Kynurenine pathway metabolomics predicts and provides mechanistic insight into multiple sclerosis progression

Sample Cohorts and Study design

All studies were carried out in accordance with the guidelines of the relevant institutional human research ethics committee and approved by St Vincent’s Hospital Sydney (HREC– H03/037) and Macquarie University (HREC – 5201300333). Our study adheres to ‘The code of ethics of the World Medical Association (Declaration of Helsinki)’ for experiments involving human subjects. Written consent was obtained from respective sources that provided the samples. In order to thoroughly profile KP metabolism in MS, Cohort 1 samples were obtained through the Accelerated Cure Project for Multiple Sclerosis (ACPMS), USA, and consisted of serum samples from patients with RRMS, SPMS, primary progressive (PPMS) and healthy controls (HC) subjects. Exclusion criteria included MS patients currently receiving disease modifying drugs or who had corticosteroid within the past 3 months, or presence of other medical conditions. Randomized samples were age and gender matched between the experimental groups (i.e. RRMS, SPMS, PPMS and HC) where possible. As repeated blood samples collected over two years were available, Cohort 2 offered the ability to track changes in the KP longitudinally and was obtained through the Tasmanian MS Longitudinal Study conducted between 2003 and 2005. To further validate our study, we sourced another cohort of MS patients (Cohort 3) with matched same-patient serum and cerebrospinal fluid (CSF) from The Human Brain and Spinal Fluid Resource Center (HBSFRC), USA. All samples were unsorted and re-labelled for blinding prior to analysis.

Profiling of the kynurenine pathway metabolites

All reagents and KP metabolites were analytical reagent grade and were purchased from Sigma-Aldrich (St Louis, MO), unless otherwise stated. Deuterated internal standards were purchased from Medical Isotopes, Inc (Pelham, NH). KP metabolites were extracted using 10% (w/v) trichloroacetic acid (TCA) with equal volume of serum samples in accordance with methods previously described¹⁵. CSF samples were prepared similarly to serum samples except that deproteinization with TCA was not performed.

Concurrent analysis of tryptophan, kynurenine, 3-hydroxykynrenine (3-HK), 3-hydroxyanthranilic acid (3-HAA), and anthranilic acid (AA) was performed with UHPLC as described by Jones et al.¹⁶, using an injection volume of 20 μL of the prepared extract from each samples. KA detection was performed using a gradient mobile phase comprise of 50 mM sodium acetate buffer supplement with 25 mM zinc acetate (dihydrate) to enhance fluorescence intensity and 2.25% acetonitrile as organic modifier (Solvent A), and 10% acetonitrile (Solvent B). Each sample (10 μL) was injected into a Poroshell RRHT C-18, 1.8 μm 2.1 × 100 mm column (Agilent Technologies, Inc, Santa Clara, CA) maintained at 38 °C for 12 min run time at a unison flowrate of 0.75 mL/min. The gradient elution consisted of 100% solvent A for 3 min and then 50% solvent A and 50% solvent B for 2 min, followed by 100% B for 2 min and 100% solvent A (run time 10 min). This gradient ensures sufficient time for KA retention while minimizing potential build-up of pressure due to precipitation of the high salt buffer. Detection of KA used fluorescence (excitation and emission wavelengths of 344 and 388 nm, respectively with a retention time of 1.5 min). Agilent OpenLAB CDS ChemStation (Edition C.01.04) was used to analyze the chromatograms (Supplementary Figure S1A and B).

For GCMS, 50 μL of the prepared extract were derivatized. Concurrent analysis of PA and QA were carried out as described by Smythe et al.¹⁷ with slight modification using an Agilent 7890 A GC system coupled with Agilent 5975 C mass spectrometry detector and Agilent 7693 A autosampler (Agilent Technologies, Inc, Santa Clara, CA) with one microliter of derivatized mixture. Separation of PA and QA were achieved with a DB-5MS column, 0.25 μm film thickness, 0.25 mm × 30 m capillary column (Agilent Technologies, Inc, Santa Clara, CA) within 7 min but the assay run time was set for 12 min to prevent sample carryover. Concentrations of PA and QA were analyzed using Agilent GC/MSD ChemStation software (Edition 02.02.1431) and interpolated from the established six-point calibration curves based on the abundance count ratio of the metabolites to their corresponding deuterated internal standards within each standards and samples (Supplementary Figure S1C and D).

The intra- and inter-assay CV was within the acceptable range of 4–8% for UHPLC assays and 7–10% for GCMS assays calculated from the repeated measures of the metabolites standards incorporated during the sequence run. Nicotinamide adenine dinucleotide (NAD⁺) was measured with 20 μL of neat serum using a previously described method¹⁸.

Profiling of the inflammatory mediators

Quantification of multiplexing cytokines, chemokines and growth factors was performed using commercial 27-plex magnetic bead based immunoassay kits (Bio-Rad, Hercules, CA) on Cohort 1. Each assay was performed in accordance to manufacturer’s instructions at the Australian Proteome Analysis Facility, as described by A. Khan¹⁹. Final readout of the sample concentration was expressed as picogram per millilitre (pg/mL) based on the standard curves integrated in the assay using the Bio-Plex Manager v5.0 software with reproducible intra- and inter-assay CV of 5–8%.

Statistical analysis and modelling

The normality of the variables, where required, was checked by Shapiro-Wilk and/or Kolmogorov Smirnov normality tests and equality of the variances by Levene’s test for equality of variances. Comparisons between different MS subtypes and control group were performed by either ANOVA or Kruskal Wallis test, depending on the normality and equal variances analyses. Correlations between variables in Cohort 1 were assessed by Pearson’s correlations; where required, log transformation of the variables was made before calculating correlations. A p-value of <0.05 was considered statistically significant. Paired t-tests or Wilcoxon Signed Ranks tests, depending on the distribution of the variables were applied to investigate the changes to the inflammatory mediators and KP variables from Cohort 2 where the patients were observed between the years 2003 (baseline, cohort enrollment) and 2005. Four classification methods were used to identify the relative predictive value of each of the variables in each MS subtype in Cohort 1 (Training set) and then validated against Cohort 3 (Test set). These methods included the Classification and Regression Tree, Support Vector Machines, Discriminant Analysis and C5.0 Decision Tree²⁰,²¹. A classification model was considered successful when it consistently returned high predictive accuracy rates across all three MS subtypes and control group. The model predictions were compared to random predictions (class specific lifts) to decide which model was best for future predictions. In addition, class-specific statistical powers were compared to identify the best classification model. All classification models were developed using IBM SPSS Modeler 14.2, and R²² incorporated with various R packages such as Rcmdr, Lawstat, MASS²³, CAR and Lattice. To investigate whether changes to the CSF KP profiles mirrored serum KP profiles in the controls and MS subtypes of Cohort 3, we developed a series of models to identify the best relationships between the CSF KP variables and serum KP variables in the experimental groups. In the first model (model 1), we investigated whether it was possible to predict the CSF value for a KP variable based on its paired serum sample variable. In the second model (model 2), we added the experimental groups as covariates to model 1 to predict the CSF variables. In the last model (model 3), we used model 2 but adjusted for potential confounding factors from all other serum KP variables. We compared these models based on Akaike information criterion (AIC) and adjusted R², to identify the best model(s).

Participants’ characteristics

Cohort 1 made up a total of 136 participants consisting of 50 RRMS, 20 SPMS, 17 PPMS and 49 HC. Age differences between the subtypes were adjusted accordingly in the subsequent analysis. Similarly a total of 59 participants consisting of 44 RRMS and 15 SPMS from Cohort 2 were included in this study. Each Cohort 2 participant provided a baseline and follow-up (1.72 years ±0.27) serum samples for analysis. A total of 36 participants from Cohort 3 met exclusion criteria and provided CSF and matching serum. Accordingly, Cohort 3 consisted of 10 patients with RRMS, 20 patients with SPMS and 6 HC. Demographic and clinical characteristics of the subjects in all cohorts are summarized in Table 1.

Activation of the kynurenine pathway in MS

KP activation occurs when the activity of indoleamine 2,3-dioxygenase (IDO-1; the first enzyme in the KP) is significantly higher, leading to the consumption of kynurenine and higher kynurenine/tryptophan (K/T) ratio. In all the MS subtype groups the K/T ratio was significantly increased compared to the HC group (p < 0.0001, Fig. 1A–C; see Supplementary Table S2 for complete KP profile). These data also emphasize the importance of analyzing both TRP and KYN to determine KP activation, as relying on a single metabolite can lead to misinterpretation of data.

Abnormal downstream KP metabolites production in MS indicating excitotoxicity

As noted above, KA is capable of preventing glutamate-induced excitotoxicity induced by QA. Significantly, we found that KA levels were highest in the RRMS group relative to HC and progressive MS groups (p < 0.0001, Fig. 1D). However, KA levels were significantly lower in the progressive MS groups relative to controls. Picolinic acid, another known neuroprotective KP metabolite²⁴, showed a similar trend to KA, being highest in the RRMS but lowest in the PPMS, groups (p < 0.0001, Supplementary Table S2).

Production of QA increased uniformly in concert with disease severity and was particularly elevated in the PPMS group (p < 0.0001, Fig. 1E). We also found decreased NAD+ in all MS patient groups (p < 0.0001, Fig. 1F) which is also indicative of net QA elevation. The QA/KA ratio is indicative of excitotoxic potential with higher QA/KA ratio value favoring excitotoxicity. The QA/KA ratio was higher in both PPMS and SPMS groups compared to controls and the RRMS group (p < 0.0001, Fig. 1G). 3-HK, another potential neurotoxin, was found to be significantly higher in MS groups compared to healthy control (p < 0.0001, Supplementary Table S2). These data support our hypothesis that toxic KP metabolites fuel neurodegeneration in MS.

KP profile changes involving innate immunity are associated with MS disease severity and subtype

Pearson’s correlation analysis showed that the KP variables correlated significantly to Expanded Disability Status Scale (EDSS) scores suggesting that perturbed KP metabolism may reflect the progression of the disease. Notably, the QA/KA ratio has the strongest correlation with EDSS (r = 0.62, p < 0.0001) underscoring the potential significance of these key KP parameters to the disability and severity of MS (Table 2). Using the Wilcoxon signed rank test for paired data analysis, we showed that the K/T ratio significantly increased over time (p = 0.029) in patients with RRMS demonstrating increased indoleamine 2,3-dioxygenase (IDO-1) activity (Table 3). We also found increased expression of the innate immunity signature proteins, macrophage inflammatory protein (MIP)-1α (p = 0.004) and MIP-1β (p = 0.001) over time in RRMS that are produced by infiltrating macrophages (Table 3). This is accompanied by decrease in adaptive immune respond evident by decreased IL-2 level. Although IL-7 was shown to be significantly increased over time, however, the degree of change is very small (from 13 to 18 pg/ml) and unlikely to be clinically relevant.

The KP and immune profiles predict the course of disease in MS

Currently, there are no validated biomarkers of MS¹⁴ but our data suggests that certain metabolic KP signatures may discriminate MS subtype. To further explore the utility of KP metabolites as MS biomarkers, we applied predictive analytics. Initially, a predictive model was developed incorporating KP metabolites, inflammatory mediators and patient demographic information. The predictive model was then used to yield classification results, i.e., successful identification of MS subtype. From a total of 37 potential predictors, we found six that were the most critical determinants successfully predicting MS subtype (Fig. 2A). These were KA, QA, tryptophan, PA, fibroblast growth factor-basic and tumour necrosis factor-α (in order of relevance). Using various classification models in Training Set, a C5.0 Decision Tree provided the best overall prediction achieving an accuracy of 91% shown in Fig. 2B. We further validated the model using a Test set (Cohort 3) and obtained a predictive accuracy of 83% (Fig. 2D). This confirmed that our panel of 6 predictors can be used as a MS subtype biomarker.

To confirm that a facile blood-based KP biomarker for MS is feasible, we correlated CSF KP variables to patient-paired serum profiles (Cohort 3). After adjusting for confounding factors (stratification of clinical grouping and presence of other KP parameters), our regression analysis indicated that serum KP metabolites were able to explain 62.9% (p < 0.001) of changes observed in patient-paired CSF KP variables (Table 4) with a moderately strong (65%) to strong (79.3%) correlation between CSF and serum KP metabolites.