NMDA-Receptor Dysfunction Disrupts Serial Biases in Spatial Working Memory

In working memory (WM) tasks, attractive biases to previous items are evidence for continuous temporal integration of memories. These serial biases have been modeled as a product of synaptic short-term plasticity, allowing WM representations to endure in a synaptic trace and interfere with the next trial even when neural activity returns to baseline values. We hypothesized that the NMDAR, a key component of both short-term potentiation (STP) and stable WM delay activity, would be of central importance to serial biases in a visuospatial WM task. Confirming this hypothesis, we found drastically reduced biases in patients with anti-NMDAR encephalitis and schizophrenia, both diseases that have been related to NMDAR hypofunction. We simulated serial biases in a spiking neural network supported by a Hebbian STP mechanism that builds up during persistent delay-activity. We found a close correspondence between patient and model behavior when gradually lowering levels of STP, suggesting a disruption of short-term plasticity in associative cortices of schizophrenic and anti-NMDAR encephalitis patients. Further, we explored the capability of the model to explain reduced biases in light of the disinhibition theory of schizophrenia.


Introduction
Continuity of mnemonic contents in time contributes to forming coherent memory representations. Recently, attractive response biases towards previously memorized features in delayed-response tasks have been reported as evidence for the continuous integration of working memory (WM) contents between trials (Fischer & Whitney, 2014). In turn, brain disorders with reported executive and memory dysfunction may be characterized by reduced WM serial bias (Lieder et al., 2019), revealing reduced temporal coherence of memory representations. To gain mechanistic insight into this effect, w e tested a unique population of patients recovering from anti-NMDAR encephalitis, an immune-mediated brain disease causing a drastic reduction of NMDAR, and a population of patients suffering from schizophrenia, another disorder associated with hypofunctional NMDARs (e.g., Javitt, 2010).
NMDAR hypofunction is a candidate mechanism for a disturbed balance between cortical excitation and inhibition (E-I balance, Belforte et al., 2010), supporting one major glutamatergic theory of schizophrenia, the cortical disinhibition hypothesis (Lewis, Hashimoto, & Volk, 2005). On the other hand, there is evidence for reduced excitation of cortical networks under the influence of NMDAR antagonists in the presence of cognitive demand (Wang et al., 2013;Driesen et al., 2013). Further, NMDARs are an essential component of synaptic plasticity, including short-term potentiation (STP), leading to reduced STP when NMDAR are inhibited (Volianskis et al., 2013).
We hypothesized that NMDAR hypofunction in anti-NMDAR encephalitis and schizophrenic patients would change the temporal dynamics of working memory, influencing how information is maintained and integrated between trials. We collected behavioral data from anti-NMDAR encephalitis patients, schizophrenic patients, and healthy controls performing a visuospatial delayed-response task, and analyzed serial biases as a behavioral read-out of information integration across time.  averaged errors e as a function of d for 0, 1 and 3 s delay lengths, left to right). Groupwise models allowed to asses the time course of biases in each population: initially slightly repulsive biases became gradually attractive with delay length for healthy controls (F(2,16) = 24.39, p = 1.4e-5) and for encephalitis patients (F(2,22.8) = 5.01, p = 0.016). In contrast, delay length did not affect serial biases in schizophrenic patients (F(2,20.5) = 1.14, p = .34). Rather, a significant repulsive bias (F(1,14.3) = 7.19, p = .018) dominated all delay lengths. Mean biases for each group and delay are shown in Figure 1c.

Reduced NMDARdependent STP causes reduced serial bias in spiking neural networks
To explore potential network mechanisms of reduced serial biases in NMDAR-related diseases, we ran simulations of consecutive WM trials in a spiking neural network model of prefrontal cortex (Compte, Brunel, Goldman-Rakic, & Wang, 2000). In this model, a pool of inhibitory and a pool of excitatory leaky integrate-and-fire (LIF) neurons are connected all-to- Left to right, bias plots after 0, 1, and 3 s delay. Error shading depicts 1 SEM. d) A reduction of excitation to inhibitory neurons (disinhibition) increases serial bias, while reducing excitation to excitatory neurons decreases serial bias. Note that decreasing or increasing NMDA-mediated excitation to both excitatory and inhibitory neurons concurrently does not affect biases (diagonal pattern), while inducing a strong E-I imbalance causes a loss of network bistability (loss of the memory attractor for I>E, lower right orange triangle, and loss of stable low baseline firing when E>I, upper left black triangle).
all. Crucially, the strength of connectivity within the excitatory pool is determined by the similarity of any two neurons' preferred stimuli, so that neurons with similar preferred locations excite each other more strongly. During the delay period, attractor dynamics cause persistent neural firing even in the absence of external input (Figure 2a). In the response period, neural activity is suppressed with an external input and the network returns to a stable baseline state of low firing rates.
We extended this network model by adding an associative STP mechanism in excitatory synapses. Synapses between neurons that participated in previous-trial delay activity are potentiated by~1% ( Figure 2b) and this potentiation decays between trials as a result of sparse, spontaneous presynaptic spikes (Volianskis et al., 2015). At the onset of the next stimulus, a location-specific synaptic trace persists from the previous trial and attracts the current trial's bump of activity progressively over the course of the delay (orange line in Figure 2a).
To investigate WM behavior, we simulated 5,000 pairs of consecutive trials with randomized stimulus locations. Network inputs were slightly transformed to mimic the repulsive bias away from previous stimulus locations after 0 s delay, reflecting adaptation effects taking place in sensory cortices (Bliss & D'Esposito, 2017). To obtain behavioral readouts from the network, the population vector was calculated from firing rates measured just after stimulus offset (0 s), after 1 s, and after 3 s. Errors ' e were calculated as the distance between the decoded angle and the stimulus angle for each trial. Figure 3a-c shows serial biases after 0, 1, and 3 s for different levels of STP (strong, intermediate, and low STP). Similarly to patients' data, serial biases were initially repulsive, and increased over the course of WM delay for intact STP. In contrast, conditions of reduced STP led to lower serial biases, even after a 3 s delay (low STP).
We further investigated whether disturbing the EIbalance of our network could account for reduced serial biases. We gradually reduced/increased the efficiency of NMDAR in excitatory and inhibitory synapses ( Figure 3d) and measured peak serial bias amplitudes for all conditions. Our analyses show that a disinhibition of the circuit, by reducing excitatory NMDA conductances onto inhibitory neurons, increased serial bias, rather than decreasing it. In contrast, a decrease in excitatory conductances onto excitatory neurons reduced bias.
Together, these simulations are inconsistent with the cortical disinhibition hypothesis of schizophrenia, at least during the mnemonic period of a WM task. Rather, they support two alternative hypotheses: (1) a reduction in NMDA-dependent STP or (2) reduced NMDA-mediated recurrent excitation between delay-active neurons.

Conclusions
We found a drastic reduction of attractive WM serial biases in patients with anti-NMDAR encephalitis and schizophrenia, as compared to healthy controls. Using a biologically-constrained model, we propose that synaptic dysfunction underlies these reductions in bias, either as a result of reduced STP or diminished recurrent cortical excitation.

Experimental Procedures
We tested 16 patients with anti-NMDAR encephalitis (enc), 16 schizophrenic patients (schz), and 18 healthy controls (ctrl) in a visuospatial WM task shown in Figure 1a. In each trial, subjects were to memorize a stimulus that appeared at a random angular location with fixed eccentricity. After a delay of 0, 1 or 3 seconds, they were prompted to respond by making a mouse click at the remembered location. Response errors e were measured as the angular displacement of responses from the target stimulus in each trial.

Data analysis
A linear mixed model of errors ij e in trial i and subject j included the factors group (ctrl, enc or schz), delay (0, 1, or 3 s) and first derivative of Gaussian (DoG)  The DoG location hyperparameter was fixed at 0º and the variance hyperparameter was determined using cross-validation.

Neural network simulations
We simulated 5,000 consecutive pairs of trials using a LIF bump-attractor network model of spatial WM as previously described by Compte, Brunel, Goldman-Rakic and Wang (2000). Synaptic conductance dynamics modeled NMDARs, AMPARs, and GABA A Rs. We extended the model by including Hebbian STP at recurrent excitatory synapses.