Computational Neuroscience Modeling Frameworks

Abstract

Computational neuroscience seeks to understand brain function through mathematical and computational models that capture neural dynamics at multiple scales -- from individual neuron spikes to whole-brain activity patterns observed in neuroimaging. The choice of modelling framework fundamentally shapes what questions can be addressed: detailed biophysical models (Hodgkin-Huxley, multi-compartment) elucidate single-neuron computation but do not scale to brain-wide simulation, while whole-brain neural mass models capture macroscopic dynamics but abstract away cellular mechanisms. Deep learning models of neural activity predict neural responses with high accuracy but provide limited mechanistic insight. This study developed and benchmarked six computational neuroscience modelling frameworks - Hodgkin-Huxley biophysical (NEURON), leaky integrate-and-fire spiking networks (NEST), neural mass models (The Virtual Brain, TVB), deep recurrent neural networks (RNN), predictive coding networks, and a proposed Multi-Scale Neural Dynamics Integration Framework (MSNDIF) combining biophysically-constrained cortical microcircuit models, mesoscale population dynamics via mean-field reduction, whole-brain connectivity from the Human Connectome Project, and task-adaptive learned priors from deep predictive coding -- across four neuroscience tasks: single-neuron response prediction (Allen Brain Observatory, 1,248 cortical neurons), cortical microcircuit dynamics (mouse V1, two-photon Ca2+ imaging, 8,400 neurons), whole-brain fMRI prediction (HCP task-fMRI, 996 subjects, 7 tasks), and disease state classification (ADNI Alzheimer's dataset, 1,842 subjects). MSNDIF achieved the highest composite accuracy across scales: single-neuron spike prediction r = 0.84 (vs Hodgkin-Huxley 0.72, LIF 0.68; p < 0.001), microcircuit correlation structure match 0.88 (vs LIF 0.72), whole-brain task prediction AUROC 0.914 (vs TVB 0.824, RNN 0.882), and Alzheimer's classification F1 0.902 (vs RNN 0.862). These results establish multi-scale integrative modelling as the most accurate framework for bridging cellular and brain-wide neural dynamics.