Распространение эпсилонного токсина Clostridium Perfringens изолируется среди пациентов с множественным склерозом и расстройством спектра нейромиелита в Иране.

Background

Clostridium perfringens is a ubiquitous Gram-positive spore-forming anaerobic microorganism which is commonly found within the gastrointestinal (GI) tract of humans and animals [ 1 ]. This bacterium can cause various diseases including food poisoning, non-foodborne diarrhea, enteritis necroticans, gas gangrene, enterotoxemia, and sudden death syndrome [ 2 ]. According to a new toxin-based typing system, C. perfringens strains are classified into seven toxinotypes (A-G) according to the production of six major toxins, alpha (α), beta (β), epsilon (ε), iota (Ɩ), CPE, and Net-B [ 3 ].

Epsilon toxin (ETX) produced by C. perfringens types B and D is the third most toxic among the clostridial toxins as determined by a 50% lethal dose ranking behind the Clostridium botulinum and Clostridium tetani neurotoxins [ 4 , 5 ]. ETX belongs to β-pore-forming toxins (β-PFTs), which are characterized by the formation of pores containing a β-barrel of 14 amphipathic β-strands into the plasma membrane of eukaryotic cells [ 6 , 7 ].

ETX secretes in the gut lumen as a relatively inactive, 32.9 kDa, epsilon protoxin (proETX), binds to the intestinal epithelium and subsequently is activated by extracellular serine-type proteases (trypsin/chymotrypsin) and carboxypeptidases or occasionally by bacterial lambda proteas [ 6 , 8 – 10 ]. Additionally, ETX absorbed via the intestine permeabilizes the blood-brain barrier (BBB) and causes a range of symptoms from BBB disruption to cerebral edema such as visual dysfunction, incoordination, and spastic paralysis [ 11 , 12 ]. ETX is an environmental trigger for demyelinating diseases that potentially initiates and worsens demyelination by disrupting immune responses and damaging myelin-producing cells [ 13 , 14 ]. This toxin binds to the myelin and lymphocyte protein (MAL) on BBB endothelial cells, which results in the formation of pores that impair barrier integrity and allow autoreactive immune cells to enter the central nervous system (CNS) [ 14 , 15 ]. Within the CNS, ETX targets mature oligodendrocytes via MAL and causes their selective death and demyelination, while sparing other neural cell types [ 16 ].

Multiple sclerosis (MS) is a chronic autoimmune disease of the CNS characterized by inflammation, demyelination, and neurodegeneration [ 17 ]. Relapsing-remitting MS (RRMS) is the most common clinical form of the disease, which affects approximately 80–85% of patients and is marked by episodes of neurological dysfunction followed by periods of remission [ 18 , 19 ]. Neuromyelitis optica spectrum disorder (NMOSD) is a distinct, an astrocytopathy that mostly targets the optic nerves and spinal cord and can lead to recurrent optic neuritis and transverse myelitis [ 20 , 21 ]. There have been few reports of detection of ETX-producing C. perfringens in some patients with MS and NMOSD [ 21 – 23 ]. ETX has been proposed as a provocative candidate for nascent lesion formation in MS based on its tropism for BBB vasculature and binding to oligodendrocytes [ 22 , 24 , 25 ]. ETX is absorbed via the intestine [ 8 ], binds to CNS endothelial cells of BBB and blood-retinal barriers through the bloodstream, and permeabilizes the BBB, leading to toxin entry into the brain parenchyma [ 22 ]. However, the role of ETX in the formation of MS lesions is not elucidated. This study aimed to investigate the prevalence of C. perfringens in stool samples of patients with MS, NMOSD, and healthy controls using culture and PCR. Additionally, anti-epsilon toxin immunoglobulin G (IgG) and immunoglobulin M (IgM) levels were measured in all subjects by using competitive enzyme-linked immunosorbent assay (ELISA). Further, multivariable prediction models, trained on all variables, were applied to assess the predictive value of sex, age, ETX IgG and IgM levels, and C. perfringens . Logistic regression (LR) was used to evaluate the independent contribution of each variable.

Methods

Patients and sample collection

During September 2019 to June 2020, 44 sera and stool samples were collected from patients with MS ( n  = 33) and NMOSD ( n  = 10), who were referred to Multiple Sclerosis Research Center of Tehran University of Medical Sciences. Forty-nine sera and stool samples were also collected from healthy subjects. Healthy subjects, aged 18–45 years, were recruited from employees referred to Safety Promotion and Injury Prevention Research Center, Shahid Beheshti University of Medical Sciences, Tehran, Iran. Healthy subjects were selected based on criteria as follow: they showed no notable medical or psychiatric disorders and were free from chronic illnesses necessitating treatment. Furthermore, they were not on any medications that could potentially interfere with the results of the study. Participants had no history of substance abuse and were classified as either non-smokers. To ensure their overall health status, a thorough physical examination was conducted, confirming their eligibility for the study.

Patients aged ≥ 18 years were evaluated by an experienced neurologist based on the definite diagnosis of MS and NMOSD, which are introduced in McDonalds 2017 criteria [ 26 ], and International Consensus Diagnostic Criteria for Neuromyelitis Optica Spectrum Disorders [ 27 ], respectively. The exclusion criteria were as follows: age < 18 years, patients with underlying disease, drug consumption that affects the GI system, occurrence of MS attack within the preceding month, and corticosteroids administration. Written informed consent was obtained from all subjects before sample collection. Clinical and demographic features were recorded for all patients on the day of admission through a questionnaire, which was developed for this study (Supplementary Information).

Bacterial culture

Stool samples

For isolation of C. perfringens , stool samples were cultured as previously described [ 28 ]. Briefly, fecal samples (0.5–1 g) were homogenized in 9 ml sterile phosphate-buffered saline (PBS), heat-treated (70 °C, 20 min), and 100 µl inoculated onto egg yolk agar (Himedia, India) supplemented with 10% sheep blood and neomycin (75 µg/ml, Sigma, USA). Plates were incubated anaerobically (85% N 2 , 10% CO 2 , 5% H 2 ) at 37 °C for 48-72 h. Suspected colonies were identified by morphology, Gram staining, motility, biochemical tests and confirmed by 16 S rRNA PCR [ 28 , 29 ].

Detection of C. perfringens epsilon toxin

Plasmid DNA was purified from the pure colonies after 48 h by using GeneJET Plasmid Miniprep Kit (Thermo Fisher Scientific, Lithuania) per the manufacturer’s instructions. The quality and purity of DNA samples were measured using NanoDrop (Thermo Scientific, Asheville, NC, USA). Primers and PCR amplification conditions are listed in Supplementary Table 1 [ 24 ].

Competitive enzyme-linked immunosorbent assay

Epsilon toxin was prepared and purified as previously described [ 30 ]. The 96-well plastic plates (Nunc, thermo Fisher Scientific, USA) were coated with purified epsilon toxin (4 mg/ml) in PBS containing 2.5% sodium deoxycholate (Sigma-Aldrich, USA) and incubated overnight at 37 °C. Plates were washed 3 times with PBST (PBS-0.1%, Tween 20). After blocking with 3% BSA in PBS during 30 min under agitation, three washes were performed, 100 ml/well of sample serums diluted in PBST containing 1% BSA were added and plates were incubated for 1 h at 37 °C with shaking. After three washes, plates were incubated with goat anti-human IgG and IgM alkaline phosphatase (Sigma-Aldrich, Cat NO. A9544 or Sigma-Aldrich, Cat NO. A3437) in PBST-1% BSA for 1 h at 37 °C. Plates were washed three times and incubated with p-nitrophenyl phosphate (1 mg/ml) in 10% diethanolamine buffer (pH 9.8) containing 0.01% MgCl 2 and 0.2% sodium azide for 30 min at 37 °C. Concentration of specific antibodies was calculated according to the curve obtained from standard immunoglobulin solution absorbance by Ascent software, Thermo Electron corp a previously reported [ 31 ]. The ELISA data are expressed in mg/ml.

Machine learning algorithm

To address concerns regarding class balance, we categorized patients (including both MS and NMOSD) into a single ‘Patients’ group. Study participants were randomly divided into 70% of samples (64 participants after excluding the missing data) to be used as training data for predicting disease outcomes and 30% as validation data to test algorithms. Depending on which sample was used, a patient could appear only in the training sample or the validation sample. A logistic regression model (LR), a naive bayes model (NB), a support vector machine model (SVM) and a neural network model (NN) were developed using the independent variables of the training data [ 32 ]. There are parameters associated with each of the methods mentioned above, and it is critical to select the appropriate parameter in order to obtain the most complete model while also maintaining the simplest model. The parameters of each algorithm were tuned to accurately predict the disease. Available statistical codes and packages have been used to set the parameters.

Performance evaluation

As part of the validation data, k-fold cross validation (k = 10) was used to evaluate the four models, followed by an analysis of receiver operating characteristics (ROC) based on their sensitivity, specificity, positive and negative predictive values, and accuracy (ACC). The area under the curve (AUC) measurement was used to compare the performance of the four models. Thus, for the purpose of clinical decision-making, the superior prediction model has been chosen. The ROC curve illustrates the trade-off between sensitivity and specificity across different classification thresholds, providing a measure of diagnostic performance. The area under the ROC curve quantifies the model’s ability to distinguish between cases and controls, with values closer to 1 indicating better performance [ 33 ]. Further details are provided in Supplementary Information. Sensitivity and specificity evaluate a model’s ability to correctly classify cases and controls, while PPV (positive predictive value) and NPV (negative predictive value) assess the reliability of positive and negative predictions [ 34 – 36 ].

Nomogram analysis

The probability of disease occurring was calculated by using a nomogram as the visual and predictive tool based upon the presence of patient risk factors, confounders, and predictor variables. A LR model was used to predict the outcome in the nomogram [ 37 ].

Statistical analysis

The descriptive statistics were presented by stratifying the patients according to disease occurrence and comparing their characteristics. Frequency (%) was used to present the categorical data. Mean ± standard deviation (SD) or median (first quartile [Q1], third quartile [Q3]) were used for summarizing the numeric variables. Fisher-Freeman-Halton exact test was used to evaluate the association between categorical variables and group. Analysis of variance (ANOVA) was used to compare mean of age between groups. Kruskal-Wallis test was used to compare the distribution of IgG and IgM between groups. Analyses were conducted using SPSS (version 26) and R (version 4.2.1). A P value of 0.05 and a 95% confidence interval (CI) were used to determine statistical significance.

Results

Baseline demographics and clinical characteristics

The demographic and clinical data of patients and healthy subjects are summarized in Table  1 . Forty-three patients with demyelinating CNS disorders consisting of 33 (77.7%) MS with mean age of 32.4 ± 8.0 years and 10 (23.3%) NMOSD (mean age of 36.0 ± 13.30) were included in the study. Treatment regimens of participants in this study are shown in Supplementary Table 2 . All MS patients are diagnosed with RRMS. Healthy subjects included 33 females (67.3%) and 16 males (32.7%) with age range of 27 to 44 years (mean 35.02 ± 4.09 years). The clinical manifestations of MS patients included paresthesias 15.2% (5/33), optic neuritis 54.5% (18/33), ataxia 6.1% (2/33), paresis 42.4% (14/33), hearing loss or tinnitus 2.9% (1/33), and bladder dysfunction 5.9% (2/33). The paresthesias 40% (4/10), optic neuritis 30% (6/10), and paresis 30% (3/10) were the most common manifestations in patients with NMOSD [ 38 , 39 ].

Table 1 Demographic and clinical characteristics of participants in this study Characteristics n (%)/mean ± SD or median (Q1, Q3) P value Healthy ( n  = 49) MS ( n  = 33) NMOSD ( n  = 10)
Age
35.02 ± 4.09 32.4 ± 8.0 36.0 ± 13.30 0.180
Sex
Female Male 33 (67.3) 16 (32.7) 30 (90.9) 3 (9.1) 9 (90) 1 (10) 0.026
Clinical manifestations n (%)
Paresthesias - 5 (15.2) 4 (40) 0.177 Optic neuritis - 18 (54.5) 6 (30) 1.000 Ataxia - 2 (6.1) 0 1.000 Paresis 14 (42.2) 3 (30) 0.714 Hearing loss or tinnitus - 1 (3) 0 1.000 Bladder dysfunction - 2 (6.1) 0 1.000 Disease duration < 1–2 year 3–5 years 6–12 years 13–17 years 18–25 years - - - - - 16 (48.5) 2 (6.1) 13 (39.4) 1 (3) 1 (3) 4 (40) 1 (10) 4 (40) 1 (10) 0 0.29
Clostridium perfringens
No Yes 44 (89.8) 5 (10.2) 29 (87.9) 4 (12.1) 8 (80) 2 (20) 0.657
Anti-ETX IgG µg/ml
0.87 (0.60, 1.43) 0.74 (0.47, 1.37) 0.52 (0.33, 0.82) 0.082
Anti-ETX IgM µg/ml
1.95 (1.28, 2.71) 2.37 (1.24, 4.34) 1.38 (0.38, 2.66) 0.146 Abbreviations: standard deviation = SD; first quartile = Q1; third quartile = Q3; Epsilon toxin = ETX; multiple sclerosis = MS; neuromyelitis optica spectrum disorder = NMOSD

Table 1 Demographic and clinical characteristics of participants in this study Characteristics n (%)/mean ± SD or median (Q1, Q3) P value Healthy ( n  = 49) MS ( n  = 33) NMOSD ( n  = 10)
Age
35.02 ± 4.09 32.4 ± 8.0 36.0 ± 13.30 0.180
Sex
Female Male 33 (67.3) 16 (32.7) 30 (90.9) 3 (9.1) 9 (90) 1 (10) 0.026
Clinical manifestations n (%)
Paresthesias - 5 (15.2) 4 (40) 0.177 Optic neuritis - 18 (54.5) 6 (30) 1.000 Ataxia - 2 (6.1) 0 1.000 Paresis 14 (42.2) 3 (30) 0.714 Hearing loss or tinnitus - 1 (3) 0 1.000 Bladder dysfunction - 2 (6.1) 0 1.000 Disease duration < 1–2 year 3–5 years 6–12 years 13–17 years 18–25 years - - - - - 16 (48.5) 2 (6.1) 13 (39.4) 1 (3) 1 (3) 4 (40) 1 (10) 4 (40) 1 (10) 0 0.29
Clostridium perfringens
No Yes 44 (89.8) 5 (10.2) 29 (87.9) 4 (12.1) 8 (80) 2 (20) 0.657
Anti-ETX IgG µg/ml
0.87 (0.60, 1.43) 0.74 (0.47, 1.37) 0.52 (0.33, 0.82) 0.082
Anti-ETX IgM µg/ml
1.95 (1.28, 2.71) 2.37 (1.24, 4.34) 1.38 (0.38, 2.66) 0.146 Abbreviations: standard deviation = SD; first quartile = Q1; third quartile = Q3; Epsilon toxin = ETX; multiple sclerosis = MS; neuromyelitis optica spectrum disorder = NMOSD

Demographic and clinical characteristics of participants in this study

Sex

Female

Male

33 (67.3)

16 (32.7)

30 (90.9)

3 (9.1)

9 (90)

1 (10)

Disease duration

< 1–2 year

3–5 years

6–12 years

13–17 years

18–25 years

-

-

-

-

-

16 (48.5)

2 (6.1)

13 (39.4)

1 (3)

1 (3)

4 (40)

1 (10)

4 (40)

1 (10)

0

Clostridium perfringens

No

Yes

44 (89.8)

5 (10.2)

29 (87.9)

4 (12.1)

8 (80)

2 (20)

Abbreviations: standard deviation = SD; first quartile = Q1; third quartile = Q3; Epsilon toxin = ETX; multiple sclerosis = MS; neuromyelitis optica spectrum disorder = NMOSD

Prevalence of C. perfringens isolates and etx gene

In total, C. perfringens was isolated from 6/43 (14%) stool samples of patients with demyelinating CNS disorders, including 4/33 (12.1%) from MS and 2/10 (20%) from NMOSD patients. In contrast, C. perfringens was detected in 5/49 (10.2%) of healthy subjects (Table  1 ). Of 11 C. perfringens obtained, 1/4 (25%) from MS and 2/2 (100%) from NMOSD patients were positive for etx -encoding gene, respectively. None of the C. perfringens isolates from healthy subjects were positive for etx gene. No statistically significant difference was found in prevalence of C. perfringens and etx gene positivity between NMOSD, or MS patients and healthy subjects.

Antibodies to epsilon toxin identified by ELISA

To investigate the prevalence of antibodies to ETX, serum samples from MS, NMOSD, and healthy subjects were examined for IgG and IgM level against ETX. The median (interquartile range [IQR]) of IgG was 0.87 (0.60, 1.43) for healthy subjects, 0.74 (0.47, 1.37) for MS, and 0.52 (0.33, 0.82) for NMOSD patients (Table  1 ). In addition, the median IgM was 1.95 (1.28, 2.71), 2.37 (1.24, 4.34), and 1.38 (0.38, 2.66) for healthy subjects, MS and NMOSD patients, respectively. No statistically significant difference was observed in medians of anti-ETX IgG level ( P  = 0.082), and anti-ETX IgM level ( P  = 0.146) between NMOSD, or MS patients and healthy subjects (Fig.  1 ).

Fig. 1 Comparing IgG, IgM and age between healthy individuals, MS and NMOSD patients. Analysis of variance (ANOVA) was used to compare mean age between groups. Mean ± standard deviation (SD) or median (first quartile [Q1], third quartile [Q3]) were used for summarizing the numeric variables. Kruskal-Wallis test was used to compare the distribution of IgG and IgM between groups. Analyses were conducted using SPSS (version 26) and R (version 4.2.1). A P value of 0.05 and 95% confidence interval (CI) were used to determine statistical significance

Comparing IgG, IgM and age between healthy individuals, MS and NMOSD patients. Analysis of variance (ANOVA) was used to compare mean age between groups. Mean ± standard deviation (SD) or median (first quartile [Q1], third quartile [Q3]) were used for summarizing the numeric variables. Kruskal-Wallis test was used to compare the distribution of IgG and IgM between groups. Analyses were conducted using SPSS (version 26) and R (version 4.2.1). A P value of 0.05 and 95% confidence interval (CI) were used to determine statistical significance

Association between IgG and IgM levels with demographic and clinical features

In order to understand whether IgG and IgM levels against ETX were related to disease severity, we further assessed the association between ETX IgG or IgM concentration with symptoms, age, disease duration, duration of hospitalization, and attack number reported in patients. However, we found no association between IgG or IgM levels against ETX with disease severity ( P  = 0.285 and P  = 0.798, respectively) and other disease parameters, suggesting that immunoglobulin intrathecal levels were not related to disease progresses (Supplementary Information).

Model performance

Five clinical and demographic features including IgG, IgM, age, C. perfringens , and sex were used to construct the model. At the first step, the prediction power of all variables for predicting patient status using logistic regression was calculated (Table  2 ). Combining the accuracy, sensitivity, specificity and the AUC of the five models presented that the highest AUC was related to sex (AUC = 0.63, 95% CI: 0.52, 0.75) and IgG (AUC = 0.57, 95% CI: 0.35, 0.80). This finding suggests 63% and 57% chance that patients and individuals were correctly classified according to their gender and IgG values, respectively. Next, we performed four models including LR, SVM, NB and NN to predict the patient status with including all five predictor variables. The highest AUC shows the most valuable model. According to Table  3 , NB showed the highest AUC (0.93, 95% CI: 0.70, 1.00), SE (0.85, 95% CI: 0.55, 0.98), SP (0.93, 95% CI: 0.68, 1.00), PPV (0.92, 95% CI: 0.63, 0.99), NPV (0.88, 95% CI: 0.60, 1.00), and ACC (0.86, 95% CI: 0.67, 0.96) for predicting patients. Accordingly, there is a 93% chance that patients and individuals were correctly classified using NB algorithm with all predictor variables. In addition, 93% of healthy subjects were correctly classified as healthy using NB model. Moreover, 92% of patients who were predicted to have the disease actually had the disease. Furthermore, the accuracy of LR, SVM, and NB was significantly higher than INR value ( P  < 0.05). It means that 71%, 82%, and 86% of patients with the disease will be correctly predicted as patients using the LR, SVM, and NB models. As shown in Tables  4 and 28 cases were included to the validation dataset. Based on NB model as the most valuable algorithm, 14 out of 17 patients were correctly predicted, while 10 out of 11 healthy subjects were correctly classified. The ROC curve for all methods of machine learning is shown in Fig.  2 . Models with the higher AUC may be considered better to distinguish patients and healthy subjects.

Table 2 The prediction power of variables for disease predicting Variables AUC (95% CI) SE (95% CI) SP (95% CI) PPV (95% CI) NPV (95% CI) Accuracy (95% CI) Anti-ETX IgG 0.57 (0.35, 0.80) 0.46 (0.19, 0.75) 0.80 (0.52, 0.96) 0.67 (0.35, 0.87) 0.63 (0.32, 0.90) 0.64 (0.44, 0.81) Anti-ETX IgM 0.51 (0.27, 0.76) 0.38 (0.14, 0.68) 0.93 (0.68, 1.00) 0.83 (0.43, 0.95) 0.64 (0.31, 0.99) 0.64 (0.44, 0.81) Age 0.54 (0.28, 0.80) 0.46 (0.19, 0.75) 1.00 (0.78, 1.00) 1.00 (0.59, 1.00) 0.68 (0.37, 1.00) 0.75 (0.55, 0.89)
C. perfringens
0.51 (0.37, 0.65) 0.15 (0.02, 0.45) 0.87 (0.60, 0.98) 0.50 (0.18, 0.82) 0.54 (0.11, 0.92) 0.54 (0.34, 0.72) Sex 0.63 (0.52, 0.75) 1.00 (0.75, 1.00) 0.27 (0.08, 0.55) 0.54 (0.22, 1.00) 1.00 (0.48, 1.00) 0.61 (0.41, 0.78) Abbreviations: area under the curve = AUC; sensitivity = SE; specificity = SP; positive predictive value = PPV; negative predictive value = NPV; confidence interval = CI. MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

Table 2 The prediction power of variables for disease predicting Variables AUC (95% CI) SE (95% CI) SP (95% CI) PPV (95% CI) NPV (95% CI) Accuracy (95% CI) Anti-ETX IgG 0.57 (0.35, 0.80) 0.46 (0.19, 0.75) 0.80 (0.52, 0.96) 0.67 (0.35, 0.87) 0.63 (0.32, 0.90) 0.64 (0.44, 0.81) Anti-ETX IgM 0.51 (0.27, 0.76) 0.38 (0.14, 0.68) 0.93 (0.68, 1.00) 0.83 (0.43, 0.95) 0.64 (0.31, 0.99) 0.64 (0.44, 0.81) Age 0.54 (0.28, 0.80) 0.46 (0.19, 0.75) 1.00 (0.78, 1.00) 1.00 (0.59, 1.00) 0.68 (0.37, 1.00) 0.75 (0.55, 0.89)
C. perfringens
0.51 (0.37, 0.65) 0.15 (0.02, 0.45) 0.87 (0.60, 0.98) 0.50 (0.18, 0.82) 0.54 (0.11, 0.92) 0.54 (0.34, 0.72) Sex 0.63 (0.52, 0.75) 1.00 (0.75, 1.00) 0.27 (0.08, 0.55) 0.54 (0.22, 1.00) 1.00 (0.48, 1.00) 0.61 (0.41, 0.78) Abbreviations: area under the curve = AUC; sensitivity = SE; specificity = SP; positive predictive value = PPV; negative predictive value = NPV; confidence interval = CI. MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

The prediction power of variables for disease predicting

Abbreviations: area under the curve = AUC; sensitivity = SE; specificity = SP; positive predictive value = PPV; negative predictive value = NPV; confidence interval = CI. MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

Table 3 The prediction power of all variables for predicting the status (healthy or patient) using different models Variables AUC (95% CI) SE (95% CI) SP (95% CI) PPV (95% CI) NPV (95% CI) Accuracy (95% CI) P value LR 0.71 (0.50, 0.91) 0.54 (0.25, 0.81) 0.93 (0.68, 1.00) 0.88 (0.52, 0.96) 0.70 (0.40, 0.99) 0.71 (0.51, 0.87) 0.042 SVM 0.89 (0.77, 1.00) 0.69 (0.39, 0.91) 0.93 (0.68, 1.00) 0.90 (0.58, 0.98) 0.78 (0.49, 0.99) 0.82 (0.63, 0.94) 0.002 NB 0.93 (0.83, 1.00) 0.85 (0.55, 0.98) 0.93 (0.68, 1.00) 0.92 (0.63, 0.99) 0.88 (0.60, 1.00) 0.86 (0.67, 0.96) < 0.001 NN 0.57 (0.34, 0.80) 0.46 (0.19, 0.75) 0.80 (0.52, 0.96) 0.67 (0.35, 0.87) 0.63 (0.32, 0.90) 0.64 (0.44, 0.81) 0.172 Abbreviations: area under the curve = AUC; sensitivity = SE; specificity = SP; positive predictive value = PPV; negative predictive value = NPV; accuracy = ACC; Logistic Regression = LR; support vector machine = SVM; naïve bayes = NB; neural network = NN. * P-value was calculated to compare the accuracy of the no information rate (NIR) and each machine learning model, CI, confidence interval. MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

Table 3 The prediction power of all variables for predicting the status (healthy or patient) using different models Variables AUC (95% CI) SE (95% CI) SP (95% CI) PPV (95% CI) NPV (95% CI) Accuracy (95% CI) P value LR 0.71 (0.50, 0.91) 0.54 (0.25, 0.81) 0.93 (0.68, 1.00) 0.88 (0.52, 0.96) 0.70 (0.40, 0.99) 0.71 (0.51, 0.87) 0.042 SVM 0.89 (0.77, 1.00) 0.69 (0.39, 0.91) 0.93 (0.68, 1.00) 0.90 (0.58, 0.98) 0.78 (0.49, 0.99) 0.82 (0.63, 0.94) 0.002 NB 0.93 (0.83, 1.00) 0.85 (0.55, 0.98) 0.93 (0.68, 1.00) 0.92 (0.63, 0.99) 0.88 (0.60, 1.00) 0.86 (0.67, 0.96) < 0.001 NN 0.57 (0.34, 0.80) 0.46 (0.19, 0.75) 0.80 (0.52, 0.96) 0.67 (0.35, 0.87) 0.63 (0.32, 0.90) 0.64 (0.44, 0.81) 0.172 Abbreviations: area under the curve = AUC; sensitivity = SE; specificity = SP; positive predictive value = PPV; negative predictive value = NPV; accuracy = ACC; Logistic Regression = LR; support vector machine = SVM; naïve bayes = NB; neural network = NN. * P-value was calculated to compare the accuracy of the no information rate (NIR) and each machine learning model, CI, confidence interval. MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

The prediction power of all variables for predicting the status (healthy or patient) using different models

Abbreviations: area under the curve = AUC; sensitivity = SE; specificity = SP; positive predictive value = PPV; negative predictive value = NPV; accuracy = ACC; Logistic Regression = LR; support vector machine = SVM; naïve bayes = NB; neural network = NN. * P-value was calculated to compare the accuracy of the no information rate (NIR) and each machine learning model, CI, confidence interval. MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

Table 4 Confusion matrix of machine learning models in the validation dataset Models Prediction All variables Reference Healthy Patients LR Healthy 14 7 Patients 1 6 SVM Healthy 15 5 Patients 0 8 NB Healthy 14 3 Patients 1 10 NN Healthy 12 7 Patients 3 6 Abbreviations: logistic regression = LR; support vector machine = SVM; naïve bayes = NB; neural network = NN. MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

Table 4 Confusion matrix of machine learning models in the validation dataset Models Prediction All variables Reference Healthy Patients LR Healthy 14 7 Patients 1 6 SVM Healthy 15 5 Patients 0 8 NB Healthy 14 3 Patients 1 10 NN Healthy 12 7 Patients 3 6 Abbreviations: logistic regression = LR; support vector machine = SVM; naïve bayes = NB; neural network = NN. MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

Confusion matrix of machine learning models in the validation dataset

Abbreviations: logistic regression = LR; support vector machine = SVM; naïve bayes = NB; neural network = NN. MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

Fig. 2 Receiver operating characteristic (ROC) curves were reconstructed for predicting of MS and NMOSD in four models including multiple logistic regression (LR), support vector machine (SVM), naïve bayes (NB), and neural network (NN). In LR Model, the estimated area under the curve (AUC) for MS and NMOSD was calculated at 0.71 (95% CI; 0.50, 0.91) by univariate logistic regression. The AUC of SVM Model was increased to 0.89 (95% CI; 0.77, 1.0). In NB Model, the AUC was increased to 0.93 (95% CI; 0.83, 1.0). Finally, the AUC of NN Model was estimated about 0.57 (95% CI; 0.34, 0.80). MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

Receiver operating characteristic (ROC) curves were reconstructed for predicting of MS and NMOSD in four models including multiple logistic regression (LR), support vector machine (SVM), naïve bayes (NB), and neural network (NN). In LR Model, the estimated area under the curve (AUC) for MS and NMOSD was calculated at 0.71 (95% CI; 0.50, 0.91) by univariate logistic regression. The AUC of SVM Model was increased to 0.89 (95% CI; 0.77, 1.0). In NB Model, the AUC was increased to 0.93 (95% CI; 0.83, 1.0). Finally, the AUC of NN Model was estimated about 0.57 (95% CI; 0.34, 0.80). MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

Nomogram analysis

To further improve the prediction model, the nomograms based on significant and independent predictors determined by LR analysis were established to predict patients’ risk of developing MS and NMOSD. The nomogram model included five independent predictors such as age, sex, IgG or IgM levels against ETX and the positive test for C. perfringens. A subset of these variables is assigned a count on the point set of the scale, and the respective scores of points of different variables are calculated and summed to produce a total score. As shown in Fig.  3 , by adding scores of the corresponding predictors in the respective nomograms, the probability of patients’ risk of developing MS and NMOSD can be predicted.

Fig. 3 The nomogram of prediction model in NMOSD and MS patients using logistic regression. Each variable is on its axis. Draw a vertical line to the points axis to determine the number of points for each variable. Then, add the points for all the variables together and positioned them on the total points axis. Draw a vertical line down to determine the possibility of NMOSD and MS at the bottom. MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

The nomogram of prediction model in NMOSD and MS patients using logistic regression. Each variable is on its axis. Draw a vertical line to the points axis to determine the number of points for each variable. Then, add the points for all the variables together and positioned them on the total points axis. Draw a vertical line down to determine the possibility of NMOSD and MS at the bottom. MS and NMOSD cases were combined into a single “patient” group and compared to healthy controls within a binary classification framework for the development and evaluation of machine learning-based predictive models

Discussion

Several factors such as genetic and environmental parameters significantly influence the risk of demyelinating CNS disorders such as MS and NMOSD [ 40 ]. Gut microbiome analyses have shown an overabundance of C. perfringens in the gut microbiota of NMOSD patients, indicating the potential role of C. perfringens in NMOSD pathogenesis [ 41 ]. Previous studies have reported that C. perfringens potentially influence the pathogenesis of NMOSD through dual mechanisms. One line of evidence suggests that this bacterium represents a proinflammatory adjuvant that enhances Th17 cell polarization and promotes autoimmune responses. Other studies identified the bacterial adenosine triphosphate binding cassette transporter permease (ABC-TP) with molecular similarity to an immunodominant epitope of aquaporin-4 (AQP4), which is the main target antigen in MS and NMOSD [ 21 , 40 , 42 – 44 ]. High homology of ABC-TP to AQP4 can result in cross-reactivity and proliferation of Th17 cells to this peptide and potentially convert a normal immune response toward C. perfringens into CNS autoimmunity [ 41 ].

In this study, the prevalence of C. perfringens and anti-ETX antibody levels were assessed in MS and NMOSD patients and compared with healthy subjects. C. perfringens was detected using culture and PCR [ 45 ], and ELISA was employed to determine the antibody response [ 46 ]. We found higher frequencies of C. perfringens in NMOSD (20%) and MS (12.1%) groups than in controls (10.2%), although these differences were not statistically significant. PCR analysis revealed that 25% of MS and all NMOSD patients harbored etx -positive isolates, while none of healthy controls carried etx -positive isolates. These findings are consistent with previous studies reporting an overabundance of C. perfringens or etx gene in MS and NMOSD patients [ 21 , 41 , 47 , 48 ].

The sensitivity and specificity of the detection assay used for the isolation of C. perfringens can vary depending on the type of method exploited—culture, PCR, and ELISA. Microbial culture, while considered the gold standard for its reliability and cost-effectiveness, is labor-intensive and time-consuming [ 45 ]. Drissi and Ward found that ELISA had sensitivity and specificity rates of 90.5% and 89.2% for β-toxin assays, and 97.4% and 94.6% for ε-toxin assays, respectively, compared to conventional methods such as mouse assays and conventional microbial culture [ 49 ]. A 2020 systematic review and meta-analysis found that nucleic acid amplification tests (NAATs) had superior diagnostic accuracy compared to both immunoassay and microbiological culture [ 50 ]. NAATs demonstrated higher sensitivity (0.53 [95% CI, 0.35–0.7]), specificity (0.97 [95% CI, 0.95–0.99]), positive likelihood ratio (PLR) (23.2 [95% CI, 3.49-153.98]), negative likelihood ratio (NLR) (0.25 [95% CI, 0-245.28]), and diagnostic odds ratio (DOR) (74.11 [95% CI, 2.11-2593.7]) than immunoassay. Microbiological culture exhibited lower values for sensitivity (0.31 [95% CI, 0.22–0.41]), specificity (0.95 [95% CI, 0.93–0.97]), PLR (11.56 [95% CI, 3.87–34.6]), NLR (0.57 [95% CI, 0.27–1.21]), and DOR (18.1 [95% CI, 4.83–67.8]).

Although ETX immunoreactivity was previously observed in a greater proportion of MS patients compared to controls [ 51 ], our study found no statistically significant differences in anti-ETX IgG or IgM levels between patient groups and controls. For example, Rumah et al. reported a 10-fold increase in immunoreactivity to ETX in MS patients compared to controls [ 22 ]. This discrepancy can be attributed to differences in assay sensitivity, patient demographics, geographic exposure, sample size, timing of sample collection, and treatment effects. For instance, immunosuppressive therapies, disease stage, and microbial variability could modulate antibody responses. The heterogeneity of MS and NMOSD, including genetic and microbiome differences, may also influence immune recognition of ETX [ 12 , 22 ].

ETX is a potent pore-forming toxin that disrupts the BBB by binding to its receptor on CNS endothelial cells and oligodendrocytes [ 4 , 5 ]. Studies on sheep, goats, and mice show that ETX produced by C. perfringens type B and D is able to induce lesions resembling MS pathology and damage CNS tissues [ 52 , 53 ]. Recently, the involvement of C. perfringens and its ETX in the development of neurological disorder s in patients with clinically isolated syndrome (CIS), MS, or NMOSD has been reported in multiple studies [ 12 , 41 , 51 ]. Several pathophysiological consequences of the exposure of animals to ETX are consistent with the role of toxin in neurological disorders [ 52 , 54 ], indicating that ETX may affect the severity of the neurological lesions and elicit the injury of nervous system [ 4 ]. Hence, C. perfringens producing ETX may cause CNS disorders by activation of auto-reactivity of Th17 cells against ABC-TP peptide and development of neurological lesions by ETX (Fig.  4 ).

Fig. 4 Schematic overview of effects of C. perfringens producing ETX to develop CNS disorders. ( A ) Gut homeostasis is a well-balanced state between intestinal microbiota and mucosal barriers, in which gut microbes and host immune cells can regulate the functions of intestinal epithelium. ( B ) Gut microbiota dysbiosis can lead to the overabundance of C. perfringens and cause symptoms similar to CNS disorder. C. perfringens produces ABC-TP peptides that show high homology with the amino acid sequence of AQP-4-specific T lymphocytes, thus, can induce immune system responses. ( C ) ETX produced by C. perfringens can act on colonic epithelial cells, disrupt tight junctions, and induce tissue damage. Additionally, ETX can cross the BBB, damage oligodendrocytes, and cause a range of symptoms from BBB disruption to cerebral edema. Therefore, ETX may cause symptoms similar to CNS disorders

Schematic overview of effects of C. perfringens producing ETX to develop CNS disorders. ( A ) Gut homeostasis is a well-balanced state between intestinal microbiota and mucosal barriers, in which gut microbes and host immune cells can regulate the functions of intestinal epithelium. ( B ) Gut microbiota dysbiosis can lead to the overabundance of C. perfringens and cause symptoms similar to CNS disorder. C. perfringens produces ABC-TP peptides that show high homology with the amino acid sequence of AQP-4-specific T lymphocytes, thus, can induce immune system responses. ( C ) ETX produced by C. perfringens can act on colonic epithelial cells, disrupt tight junctions, and induce tissue damage. Additionally, ETX can cross the BBB, damage oligodendrocytes, and cause a range of symptoms from BBB disruption to cerebral edema. Therefore, ETX may cause symptoms similar to CNS disorders

We developed predictive models using sex, age, ETX IgG and IgM levels, and C. perfringens status. Our NB model achieved 84% accuracy, with sex (AUC = 0.63) and IgG levels (AUC = 0.57) as key predictors. Prediction models are easily accessible for identifying individuals at risk of developing MS and demonstrate promising predictive accuracy. However, given the relatively small sample size of this study, caution should be exercised when interpreting their generalizability. Larger, multi-center studies with independent validation cohorts are required to confirm the reliability and applicability of these models in broader populations [ 55 ]. Accordingly, some studies have demonstrated that the probability of getting MS could be linked to gender since MS is more likely to occur in women than men [ 56 ]. This probability may be due to the difference in the nervous and immune system between women and men [ 56 , 57 ]. Although there is no report about the role of ETX IgG levels in MS prediction, this factor was a powerful predictor in our study. Similar studies have been conducted in MS prediction. For example, a predictive model based on parameters like sex, birth date, and disability score recorded at early stages of the disease, is proposed by Manouchehrinia et al. and estimated the risk of disease progression within 10, 15, and 20 years of disease onset [ 58 ]. Darvishi et al. worked on a group of MS patients ( n  = 100), who were diagnosed via neurological evaluation and brain or total spinal MRI, and infectious disease patients ( n  = 100) with no prior neurological history seeking medical care served as control [ 59 ]. Their study identified age, season of birth, and gender as risk factors for MS disease. They indicated control’s mean age was higher than MS patients; 41.2 years vs. 36.1 years. This study indicated age as the highest risk factor associated with MS prediction. This factor can be considered since MS is more likely to occur in 20 to 40 years age groups. Also, partial dependence plot (PDP) in this study revealed a higher probability of MS diagnosis in females, who comprised 80% of MS patients compared to 48% of controls. In another study, Talebi et al. introduced a prediction model based on neurophysiological indices related to physical activity such as static balance, strength, dynamic balance, body composition, and flexibility. The model, tested on 110 volunteers (38 MS patients and 72 healthy subjects), achieved a high prediction accuracy (90%) in predicting patients with MS [ 60 ].

Although our sample size was limited, the variables used were found to be promising accessible screening tools. Of note, we constructed a nomogram for MS prediction with an AUC of 0.79, which represents clinical utility for early risk stratification. Incorporation of these models into electronic health records (EHRs) could support real-time decision making and patient triage [ 61 – 63 ]. External validation in larger, diverse populations remains necessary to confirm these findings and optimize model performance [ 64 ]. Although the detection rate of the etx gene appeared higher in NMOSD patients, the small number of positive cases (2 out of 10) warrants cautious interpretation. It should be noted that small sample sizes may exaggerate associations and lack statistical power, however, our main analysis emphasized predictive modeling, which integrates multivariate patterns rather than relying on univariate statistical associations [ 65 , 66 ]. The etx variable was included among several predictors in multivariable prediction models and does not independently drive the outcomes. Additionally, to address the risk of bias from small subgroup sizes, we used Fisher’s exact test for categorical comparisons and constructed patient-versus-control multivariable prediction models that combined MS and NMOSD groups to ensure model balance. We validated performance through 10-fold cross-validation and independent testing to improve generalizability. Furthermore, confounding variables, including antibiotic use, diet, immune status, and infection history have the potential to influence gut colonization and overall gene carriage [ 67 ]. Despite the collected medication history in this study, larger multi-center studies, particularly with expanded NMOSD cohorts are suggested for substantiating these preliminary findings and evaluating the biological significance of etx gene carriage.

Conclusions

Our study highlights a potential link between ETX-producing C. perfringens and demyelinating diseases including MS and NMOSD. Although anti-ETX IgG and IgM levels did not significantly differ between patients and healthy controls, etx -positive C. perfringens isolates were found exclusively in patients with MS and NMOSD. This suggests a possible microbiological association that warrants further investigation. While we did not identify a statistically significant correlation between ETX immunoreactivity and disease status, the integration of microbiological, serological, and machine-learning data enabled a multifactorial analysis of disease risk. A key strength of the present study is the combined use of molecular detection, antibody profiling, and predictive modeling, which provides a holistic approach to understanding ETX involvement in CNS demyelination. Moreover, the inclusion of MS and NMOSD patients enabled a broader exploration of ETX-related pathology across these related neuroinflammatory conditions. Future studies with larger cohorts are required to validate the association of ETX and demyelination and clarify the temporal relationship between ETX exposure and disease activity.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Supplementary Material 1

Supplementary Material 1