Multimodal electrophysiological analyses reveal that reduced synaptic excitatory neurotransmission underlies seizures in a model of NMDAR antibody-mediated encephalitis

Seizures are a prominent feature in N-Methyl-D-Aspartate receptor antibody (NMDAR antibody) encephalitis, a distinct neuro-immunological disorder in which specific human autoantibodies bind and crosslink the surface of NMDAR proteins thereby causing internalization and a state of NMDAR hypofunction. To further understand ictogenesis in this disorder, and to test a potential treatment compound, we developed an NMDAR antibody mediated rat seizure model that displays spontaneous epileptiform activity in vivo and in vitro. Using a combination of electrophysiological and dynamic causal modelling techniques we show that, contrary to expectation, reduction of synaptic excitatory, but not inhibitory, neurotransmission underlies the ictal events through alterations in the dynamical behaviour of microcircuits in brain tissue. Moreover, in vitro application of a neurosteroid, pregnenolone sulphate, that upregulates NMDARs, reduced established ictal activity. This proof-of-concept study highlights the complexity of circuit disturbances that may lead to seizures and the potential use of receptor-specific treatments in antibody-mediated seizures and epilepsy.


Dynamic causal modeling
Feature extraction from whole-cell patch clamp recordings We quantified amplitude, decay time, and variability of sEPSCs measured from the whole-cell hippocampal patch clamp recordings described above. We selected 10 minutes of continuous, artefact free recordings from both control antibody and NMDAR antibody conditions, and bandpass filtered these epochs between 0.3Hz and 2000Hz. In these traces, we identified peaks with a local prominence of at least 20mV (n=1624 in control antibody condition; n = 374 in NMDAR antibody condition), and segmented windows of 30ms prior to the peak, and 60ms following the peak, of which we considered the first 20ms as baseline. We normalised each segment to the baseline mean and extracted the mean amplitude of these normalised peaks.
For each segment we identified the time elapsed from the peak to when the signal had decayed to half-peak values, and we then calculated the mean of this period as the time constant of sEPSCs. To quantify variance in the responses, we calculated the cumulative frequencies of sEPSCs of given amplitudes (normalized to the maximal amplitude observed in each condition, respectively). We then quantified the gradient of this distribution at the 50 th centile to approximate the population parameter σ used in DCM. Log-differences between these average quantitative features in the control antibody and NMDAR antibody conditions from the wholecell patch clamp recordings were subsequently used to inform priors of DCM inversions of LFP data.

LFP segmentation and data extraction
We selected six 1 hour segments for 2 animals treated with control IgG, and three 1 hour segments for 1 treated with NMDAR antibody IgG, at the peak of the observed epileptogenesis (that is, at 48h). Each trace was z-scored and sections exceeding an absolute z-score of 5.5 were coded as seizure; else they were coded as interictal. We then added the 45s before and after the automatically identified seizure-segments to the seizure segments to capture seizure onset and seizure offset transitions. This process served to automate the LFP ictal/interictal classification confirmed by visual analysis, and we then divided all segments into 45s sections for subsequent DCM analysis. In total this process resulted in 514 control antibody interictal segments, 0 control antibody seizure segments, 114 NMDAR antibody interictal segments, and 228 NMDAR antibody seizure segments after exclusion of artefact. For each of the LFP traces, average power spectra were estimated using a multivariate autoregressive model implemented in the DCM software 1 .

DCM fit to control data
Assuming that data recorded from control antibody injected mice are representative of the 'baseline' state, we first fitted a single canonical microcircuit model (CMC) 2 using standard DCM inversion techniques (EM algorithm performing gradient descent on a free energy approximation of the negative log likelihood). This process provides us with posterior densities over neuronal parameters for the CMC (as summarized in Table 1).

DCM fit to NMDAR antibody interictal data
In order to test whether the synaptic changes identified in the whole-cell patch clamp recordings contribute to the LFP features recorded in NMDAR antibody injected mice, we fitted single CMC models under different prior parameter sets to the NMDAR antibody interictal LFP data and compared their relative evidence. The null model used control antibody derived parameter values exactly as priors, without any additional changes. The remainder of the model space was defined by altering prior parameter values quantitatively based on the expected control antibody vs. NMDAR antibody differences derived from the whole-cell patch clamp recordings. Specifically, we added the log difference between NMDAR antibody and control antibody sEPSC amplitude to excitatory coupling parameters γ1-3; the log difference between NMDAR antibody and control antibody sEPSC half-life to the excitatory time constants τ2,3; and the log difference in cumulative variance between NMDAR antibody and control antibody size distribution to the population variance parameter σ. We divided the model space to compare models where only subsets of these changes were made on the priors. These were divided along two main design features: parameter type (γ, τ, σ, or their combinations) and location in the microcircuit (superficial, fast oscillator pair; deep, slow oscillator pair; or both), resulting in a total of 7 x 3 models that carried some microscale information in their priors; and one null model without the microscale information. Comparison between these models was then made based on the free energy approximation of their respective model evidence 3 . The winning model was used for subsequent analyses.
DCM fit to NMDAR antibody seizure data To identify parameter changes associated with the transition into epileptic seizures, we fitted a single CMC to NMDAR antibody seizure LFP segments using parameters from the winning model identified in the previous step. We considered the absolute parameter differences between the DCM fit to NMDAR antibody interictal data, and the DCM fit to NMDAR antibody seizure data as the parameter change associated with ictogenesis for the subsequent simulations.

Simulations
The parameters inferred by fitting the microcircuit models to LFP data are fully generative and can therefore be used to simulate novel data. We exploited this 'simulation mode' to test the effects of different gradual changes in parameters on the simulated LFP power spectrum. We ran simulations along two sets of parameter changes: 'epileptogenesis' parameters θE (i.e. the difference between control antibody and NMDAR antibody interictal estimated parameters); and 'ictogenesis' parameters θI (i.e. the difference between MMDAR antibody interictal and NMDAR antibody seizure estimated parameters). Simulations were in n = 200 steps along each of those axis, so that , = + + with θbase representing the empirical parameters inferred from the control antibody model inversion (Equation 1).
We classified each simulated power spectrum across this parameter space based on the least squares difference to any of the three empirically observed states into control antibody-like, NMDAR antibody interictal-like, and NMDAR antibody seizure-like states. To quantify local changes induced by small changes in ictogenicity, we calculated the mean squared difference between (i) the power spectra derived at a given simulation parameterization, and (ii) the power spectra derived at neighbouring parameterizations along the θI direction.

Supplementary Figure 1. Immunohistochemistry confirming NMDAR antibody binding to juvenile Wistar rat hippocampus and morphology of pyramidal cells in CA3 used for wholecell patch clamp recordings.
a Representative confocal image of hippocampus from sagittal brain slice prepared after acute ICV injection of monoclonal NMDAR antibody SSM5 shows typical staining pattern with secondary anti-human IgG (green). Scale bar = 250µm. b Magnification of panel (a) shows the typical binding pattern of NMDAR antibodies with relative sparing of the granular cell layer (GCL) compared to the molecular cell layer (ML). Scale bar = 50µm. c Representative confocal image of hippocampus from sagittal brain slices prepared after chronic infusion of monoclonal NMDAR antibody 12D7 after application of secondary anti-human IgG (green). Scale bar = 250µm. d Representative confocal image of hippocampus from sagittal brain slice prepared after chronic infusion of NMDAR antibody positive IgG also shows typical staining pattern with secondary anti-human IgG (green). Scale bar = 250µm. e Representative confocal image of hippocampus from sagittal brain slices prepared after chronic infusion of healthy control IgG after application of secondary anti-human IgG (green). There is no specific binding of these antibodies. Scale bar = 250µm. f Representative confocal image of hippocampus from sagittal brain slices prepared after acute injection control monoclonal antibody mG053 after application of secondary anti-human IgG (green). There is no specific binding of these antibodies. Scale bar = 250µm. g The NMDAR antibody treated brain slices showed increased median fluorescent intensity log EC50 ratios (GCL vs. ML) compared to controls (control group contains: healthy control IgG n=2 animals, 6 brain slices and control monoclonal 12D7 n=3 animals, 9 brain slices; NMDAR group contains: SSM5 monoclonal n=8 animals, 22 brain slices, 003-102 n=2 animals, 2 brain slices and patient IgG n=3 animals 5 brain slices; p=0.035, Mann-Whitney). h Cells used for spontaneous excitatory and inhibitory current recordings were from putative pyramidal cells with the CA3 region; an example cell injected with neurobiotin shows clear features of a pyramidal cell (green, anti-streptavidin IgG). Scale bar 150 µm.

Supplementary Figure 2. Pregnenolone sulphate increases synaptic levels of NMDARs in naïve rat hippocampal neurons and glutamatergic neurotransmission in rat brain slices in vitro.
a Representative whole-cell patch clamp sEPSC recordings from CA3 pyramidal cells in untreated (no antibody) hippocampal slices.