The secret to human cognition might lie in the complex computing power of individual brain cells

A recent study published in the Proceedings of the National Academy of Sciences suggests that individual cells in the human brain possess significantly greater computational power than those found in other mammals. By applying artificial intelligence to model these brain cells, scientists found that human neurons are highly sophisticated information-processing units on their own. These findings provide evidence that the unique cognitive abilities of humans might stem from the complex structure and function of individual cells, rather than just the vast number of cells in the brain network.

The brain is composed of billions of individual cells called neurons, which communicate with one another to process information. Most of the advanced cognitive functions in humans, such as language and problem-solving, take place in the cerebral cortex. This is the wrinkled outer layer of the brain. Within the cortex, the primary cells responsible for transmitting excitatory signals are called pyramidal neurons. These cells are named for their distinctive cone-shaped cell bodies.

A neuron receives incoming electrical signals through branch-like structures called dendrites. The signals travel down the dendrites to the main cell body. If the combined signals reach a certain threshold, the neuron fires an electrical pulse, known as a spike or action potential, to pass the message along to other cells. The way a neuron integrates these incoming signals and decides whether or not to fire is essentially a form of microscopic computation. These tiny cellular decisions form the biological basis for all human thought and behavior.

Previous anatomical observations have shown that human cortical pyramidal neurons look physically different from those of rodents. Human neurons tend to be larger, with much more extensive and elaborately branched dendritic trees. However, scientists lacked a standardized way to measure exactly how these physical differences affect the cell’s ability to process information.

The research team, led by scientists at the Hebrew University of Jerusalem and the Vrije Universiteit Amsterdam, aimed to measure the functional complexity of these microscopic brain cells. They sought to determine whether the unique physical traits of human neurons actually translate into greater computational power compared to the neurons of a rat. To do this, they needed a tool that could quantify how well a neuron translates multiple incoming signals into a single outgoing spike.

Free daily newsletter

To solve this problem, the scientists developed a new metric called the Functional Complexity Index. This index relies on machine learning concepts. While machine learning is often used to find patterns in massive consumer datasets, here it is used as a ruler to measure biological complexity. The core idea is to train a standard artificial neural network to mimic the behavior of a biological brain cell. An artificial neural network is a computer system designed to recognize patterns, loosely inspired by the structure of the brain.

The researchers reasoned that if a biological neuron acts like a simple switch, a small artificial network will easily learn to predict its behavior. If the biological neuron performs highly complex computations, the same artificial network will struggle to replicate its output. A worse performance by the artificial network results in a higher Functional Complexity Index score for the biological cell.

The scientists conducted their experiments using detailed digital models of biological neurons. They utilized three-dimensional reconstructions of 24 specific cells. This sample included 12 human cortical pyramidal neurons and 12 rat cortical pyramidal neurons. The selected cells represented different depths of the brain’s cortex, specifically spanning layers two, three, four, five, and six.

For each of the 24 digital neurons, the researchers generated an enormous dataset. They ran simulations exposing the digital cell to random incoming electrical signals spread across its dendritic branches. Each simulation lasted ten seconds, and they ran 12,000 simulations per neuron. This generated the equivalent of over a day of continuous neural activity data for each individual cell model.

Next, the researchers built a standard artificial neural network featuring three internal processing layers, each containing 128 computational units. They fed this network the exact same incoming signals used in the simulations. The artificial network was then tasked with predicting the exact millisecond timing of the electrical spikes that the biological cell model had produced.

The authors found that human cortical neurons scored significantly higher on the complexity index than their rat counterparts. The artificial network had a much harder time predicting the spike timing of the human cells. This suggests that human neurons perform a much more complex translation of incoming signals than the neurons of a rat.

To understand what drives this difference, the team analyzed 58 separate physical measurements of the cells’ dendritic branches. They found that the total surface area of the dendrites was the single strongest predictor of a cell’s complexity score. The length of the branches that split off into other branches was also a major factor. This provides evidence that a larger, more sprawling dendritic structure allows different parts of the cell to process information somewhat independently, which greatly increases overall computational power.

The researchers also investigated the role of synapses, which are the tiny connection points where signals enter the dendrites. Specifically, they looked at NMDA receptors. These are specialized proteins located at the synapses that respond to incoming electrical signals in a non-linear way. This means that if enough signals arrive at once, the NMDA receptors amplify the electrical current dramatically, rather than just adding the signals together simply.

In their digital simulations, the team tested both rat-like and human-like synaptic properties. Scientific evidence suggests that human excitatory synapses contain a larger number of NMDA receptors and react more sharply to voltage changes. When the researchers applied these human-like synaptic traits to the models, the functional complexity of the cells increased significantly. The combination of sprawling dendrites and highly reactive NMDA receptors tends to push the human neuron into a much higher tier of processing power.

The data also revealed an interesting shift in how complexity is distributed across the layers of the cortex. In the rat models, the neurons located in layer five were the most complex. In the human models, the neurons in layers two and three were significantly more complex than those in other layers. Layers two and three are known to be particularly expanded in the human brain, which suggests an evolutionary adaptation in how the human brain allocates its computational resources.

While the study provides a detailed look at single-cell computation, it does have a few limitations. The research relied entirely on computer simulations of neurons rather than living tissue in an active brain. Because there is currently a lack of experimental data regarding certain electrical properties in human dendrites, the models did not include every possible active ion channel found in a living cell. This means the digital cells might behave slightly differently than biological cells in a real human brain.

Additionally, the Functional Complexity Index is heavily dependent on the specific design of the artificial neural network used for the testing. If the artificial network is too shallow or too deep, it can compress the differences in scores between the cells. The researchers selected a three-layer network as a middle ground, but different computer architectures could yield different specific numbers.

Future research directions might involve exploring other anatomical features, such as the tiny protrusions on dendrites known as spines, to see how they alter signal processing. The researchers also hope to apply this new measurement tool to other types of brain cells and to different species, such as nonhuman primates. Eventually, gathering data from living human brain cells in a laboratory setting could help scientists verify the computational patterns observed in these digital simulations.

The study, “Dendritic morphology and synaptic nonlinearities enhance functional complexity in human cortical neurons,” was authored by Ido Aizenbud, Daniela Yoeli, David Beniaguev, Christiaan P. J. de Kock, Michael London, and Idan Segev.