CNS*2026 Halifax

Introduction
Biophysical models impose substantial computational burdens, limiting large-scale simulation of complex neuronal dynamics, particularly with morphologically detailed neuron models. Previous evidence [1] suggests that the neuronal spiking behavior is primarily constrained by recent causal stimulus events rather than continuous full-timescale integration, which makes event-driven dynamical computation possible. We developed a machine-learning framework with a recurrent architecture for sustained spike prediction. The framework replaces computationally expensive continuous differential equation solving with an event-based mechanism, enabling temporal computation without requiring timestep-level simulation.

Methods
The framework uses recent excitatory and inhibitory events as input to recurrent neural architectures (LSTM/GRU) to encode temporal neuronal dynamics and learn a reduced representation of the latent state. Then we utilize a downstream multilayer perceptron to predict whether or not the neuron will spike, and if it does, the next-spike time (NST). Following individual spike evaluation, the trained framework was further tested under 40000 ms sustained neuronal activity driven by excitatory and inhibitory event streams at 200 Hz and 67 Hz, respectively, where predicted spikes constantly influenced subsequent neuronal activity, to estimate long-term temporal stability and dynamic spike prediction performance.


Results
Trained on a dataset of over one million stimulation trials and tested on 38k trials, the proposed event-driven framework achieved an F1 score and AUC of > 0.99, with a next-spike timing (NST) mean absolute error (MAE) of 0.07 ms, approaching the intrinsic temporal resolution of the NEURON simulation environment. Under sustained neuronal activity over a 40,000 ms simulation window, the framework reproduced 929 of 947 ground-truth spikes with only 51 missed spikes and 33 false-positive predictions, exceeding the performance of the baseline event-based model, which reproduced 135 false positives and missed 76 spikes under the same conditions.

Discussion
The stable performance observed under sustained neuronal activity suggests that the framework can preserve long-term temporal consistency beyond isolated spike prediction tasks, potentially including in network models, with potential run-time improvements for large cell models. However, the resulting errors differ from those of traditional biophysical models, so further work is needed to understand their effects on system behavior. Studying parameter sweeps or heterogeneous models would require incorporating parameters of interest into the machine learned model, re-introducing complexity. Extending the framework to generalize across broader biophysical conditions remains an important direction for future work.

References
Cudone, E., Lower, A. M., & McDougal, R. A. (2023). Reproducibility of biophysical in silico neuron states and spikes from event-based partial histories. PLOS Computational Biology, 19(10), e1011548.


Acknowledgement
Research was supported by the National Institute of Neurological Disease and Stroke of the National Institutes of Health under award number R01NS011613. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.