Whole Brain Emulation

Whole Brain Emulation (WBE) is a research framework in computational neuroscience, nanoscale connectomics, and high-performance scientific computing aimed at measuring the complete structural wiring diagram and physiological state parameters of a biological brain and executing its coupled dynamical equations in silico so the simulated circuit reproduces native sensorimotor and neural input–output transformations. Unlike abstract deep-learning architectures that optimize arbitrary weight matrices via backpropagation without anatomical correspondence, whole brain emulation seeks structural and mechanistic isomorphism: mapping individual biological Neuron morphologies, pre- and postsynaptic active zones, electrical gap junctions, glial interactions, and membrane conductance distributions directly onto coupled systems of differential equations. Roadmapped systematically by Anders Sandberg and Nick Bostrom in 2008 and grounded empirically in electron-microscopy connectomics, the field has progressed from 302-neuron nematodes to complete adult insect brains containing over 139,000 neurons.

Three-Dimensional Electron Microscopy and Automated Segmentation

Optical fluorescence microscopy—even with stimulated emission depletion (STED) or structured illumination—cannot densely trace packed axonal arbors across macroscopic brain volumes because unmyelinated cortical axons narrow to 50–100 nanometers, dendritic spine necks constrict to 40–60 nm, and chemical Synapse clefts span ~20 nm with 35–50 nm vesicles. Resolving synaptic adjacency across every process in neuropil requires three-dimensional volume electron microscopy (vEM) at voxel dimensions of 4 × 4 × 8 nm³ to 8 × 8 × 40 nm³:

• Serial Block-Face Scanning Electron Microscopy (SBEM): An in-chamber diamond ultramicrotome shaves 25–30 nm slices from the top of a heavy-metal-stained (osmium–thiocarbohydrazide–osmium, uranyl, lead) resin-embedded tissue block while backscattered electrons image the freshly exposed block face in automatic registration.
• Focused Ion Beam SEM (FIB-SEM): A gallium or plasma ion beam mills 4–8 nm layers off the block face, achieving truly isotropic nanoscale voxels ideal for tracing tortuous spine necks in Drosophila hemibrain volumes.
• Automated Tape-Collecting Ultramicrotomy (ATUM-SEM) and GridTape TEM: Thousands of ultrathin serial sections (30–45 nm) are collected non-destructively onto conveyorized kapton or slotted film reels and imaged in parallel across multi-beam scanning or camera-array transmission electron microscopes, scaling throughput to petavoxel cubic-millimeter specimens.

Because manual tracing of 1 mm³ of mammalian cortex would require millions of person-hours, automated 3D Flood-Filling Networks (FFNs) and U-Net affinity classifiers segment membranes and detect synaptic cleft polarities, followed by targeted human proofreading along thin axons.

Empirical Connectome Scaling from C. elegans to Mammalian Cortex

Empirical whole-nervous-system and cortical-volume reconstructions span five orders of magnitude in neuronal count:

1. Caenorhabditis elegans (302 Neurons): Completed in 1986 by John G. White, Eileen Southgate, J. Nichol Thomson, and Sydney Brenner from ~8,000 serial TEM micrographs (and refined across developmental stages by Witvliet et al. in 2021), mapping all 302 hermaphrodite neurons, ~6,400–7,000 chemical synapses, and ~890 electrical gap junctions in Caenorhabditis elegans.
2. Larval and Adult Drosophila melanogaster (3,016 to 139,255 Neurons): Following the 3,016-neuron larval brain connectome (Winding et al., 2023), the FlyWire Consortium (Dorkenwald et al., 2024; Schlegel et al., 2024) completed the entire adult female Drosophila melanogaster brain from 21 million serial TEM images (~106 TB), resolving 139,255 neurons, ~5.45 × 10⁷ chemical synapses, and 8,453 annotated cell types; computational Leaky Integrate-and-Fire simulations built directly on the FlyWire graph (Shiu et al., 2024) accurately predicted gustatory and grooming motor sensorimotor cascades without free weight training.
3. Cubic-Millimeter Mammalian Cortical Volumes (~50,000–75,000 Neurons): The MICrONS mouse visual cortex dataset and the Harvard/Google H01 human temporal cortex 1 mm³ reconstruction (Shapson-Coe et al., 2024; ~57,000 cells, ~1.5 × 10⁸ synapses, 1.4 petabytes of raw 4 × 4 × 33 nm³ EM data) demonstrated petascale mammalian connectomics. Extrapolating this voxel density implies ~500–1,000 PB (~0.5–1 exabyte) for an entire 500 mm³ mouse brain (~7 × 10⁷ neurons) and ~1–1.5 zettabytes for a 1,200 cm³ human brain (~8.6 × 10¹⁰ neurons, ~10¹⁴–10¹⁵ synapses).

Multi-Compartment Cable PDEs Versus Spiking Point Neurons

Translating a static structural graph into dynamical Action Potential propagation requires choosing a level of biophysical granularity along a steep accuracy–compute tradeoff. At high fidelity, multi-compartment Hodgkin–Huxley cable models discretize dendritic trees and axonal arbors into 10²–10³ cylindrical compartments per neuron governed by the transmembrane cable equation:

C_m (∂V / ∂t) = (d / 4R_a)(∂²V / ∂x²) − g_L(V − E_L) − ∑_k ḡ_k m_k^p h_k^q (V − E_k) − I_syn(t)

where C_m ≈ 1 μF/cm² is specific membrane capacitance, R_a is axial cytoplasmic resistivity, and gating variables (m_k, h_k) track voltage-dependent conformational kinetics of Na⁺, K⁺, Ca²⁺, and HCN Ion Channels. Stiff branched cable systems are integrated at Δt ≈ 25 μs via O(M) Hines tridiagonal implicit solvers or adaptive Runge–Kutta Methods, demanding ~10³–10⁵ floating-point operations per compartment per millisecond (~10¹⁸–10²² FLOP/s for human-scale multi-compartment emulation). Conversely, Leaky Integrate-and-Fire (LIF) and two-variable Izhikevich point-neuron models collapse dendritic geometry into somatic state updates at Δt = 0.1–1.0 ms (~10¹⁴–10¹⁶ FLOP/s at human scale), sacrificing nonlinear N-methyl-D-aspartate (NMDA) dendritic plateau spikes unless augmented by two-layer cascaded subunits.

The Connectome-to-Function Gap and Extrasynaptic Neuromodulation

Decades of electrophysiology—most famously Eve Marder's analyses of the 30-neuron crustacean stomatogastric ganglion and ongoing C. elegans modeling—demonstrated that anatomical wiring alone does not uniquely specify circuit dynamics. Identical synaptic graphs switch across distinct rhythmic states depending on invisible molecular parameters absent from heavy-metal EM contrast: excitatory versus inhibitory neurotransmitter identity (glutamate, GABA, acetylcholine, glycine), AMPA/NMDA receptor ratios, subcellular ion-channel density gradients, electrical gap-junction innexin/connexin conductances, astroglial tripartite uptake kinetics, and extrasynaptic volume transmission (monoamines and neuropeptides diffusing across tens of micrometers without anatomical synaptic junctions, mapped separately in C. elegans as the wireless neuropeptidome). Closing this parameter-identifiability gap motivates multimodal expansion microscopy (physically swelling hydrogels 4×–20× for multiplexed antibody barcoding of ion channels and receptors), spatial transcriptomics, and pre-fixation two-photon calcium or genetically encoded voltage-indicator recording.

Why hasn't the 302-neuron Caenorhabditis elegans connectome alone yielded a complete predictive emulation?

Heavy-metal electron micrographs measure contact areas and vesicle counts but do not indicate whether a synapse is excitatory or inhibitory, which receptor isoforms populate the postsynaptic membrane, or how >100 neuropeptides and monoamines (serotonin, dopamine, octopamine, tyramine) diffuse extrasynaptically to switch behavioral states. Recent neurotransmitter atlases, optogenetic perturbation maps, and neuropeptide GPCR networks have begun supplying these missing sign and gain parameters.

What dominates hardware cost in large-scale neural simulations: arithmetic compute or memory bandwidth?

Memory-access bandwidth and inter-node latency dominate over raw floating-point throughput. Cortical neurons fire sparsely (~1–5 Hz) yet project to 1,000–10,000 irregular targets across distant regions; routing asynchronous spikes and updating S ≈ 10¹⁴ synaptic states generates irregular memory traffic that underutilizes dense GPU pipelines and motivates event-driven neuromorphic hardware.

1. White, J. G., Southgate, E., Thomson, J. N., & Brenner, S. (1986). The structure of the nervous system of the nematode Caenorhabditis elegans. Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences, 314(1165), 1–340. https://doi.org/10.1098/rstb.1986.0056
2. Marder, E. (2012). Neuromodulation of neuronal circuits: Back to the future. Neuron, 76(1), 1–11. https://doi.org/10.1016/j.neuron.2012.09.010
3. Dorkenwald, S., Matsliah, A., Sterling, A. R., Schlegel, P., Yu, S.-C., McKellar, C. E., ... & Murthy, M. (2024). Neuronal wiring diagram of an adult brain. Nature, 634(8032), 124–138. https://doi.org/10.1038/s41586-024-07558-y
4. Shiu, P. K., Sterne, G. R., Spiller, N., Franconville, R., Sandoval, A., Zhou, J., ... & Scott, K. (2024). A Drosophila computational brain model reveals sensorimotor processing. Nature, 634(8032), 210–219. https://doi.org/10.1038/s41586-024-07763-9
5. Shapson-Coe, A., Januszewski, M., Berger, D. R., Pope, A., Wu, Y., Blakely, T., ... & Lichtman, J. W. (2024). A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution. Science, 384(6696), eadk4858. https://doi.org/10.1126/science.adk4858