Proteome profiling of cerebrospinal fluid-derived extracellular vesicles reveals potential biomarkers for drug-resistant epilepsy

Abstract

Background

Epilepsy is one of the most common neurological disorders in humans and in dogs. Treatment currently focuses on alleviating symptoms, and a wide range of anti-seizure medications (ASMs) is available. Still, over one-third of patients have an inadequate response to ASM. The proteome of cerebrospinal fluid (CSF)-derived extracellular vesicles (EVs) offers a potential source of biomarkers for drug-resistant epilepsy (DRE).

Methods

We utilised a spontaneous canine epilepsy model to study the proteomic content of CSF-derived EVs as a source of biomarkers for DRE. We included 37 drug-naïve dogs with recent onset epilepsy and confirmed diagnosis of idiopathic epilepsy. CSF samples were collected at the onset of epilepsy. After the first visit, ASM treatment was started in all dogs and they were followed up for at least 12 months. After the follow-up period, based on their response to ASM treatment, dogs were grouped as either drug-responsive or drug-resistant. We isolated CSF-derived EVs with ultrafiltration combined with size-exclusion chromatography and then performed proteomic analysis with liquid chromatography-tandem mass spectrometry. A comparison between the drug-responsive and drug-resistant dogs was conducted regarding clinical factors and CSF-derived EV proteomic data.

Results

Younger age at seizure onset and occurrence of cluster seizures were identified as risk factors for drug-resistance. The proteomic analysis of normalised data identified five proteins with differential abundance between the two groups: KRT4, an uncharacterised immunoglobulin-like domain-containing protein (IgDCPa), F2, DSC1b, and LOC607874. A receiver operating characteristic analysis was performed, revealing a predictive value of ≥ 0.90 for two combinations of three proteins (KRT4, IgDCPa, and F2 (area under curve (AUC) = 0.91, confidence interval (CI) = 0.78-1.00); DSC1b, F2, and IgDCPa (AUC = 0.90, CI = 0.78-1.00)).

Conclusions

Proteins with differential abundance studied here are associated with epilepsy due to their potential involvement in critical processes such as neuroprotection, inflammation, cell integrity, and immune response. The observed reduction in the abundance of these proteins in drug-resistant dogs suggests that disruptions in these processes may contribute to the severity of the condition and its resistance to treatment. Results from this pilot study warrant further study in a larger cohort.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12014-025-09569-x.

Background

Epilepsy is one of the most common neurological disorders worldwide in both humans and in dogs [1, 2]. The main emphasis on epilepsy treatment is currently still on alleviating the symptoms of the disease i.e., seizures, but despite the wide range of antiseizure medications (ASMs) available, approximately one third of canine and human patients remain pharmacoresistant (drug-resistant epilepsy (DRE)) [2–5]. Diagnosis of DRE may currently require a long treatment cycle of several years potentially aggravating patients’ cerebral function and leading in delays of optimal treatment [6]. The identification of robust biomarkers for DRE would provide a basis for tailored individualised treatment at the onset of epilepsy. Here, we utilised a spontaneous canine idiopathic epilepsy (IE) model to study biomarkers for DRE. Canine IE is defined as ‘two or more unprovoked seizures at least 24 h apart with no identifiable underlying aetiology other than a suspected genetic origin’ [7]. Epilepsy in dogs closely resembles epilepsy in humans for instance in symptoms, prevalence, genetic background, environmental risk factors, and occurrence of comorbidities [8]. Thus, a spontaneous canine epilepsy model offers a reliable model for epilepsy research.

A biomarker is an objectively measured characteristic of a normal or pathological biological process. Biomarkers for drug resistance in epilepsy could predict patients who are likely to develop DRE, diagnose drug resistance at an early stage to guide optimal treatment, and they can also be used as novel targets to treat, and even prevent or cure drug resistance [9]. Due to their high stability and prevalence in all biofluids, extracellular vesicles (EVs) have been a popular focus in biomarker studies [10]. EVs are small membrane-covered particles that are secreted by cells. They are involved in intercellular communication by transporting components and acting as signalling vehicles involved in both normal cellular functions and pathological developments. The cargo of EVs consists of proteins, microRNA, and lipids among others [11]. Studying the cargo of cerebrospinal fluid (CSF)-derived EVs provides a promising way of studying brain-related diseases without brain biopsies, as CSF is in close contact with the central nervous system (CNS) and the cargo of EVs reflects the cells of their origin [12–14]. While the proteomic content of for example blood, tissue samples, and whole CSF has been studied in connection to DRE [15–17], to our knowledge, CSF EV proteomics have not been explored in this condition. However, changes in CSF EV cargo have been detected for example in patients with Alzheimer’s disease and amyotrophic lateral sclerosis, thus providing evidence that CSF derived EVs can be used as a source of biomarkers for diseases of the brain [18, 19].

Here, we recruited client-owned dogs with recent onset IE and classified them as either drug-responsive or drug-resistant based on their response to ASM. A comparison was then conducted between the groups regarding clinical factors and CSF-derived EV proteomic data, with the aim to find novel biomarkers for DRE.

Materials and methods

Ethical approval

We recruited client-owned dogs for the study, which was approved by the Viikki Campus Research Ethics Committee at the University of Helsinki, Finland. All owners of the dogs participating in the study provided a written consent and were made aware that they could exit the study at any time without consequences.

Pilot samples

CSF was obtained from two euthanised dogs for general characterisation of canine CSF-derived EVs. The carcasses were donated for teaching and research purposes with a written consent from the owners. The reasons for euthanasia were not related to the study.

The CSF was collected immediately after euthanasia via a cisternal tap with a sterile puncture (40 mm/22G needle, #4507401, B. Braun Melsungen AG, Melsungen, Germany) into 2 ml plain tubes (#72.609.001, Sarstedt AG & Co. KG, Nümbrecht, Germany). Samples were visually assessed to be clear of any visible red blood cell contamination. An aliquot of 160 µl of the CSF sample was taken directly after collection for the CSF’s total nucleated cell count and total protein concentration determinations to ensure these were in normal ranges in all samples. The reference value for cell count was < 5 cells/µl and for protein concentration < 300 mg/l [20]. The remainder of the sample was cleared by pelleting any cellular material by centrifugation at 4000xg for 10 min at room temperature. The supernatant was then divided into 1000 µl aliquots into Protein LoBind tubes (#022431064, Eppendorf AG, Hamburg, Germany), snap-frozen in liquid nitrogen, and stored at − 80 °C until further use.

Clinical characterisation of study population

Dogs with recent onset epilepsy were enrolled to the study; their first seizure had occurred a maximum of six months before the initial visit. All included dogs had had at least two generalised tonic-clonic seizures (evaluation based on video-footage taken by the owner) more than 24 h apart. The included dogs had not received any ASMs prior to the study. Only dogs weighing > 10 kg were enrolled.

The dogs were diagnosed with IE at the first visit based on International Veterinary Epilepsy Task Force (IVETF) criteria and at least Tier II confidence level [21]: (1) onset of seizures was between the age of 6 months and 6 years (2), abnormalities were not detected in clinical and neurological examination (3), abnormalities were not detected in routine blood samples (complete blood cell count (CBC) and basic serum biochemistry) or urinalysis, and (4) magnetic resonance imaging (MRI) (1.5T) of the brain and CSF analysis revealed no abnormalities.

After the first visit, same ASM (i.e., phenobarbital (PB)) treatment was initiated in all dogs. The included dogs were followed up for at least 12 months and control visits took place 3 and 12 months after the initial visit. During the follow-up the owners of dogs kept a detailed diary about seizure occurrence. During control visits general clinical and neurological examination, CBC, and serum biochemistry analysis were performed. Also, the serum PB level was measured to confirm that it is at therapeutic level (≥ 20 µg/ml) [22]. During the follow-up period PB dosage was titrated to the maximum tolerated effective dose with serum concentration of ≤ 30 µg/ml.

The clinical examinations were performed at the Veterinary Teaching Hospital of the University of Helsinki, Finland.

Cerebrospinal fluid collection

From the dogs with IE, CSF was collected at the first visit under general anaesthesia after the MRI. A maximum of 1 ml of CSF per 10 kg of body weight was collected via a cisternal tap with a sterile puncture (40 mm/22G needle, #4507401, B. Braun Melsungen AG, Melsungen, Germany) into 2 ml plain tubes (#72.609.001, Sarstedt AG & Co. KG, Nümbrecht, Germany). The samples from dogs with IE were assessed and processed in the same manner as the pilot samples. However, they were divided into 200 µl aliquots using the same Protein LoBind tubes (#022431064, Eppendorf AG, Hamburg, Germany) before being snap-frozen and stored at − 80 °C until further use.

Classification

Phenobarbital treatment was initiated in all dogs after the first visit. After completing the follow-up period of 12 months, based on their response to PB treatment, the dogs were grouped to either drug-responsive or drug-resistant. Treatment success was defined as a 3-fold extension of the longest baseline interseizure interval and a minimum of three months [23]. In case of intolerable side-effects due to PB therapy at therapeutic serum levels, the patient was excluded from the study because treatment response to PB monotherapy could not be assessed.

Extracellular vesicle isolation and characterisation

Extracellular vesicle isolation

For general characterisation, EVs were isolated from a 4 ml pilot sample consisting of pooled CSF from two dogs. For the epileptic dogs, individual samples with the starting volume of 500 µl were used. Essentially, the same protocol was used as described in our previous publication [24], but the buffer volume was adjusted to suit Gen 2 columns and an automatic fraction collector was used.

The aliquoted CSF samples were thawed on ice, pooled, and 1/100 volume of protease inhibitor (#P8340, Sigma-Aldrich, St. Louis, Missouri, United States) was added to the sample. The sample was centrifuged at 4,000xg for 20 min at 4 °C to remove any remaining debris. The sample was concentrated by centrifugation at 4,000xg at 22 °C using Amicon^® Ultra-4 Centrifugal Filter Unit with 100 kDa molecular weight cutoff (UFC810024, Merck Millipore, Burlington, Massachusetts, United States). The final volume of each sample was adjusted to 150 µl with PBS and subjected to size-exclusion (SEC) isolation with qEV single/70nm Gen 2 columns (ICS-70, Izon, Christchurch, New Zealand) according to the manufacturer’s protocol. Briefly, the column was attached to The Automatic Fraction Collector (AFC) V2 (AFC-V2, Izon, Christchurch, New Zealand) and rinsed with PBS. The concentrated sample was loaded into the column and PBS was added as needed. The buffer volume was set to 300 µl and ten (pilot sample) or six (samples from dogs with IE) fractions of 500 µl were collected. Fraction 2 was the calculated EV fraction.

A 20 µl aliquot of the fractions (fractions 1–10 for the pilot sample and fraction 2 for the samples from dogs with IE) was reserved for nanoparticle tracking analysis (NTA). Also, from the pilot sample, 25 µl of the fractions 1–10 were aliquoted for BCA protein assay, 90 µl of the fractions 1–9 for Western blot, and 30 µl of the fraction 2 for transmission electron microscopy (TEM). Samples were stored at −80 °C until analyses. Rest of the fraction 2 (335 µl for the pilot sample and 480 µl for the samples from dogs with IE) was used for proteomic analysis with liquid chromatography-tandem mass spectrometry (LC-MS/MS). For the pilot sample, LC-MS/MS was used for identifying EV markers.

Nanoparticle tracking analysis

Particle concentration and size in the SEC fractions (fractions 1–10 for the pilot sample, fraction 2 for the samples from dogs with IE) was measured with the ZetaView PMX-120 NTA instrument (Particle Metrix GmbH, Ammersee, Germany) equipped with a Z NTA cell assembly, a blue (488 nm, 40mW) laser and a complementary metal-oxide-semiconductor camera with 640 × 480-pixel resolution. Samples were diluted in a total volume of 1 ml of particle-free ultra-pure milli-Q water to obtain 50–200 particles per frame. Videos in NTA mode were recorded at 11 positions across the measurement chamber in 2-second increments at 30FPS frame rate with camera shutter speed at 100s-1 and sensitivity at 85. Temperature was controlled at 22 °C for NTA. Videos were processed and outliers (> 10% coefficient of variation) were removed using the Grubbs method with the built-in ZetaView software (version 8.05.12 SP2). Particles between 10 and 1000 nm in diameter with a minimum trace length of 15 frames and a minimum brightness of 20 were included in the analysis.

Bicinchoninic acid assay

The Pierce BCA Protein Assay Kit (#23227, Thermo Scientific™, Waltham, Massachusetts, United States) was used according to the manufacturer’s instructions to measure the protein concentration of the SEC fractions 1–10 from the pilot sample. Briefly, two replicates of 25 µl of standard solution and one set of each SEC fraction were pipetted on the well plate and mixed with 200 µl of working reagent. The well plate was incubated at 37 °C for 30 min and read with the Multiskan GO instrument and the Thermo Scientific SkanIt software (Thermo Scientific™, Waltham, Massachusetts, United States) at a wavelength of 562 nm.

Western blot

Western blot analysis was performed as a negative control for SEC fractions 1–9 of the pilot sample with two antibodies: albumin and ApoA1. An aliquot of each fraction was mixed with a self-made 10X sample buffer (sodium dodecyl sulfate (4%), tris hydrochloride (0.125 M), glycerol (20%), bromophenol blue (0.006%), 2-mercaptoethanol (10%)) and boiled for 10 min. Samples of 10 µl were then loaded into 4–20% Mini-PROTEAN^® TGX™ Precast Protein gels (#4561093, Bio-Rad Laboratories, Hercules, California, United States) and the gels were run at 200 V for 35 min. The proteins were transferred to nitrocellulose membranes with semi-dry protocol using Power Blotter Select Transfer Stacks (#PB3310, Invitrogen, Waltham, Massachusetts, USA) and Power Blotter Station (Invitrogen, Waltham, Massachusetts, USA). The membranes were blocked for 1 h in TBS-Tween (0.1% Tween-20) 5% milk. The membranes were incubated overnight with the primary antibodies against albumin (dilution 1:40,000, #ab194215, Abcam, Cambridge, UK) or ApoA1 (dilution 1:3,000, #ab227455, Abcam, Cambridge, UK) in the blocking buffer. The following day, the membranes were washed with TBS-Tween three times and incubated with secondary antibodies: Donkey anti Sheep/Goat IgG (albumin, dilution 1:100,000, #STAR88P, Bio-Rad Laboratories, Hercules, California, United States), Goat Anti-Rabbit Immunoglobulins/HRP (ApoA1, dilution 1:100,000, #P0448, Dako Denmark A/S, Glostrup, Denmark) for 1 h in RT. The washing steps were repeated, SuperSignal™ West Atto Ultimate Sensitivity Substrate (#A38554, Thermo Scientific™, Waltham, Massachusetts, United States) was applied on the membranes, and the chemiluminescence was detected using FujiFilm LAS-3000 imager.

Transmission electron microscopy

For EV characterisation, SEC fraction 2 of the pilot sample was visualised with TEM. For this, 200 square Mesh copper 3.05 mm (G2200C, Agar Scientific, Stansted, UK) grids were used. The film was handmade in 2% biofoform in chloroform solution (SPI-Chem™ Pioloform^® Resin, #2466, SPI Supplies, West Chester, PA, USA). Bal-Tec CED030 carbon thread evaporator (Bal-Tec Union Ltd., Liechtenstein) was used for the carbon coating and Emitech K100X Glow discharge unit (Emitech Ltd., UK) as the glow discharge system. A grid was placed on a 10 µl droplet of sample on parafilm for 2 min, and the buffer salts were removed by transferring the grids twice to a fresh drop of distilled water for 5 s each. The excess fluid was removed with filter paper and the grids were transferred on a drop of uranyl acetate (1.5% in distilled water) for 1 min. Excess fluid was removed with filter paper and the grids were air dried prior to imaging. The samples were imaged using a Jeol JEM-1400 transmission electron microscope with 80,000 V and magnifications of 4,000X and 10,000X.

Proteomic analysis

The EV sample (335 µl for the pilot sample and 480 µl for the samples from dogs with IE) was lysed, and using a previous protocol for protein aggregation capture [25], proteins were precipitated and digested into peptides with trypsin. Briefly, acetonitrile (ACN) was added to the samples resulting in a final ACN concentration of 70%. MagReSynAmine beads (ReSyn Biosciences) were used to precipitate the proteins, which were then captured with a magnetic rack and washed with 100% ACN and 70% ethanol. Finally, 50mM NH4HCO3 was added to the beads and dithiothreitol was used to reduce the proteins, iodoacetic acid to alkylate them, and trypsin (Promega) to on-beads digest them overnight.

Self-made C18 Stage tips-columns were used to purify the resulting tryptic peptides, which were then subjected to nanoLC-MS/MS analysis with nanoElute coupled to timsTOF fleX (Bruker). For peptide separation, a 60-min linear separation gradient with 0–35% ACN using 25 cm Aurora C18 column (Ion Optics) was used. The timsTOF fleX was operated in PASEF mode and mass spectra for MS and MS/MS scans were recorded between m/z 100 and 1700. Ion mobility resolution was set to 0.85–1.35 V·s/cm over a ramp time of 100ms, and data-dependent acquisition was done using four PASEF MS/MS scans per cycle with a near 100% duty cycle. To exclude low m/z, singly charged ions from PASEF precursor selection, a polygon filter was applied in the m/z and ion mobility space. To precursors that attained 20,000 intensity units, an active exclusion time of 0.4 min was applied and collisional energy was ramped stepwise as a function of ion mobility.

The data from LC–MS/MS was analysed by MaxQuant ver 2.4.3.0 against dog (Canis lupus) database downloaded from Uniprot in June 2024 for protein identification. The searches were done with ‘match between runs’ active and using MaxQuant’s default parameters with a 1% false discovery rate on protein and peptide level. The differential abundance analysis of the proteomic data was done in Perseus (version 1.6.15.0). Both normalised data (LFQ) and data from raw intensities (INTENSITY) was used in the analysis. The data was log₁₀ transformed, filtered with at least 50% valid values in at least one of the groups, missing values were imputed with constant = 0, and t-test with p < 0.05 as the criteria was done.

The datasets generated and/or analysed during the current study are available in the ProteomeXchange Consortium via the PRIDE [26] partner repository with the dataset identifier PXD061528. We performed proteomic analysis also from dogs excluded from the present study and this data is also available in the repository. In the dataset, one canine sample is entirely absent due to its status as an outlier, characterised by a substantially lower intensity compared to the other samples. However, this particular dog was already initially categorised as excluded.

Additional data analysis

Frequency tables and descriptive statistics were provided for potential clinical variables by responsiveness-status (drug-responsive vs. drug-resistant). Here, cluster seizures were defined as two or more seizures occurring within a 24-hour period, and status epilepticus as a condition involving prolonged seizure lasting more than five minutes or multiple seizures with incomplete recovery of consciousness between them [1]. The associations of clinical variables and the responsiveness to treatment were first assessed with univariate logistic regression analyses, where each clinical variable was assessed separately. Secondly, variables shown statistically significant (p < 0.05) in the univariate analyses were included in a multivariate model. Models were constructed to assess the probability of non-responsiveness to treatment. Odds ratios (ORs) and their 95% confidence intervals (CI) were used to quantify the results.

Mann-Whitney U-test on NTA particle concentration data for comparison between drug-responsive and drug-resistant dogs was performed and plots were created in RStudio 2024.04.2, Build 764. The heatmap was done on normalised proteomic data without clustering and also using the ‘pheatmap’ package’s own row-wise data normalisation. The Venn diagram was created in FunRich version 3.1.4 [27–29]. The biological functions of proteins with significant differential abundance between the groups were analysed using Uniprot [30] and by reviewing relevant literature.

To evaluate the discriminative performance of proteins with differential abundance, we performed the receiver operating characteristic (ROC) curve analysis of both individual proteins and combined sets of proteins. First, individual proteins with differential abundance between the groups were modelled separately. Youden’s index (maximising the sum of sensitivity and specificity) was used to define the optimal cut-off value for each protein. Sensitivities and specificities (together with 95% CIs) were calculated using the optimal cut-offs for each protein. In addition, ROC-curves were constructed using combinations of several proteins [2–4] as predictors. Areas under the curve (AUCs), together with 95% CIs were calculated for each ROC-curve to compare the different models. These statistical analyses were done using SAS System for Windows, version 9.4 (SAS Institute Inc., USA).

Results

Descriptive data of the study population

Overall, 50 dogs fulfilled the study inclusion criteria and out of the total, 37 dogs were included in the analyses. Of these, 24 were classified as drug-responsive and 13 as drug-resistant. Of the drug-responsive dogs, 15 were seizure free after the start of ASM therapy when the serum PB level was adequate. Altogether 20 different dog breeds were represented in the study, with Border Collies (n = 9) and Labrador retrievers (n = 5) being the most common ones. Among a variety of clinical factors compared between drug-responsive and drug-resistant dogs, a younger age at seizure onset and the occurrence of cluster seizures were identified as risk factors for drug-resistance (Tables 1 and 2). Importantly, these factors were found significant in both univariate and multivariate analysis (Tables 1 and 2). Furthermore, a trend was observed between having a family history of epilepsy and being drug-resistant, although the p-value fell slightly short of statistical significance.

Table 1.

Logistic regression for the association of clinical variables and drug-responsiveness; univariate analysis of variables with significant differences, odds ratios for the probability to be drug-resistant

Variable	Comparison	OR	Lower 95% CI	Upper 95% CI	p-Value
Age at onset (−100 days)	Age at onset (−100 days)	1.37	1.10	1.72	0.006**
Breed	OB vs. BC	1.54	0.31	7.72	0.601
Family history (epilepsy)	Yes vs. No	3.89	0.94	16.11	0.061
Frequency of seizures per 30 days pre-treatment (category)	At least 5 vs. < 5	0.58	0.05	6.25	0.656
Mixed seizure types	Yes vs. No	0.89	0.21	3.80	0.874
Cluster seizures	Yes vs. No	11.20	2.16	58.13	0.004**
Neutering status	Yes vs. No	1.69	0.36	7.84	0.504
Number of seizures pre-treatment	Number of seizures pre-treatment	1.22	0.69	2.15	0.495
Sex	M vs. F	0.45	0.11	1.78	0.253
Status epilepticus	Yes vs. No	0.92	0.08	11.18	0.946

PB = phenobarbital, OR = odds ratio, CL = confidence limit, OB = other breed, BC = Border collie. **p ≤ 0.01

Table 2.

Logistic regression for the association of clinical variables and drug-responsiveness; multivariate analysis, odds ratios for the probability to be drug-resistant

Variable	Comparison	OR	Lower 95% CL	Upper 95% CL	p-Value
Age at onset (−100 days)	Age at onset (−100 days)	1.34	1.06	1.70	0.013*
Cluster seizures	Yes vs. No	11.24	1.49	84.97	0.019*

PB = phenobarbital, OR = odds ratio, CL = confidence limit. *p ≤ 0.05

In total, 13 dogs were excluded from the analyses. In five cases, the ASM therapy was not started despite IE diagnosis; in seven cases the side-effects of PB were intolerable, therefore preventing the usage of PB at a therapeutic level; and one dog was diagnosed with structural epilepsy, thus excluded from the analyses.

Extracellular vesicle characterisation

Prior to the proteomic analysis, the pilot sample was used to characterise EVs following the principles of the MISEV2023 guidelines [31]. First peak of particles was detected with NTA in fraction 2, and the peak was separate from the highest measured protein concentration (Fig. 1a). Another peak in fraction 4 is likely to contain more lipoproteins as we showed in our earlier study [24], and here we wanted to exclude those from the analysis. The NTA revealed also that the modal particle size of particles in fraction 2 was approximately 100 nm, which is in the correct size range for EVs [32] (Fig. 1b). Particles with typical EV morphology were visualised with TEM in fraction 2 (Fig. 1c). Importantly, when analysing EV markers with LC-MS/MS, 21 proteins identified in the fraction 2 aligned with Vesiclepedia’s top 100 protein list. These included for example the heterotrimeric G protein GNAI2 and actin ACTB (Fig. 1 d). Additionally, Western blots showed that negative controls Albumin and ApoA1 were separated from fraction 2 (Fig. 1e). The basic characterisation identified SEC fraction 2 as our EV fraction, thereby justifying its selection for proteomic analysis on the samples from dogs with IE.

Fig. 1.

Extracellular vesicle characterisation performed on EVs isolated by UF-SEC from pooled (4 ml) canine CSF. (a) Particle count by NTA and protein quantity by BCA of SEC fractions (500 µl each). (b) Particle size distribution of SEC fraction 2 measured with NTA. (c) TEM images of SEC fraction 2 showing cup-shaped particles typical for EVs [32]. (d) Proteins identified with LC-MS/MS in SEC fraction 2 compared with Vesiclepedia’s top 100 EV proteins. (e) Western blot analysis of negative controls albumin and ApoA1 of SEC fractions 1–9. Abbreviations: EV = extracellular vesicle, UF-SEC = ultrafiltration combined with size-exclusion chromatography, CSF = cerebrospinal fluid, NTA = nanoparticle tracking analysis, BCA = bicinchoninic acid assay, TEM = transmission electron microscopy, LFQ = label-free quantification, LC-MS/MS = liquid chromatography-tandem mass spectrometry

Nanoparticle tracking analysis

In NTA, no significant difference was found in the particle concentration (particles/ml) of the EV samples between drug-responsive and drug-resistant groups (p = 0.86), indicating that the samples contained similar numbers of particles (Fig. 2a).

Fig. 2.

Differences in CSF EVs between drug-responsive and drug-resistant dogs. (a) Particle concentration in CSF EV fraction in drug-responsive and drug-resistant groups. (b) Results of PCA done with normalised proteomic data. (c) Volcano plot of the normalised proteomic dataset, significant proteins are labelled. (d) Heatmap of the proteins with differential abundance using normalised proteomic data. Colour scale represents row-wise normalised intensities. (e) Comparisons of the significantly differing proteins’ normalised intensities between responders and non-responders. (f) Volcano plot of the raw data, significant proteins with log₁₀ difference > 1 are labelled. (g) Comparisons of the significantly differing proteins’ with log₁₀ difference of > 1 in the raw data between responders and non-responders. Abbreviations: NR = drug-resistant group, R = drug-responsive group, PC = principal component, LFQ = label-free quantification, CSF = cerebrospinal fluid, EV = extracellular vesicle, PCA = principal component analysis, IgDCPa = immunoglobulin-like domain-containing protein a, IgDCPb = immunoglobulin-like domain-containing protein b. *p ≤ 0.05, **p ≤ 0.01

Proteomic analysis

Altogether 687 proteins were identified in the samples with LC-MS/MS. The proteomic data was then analysed for quantitative differences between drug-responsive and drug-resistant dogs. Principal component analysis using the normalised proteomic dataset showed that both groups’ results overlapped, with only a slight separation from one another (Fig. 2b). Still, differential abundance analysis revealed significant differences between the groups for five individual proteins in the normalised data: KRT4, an uncharacterised immunoglobulin-like domain-containing protein (shortened as IgDCPa here), F2, DSC1b, and LOC607874 (Table 3, the whole normalised proteomic data with dog classifications used for analyses is available in Supplementary Table 1). All these five proteins were downregulated in the drug-resistant group (Fig. 2c and e). Normalisation in our data was based on ‘MaxLFQ-algorithm’ which often normalises low intensity values from MaxQuant output into zero, and data from these low abundant proteins is lost in LFQ. Thus, we also did quantitative comparison analysis based on non-normalised intensities, focusing only on proteins which had a log₁₀ difference of > 1, meaning over ten-fold difference in abundance between the groups. Eight proteins fulfilled these criteria, namely NEGR1, DSC1b, LOC476816, another uncharacterised immunoglobulin-like domain-containing protein (shortened as IgDCPb here), OGN, F2, CACNA2D1, and LSAMP. Out of these, F2 and DSC1b also showed significant differences in the normalised data. The other six were not detected in the normalised data (Fig. 2f g, Table 3) (The whole raw intensity data with dog classifications used for analyses is available in Supplementary Table 2).

Table 3.

Proteins with significant differences between drug-responsive and drug-resistant groups in the normalised and significant differences with > 1 log₁₀ difference in the Raw proteomic data

LFQ					Intensity
Protein	Unique (+ razor) peptides	Total intensity	Log10 diff., R vs. NR	p-Value	Unique (+ razor) peptides	Total intensity	Log10 diff., R vs. NR	p-Value
KRT4	0 (2)	1,00E + 08	2.70	0.005**	0 (2)	1,06E + 08	0.58	0.013*
IgDCPa	1 (7)	2,00E + 07	1.25	0.006**	1 (7)	1,88E + 07	0.83	0.012*
F2	6 (6)	2,00E + 06	1.70	0.012*	6 (6)	2,44E + 06	21,551	0.041*
DSC1b	2 (2)	6,00E + 06	1.83	0.019*	2 (2)	6,08E + 06	25,204	0.017*
LOC607874	6 (6)	9,00E + 06	1.60	0.031*	6 (6)	8,61E + 06	1.18	0.120
NEGR1	-	-	-	-	3 (3)	1,30E + 06	1.93	0.01**
LOC476816	-	-	-	-	3 (3)	7,74E + 05	1.69	0.018*
IgDCPb	-	-	-	-	1 (1)	1,72E + 06	1.67	0.027*
OGN	-	-	-	-	4 (4)	9,30E + 05	1.61	0.033*
CACNA2D1	-	-	-	-	3 (3)	4,78E + 05	1.43	0.043*
LSAMP	-	-	-	-	3 (3)	8,85E + 05	1.45	0.048*

LFQ = label free quantification, R = drug-responsive group, NR = drug-resistant group, IgDCP = immunoglobulin-like domain-containing protein. *p ≤ 0.05, **p ≤ 0.01

Numbers of peptides are given as unique peptides as well as unique + razor peptides

Receiver operating characteristic analysis

To assess the predictive value of individual proteins and their combinations for drug-resistance, ROC analysis was performed on the normalised data. Univariate analysis yielded predictive values of ≥ 0.70 for three proteins: F2 (AUC = 0.70, confidence interval (CI) = 0.55–0.84), KRT4 (AUC = 0.72, CI = 0.58–0.86), and IgDCPa (0.71, CI = 0.52–0.90) (Fig. 3a; Table 4). Multivariate analysis was then performed to analyse predictive values of combinations of two, three, or four proteins. Adding a second protein increased the predictive power compared to a single protein (Fig. 3a-b; Table 4). Also, adding a third protein increased the predictive power compared to models with two proteins (Fig. 3b-c; Table 4). Two combinations of three proteins resulted in predictive values of ≥ 0.90, these being KRT4, IgDCPa, and F2 (AUC = 0.91, CI = 0.78–1.00.78.00) as well as DSC1b, F2, and IgDCPa (AUC = 0.90, CI = 0.78–1.00.78.00) (Table 4; Fig. 4a and b). Adding a fourth protein to the model resulted in no remarkable improvement compared to the models with three proteins, as the combination with the highest predictive value – KRT4, DSC1b, F2, and IgDCPa – achieved an AUC of 0.92 (CI = 0.80–1.00.80.00) (Fig. 3 d; Table 4)

Fig. 3.

ROC curves with AUCs of the proteomic data assessing predictive values of individual proteins and their combinations for DRE. (a) Single proteins. (b) Combinations of two proteins. (c) Combinations of three proteins. (d) Combinations of four proteins. Abbreviations: ROC = receiver operating characteristic, AUC = area under the curve, DRE = drug-resistant epilepsy, IgDCP = immunoglobulin-like domain-containing protein

Table 4.

ROC analysis for proteins and combinations of proteins

Protein	AUC	Lower 95% CL	Upper 95% CL
Univariate
KRT4	0.72	0.58	0.86
IgDCPa	0.71	0.52	0.90
F2	0.70	0.55	0.84
DSC1B	0.67	0.49	0.86
LOC607874	0.61	0.38	0.84

Models with two proteins
KRT4 & DSC1b	0.80	0.66	0.95
KRT4 & IgDCPa	0.87	0.71	1.00
KRT4 & F2	0.81	0.68	0.94
KRT4 & LOC607874	0.76	0.58	0.93
DSC1B & IgDCPa	0.83	0.68	0.99
DSC1B & F2	0.81	0.67	0.95
DSC1B & LOC607874	0.74	0.57	0.91
F2 & IgDCPa	0.81	0.67	0.95
F2 & LOC607874	0.72	0.54	0.90
IgDCPa & LOC607874	0.73	0.54	0.92

Models with three proteins
KRT4 & DSC1b & F2	0.87	0.75	0.99
KRT4 & IgDCPa & F2	0.91	0.78	1.00
KRT4 & DSC1B & IGDCP	0.88	0.72	1.00
DSC1B & F2 & IgDCPa	0.90	0.78	1.00
DSC1B & F2 & LOC607874	0.83	0.70	0.97
DSC1B & IgDCPa & LOC607874	0.86	0.71	1.00
DSC1B & KRT4 & LOC607874	0.81	0.66	0.97
F2 & IgDCPa & LOC607874	0.81	0.67	0.95
KRT4 & F2 & LOC607874	0.81	0.66	0.95
KRT4 & IgDCPa & LOC607874	0.88	0.73	1.00

Models with four proteins
KRT4 & DSC1b & F2 & LOC607874	0.87	0.75	1.00
KRT4 & IgDCPa & F2 & LOC607874	0.91	0.79	1.00
KRT4 & DSC1b & IgDCPa & LOC607874	0.89	0.73	1.00
KRT4 & DSC1b & F2 & IgDCPa	0.92	0.80	1.00
IgDCPa & DSC1b & F2 & LOC607874	0.90	0.79	1.00

ROC = receiver operating characteristic, AUC = area under the curve, CL = confidence limit, IgDCP = immunoglobulin-like domain-containing protein

Fig. 4.

ROC curves with AUCs of the proteomic data assessing two combinations of three proteins with the highest predictive value for drug-resistance. (a) KRT4, IgDCPa, and F2. (b) DSC1b, F2, and IgDCPa. Abbreviations: ROC = receiver operating characteristic, AUC = area under the curve, CI = confidence interval, DRE = drug-resistant epilepsy, IgDCP = immunoglobulin-like domain-containing protein

Discussion

In the present study, we explored the CSF-derived EV protein content to reveal biomarkers for DRE. The proteomic analysis identified five proteins with differential abundance between the two groups in the normalised data: KRT4, IgDCPa, F2, DSC1b, and LOC607874. ROC analysis revealed that the best predictive model was achieved with combinations of three proteins. NTA did not demonstrate significant difference in the total number of particles between the groups, indicating that the observed differences in the proteomic analysis were not affected by a differing number of EVs. Additionally, we analysed clinical factors and their association with drug-resistance. We found younger age at onset of seizures as well as experiencing cluster seizures to be significant predictors for drug-resistance, aligning with previous research as both factors have been reported to predict DRE in both humans and dogs [5, 33–35].

All five significant proteins in the normalised data were downregulated in the non-responders. ROC analysis was then used to study the proteins’ potential to predict DRE. The analysis revealed three AUC values of ≥ 0.70, these being the values of F2, KRT4, and IgDCPa. However, the best predictive performance was achieved using combinations of three proteins: KRT4, IgDCPa, and F2 as well as DSC1b, F2, and IgDCPa, both of which demonstrated AUC values of ≥ 0.90. Adding a fourth protein didn’t result in any noteworthy improvement in AUC.

The protein F2, also known as prothrombin, is a precursor to thrombin [30, 36]. It is converted into thrombin through a proteolytic cleavage [36]. Thrombin is a serine protease known to be involved in coagulation, inflammation, cell protection, and apoptosis. While the liver is the main site of prothrombin production, growing evidence suggest that small amounts are also produced locally in the brain [36]. Low concentration of thrombin protects the CNS, but abnormally high concentrations, as a result of BBB breakdown, can lead to pathological effects such as increased sensitivity to seizures [36, 37]. In our study, CSF samples were taken >24 h after the last seizure, and neither the number of seizures nor seizure frequency differed significantly between the groups. Also, the concentration of F2 was lower in drug-resistant dogs than in drug-responsive ones. This suggests that the differences in F2 were probably not due to BBB breakdown caused by seizures. Our study suggests that the reduced prothrombin levels observed in drug-resistant dogs may not be sufficient for its neuroprotective role, potentially contributing to their increased susceptibility to seizures.

Keratins – such as KRT4 or keratin 4 in the present study [30] – are intermediate filament proteins of epithelia [38]. While important to the structure of epithelial cells and tissues, there are also studies connecting keratin to functions of the brain. For example, keratin 28 has been identified to regulate demyelination in forebrain slice cultures [39]. Also, differences in the expression of both keratin 1 and keratin 9 have been found in the CSF of multiple sclerosis, neuromyelitis optica, and control individuals [40]. Additionally, keratin 9 has been suggested as a part of a biomarker panel Alzheimer’s disease, probably due to dysregulated signalling pathways and break-down in the BBB [41]. However, it remains currently unknown how KRT4 is associated with DRE. It could potentially involve changes in cellular structure, stress response, or signalling in cells, possibly affecting neuronal function or drug metabolism [38].

Proteins containing immunoglobulin-like domains are part of the immunoglobulin superfamily (IgSF), which is one of the most common protein classes [42, 43]. Here, a significant difference was detected in the expression of an IgDCP with an unknown gene name (shortened as IgDCPa in the normalised data). Many proteins in the IgSF are known to affect the brain’s synapse wiring as well as play a role in the immune system [42, 43]. Due to their role in regulating synapse formation, the proteins in the IgSF have been connected to several different neuropsychiatric and neurodevelopmental disorders [42]. Since epilepsy involves abnormal excessive or synchronous neuronal activity [44], further research on the potential role of IgSF proteins in seizure development is warranted.

Desmosomes are intercellular adhesive junctions that strengthen solid tissues in vertebrae by linking the intermediate filament networks of adjacent cells. They can be divided into two subgroups: desmocollins (DSCs) – such as DSC1b in our data – and desmogleins (DSGs). Loss of either has been shown to lead to worsened desmosomal adhesion in genetic experiments, and dysfunction of desmosomes can lead to defects in heart muscle and skin [45]. Though little is known about the connection of desmosomes and the brain, DSG1 and DSG2 as well as DSC1 have been identified in corpus callosum of the mouse brain. Furthermore, subtype DSG1c was upregulated in oligodendrocytes after chronic stress exposure in the same study [46]. While the direct association between desmosomes and epilepsy is not well-established, alterations in cell adhesion and synaptic connectivity could contribute to the pathophysiology of epilepsy.

In addition to the proteins included in the combined ROC analyses with the highest AUCs, a significant difference was detected between drug-responsive and drug-resistant dogs regarding the protein LOC607874 in the normalised data. It is a cystatin domain-containing protein which belongs to the cystatin family [30]. Cystatin inhibitory activity is important for regulating normal physiological processes [47]. Cystatin C, for example, has been linked to epilepsy in previous studies. Cystatin C is a neuroprotective protein with immunomodulatory effects on CNS, and it has been shown to be altered during the chronic phase of epilepsy, suggesting that it is involved in network reorganisation in the epileptic dentate gyrus [48, 49]. Another example is the cystatin B gene: variants of the gene have been shown to be the primary reason for developing Unverricht–Lundborg-type epilepsy [50, 51]. The existing data on cystatins thus suggests that they could also contribute to the development of DRE, possibly due to their role in neuroprotection and homeostatic functions.

Raw intensity data from protein identification and quantification results was also further analysed since normalisation based on ‘MaxLFQ-algorithm’ often turns low-intensity values to zero, resulting in loss of data from low-abundant proteins. In our data, after filtering the identification results to focus on the reproducibly quantified proteins, there were 156 proteins left in the raw intensity data in comparison to 79 in the normalised data. This analysis revealed six additional proteins of interest: NEGR1, LOC476816, IgDCPb, OGN, CACNA2D1, and LSAMP. While these proteins were not as prevalent in the normalised data, they remain interesting due to their statistically significant p-values and substantial log₁₀ differences in the raw data. These proteins and their physiological mechanisms have in earlier studies been connected to functions and disorders of the brain [30, 42, 43, 52–61], making them noteworthy targets also for epilepsy research. For instance, abnormalities of CACNA2D1 have been associated with epilepsy and neuropsychiatric disorders [60, 61], while folate deficiency has been linked to poor seizure control in epileptic children, highlighting the relevance of LOC4768, a folate receptor-like domain-containing protein, in epilepsy research [30, 54].

DRE is determined as failure to achieve treatment success with ≥ 2 ASMs with adequate serum concentrations [23, 62]. However, to avoid heterogeneity between groups, our grouping was based entirely on the dog’s response to PB. It is shown to be an effective ASM in human patients with epilepsy but due to side effects, PB is relatively seldom used in developed countries in human patients [63]. On the other hand, PB is the most commonly used ASM in dogs with epilepsy due to its efficacy, low cost, and safety in this species [64]. Studies in human epilepsy patients have found that the response to the first ASM is the strongest predictor of long-term outcome [65]. Furthermore, it has been reported that patients who did not respond to first ASM had a greater likelihood of uncontrolled epilepsy for each additional ASM tried. If the first ASM treatment failed, the second ASM regimen only provided an additional 11.6% probability of treatment success. If the first two drugs failed to control seizures, the third ASM regimen offered only a 4.1% additional probability of treatment success [4]. Consequently, in the current study, canine patients who did not respond to PB with therapeutic serum levels were classified as drug-resistant.

Different starting materials are being explored in the field of protein quantification techniques in biomarker studies. Potential proteomic biomarkers for DRE have been identified notably in tissue samples, blood, and even CSF [15–17]. Despite this, the results remain highly variable, highlighting the need for further research. According to our knowledge, no previous studies have explored CSF-derived EV proteomics as a source of prognostic biomarkers for epilepsy. By studying EVs, unique proteins may be identified compared to whole CSF [24]. Our prospective longitudinal study, with sampling at epilepsy onset before ASM treatment, provides a reliable way to study biomarkers for DRE and our results provide valuable additional insights to this field.

Here, we detected differences in several proteins with reported links to brain-specific functions. These proteins have been implicated in critical neurological processes, including BBB integrity, inflammatory signalling, neuroprotection, and synaptic plasticity, which are relevant to the pathophysiology of epilepsy [66–69]. There are several theories about the mechanisms resulting in DRE, and the cause for this condition is most likely multifactorial: this includes genetic, external, as well as drug- and disease-related aspects [68, 70, 71]. Molecular, cellular, and network variations have also been reported to affect severeness and drug-responsiveness of epilepsy [68]. Given the multifactorial nature of DRE, it is unlikely that a single protein can fully explain the variability in drug response, and ROC analysis revealed that with our data, a combination of three proteins was the most promising prognostic model for drug-responsiveness. To further validate the findings of this study, additional research regarding both clinical factors and proteomic analysis is needed with a separate population of individuals.

Conclusions

The proteomic analysis of normalised data identified five proteins with differential abundance when comparing CSF-derived EVs of drug-responsive and drug-resistant dogs with IE: KRT4, IgDCPa, F2, DSC1b, and LOC607874. These proteins are related to epilepsy through their potential role in processes such as neuroprotection, inflammation, cell structure, and immune response. Their reduced abundance in dogs with DRE suggests that these processes may be disrupted, contributing to the severity and treatment resistance of the condition. While validation of the results of this pilot study is needed by further research in a larger cohort, our work provides additional insights and a strong foundation for further research on biomarkers for DRE.

Abbreviations

ACN

Acetonitrile

ASM

Anti-seizure medication

AUC

Area under the curve

BC

Border collie

CI

Confidence interval

CL

Confidence limit

CNS

Central nervous system

CSF

Cerebrospinal fluid

DRE

Drug-resistant epilepsy

DSC

Desmocollin

DSG

Desmoglein

EV

Extracellular vesicle

IE

Idiopathic epilepsy

IgDCP

Immunoglobulin-like domain-containing protein

IgSF

Immunoglobulin superfamily

LC-MS/MS

Liquid chromatography-tandem mass spectrometry

LFQ

Label-free quantification

MRI

Magnetic resonance imaging

NR

Drug-resistant group

NTA

Nanoparticle tracking analysis

OB

Other breed

OR

Odds ratio

PB

Phenobarbital

PC

Principal component

PCA

Principal component analysis

R

Drug-responsive group

ROC

Receiver operating characteristic

SEC

Size-exclusion chromatography

TEM

Transmission electron microscopy

UF-SEC

Ultrafiltration combined with size-exclusion chromatography

Author contributions

Conceptualisation: P.K., J.K., T.S.J., T.A.N., L.M., and J.J. Methodology: P.K., J.K., T.S.J, T.A.N., and J.J. Writing of the original manuscript: P.K. Data analysis: P.K., T.A.N, J.J. Writing (review and editing): P.K., J.K., T.S.J., T.A.N., L.M., and J.J. Visualisation: P.K. and J.J. Project administration and funding acquisition: T.S.J. All authors have read and agreed to the published version of the manuscript.

Funding

Open Access funding provided by University of Helsinki (including Helsinki University Central Hospital). This study was supported by the Academy of Finland (grant) and Finnish Foundation of Veterinary Research (scholarship).

Data availability

The datasets generated and/or analysed during the current study are available in the ProteomeXchange Consortium via the PRIDE (26) partner repository with the dataset identifier PXD061528.

Declarations

Ethics approval and consent to participate

The study was approved by the Viikki Campus Research Ethics Committee at the University of Helsinki, Finland. All owners of the dogs participating in the study provided a written consent and were made aware that they could exit the study at any time without consequences.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.