Synaptic Function and Sensory Processing in ZDHHC9-Associated Neurodevelopmental Disorder: A Mechanistic Account

Abstract

Loss-of-function ZDHHC9 variants are associated with X-linked intellectual disability (XLID), rolandic epilepsy (RE) and developmental language difficulties. This study integrates human neurophysiological data with a computational model to identify a potential neural mechanism explaining ZDHHC9-associated differences in cortical function and cognition. Magnetoencephalography (MEG) data was collected during an auditory roving oddball paradigm from eight individuals with a ZDHHC9 loss-of-function variant (ZDHHC9 group) and seven age-matched individuals without neurological or neurodevelopmental difficulties (control group). Auditory-evoked fields (AEFs) were larger in amplitude and showed a later peak latency in the ZDHHC9 group but demonstrated normal stimulus-specific properties. Magnetic mismatch negativity (mMMN) amplitude was also increased in the ZDHHC9 group, reflected by stronger neural activation during deviant processing relative to the standard. A recurrent neural network (RNN) model was trained to mimic group-level auditory-evoked responses, and subsequently perturbed to test the hypothesised impact of ZDHHC9-driven synaptic dysfunction on neural dynamics. Results of model perturbations showed that reducing inhibition levels by weakening inhibitory weights recapitulates the observed group differences in evoked responses. Stronger reductions in inhibition levels resulted in increased peak amplitude and peak latency of RNN prediction relative to the pre-perturbation predictions. Control experiments in which excitatory connections were strengthened by the same levels did not result in consistently stable activity or AEF-like RNN predictions. Together, these results suggest that reduced inhibition is a plausible mechanism by which loss of ZDHHC9 function alters cortical dynamics during sensory processing.

Keywords: MEG; ZDHHC9; epilepsy; intellectual disability; language; recurrent neural networks.

FIGURE 1

Roving oddball paradigm. A standard tone (S) of either 250 Hz, 500 Hz, or 1000 Hz is presented randomly between 3 and12 times (n). After this sequence of repetitions, the frequency of the tone changes (deviant tone, D), which then becomes the new standard through repetitions (R₁ denotes the first repeat). In the right part of the figure, the first tones (deviants) of new series of tones (10, 11, or 12 presentations) are depicted (subsequent repeats are not drawn for simplicity).

FIGURE 2

Cluster‐based permutation testing. Permutation‐based clusters of statistically significant differences between standard‐evoked and deviant‐evoked AEFs in control and ZDHHC9 groups. Time‐windows where these differences were identified are highlighted in yellow.

FIGURE 3

Overview of modelling workflow. (a) Spectrograms of example standard and deviant inputs used for the RNN. Three frequencies were used as in the roving oddball paradigm: 250 Hz, 500 Hz and 1000 Hz. A standard (S) input to the RNN model refers to a spectrogram with three identical tones. A deviant (D) input to the RNN model refers to a spectrogram with the third tone of a different frequency from the previous two. The train set consisted of S inputs of 250 Hz and 500 Hz, as well as the following sequences of tones, of which the third represented the D input: 250–250‐500 (Hz), 250–250‐1000 (Hz), 500–500‐1000 (Hz). The test set consisted of standard inputs of 1000 Hz and tone triads including: 500–500–250 (Hz), 1000–1000–250 (Hz), and 1000–1000–500 (Hz). Spectrograms spanned the same time range (548 ms) as the RNN labels (post‐trigger AEFs with a stimulus offset at 50 ms) and RNN outputs. Each tone in the spectrograms was 50 ms long, and inter‐tone intervals were 199 ms each. The RNN was trained to respond to the third tone in the sequence of three, depending on whether it is a standard or deviant. (b) Simplified diagram of the hierarchical RNN architecture. Input layer (green) had 63 recurrent units, each hidden layer had 64 units and the output layer had 1 recurrent unit. (c) Targets were 1200 simulated AEFs obtained by adding Gaussian white noise (standard deviation 0.6) to the control group level post‐stimulus AEF in response to standard tones. The same was done for deviant AEFs, resulting in 1200 simulated deviant AEFs. d. The RNN was trained for 10 epochs (i.e. iterations through the entire training dataset). The lower validation loss reflects the absence of dropout regularisation during validation, as opposed to training. e. RNN predictions to S and D inputs.

FIGURE 4

Hidden layer activations and output unit predictions for standard and deviant inputs. (a) Model predictions and hidden layer activations plotted by layer (rows) and input condition (columns). The activations were computed as an average across all S input types and all D input types, respectively. The x‐axis represents time. Amplitudes of hidden unit activations, shown in the four upper rows, increase toward the output. Patterns of hidden unit activations also spread out and become more complex with increasing layer depth. Visible differences in hidden unit response magnitudes between input conditions were also found. Model outputs and corresponding grand‐average AEFs (with Gaussian white noise; mean = 0 and s.d. = 0.6) are plotted in the bottom row. All data displayed in the panels represents the RNN response to the three‐tone input sequence (S or D). b. Hidden layer activations over time, averaged across the 64 hidden units.

FIGURE 5

RNN predictions before and after each perturbation experiment. Predicted AEFs for standard and deviant inputs (left panel) and relative increases from initial RNN predictions (pre‐perturbation) after each perturbation experiment (right panel). (a) Experiment 1 (negative weight perturbation). (b) Experiment 2 (positive weight perturbation). (c) Experiment 3 (random weight perturbation).