISR inhibitor ISRIB to mimic DYT-TOR1A phenotypes (4, 20, 24). We therefore hypothesized that ISRIB-induced EV composition changes in WT MEF EVs would reproduce DYT-TOR1A genotype differences that were related to ISR dysregulation and that salubrinal treatment of DYT-TOR1A MEF EVs would cause normalizing shifts in genotype differences if they were related to ISR dysregulation.

WT MEFs were treated with 50 nM ISRIB to inhibit ISR pathway output for 24 h prior to EV harvest from the conditioned media. ISRIB treatment of WT cells disrupted fewer proteins at the statistical threshold of p ≤ 0.05 than were observed between DYT and WT samples (103/1974 (5%) vs. 363/1974 (18%)) and only 7% of the genotype-disrupted proteins (26/363) were reproduced by ISRIB at the statistical threshold (p ≤ 0.05). However, an examination of proteome-wide effects independent of p-value thresholds showed protein abundance directionality (greater or lesser) to be non-randomly distributed (Fisher’s exact test, p < 0.0001) and in a directionality similar to the DYT genotype effects (Figure 3A). A Pearson’s correlation analysis showed a positive correlation between DYT genotype effects and ISRIB effects, supporting the hypothesis that ISRIB treatment of WT cells mimics DYT genotype effects (Pearson r = 0.40, p < 0.0001)(Figure 3B). Interestingly, as was observed with ritonavir effects, this correlation was also maintained when the entire proteome was evaluated (Pearson r = 0.37, p < 0.0001).

To augment ISR activity in DYT-TOR1A MEFs, cell cultures were treated with 20 μM salubrinal during the 24 h conditioning period prior to EV harvest from the media. Salubrinal is a specific inhibitor of eIF2α phosphatases, CReP and GADD34 (25). Salubrinal treatment of DYT samples significantly modified 9% of the total proteins (169/1974) and caused significant corrective effects on 13% of DYT disrupted proteins (46/363) (Figure 3C). Like ISRIB, secondary analyses of effects independent of p-value thresholds showed that DYT disrupted proteins were not randomly distributed (Fisher’s exact test, p < 0.0001) and showed directionality biases supporting the hypothesis that salubrinal has corrective effects on DYT disruptions (Figure 3C). Lastly, we examined the concordance of drug effects between salubrinal and ritonavir on DYT-TOR1A MEF EV protein abundances, given that both drugs augment ISR activity (4, 25, 26). Ritonavir and salubrinal effects on the putative DYT-TOR1A biomarker proteins were positively correlated (Pearson r = 0.47, p < 0.0001) (Figure 3D). This result is consistent with a degree of shared mechanism of action between salubrinal and ritonavir.

In summary, ISR tool compound experiments demonstrate that ISR activity effects correlate with DYT-TOR1A genotype disruptions and ritonavir corrective effects on MEF EV protein composition. These results support the hypothesis that DYT-TOR1A genotype disruptions of MEF EV protein abundances and the corrective effects of ritonavir treatment are related, at least in part, to ISR pathway activity.

In this study, we have taken advantage of the benefits of control over biological variables that an animal model system affords to generate initial proteomic discovery datasets for putative biomarkers of DYT-TOR1A. To guide translation to dystonia biomarker discovery in future patient-derived cell line or human plasma and CSF samples, we considered the results from our three experimental tests alongside human biospecimen datasets to prioritize candidates with the greatest potential.

Our stratification process considered the following features. First, we identified protein candidates that have been previously detected in human plasma (27). This criterion identified 164 of the 363 genotype-disrupted proteins. Second, we identified candidates that showed conserved directionality of effects across two drug perturbations, independent of effect size or p-value (Ritonavir-DYT, ISRIB-WT) (Supplementary Data Sheet S1 Columns 5,6). This criterion identified 121 of 164 proteins. Then, we created a composite score of genotype and ritonavir effect sizes by summing the absolute value of their respective Cohen’s d score. The results of this analysis are compiled in Supplementary Data Sheet S1. Overall, a third of the DYT-TOR1A genotype disrupted proteins show favorable characteristics according to these prioritizations.

Sample size was arbitrarily set a priori at three samples per group. Experimenters were blinded to MEF cell line genotype and drug treatment prior to cell culture experiments, and proteomics were performed on these blinded sample groups. Experimenters were unblinded after initial differential abundance analyses were completed. For differences in protein abundances, statistical testing used unpaired Student’s t tests between n = 3 WT and n = 3 DYT samples without correction for multiple hypothesis testing or with Bonferroni correction where noted (23). Two-tailed binomial sign test was performed using a null probability p = 0.5. Fisher’s Exact Test was performed on contingency tables for overlaps of the 363 significantly different proteins between conditions using abundances greater than or less than zero log₂ fold change.

ΔE Torsin1a knockin (courtesy of Dr. W. Dauer, UTSW; IMSR_JAX:025637) (19) mice on C57BL/6 background were bred in standard housing conditions with food and water provided ad libitum. All procedures were approved by the Duke University Institutional Animal Care and Use Committee (IACUC).

Mouse embryonic fibroblasts were harvested as previously described (37) from E14 TOR1A ^ΔE/+ mice and immortalized via SV40 transfection. MEFs were maintained in sterile-filtered MEF media [DMEM (Thermo Fisher Scientific, #11995-065) + 10% fetal bovine serum (Hyclone, #SH0071.03) + 1X GlutaMAX (Gibco, #35050-061) + 1% penicillin/streptomycin/amphotericin (Gibco, #15240062) + 1% Non-Essential Amino Acids (Gibco, #11140050)+ 55 nM β-Mercaptoethanol (Gibco, #21985023)] at 37°C/5% CO₂.

EV-depleted (dEV) media was prepared by spinning 10% FBS MEF media for 18 h at 100,000 x g (Beckman L8-55M ultracentrifuge; SW27 rotor; 23,600 rpm; 4°C) (47). MEFs were seeded at 5.8 × 10⁵ cells into one 15 cm dish per line. At 90% confluence, cells were passaged 1:10 into four 15 cm dishes per line and when each line reached ∼50% confluence, media was exchanged for dEV media containing 1% dEV FBS and the given drug treatment. Ritonavir (Tocris Biosciences, #5856), ISRIB (Sigma, #SML0843), and salubrinal (Tocris Biosciences, #2347) were dissolved in DMSO (100 mg/mL) and frozen in aliquots at −20°C. On the day of each treatment, these aliquots were thawed and added to dEV media containing 1% FBS to final concentrations (0.04% DMSO vehicle, 50 nM ISRIB, 20 μM ritonavir, or 20 μM salubrinal). After 24 h in dEV media, the EV-conditioned media and cells were collected separately.

EV-conditioned media was centrifuged at 4°C for 20 min at 2000 x g (Sorvall HS-4, 3500 rpm). Supernatant was transferred to a new tube and centrifuged at 4°C for 30 min at 8000 x g (Sorvall HS-4, 6500 rpm). Final clarified supernatant was stored at −80°C. Media was thawed in room temperature water bath and 36 mL per sample was ultra-centrifuged in Ultra-Clear tubes (Beckman Coulter, #344058) for 16 h at 100,000 x g (Beckman L8-55M ultracentrifuge; SW27 rotor; 23,600 rpm; 4°C) to isolate EVs (21). The supernatant was discarded and protein was extracted from the pellet. Protein was extracted by adding 50 µL modified RIPA buffer [1% Triton X-100, 0.5% SDS, 0.5% deoxycholic acid, 50 mM NaPO₄ at pH 7.4, 150 mM NaCl, 2 mM EDTA, 50 mM NaF, 10 mM sodium pyrophosphate, 1 mM sodium orthovanadate, and protease inhibitor cocktail (Roche, #04693159001)], vortexing on low speed for 15 s, and shaking on an orbital shaker for 1 h at 4°C. Cell lysates were prepared in 1 mL modified RIPA buffer by rotating on a Nutator for 2 h. Lysates were then sonicated and centrifuged for 10 min at 10,000 x g to remove insoluble material, and the supernatant was taken as the whole cell lysate protein fraction.

Total EV protein concentrations were quantified using a Micro BCA™ Protein Assay Kit (Thermo Fisher Scientific, #23235) and cell lysate protein concentrations were quantified by Pierce™ BCA assay (Thermo Fisher Scientific, #23225). Proteins were resolved on 4%–15% TGX gels (BioRad, #5671085), transferred to nitrocellulose membrane, blocked in TBS-T (0.1% Tween-20) with 5% BSA, and probed as indicated. Densitometry was quantified using ImageJ (48). The following primary antibodies and dilution ratios were used for immunoblotting experiments: anti-Actin—1:5000 (Millipore, #MAB1501); anti-TSG101—1:1000 (Abcam, #ab30871); anti-calnexin—1:1000 (Proteintech, #10427-2-AP).

Sample Preparation: The Duke Proteomics and Metabolomics Core Facility (DPMCF) received 18 samples (3 biological replicates each of six conditions). Methods are as described in (37) with minor modifications and restated here for convenience: “Samples were first normalized to 20 μg and spiked with undigested casein at a total of 40, 80, or 160 fmol/μg, then reduced with 10 mM dithiolthreitol for 30 min at 80°C and alkylated with 20 mM iodoacetamide for 30 min at room temperature. Next, they were supplemented with a final concentration of 1.2% phosphoric acid and 741 μL of S-Trap (Protifi) binding buffer (90% MeOH/100 mM triethylammonium bicarbonate). Proteins were trapped on the S-Trap, digested using 20 ng/μL sequencing grade trypsin (Promega) for 1 h at 47°C, and eluted using 50 mM triethylammonium bicarbonate, followed by 0.2% formic acid, and lastly using 50% acetonitrile/0.2% formic acid All samples were then lyophilized to dryness and resuspended in 40 μL 1% trifluoracetic acid/2% acetonitrile containing 12.5 fmol/μL yeast alcohol dehydrogenase (ADH_YEAST). A Sample Pool QC (SPQC) was created from 3 uL of each sample. SPQCs were run periodically throughout the acquisition period.

Quantitative Analysis Methods: Quantitative LC-MS/MS was performed on 2 μL of each sample, using a nanoAcquity UPLC system (Waters Corp) coupled to a Thermo Orbitrap Fusion Lumos high resolution accurate mass tandem mass spectrometer (Thermo) via a nano-electrospray ionization source. Briefly, the sample was first trapped on a Symmetry C18 20 mm × 180 μm trapping column (5 μL/min at 99.9/0.1 v/v water/acetonitrile), after which the analytical separation was performed using a 1.8 μm Acquity HSS T3 C18 75 μm × 250 mm column (Waters Corp.) with a 90-min linear gradient of 5%–30% acetonitrile with 0.1% formic acid at a flow rate of 400 nL/min with a column temperature of 55°C. Data collection on the Fusion Lumos mass spectrometer was performed in a data-dependent acquisition (DDA) mode of acquisition with a r = 120,000 (at m/z 200) full MS scan from m/z 375 – 1500 with a target automatic gain control (AGC) value of 2e5 ions. MS/MS scans were acquired at Rapid scan rate (Ion Trap) with an AGC target of 5e3 ions and a max injection time of 25 m. The total cycle time between full MS scans was 2 s. A 20 s dynamic exclusion was employed to increase depth of coverage.

Proteomics Data Analysis: Following 22 total UPLC-MS/MS analyses (including 4 SPQC injections) were imported into Proteome Discoverer 2.3 (Thermo Scientific Inc.), and analyses were aligned based on the accurate mass and retention time of detected ions (“features”) using Minora Feature Detector algorithm in Proteome Discoverer. Relative peptide abundance was calculated based on area-under-the-curve (AUC) of the selected ion chromatograms of the aligned features across all runs. The MS/MS data was searched against the SwissProt M. musculus database, SwissProt bovine database (downloaded Sept 2019) and an equal number of reversed sequence “decoys” for false discovery rate determination. Mascot Distiller and Mascot Server (v 2.5, Matrix Sciences) were utilized to produce fragment ion spectra and to perform the database searches. Database search parameters included fixed modification on Cys (carbamidomethyl) and variable modifications on Meth (oxidation) and Asn and Gln (deamidation). Full trypsin enzyme rules were selected with 2 ppm precursor and 0.8 Da product ion mass tolerances. Peptide Validator and Protein FDR Validator nodes in Proteome Discoverer were used to annotate the data at a maximum 1% protein false discovery rate.

Following data alignment and AUC quantitation, missing values were imputed in the following manner. If less than half of the values are missing within any one treatment group, values are imputed with an intensity derived from a normal distribution defined by measured values within the same intensity range (20 bins). If greater than half values are missing for a peptide in a group and a peptide intensity is > 5e6, then it was concluded that peptide was misaligned and its measured intensity is set to 0. All remaining missing values are imputed with the lowest 5% of all detected values. These data were then subjected to a sample loading normalization in which the total signals were summed and those summed values were used as normalizing factors across all samples. All peptide AUCs belonging to the same protein were then summed together to generate a protein level intensity” (37).

Data were analyzed using GraphPad Prism v9 and R v4.2.0. Hierarchical clustering was performed in R using Euclidean distance measures and average-linkage clustering (49).

For proteins, our DYT-TOR1A genotype-dependent subset of 363 differential proteins were annotated as “In Human Plasma” based on their presence in a public database, the Human Plasma Proteome Project (HPPP) (50, 51). The HPPP is a set of >3500 proteins that have been detected with varying degrees of evidence in different mass spectrometry studies. We focused on HPPP proteins that were detected in a minimum of 3 distinct studies. This criterion identified 164 of the 363 genotype-disrupted proteins.

We next used Cohen’s d as a standardized effect size for each genotype and drug treatment condition. This was calculated using the formulas below (52), where n ₁ and n₂ are group sample sizes, s ₁ and s₂ are group standard deviations, and s ² _pooled is a pooled variance calculated using both groups’ features. C o h e n ′ s   d = X ¯ Exp . − X ¯ C o n t r o l s p o o l e d s p o o l e d 2 = n 1 − 1 s 1 2 + n 2 − 1 s 2 2 n 1 + n 2 − 2

To rank candidate proteins by their genotype and ritonavir-treatment effect sizes, the absolute value of Cohen’s d for each condition was summed to make a combined score, “Absolute Cohen’s d Sum (Geno+RTV).” Candidate biomarkers were filtered based on the directionality of their pharmacodynamic response to ritonavir and ISRIB being concordant with genotype predictions (ritonavir opposing genotype directionality, ISRIB reproducing genotype directionality). These criteria were then combined to stratify biomarker subsets as displayed in Supplemental Data Sheet S1.