Distinct plasma metabolomic signatures differentiate autoimmune encephalitis from drug‐resistant epilepsy

Abstract Objective Differentiating forms of autoimmune encephalitis (AE) from other causes of seizures helps expedite immunotherapies in AE patients and informs studies regarding their contrasting pathophysiology. We aimed to investigate whether and how Nuclear Magnetic Resonance (NMR)‐based metabolomics could differentiate AE from drug‐resistant epilepsy (DRE), and stratify AE subtypes. Methods This study recruited 238 patients: 162 with DRE and 76 AE, including 27 with contactin‐associated protein‐like 2 (CASPR2), 29 with leucine‐rich glioma inactivated 1 (LGI1) and 20 with N‐methyl‐d‐aspartate receptor (NMDAR) antibodies. Plasma samples across the groups were analyzed using NMR spectroscopy and compared with multivariate statistical techniques, such as orthogonal partial least squares discriminant analysis (OPLS‐DA). Results The OPLS‐DA model successfully distinguished AE from DRE patients with a high predictive accuracy of 87.0 ± 3.1% (87.9 ± 3.4% sensitivity and 86.3 ± 3.6% specificity). Further, pairwise OPLS‐DA models were able to stratify the three AE subtypes. Plasma metabolomic signatures of AE included decreased high‐density lipoprotein (HDL, −(CH 2)n−, –CH 3), phosphatidylcholine and albumin (lysyl moiety). AE subtype‐specific metabolomic signatures were also observed, with increased lactate in CASPR2, increased lactate, glucose, and decreased unsaturated fatty acids (UFA, –CH 2CH=) in LGI1, and increased glycoprotein A (GlycA) in NMDAR‐antibody patients. Interpretation This study presents the first non‐antibody‐based biomarker for differentiating DRE, AE and AE subtypes. These metabolomics signatures underscore the potential relevance of lipid metabolism and glucose regulation in these neurological disorders, offering a promising adjunct to facilitate the diagnosis and therapeutics.

This entire validation process was repeated 100 times, resulting in 1000 models in total.In parallel, permutation testing was used to assess whether the model performed significantly better than random chance.The null distribution was generated by randomly permuting class assignments and building OPLS-DA models with the same 10-fold cross-validation with repetition scheme.The accuracy of the true models was compared to the null distribution using the Kolmogorov-Smirnov test.Models were considered significant only if their accuracies were significantly better than random chance (~50%).Discriminatory variables were identified by calculating the average of the variable importance in projection (VIP) scores of the ensemble of models, which indicated the contribution of a variable to the model.LGI1 vs NMDAR) in DRE (grey, n=169), CASPR2 (blue, n=27), LGI1 (green, n=30), NMDAR (orange, n=23).Significance was determined using one-way ANOVA with post hoc Tukey's HSD tests.Holm-Bonferroni method was applied for adjustment of p values due to multiple comparisons.* q < 0.05, ** q < 0.01, *** q < 0.001.HDL, high density lipoprotein.VLDL, very low-density lipoprotein.CM, chylomicrons.GlycA/B, glycoprotein A/B.UFA, unsaturated fatty acids.UFA, unsaturated fatty acids.PUFA, polyunsaturated fatty acids.Values were presented in mean ± SD.OPLS-DA models were validated on independent test data (10%) using an external 10-fold cross-validation strategy with repetition coupled with permutation testing.Accuracy, sensitivity, and specificity were calculated from the external test set to assess the robustness and predictive ability of the models.Accuracy/Sensitivity/Specificity (random) indicated the metrics calculated using the permutated dataset.R 2 and Q 2 were calculated from the OPLS-DA model built using the full data set, where R 2 provides a measure for how much variation is represented by the model and Q 2 for the goodness of prediction.Levels of metabolites are normalised to the mean of DRE group.CL, CN, LN refers to the pairwise AE subtype model in which the metabolite is selected as discriminatory metabolite.One-way ANOVA was used to determine the significance among AE subtypes.Benjamini-Hochberg method was used to control the false discovery rate at 0.05.
Figure S1.Schematic representation of model optimisation, cross-validation and permutation strategy (A) details for the OPLS-DA model building and cross-validationOPLS-DA model was built using the 'ropls' package, employing q2 and internal 7-fold cross-validation to optimise the number of orthogonal components (OrthoI).To address the instability of the OrthoI selected by the built-in 'ropls' package for this specific dataset, a 10-time repetition was introduced (Panel B).This involved shuffling the dataset and constructing OPLS-DA models ten times.The median OrthoI derived from the ten models was considered the optimum choice for building the OPLS-DA model.The optimum orthogonal number (The optimum OrthoI A) for the full dataset was recorded as optimum OrthoI A, and would be used for model optimisation for the training set in the cross-validation (Panel C).OPLS-DA models were subjected to rigorous validation using a 10-fold external cross-validation with 100 repetitions and permutation testing.In brief, this involves shuffling the dataset and splitting into a training set (90%) and a test set (10%), with equalised class sizes.The OPLS-DA model was trained exclusively on the training set with model optimisation (as illustrated in panel C) and evaluated on the test set to measure accuracy, sensitivity, and specificity.

(Figure S2 .
Figure S2.Prediction of subjects with post-AE epilepsy.The OPLS-DA scores plot illustrates the application of the OPLS-DA model, differentiating AE from DRE, for the prediction of three individuals with post-AE epilepsy.The plot includes the projection of these three subjects with post-AE epilepsy onto the DRE vs AE separation.Patients situated to the left of the dashed vertical line are anticipated to have DRE.

Figure S10 .
Figure S10.Correlation plot of discriminatory metabolites from the OPLS-DA models.The correlation plot encompasses all discriminatory metabolites extracted from the OPLS-DA models involving different comparisons (AE vs DRE, CASPR2 vs LGI1, CASPR2 vs NMDAR, LGI1 vs NMDAR).Pearson correlation coefficient r is visualised with different sized circles in the upper triangular of the correlation matrix, and the original r values are shown in the lower triangular.Holm-Bonferroni method was applied for adjustment of p values due to multiple comparisons.The plot includes only significant correlations.

Table S2 . Detailed case information of 3 post AE patients Post AE Case 1-
Considered to have had a probable encephalitis with seizures, headaches and fever in June 2016 and was admitted to hospital.At that time didn't have immunotherapy.-Returned to hospital 6 weeks later where they found a Glycine receptor antibody in her CSF but not serum and gave her steroids.Her MRI and CSF were normal and her antibodies in March 2014 and August 2016 were negative (NMDA/AMPA/GABAb/VGKC) -Continue to have focal and generalised seizures every few months -Seen in epilepsy clinic in October 2015 where no further mention is made of the possible encephalitis and she is only treated with anti-seizure medications -Time interval: November 2013 to October 2015 = 23 months

Table S3 . Discriminatory metabolites identified in the OPLS-DA models of AE vs DRE with fold changes.
in each key metabolite.Benjamini-Hochberg method was used to control the false discovery rate at 0.05.VIP, variable importance in projection.