2026 ECNP Congress

PS03-3039

Towards in silico modeling of psychosis: NMDA-Mg2+ dynamics and auditory sequence learning in a spiking neural network

Francesco Achilli — A. Muscettola, G.A. De Bellis, G.M. Gobbi, L. Grassi, G. Dumas, M. Belvederi Murri · University of Ferrara, Department of Neuroscience and Rehabilitation, Ferrara, Italy

2026 ECNP Congress

Introduction

Psychotic disorders have been influentially characterized under the predictive processing framework, distinguishing bottom-up disruptions (overly precise sensory signals and weak priors) and top-down disruptions (overly strong priors) across different illness stages [1]. However, mechanistic demonstrations relating biological alterations (such as NMDA receptor hypofunction or dopaminergic dysregulation) to network disruptions and symptom emergence remain lacking. Spiking Neural Networks (SNNs) offer a biologically grounded framework for testing such links [2].

Here, we present an SNN that features an auditory cortex capable of learning melodic sequence patterns, designed as input to a Global Neuronal Workspace (GNW) architecture [3] for multi-scale analysis of psychosis.

Methods

We developed an SNN including a cochlear layer (neurons tuned with tonotopic Gaussian curves) and projecting to an auditory cortex with four biologically inspired neuron populations: Regular-Spiking (RS, fast excitatory), Parvalbumin-positive (PV, fast inhibitory), Intrinsically-Bursting (IB, slow excitatory), and Somatostatin-positive (SOM, slow inhibitory), with an overall 80:20 excitatory-to-inhibitory ratio. Neurons learn via spike-timing-dependent plasticity (STDP), with an eligibility-trace variant for IB recurrent connections [4]. These neurons combine fast AMPA-like feedforward currents and slow NMDA-like currents gated by a voltage-dependent Mg2+ block [5]. During silence, accumulated slow current remains sub-threshold (97% attenuation at resting potential); subsequent feedforward input depolarizes the membrane and relieves the block, allowing the stored slow current to boost the response. The network receives two three-note sequences (C4→E4→A4 vs C4→E4→G3). After training (10 episodes, 50 parallel batches), plasticity was frozen and three validation trials were conducted: (1) context removal (full sequence vs isolated final note), (2) order scrambling (standard vs shuffled sequence), and (3) incomplete sequence (silence replacing the final note).

Results

IB neurons showed a 34% stronger response to the final note when it was preceded by the full learned sequence compared with isolated presentation (Cohen’s d = 3.8 and 3.3 for target and deviant, respectively, across 50 batches), confirming context-dependent modulation via NMDA-like slow dynamics. The order of sequence mattered: the activation ratio at note 3 relative to note 1 was 1.6 for standard sequences versus 0.5 for scrambled sequences. No autonomous IB firing occurred during silence in incomplete sequences, consistent with Mg2+ block preventing autonomous activation in the absence of feedforward input.

Conclusions

An SNN model learned sequence-dependent temporal context through plasticity and NMDA-like dynamics with Mg2+voltage gating. It provides a biologically constrained baseline for modeling aberrant sensory signaling in predictive processing accounts of psychosis. As the second part of the project, two biophysical perturbations will be introduced to simulate psychotic symptoms: bottom-up disruptions via increased feedforward gain for simulating aberrant salience at the sensory level, and top-down disruptions via altered recurrent excitatory connections at the GNW level. This framework will enable mechanistic test of how the alterations of priors and sensory precision influence conscious access, with potential application for in-silico evaluation of circuit-level therapeutic interventions.

References

  1. I. Goodwin, K. M. J. Diederen, E. J. Hird, V. Weilnhammer, M. I. Garrido, and F. Knolle, ‘Predictive processing accounts of psychosis: bottom-up or top-down disruptions’, Nat. Mental Health, vol. 4, no. 1, pp. 60–84, Jan. 2026, doi: 10.1038/s44220-025-00558-5.
  2. K. Volzhenin, J.-P. Changeux, and G. Dumas, ‘Multilevel development of cognitive abilities in an artificial neural network’, Proceedings of the National Academy of Sciences, vol. 119, no. 39, p. e2201304119, Sep. 2022, doi: 10.1073/pnas.2201304119.
  3. G. A. Mashour, P. Roelfsema, J.-P. Changeux, and S. Dehaene, ‘Conscious Processing and the Global Neuronal Workspace Hypothesis’, Neuron, vol. 105, no. 5, pp. 776–798, Mar. 2020, doi: 10.1016/j.neuron.2020.01.026.
  4. N. Frémaux and W. Gerstner, ‘Neuromodulated Spike-Timing-Dependent Plasticity, and Theory of Three-Factor Learning Rules’, Front. Neural Circuits, vol. 9, Jan. 2016, doi: 10.3389/fncir.2015.00085.
  5. C. E. Jahr and C. F. Stevens, ‘Voltage dependence of NMDA-activated macroscopic conductances predicted by single-channel kinetics’, J Neurosci, vol. 10, no. 9, pp. 3178–3182, Sep. 1990, doi: 10.1523/JNEUROSCI.10-09-03178.1990.