The standard neuron model used in large-scale brain simulation treats each neuron as a single integrating compartment. Inputs arrive, get summed, and when the total crosses a threshold, an action potential fires. This approximation has been computationally useful for decades and captures enough biology to reproduce many network-level phenomena. A paper in Science from March 2026 now presents evidence that this simplification misses something essential: different parts of the same neuron learn by different rules.

The research, conducted in motor cortex during active skill learning, shows that apical dendrites and basal dendrites operate under distinct plasticity mechanisms. Inputs to apical dendrites are strengthened when local dendritic activity coincides with other signals at the apical compartment. Inputs to basal dendrites follow a more conventional rule driven by postsynaptic spiking at the soma. These are not two variants of the same mechanism. They are computationally distinct operations happening simultaneously in the same cell.

For whole brain emulation, this finding imposes a requirement that current mainstream neuron models do not meet: a single simulated neuron must contain at least two spatially separated computational zones, each with its own learning rule and its own input space. The Izhikevich spiking neuron model and related reduced-form models used in large-scale simulations lack this architecture.

The Experiment

The researchers recorded from motor cortex neurons in mice learning a skilled forelimb task. Two-photon calcium imaging allowed simultaneous monitoring of apical dendritic activity and somatic spiking in the same neurons over multiple learning sessions. This is technically demanding: apical dendrites of cortical pyramidal neurons extend to the superficial layers of cortex while their cell bodies sit deeper, requiring imaging across different focal planes in the same experiment.

During learning trials, synaptic inputs to apical dendrites were strengthened when dendritic calcium events, a signature of apical dendritic activation, coincided with those inputs. The timing window for this strengthening was measured in hundreds of milliseconds, not the sub-millisecond windows associated with spike-timing-dependent plasticity. Apical inputs that were active without coincident dendritic calcium signals were not reliably strengthened.

Basal dendritic synapses followed a different rule. Their strengthening correlated with postsynaptic action potentials at the soma, the conventional association between presynaptic input and output firing. Two sets of inputs to the same neuron, during the same behavioral task, were changing their weights according to different criteria.

Why Compartment-Specific Plasticity Exists

The functional division between apical and basal dendrites reflects a broader organizational principle in cortical circuits. Basal dendrites primarily receive local inputs: recurrent connections from nearby neurons, feedforward inputs from subcortical structures, and inputs from the same cortical area. Apical dendrites receive long-range feedback signals: projections from higher cortical areas, neuromodulatory inputs, and contextual signals that reflect the animal’s current behavioral state.

The two compartments are, in effect, integrating two different classes of information simultaneously. The somatic plasticity rule strengthens connections from whichever local inputs consistently predict firing. The apical plasticity rule strengthens connections that arrive in the context of top-down feedback signals.

This architecture implements something more sophisticated than Hebbian learning: it allows a neuron to strengthen different inputs selectively depending on whether they are providing bottom-up sensory drive or top-down contextual modulation. The distinction matters for understanding how motor skills are acquired and retained, and it matters for understanding what a neuron is actually computing.

Comparison to Behavioral Timescale Synaptic Plasticity

The March 2026 Science finding parallels the concurrent discovery about behavioral timescale synaptic plasticity in hippocampus. Both results implicate dendritic calcium events as drivers of synaptic change over behavioral timescales. Both challenge the assumption that spike coincidence is the primary currency of synaptic learning.

The hippocampal BTSP finding focuses on CA1 pyramidal neurons and episodic memory encoding. The motor cortex finding focuses on the broader question of compartment specificity across the entire dendritic tree. Together, they build a picture in which dendrites are not passive conductors but active computational elements that implement distinct algorithms depending on their location in the neuron’s geometry.

The Markram neocortical microcircuitry framework introduced the concept of detailed biophysical modeling of cortical neurons and circuits. Markram’s Blue Brain Project used multi-compartment models with realistic dendritic morphology, which was computationally expensive but biologically necessary for capturing the range of electrophysiological behaviors observed in real neurons. The March 2026 finding provides additional justification for that level of anatomical detail: compartment-specific plasticity cannot be replicated without a model that distinguishes apical from basal dendrites.

Consequences for Large-Scale Brain Emulation

The scale tension in whole brain emulation is real. Simulating 9 million neurons at the Allen Institute required a supercomputer with substantial resources and still used simplified neuron models for tractability. Implementing full multi-compartment models with biophysically realistic dendrites for billions of neurons pushes computational requirements further.

The Science finding does not resolve this tension, but it sharpens the question. If the goal is a faithful emulation that reproduces motor learning, the simulated neurons must implement compartment-specific plasticity. If they do not, the emulated motor cortex will learn through a mechanism the biological system does not use, and the resulting motor behavior will diverge from what the original brain would have produced.

The deep learning cortical circuit simulation from March 2026 demonstrated that biophysically constrained models can be trained on single GPUs in reasonable time for 67,000-neuron networks. The question is how that approach scales to full cortex while maintaining the dendritic detail required by this finding.

Implications for the Emulation Validation Problem

One practical consequence of compartment-specific plasticity concerns how an emulated brain would be validated. Standard validation approaches compare input-output relationships: given a stimulus, does the emulated network produce the same firing patterns as a biological network?

Compartment-specific plasticity introduces a deeper validation challenge. The same input-output behavior during initial testing can be produced by different underlying weight distributions and learning rule configurations. Two networks that produce identical outputs under test stimuli may diverge significantly when presented with novel experiences that require learning. If the simulated motor cortex uses a single-compartment learning rule, it will initially match the biological network’s behavior but will learn differently in response to new skill demands.

This means behavioral validation of a brain emulation is not sufficient to confirm that the underlying plasticity mechanisms are correct. Structural validation of dendritic morphology and plasticity rule verification become necessary additional checks.

Technology Readiness Level: 2-3 (basic science). Validated in mouse motor cortex during skill learning. Human motor cortex has comparable pyramidal neuron anatomy, making the finding likely to apply, but direct human validation has not been published.

Official Sources

• Science (March 2026): DOI 10.1126/science.ads4706
• https://www.science.org/doi/10.1126/science.ads4706
• Related Blue Brain Project (Markram 2006): DOI 10.1126/science.1127241
• Related BTSP hippocampus (Nature Neuroscience 2026): DOI 10.1038/s41593-026-02214-2