The Language of Dendrites

Animal survival depends on the ability to analyze the environment and act on it: escape predators, find food, select a mate. Understanding how the brain achieves this is one of the most fascinating and challenging problems in neuroscience. What sequence of steps converts sensory cues into behavior? In other words, how does the brain compute? In 1943, McCulloch and Pitts noted that neurons firing action potentials act like binary devices that can either be on or off. In a seminal paper ( 1 ), they showed how connected networks of neurons could represent any logical expression, and 60 years of subsequent work in theoretical neuroscience has devised models of neuronal networks that can implement computational tasks. The basic operation in these models is the conversion of input into action potentials—the process of turning a neuron on. How is this conversion achieved? What are the rules for integrating input, and what kind of information can a single neuron interpret and process? The traditional view is that neurons sum input and, if the sum reaches a certain threshold, an action potential is triggered. The important variable is thus the amount of synaptic input—if neurons had a language and each synapse were a letter, they would only care about how long a word is. However, most neurons seem to have the ability to be more powerful than this. Contrary to the assumption of McCulloch and Pitts, synapses are not made onto the soma (cell body) but onto dendrites, protrusions from the cell body separating the input and the action potential initiation zone. Dendrites filter, transform, and compute thresholds of synaptic input and can, in theory, implement basic arithmetic operations by themselves ( 2 ). I first became interested in dendrites during my Ph.D. work. Monitoring the properties of single synapses in hippocampal neurons, I found that dendrites can implement a negative feedback that regulates the amount of input each branch receives ( 3 ). Dendrites can thus independently process and regulate input information. Can these properties be used by single neurons to perform high-order computations?