Multi Scale Brain Models

The Scale Problem in Neuroscience

The brain operates across an enormous range of spatial and temporal scales. At the molecular level, ion channels open and close in microseconds, and the binding and unbinding of neurotransmitters at synaptic receptors occurs on a timescale of milliseconds. At the cellular level, individual neurons integrate thousands of synaptic inputs over tens of milliseconds before deciding whether to fire a spike. At the circuit level, local networks of hundreds to thousands of neurons generate oscillatory rhythms and population dynamics on timescales of tens to hundreds of milliseconds. At the systems level, interactions between brain regions unfold over hundreds of milliseconds to seconds. And at the behavioral level, learning, memory consolidation, and development occur over hours, days, and years.

Each of these scales has its own appropriate modeling framework. Molecular dynamics simulations capture the physics of individual ion channels but cannot scale beyond a single synapse. Compartmental neuron models capture the electrical behavior of individual cells but are too computationally expensive to simulate entire brain regions. Point neuron models can scale to millions of neurons but sacrifice the dendritic computations that may be essential for certain cognitive functions. Mean-field models can capture whole-brain dynamics but compress the activity of millions of neurons into a few averaged variables, losing the individual-neuron-level detail that determines circuit computation.

Multi-scale modeling attempts to bridge these levels, using detailed models where detail matters and simplified models where it does not, with principled methods for coupling the scales together.

Mathematical Frameworks for Scale Bridging

Mean-field theory. Borrowed from statistical physics, mean-field theory derives the macroscopic behavior of a neural population from the statistical properties of its constituent neurons. Rather than simulating every neuron individually, mean-field models describe the population in terms of its average firing rate, mean membrane voltage, and the distribution of these variables across the population. The key insight is that if the population is large enough and the neurons are sufficiently similar, the macroscopic dynamics can be described by a small number of differential equations rather than by simulating millions of individual neurons.

Mean-field reductions have been derived for several common neuron models, including integrate-and-fire neurons, Izhikevich neurons, and conductance-based neurons. These reductions are exact in certain mathematical limits (large populations, weak correlations) and approximate in more realistic conditions. The accuracy of the approximation depends on the specific circuit being modeled and must be validated by comparison with detailed simulations.

Neural mass models. These represent the activity of a neural population by a single variable (typically the average membrane potential or firing rate) governed by a differential equation. Neural mass models are the simplest form of population-level description and are computationally inexpensive enough to simulate entire brain networks in real time. The Jansen-Rit model and the Wilson-Cowan model are widely used neural mass models that capture the basic dynamics of excitatory-inhibitory interactions within a cortical column.

Hybrid multi-scale coupling. The most sophisticated multi-scale approaches embed detailed models within coarser models, using the detailed models to parameterize the coarser ones. For example, a multi-scale model might simulate a specific cortical column with biophysically detailed neurons while representing the rest of the brain with mean-field models. The detailed and coarse regions exchange information at each time step: the mean-field model provides context (input from other brain regions) to the detailed model, while the detailed model provides more accurate dynamics back to the mean-field model. This approach allows researchers to study the interaction between cellular-level mechanisms and systems-level dynamics without the computational cost of simulating the entire brain at cellular resolution.

The Virtual Brain Project

The Virtual Brain (TVB) is the most prominent whole-brain simulation platform based on multi-scale modeling principles. TVB simulates the dynamics of the entire human brain by coupling neural mass or mean-field models for each of approximately 70 to 200 brain regions through the structural connectivity measured by diffusion MRI tractography.

Each brain region in TVB is represented by a neural mass model that captures the basic excitatory-inhibitory dynamics of a cortical column. The regions are connected according to the empirically measured white matter tracts, with connection strengths and conduction delays derived from the tractography data. The resulting network of coupled oscillators generates rich spatiotemporal dynamics that can be compared to empirical brain activity measured by fMRI, EEG, or MEG.

TVB has been used to study a range of neuroscience questions, including the origins of resting-state functional connectivity, the dynamics of epileptic seizures, the effects of focal brain lesions on network function, and the changes in brain dynamics during anesthesia. Its clinical applications include pre-surgical planning for epilepsy patients, where virtual lesion experiments can predict the effects of surgical interventions on seizure dynamics.

Challenges in Multi-Scale Modeling

Multi-scale brain modeling faces several fundamental challenges. The parameter estimation problem is particularly acute: as models span more scales, the number of parameters that must be estimated from experimental data increases rapidly, and the parameters at different scales may interact in complex ways that are difficult to characterize. A small change in a cellular-level parameter (like a synaptic time constant) can have large effects on network-level dynamics, making it essential to get these parameters right but difficult to do so given the limited experimental data available.

The computational cost of multi-scale simulation remains substantial, even with the efficiency gains from using simplified models at larger scales. A hybrid simulation that embeds a detailed cortical column within a whole-brain mean-field model may require hours of supercomputer time for each second of simulated brain activity. Scaling to larger detailed regions or incorporating additional biophysical detail further increases the computational demands.

Validation is challenging because different scales require different types of experimental data. Cellular models are validated against electrophysiological recordings from individual neurons, network models against local field potential recordings, and whole-brain models against functional imaging data. Ensuring consistency across all these validation levels, so that a model that is accurate at the cellular level also produces accurate predictions at the network and systems levels, is a formidable scientific challenge that has not yet been fully met.

Despite these challenges, multi-scale brain modeling represents the most scientifically rigorous approach to building computational models that respect the full complexity of biological neural systems. As experimental data becomes more comprehensive, computational resources more powerful, and mathematical frameworks more sophisticated, multi-scale models are becoming increasingly accurate and increasingly useful tools for both basic neuroscience and artificial brain research.