Brain

1.1Overview

Before looking at any structure, it helps to have a sense of scale. Table 1.1 collects some numbers worth remembering. Treat them as orders of magnitude: most are estimates with real uncertainty, and all vary between individuals.

Three things stand out to an engineer. The components are slow: a spike lasts about a millisecond, and even the fastest axons carry it at about the speed of a racing car — a million times slower than a signal in a copper wire. The system is frugal: the whole brain runs on the power of a dim light bulb. And the network is huge but sparse: each neuron connects to thousands of others out of tens of billions. Whatever the brain’s advantages are, they are not raw component speed. They are organization, which is what this chapter is about.

1.2History

For most of history, two views of the brain competed. One held that the brain works as a whole, every part able to do everything. The other held that it is a collection of specialized parts, made of individual cells. The evidence that settled the question also produced the field’s founding principles.

Electricity and chemistry

The idea that nerves work electrically came first. In the 1790s Luigi Galvani showed that frog nerves and muscles not only respond to electricity but produce it themselves. Emil du Bois-Reymond later recorded the electrical activity of active nerves, and in 1850 Hermann von Helmholtz measured how fast a nerve impulse travels: tens of meters per second — fast, but finite. The nerve signal was a physical process that could be timed, not the instantaneous flow of “animal spirits” imagined since antiquity. Pharmacology added chemistry. Claude Bernard, Paul Ehrlich, and John Langley showed that drugs act on cells by binding to specific receptors, the first step toward our understanding of chemical communication between neurons.

Localization

In the early 1800s Franz Joseph Gall argued that the brain is not one organ but many — dozens, each responsible for a mental faculty such as generosity or secretiveness — and that the strength of a faculty could be read from bumps on the skull. Phrenology was wrong in its method and in its list of faculties, but its central claim, that mental functions are localized, was radical. Pierre Flourens tested it in the 1820s by removing parts of animals’ brains. Behavior degraded with how much tissue he removed, not with which part, and he concluded that the cerebral hemispheres work as a whole. This became the aggregate-field view.

Clinical observation slowly overturned it. In 1861 Pierre Paul Broca described a patient who understood language but could barely speak — though he could still whistle and sing a melody, so the muscles of speech worked. After the patient died, Broca found damage to the left frontal lobe, and after seeing similar cases he concluded that “we speak with the left hemisphere.” In the same decade John Hughlings Jackson noticed that some epileptic seizures start in one part of the body and spread to others in a fixed order, and inferred that different parts of the cortex control different parts of the body. In 1870 Gustav Fritsch and Eduard Hitzig showed that electrically stimulating specific spots of a dog’s cortex moves specific limbs. And in 1874 Carl Wernicke described the complementary language disorder — fluent speech that makes little sense, with poor comprehension — caused by damage farther back, in the temporal lobe (Section 1.4).

The neuron

The argument about cells ran in parallel. By 1839 Matthias Schleiden and Theodor Schwann had established that all tissue is made of cells, but nervous tissue seemed to be an exception: under the microscope it looked like a continuous tangle. In 1873 Camillo Golgi discovered a silver stain that fills only a small, random fraction of neurons — but fills each of them completely, so that a single cell stands out with all its branches. Golgi concluded that nerve cells are fused into one continuous network. Santiago Ramón y Cajal, using Golgi’s own stain, concluded the opposite: every neuron is a discrete cell, and neurons contact one another only at specific points. Golgi and Cajal shared the 1906 Nobel Prize and used their lectures to disagree. Cajal was right. In 1907 Ross Harrison watched axons grow out of individual nerve cells kept alive in a dish, and in the 1950s the electron microscope finally showed the narrow gap that separates neurons at their contacts.

Cajal went further. He and Charles Sherrington, who named the synapse in 1897, carried the clinicians’ regional findings down to the level of cells, championing what Kandel calls cellular connectionism: behavior is produced by specific groups of neurons connected in specific ways.

Even so, the aggregate-field view dominated experiments and clinical practice until about 1950. Its champions included the neurologist Henry Head and the physiologist Ivan Pavlov, but the most influential was Karl Lashley. In the 1920s Lashley trained rats to run mazes, removed varying amounts of their cortex, and found — as Flourens had — that performance depended on how much cortex was removed, not where. His theory of mass action held that what matters for learning and memory is the total mass of cortex, not any particular region. The explanation turned out to be the task: a rat solves a maze with vision, touch, smell, and movement all at once, so no single lesion abolishes the skill, but every lesion degrades it. Meanwhile, electrical recording and stimulation in the 1930s and 1940s revealed orderly maps of the body in the cortex of cats, monkeys, and people (Section 1.8). Localization won — with a twist that we return to in Section 1.4.

The milestones behind this chapter, in order:

1.3Components

The nervous system has two divisions. The central nervous system (CNS) is the brain and spinal cord. The peripheral nervous system (PNS) is everything outside them: the nerves and clusters of nerve cells (ganglia) that connect the CNS to the skin, muscles, and internal organs (Section 1.8). Kandel divides the CNS into seven major parts (Figure 1.1). From the bottom up:

• Spinal cord. The lowest part of the CNS, running inside the vertebral column. It receives sensory information from the skin, joints, and muscles of the trunk and limbs, and it contains the motor neurons that command their muscles. It is not just a cable: reflexes, and even the basic rhythm of walking, are generated by circuits inside it. Top to bottom, it has four regions — cervical, thoracic, lumbar, and sacral — made of 8, 12, 5, and 5 segments, each with its own pair of nerve roots on either side.
• Medulla oblongata. The lowest part of the brain, continuous with the spinal cord. It contains centers for vital autonomic (automatic) functions such as breathing, heart rate, blood pressure, and digestion.
• Pons. Above the medulla. It relays information about movement from the cerebral hemispheres to the cerebellum, and it takes part in controlling breathing and sleep.
• Midbrain. Above the pons. It controls many sensory and motor functions, including eye movements and the coordination of visual and auditory reflexes.
• Cerebellum. Behind the pons and medulla, attached to the brain stem by three pairs of thick fiber bundles, the cerebellar peduncles. It modulates the force and range of movements, refines their timing, and is central to learning motor skills. Although it is small, it holds about 80% of the brain’s neurons (Section 1.9).
• Diencephalon. Two structures above the midbrain. The thalamus processes and relays most of the information that reaches the cerebral cortex from the rest of the CNS. The hypothalamus, below it, regulates autonomic, endocrine, and visceral functions — body temperature, hunger, thirst, sleep, and reproduction — partly through the hormones of the pituitary gland and partly through its projections to autonomic centers in the brain stem and spinal cord.
• Cerebrum. The two cerebral hemispheres: the cerebral cortex, the white matter beneath it, and three deep structures. The basal ganglia regulate motor performance, select actions, and learn habits; the hippocampus is essential for forming memories of people, places, things, and events; the amygdala coordinates the autonomic and hormonal responses of emotional states, including memories of threats. A thick bundle of axons, the corpus callosum, connects the two hemispheres.

The medulla, pons, and midbrain together form the brain stem. Besides its own functions, the brain stem carries all the traffic between the brain and the spinal cord, receives sensory information from the head and controls the muscles of the face and neck through the cranial nerves, and contains networks of neurons that regulate arousal and awareness.

The parts differ in how their neurons are arranged. In the brain stem and diencephalon, neurons are grouped in clusters called nuclei, each with its own connections. In the cerebrum and cerebellum, the surface is a folded sheet of neurons arranged in layers — a cortex — with stereotyped patterns of connection.

1.4Cortex

The cerebral cortex, the part of the brain most highly developed in humans, computes our most elaborate abilities: perceiving objects, using language, planning, reasoning. It is a sheet of neurons only 2–4 mm thick — the gray matter — folded so that about 0.2 m² of it fits inside the skull (Figure 1.3). Beneath the sheet lies the white matter: bundles of axons that connect each part of the sheet with other parts and with the structures below. To an engineer the layout is familiar: the processing elements sit in a thin layer, and much of the volume is taken up by the cabling between them.

Folding is how evolution packs a large sheet into a limited space. The ridges of the folds are called gyri and the grooves sulci; about two-thirds of the cortex lies hidden in the sulci, and the deepest grooves are called fissures. Two of them serve as landmarks. The central sulcus runs from the top of each hemisphere down its side, and the lateral (or Sylvian) fissure separates the temporal lobe, below it, from the frontal and parietal lobes above.

These landmarks divide each hemisphere into four lobes, named after the bones of the skull that cover them (Figure 1.4):

• The frontal lobe is concerned with planning future actions, working memory, and the control of movement. The primary motor cortex occupies the precentral gyrus, just in front of the central sulcus.
• The parietal lobe handles somatic sensation (touch, pressure, the position of the body), the body’s image, and its relation to the space around it. The primary somatosensory cortex occupies the postcentral gyrus, just behind the central sulcus.
• The occipital lobe, at the back, is devoted to vision. From there, visual information flows along two main streams: a dorsal stream into the parietal lobe, concerned with where objects are and how to act on them, and a ventral stream into the temporal lobe, concerned with what they are — including whose face you are looking at.
• The temporal lobe is concerned with hearing and — through the hippocampus and amygdala, deep inside it — with learning, memory, and emotion. Its underside also performs high-level vision, such as recognizing objects and faces.

Two more cortical regions are hidden from the outside: the insula, folded inside the lateral fissure, which takes part in emotion, the sense of the body’s internal state, and taste; and the cingulate cortex, on the inner surface of each hemisphere above the corpus callosum, which takes part in emotion, pain, and the control of behavior.

Three facts of organization apply throughout. Each hemisphere is concerned mostly with the opposite side of the body: the left hemisphere feels and moves the right hand, because the pathways cross the midline on the way (Section 1.8). The two hemispheres are neither perfectly symmetrical in structure nor equivalent in function. And each lobe contains several specialized areas: primary areas are the first cortical stop of a sense or the last cortical stage of motor output, and the association areas around them combine information into more abstract representations.

Language: the first evidence for localization

Language provided the first strong evidence that specific mental abilities depend on specific regions, through the study of aphasias — disorders of language caused by brain damage.

Broca’s patients had what is now called Broca’s aphasia (expressive, or nonfluent, aphasia): they understood language but spoke slowly and with great effort, in short, ungrammatical phrases. The damage was in the lower rear part of the left frontal lobe, in Broca’s area, next to the motor cortex that controls the face and mouth. Wernicke’s patients showed the opposite pattern, Wernicke’s aphasia (receptive, or fluent, aphasia): they spoke fluently, but their speech made little sense and they could not understand what they heard. The damage was at the back of the left temporal lobe, in what is now Wernicke’s area. Together, the two syndromes form a double dissociation — some patients can understand but not speak, others can speak but not understand — strong evidence that the two abilities depend on separate components.

Wernicke went beyond localization: he proposed how the components work together. Simple perceptual and motor functions, he argued, are localized in single areas, but complex abilities arise from the interconnections between several functional sites. In his model, as later extended by Norman Geschwind, a heard word is processed by the auditory cortex, recognized in Wernicke’s area, and sent through a bundle of axons, the arcuate fasciculus, to Broca’s area, which turns it into a program of articulation that the motor cortex executes. A written word enters by a different route, proposed by Jules Dejerine in the 1890s: from the visual cortex through the angular gyrus, where the visual and auditory forms of a word are associated, so that written and spoken words converge on a common code (Figure 1.5).

The model made a prediction: damage to the connection, sparing both areas, should spare comprehension and articulation but break the link between them. That disorder exists. Patients with conduction aphasia understand what they hear and read, and have no trouble moving the muscles of speech. Yet their speech is full of errors — dropped syllables, wrong sounds — they struggle to repeat words and phrases verbatim, and their verbal working memory is severely limited.

The Wernicke–Geschwind model is now seen as a simplification. Language engages a much broader network that includes more of the frontal, temporal, and parietal cortex and structures beneath it, and imaging has revised the nodes themselves: Broca’s area takes part in comprehension as well as speech, and reading also relies on a specialized region on the underside of the occipital and temporal lobes. But its central idea — that a complex ability is a network of specialized components, and that a broken connection is a failure of its own kind — became a pillar of neuroscience, now called distributed processing.

What is localized

If the language areas were specialized for sound, sign languages, which are visual and spatial, should depend on other regions — perhaps on the right hemisphere, which dominates spatial processing. They do not. Deaf people who use sign language develop sign-language aphasias after damage to the left hemisphere, and these resemble spoken-language aphasias: damage near Broca’s area impairs producing signs, and damage near Wernicke’s area impairs understanding them. Damage to the right hemisphere impairs their spatial abilities but not their signing.⁵5 Hickok G, Bellugi U, Klima ES (1996). The neurobiology of sign language and its implications for the neural basis of language. Nature 381:699–702. The left-hemisphere language areas are specialized for language itself, whatever its sensory form.

Imaging refines the picture. When people who learned a second language as adults speak each of their languages, the two activate adjacent but separate regions within Broca’s area; in people who learned both languages in infancy, the regions overlap. Wernicke’s area shows overlap in both groups.⁶6 Kim KHS, Relkin NR, Lee K-M, Hirsch J (1997). Distinct cortical areas associated with native and second languages. Nature 388:171–174.

Nor is all of language left-sided. The emotional tone of speech, its prosody, depends mostly on the right hemisphere, in regions that mirror the classic language areas. Damage to the right-hemisphere counterpart of Wernicke’s area impairs recognizing the emotion in other people’s voices; damage to the counterpart of Broca’s area produces flat, emotionless speech.⁷7 Ross ED (1981). The aprosodias: functional-anatomic organization of the affective components of language in the right hemisphere. Archives of Neurology 38:561–569.

So is function localized or distributed? Both, at different scales. The specialized units are not whole faculties such as “language” or “memory,” as phrenology imagined; they are elementary operations. A faculty is assembled from many such operations, carried out in many regions, working together.

The brain also carves up tasks differently than introspection suggests. What you know about an apple — how it looks, how you hold it, what it is called — is stored in separate regions and bound together when you recall it, and the color and the motion of a single object are analyzed in different pathways. Experience feels unified; the machinery behind it is not.

Two hemispheres, one mind

Distributed processing raises a puzzle: if a behavior is computed in many regions working in parallel, how do the pieces come together into a single experience? A striking clue comes from patients whose corpus callosum — some 200 million axons linking the hemispheres — was cut to stop the spread of severe epileptic seizures. In the 1960s Roger Sperry, Michael Gazzaniga, and Joseph Bogen showed that in these split-brain patients each hemisphere can perceive, learn, and remember on its own, largely unaware of the other.⁸8 Gazzaniga MS (2005). Forty-five years of split-brain research and still going strong. Nature Reviews Neuroscience 6:653–659. A word shown only to the right hemisphere (by flashing it in the left half of the visual field) cannot be read aloud, because speech depends on the left hemisphere; yet the left hand, which the right hemisphere controls, can pick out the named object by touch. Asked why the hand chose that object, the speaking left hemisphere, which never saw the word, confidently invents a reason. Occasionally the two hemispheres act at cross purposes, as when one hand buttons a shirt while the other unbuttons it. Our sense of a single, unified mind depends on continuous communication between the parts.

Split-brain patients also show how much of the brain’s work never reaches awareness. Most of its processing is unconscious, and Kandel warns that describing perception and reasoning in terms of conscious experience can mislead: we perceive an object less as a picture than as an opportunity — something we might eat, sit on, or pick up. One consequence is encouraging for science. If much of cognition does not depend on awareness, then studying simpler behaviors in animals can illuminate it.

1.5Neurons

The brain has two classes of cells: neurons (nerve cells), which carry signals, and glial cells (glia), which support them in many ways. In most respects a neuron is an ordinary cell: it has a nucleus that holds its genes and machinery that makes its proteins and energy. What sets it apart is its shape and its membrane. A neuron is polarized — it has a distinct input end and output end — and its membrane can generate and conduct electrical signals.

A typical neuron has four regions, each with its own job (Figure 1.6):

• Cell body (soma). The metabolic center. It contains the nucleus and the endoplasmic reticulum, where the cell’s proteins are made. Two kinds of processes grow out of it: several short dendrites and a single long axon.
• Dendrites. Branching like a tree (Greek déndron), they are the neuron’s main receiving surface: most of the synapses from other neurons land on them.
• Axon. The output cable. It begins at a tapered region of the cell body, the axon hillock, followed by the initial segment, where outgoing signals are triggered. It carries action potentials — brief electrical impulses — over distances from a fraction of a millimeter to more than a meter. Many axons are wrapped in myelin, an insulating sheath made by glial cells, which is interrupted at regular intervals by short gaps called nodes of Ranvier. The action potential is regenerated at each node as it travels.
• Presynaptic terminals. Near its end, the axon divides into fine branches that contact other cells at synapses. The transmitting cell is called presynaptic and the receiving cell postsynaptic; between them is a gap of 20–40 nanometers, the synaptic cleft. Most terminals contact the dendrites or the cell body of another neuron, and a few its axon; others contact muscle or gland cells. Some neurons, called neuroendocrine cells, release their transmitter into the bloodstream, where it acts as a hormone.

From the shapes of neurons, Cajal deduced two more principles. Studying circuits in the retina, the olfactory bulb, the spinal cord, and elsewhere, he noticed that dendrites always face the direction from which signals arrive and axons the direction in which they leave.

Kinds of neurons

The classic way to classify neurons is by shape — specifically, by the number of processes that leave the cell body (Figure 1.7):

• Unipolar neurons have a single process that branches into receiving segments (dendrites) and a transmitting segment (axon). They predominate in invertebrate nervous systems; in vertebrates they occur in the autonomic nervous system.
• Bipolar neurons have two processes: a dendrite that receives signals from other cells and an axon that carries them toward the CNS. Many sensory cells are bipolar, including those of the retina and the sensory neurons of the nose.
• Pseudo-unipolar neurons start out as bipolar cells during development, but their two processes fuse into a single stem that splits into two long branches, both of which work as axons: one runs to the skin, a joint, or a muscle; the other runs into the spinal cord. These sensory neurons, whose cell bodies lie in the dorsal root ganglia beside the spinal cord, carry touch, pressure, body position, and pain. Their signals can travel from one branch to the other without passing through the cell body.
• Multipolar neurons have one axon and many dendrites, and they make up most of the vertebrate nervous system. Their shapes vary with the number of inputs they integrate. A spinal motor neuron, with a relatively modest dendritic tree, receives about 10,000 synapses — some 1,000 on its cell body and 9,000 on its dendrites. A pyramidal cell of the cortex or hippocampus, named for the shape of its cell body, has two dendritic trees, one at its apex and one at its base; a single pyramidal cell in the hippocampus receives about 30,000 synapses.⁹9 Megías M, Emri Z, Freund TF, Gulyás AI (2001). Total number and distribution of inhibitory and excitatory synapses on hippocampal CA1 pyramidal cells. Neuroscience 102:527–540. A Purkinje cell of the cerebellum spreads a flat, fan-shaped dendritic tree that receives as many as a million synapses, by Kandel’s estimate; careful counts in the rat find about 175,000 from a single class of input alone.¹⁰10 Napper RMA, Harvey RJ (1988). Number of parallel fiber synapses on an individual Purkinje cell in the cerebellum of the rat. Journal of Comparative Neurology 274:168–177.

Neurons can also be classified by their role. Sensory (afferent) neurons carry information from the body’s periphery into the CNS. Motor (efferent) neurons carry commands out to muscles and glands. Interneurons, by far the most numerous, do everything in between: projection (or relay) interneurons have long axons that connect distant regions, and local interneurons have short axons that work within a small circuit. Today neurons are increasingly classified by the genes they express, which has revealed thousands of types.

Glia

Glia are often said to vastly outnumber neurons — Kandel gives 2 to 10 glial cells per neuron in the vertebrate CNS — but careful counts of the human brain find roughly one per neuron.³3 von Bartheld CS, Bahney J, Herculano-Houzel S (2016). The search for true numbers of neurons and glial cells in the human brain: a review of 150 years of cell counting. Journal of Comparative Neurology 524:3865–3895. Glia are not electrically excitable and do not carry fast signals, yet the brain cannot work without them. They also guide wiring during development and help stabilize connections changed by learning; once seen as mere support, they are now regarded as partners of neurons. They fall into two classes. Microglia are the brain’s resident immune cells: they respond to injury and infection, and during development they help remove unneeded synapses. Macroglia comprise three main types (Figure 1.8):

• Oligodendrocytes, in the CNS, and Schwann cells, in the PNS, make myelin. An oligodendrocyte extends many arms, each wrapping a segment of a different axon; a Schwann cell wraps a single segment of a single axon. Diseases that destroy myelin — multiple sclerosis in the CNS, Guillain–Barré syndrome in the PNS — slow or block conduction.
• Astrocytes, star-shaped cells, contact both blood vessels and synapses. They wrap synapses and clear away the neurotransmitter released there, which keeps neighboring synapses isolated from one another; absorb the potassium that active neurons release; supply neurons with fuel and release growth factors; help match blood flow to neural activity; and help regulate how synapses form and work.

1.6Signaling

Almost every neuron, whatever its shape or chemistry, signals with the same sequence of events. Understanding it is the key to reading any circuit.

At rest, the inside of a neuron is about 65 millivolts more negative than the outside. This resting membrane potential is maintained by pumps that keep sodium and potassium ions at different concentrations inside and outside the cell, and by channels that let potassium leak out (the next chapter explains the mechanism). A signal is a change in this potential. Making the inside less negative, depolarization, brings a neuron closer to firing; making it more negative, hyperpolarization, moves it away.

Four components, four signals

Kandel’s model of the neuron has four functional components, each producing a different kind of signal (Figure 1.9):

1. Input. A stimulus, for a sensory neuron, or a neurotransmitter, for any other neuron, opens ion channels and produces a small, local change in the membrane potential: a receptor potential or a synaptic potential. These signals are graded — their amplitude and duration track the strength and duration of their cause — and they spread passively, fading within a few millimeters at most.
2. Integration. The input signals spread to the trigger zone at the axon’s initial segment, where they add up. The trigger zone has the highest density of the voltage-gated sodium channels that generate action potentials, and it makes the neuron’s one decision: if the summed potential crosses a threshold, typically about 10 mV above rest, the neuron fires.
3. Conduction. The action potential is a stereotyped pulse of about 100 mV that lasts about a millisecond. It is produced by positive feedback: depolarization opens sodium channels, sodium ions rush in and depolarize the membrane further, and the potential swings explosively from about −65 to about +40 mV before other channels restore it. It is all-or-none: a stimulus below threshold produces none, and every stimulus above threshold produces the same full-size pulse. It regenerates itself as it travels, so it reaches the end of the axon undiminished, at 1 to 100 meters per second. Because every pulse is the same, information is carried by how many pulses there are and when they occur: a stronger stimulus produces a higher firing rate, and a longer stimulus a longer train.
4. Output. At the presynaptic terminals, each action potential triggers the release of neurotransmitter, packaged in tiny synaptic vesicles. The amount released depends on the number and frequency of the arriving action potentials, so the signal becomes graded again — and starts the sequence over in the next cell.

Not every neuron uses all four components. Some local interneurons, with no axon or only a very short one, never fire action potentials: their graded potentials spread passively to their release sites.

The meaning is in the pathway

Recording from single sensory fibers in the 1920s, Edgar Adrian noticed something puzzling: action potentials look the same in every nerve. A spike from the eye, a spike from the ear, and a spike commanding a muscle to contract are virtually indistinguishable. How, then, does the brain know whether a spike means light or sound? By its pathway. The meaning of a signal is set by where the fiber that carries it comes from and where it goes.

You can check this yourself. Press gently on the corner of your closed eye and you will see a spot of light: the pressure excites the retina, and because the signals arrive along the visual pathway, they are experienced as light. Cochlear implants exploit the same principle: electrically stimulating fibers of the auditory nerve produces sound, and stimulating different positions along the cochlea produces different pitches.

Same input, different output

Neurons differ in how they respond to the same input. Many are silent until they are driven. Others fire spontaneously, with no input at all: regularly — beating, or pacemaker, neurons — or in brief, repeated bursts — bursting neurons. The same excitatory input may make a silent neuron fire a spike or two but only speed up one that is already active. Inhibition differs even more: it tells a silent cell little, but in an active cell it can carve periods of silence into ongoing firing, creating a pattern where none existed. These differences come from the particular mix of ion channels each neuron expresses.

Synapses

Most synapses are chemical: the presynaptic terminal releases a neurotransmitter that diffuses across the cleft and binds receptors on the postsynaptic cell. A minority are electrical: protein channels called gap junctions connect the two cells directly, so current flows between them almost instantly, which helps groups of neurons fire in synchrony. Chemical synapses are slower but far more versatile. They can amplify or invert a signal, they can be modulated, and their strength can change.

Whether a chemical synapse excites or inhibits its target is decided by the postsynaptic receptor, not by the transmitter. The same molecule can do both: acetylcholine excites skeletal muscle through one kind of receptor and slows the heart through another. In practice, the brain’s main excitatory transmitter is glutamate and its main inhibitory transmitter is GABA. A neuron receiving both kinds of input sums them at its trigger zone, so inhibition can veto excitation.

Neurons differ most in their molecules

Neurons that look alike can be very different molecularly: they use different transmitters, receptors, and ion channels. This molecular diversity is why drugs and diseases can be so selective. In Parkinson’s disease, for example, a small population of dopamine-releasing neurons in the midbrain degenerates and movement suffers; supplying the missing dopamine’s chemical precursor, L-DOPA, can relieve the symptoms for years. Other diseases strike a single class of cells: amyotrophic lateral sclerosis (ALS) and polio destroy motor neurons, while in myasthenia gravis the immune system attacks the acetylcholine receptors through which motor neurons excite muscle.

1.7Circuits

Neurons never act alone. Each is part of a circuit, and what a neuron does for behavior is determined by its connections. Kandel counts this among the key organizational principles of the brain: nerve cells with similar properties can produce very different actions because of the way they are interconnected. The simplest circuits show how the pieces fit together.

The knee-jerk reflex

When a doctor taps the tendon below your kneecap, your leg kicks. The tap stretches the quadriceps, the muscle at the front of the thigh, and activates stretch receptors inside it, the muscle spindles. Sensory neurons carry the signal into the spinal cord (Figure 1.10). In cross-section the spinal cord is the cortex turned inside out: a butterfly-shaped core of gray matter, holding cell bodies, surrounded by white matter, made of axons running up and down. Sensory axons enter through the dorsal roots; motor neurons sit in the ventral horns of the gray matter and send their axons out through the ventral roots. Inside the cord, the reflex circuit does three things at once:

• Each sensory neuron excites the motor neurons of the quadriceps directly, through a single synapse, and the muscle contracts. This monosynaptic reflex is the simplest circuit in the body.
• The same sensory neurons excite inhibitory interneurons that suppress the motor neurons of the opposing muscle, the hamstring at the back of the thigh, so that it relaxes instead of resisting. The reflex would work without them, but less stably.
• Other branches carry information about the reflex up to higher centers, so that the brain can coordinate it with other behaviors, at the same time or in sequence.

The reflex shows wiring patterns that recur everywhere. Stretching the quadriceps activates several hundred sensory neurons, and each contacts 45 to 50 motor neurons — divergence, typical of the input stages of the nervous system. Each motor neuron, in turn, collects 200 to 450 contacts from about 130 sensory neurons — convergence, typical of its output stages. No single input is strong enough to fire a motor neuron; it fires only when many sensory neurons are active together. The inhibitory interneuron implements feed-forward inhibition: the signal that excites one pathway also inhibits its competitor, a pattern known in the spinal cord as reciprocal innervation. The opposite arrangement, feedback inhibition, has a neuron excite an interneuron that inhibits the neuron itself, limiting its own activity; spinal motor neurons do this through interneurons called Renshaw cells.

Circuit motifs

Larger circuits reuse a small set of wiring patterns, or motifs (Figure 1.11). Divergence broadcasts one signal to many targets, and convergence combines many signals into one decision. Feed-forward excitation relays information from stage to stage, building hierarchies (Section 1.9). Feed-forward inhibition sharpens timing and arbitrates between competing actions. Feedback inhibition stabilizes activity and controls gain. Recurrent excitation — neurons exciting each other — amplifies signals and can sustain activity after its input has ended, a basis for short-term memory.

Because a neuron can receive excitation from hundreds or thousands of synapses, circuits must keep excitation in check with inhibition. Many operate near a balance between the two, which may increase their computing power but leaves them prone to runaway activity: an epileptic seizure is excitation escaping control.

Complex behaviors are built from the same components — sensory neurons, interneurons, motor neurons — and the same motifs, but with many more stages between input and output, many parallel paths, and loops that run back from later stages to earlier ones. Kandel suggests that the reflex’s basic plan often survives. Just as the firing rate of a stretch receptor encodes how far a muscle is stretched, the firing rates of neurons in sensory cortex encode the intensity of a feature, those in motor cortex shape the movement that follows, and those in association cortex encode quantities such as the evidence for a choice. What thought adds are two elaborations. One is flexible routing: depending on context and goals, the same input can lead to many actions, and the same action can follow from many inputs. The other is freedom from immediacy. A reflex acts as soon as its signals arrive — the ankle jerk lags the knee jerk only because its nerves are longer — while neurons in association cortex can sustain graded firing for many seconds, so that action can wait for more evidence or the right moment (Section 1.9).

Circuits are modified by experience

Circuits are not fixed, which raises a puzzle: if the connections between neurons are set so precisely during development, how can learning change behavior? The answer that has proved most farsighted is the plasticity hypothesis, proposed by Cajal around 1900 and given its modern form by the Polish psychologist Jerzy Konorski in 1948: synapses change with use. The strength of a synapse can change for seconds to hours, through modifications of proteins already present at the synapse, and with repeated experience the change can last days or longer, as neurons switch on new genes and grow new synaptic connections or prune old ones. Kandel’s own work on the marine snail Aplysia showed how a simple reflex, the withdrawal of its gill, weakens with repetition (habituation) and strengthens after a noxious stimulus (sensitization) through exactly such changes at identified synapses. In 1949 Donald Hebb proposed a rule for which synapses should change (Section 1.9).

1.8Systems

Circuits are assembled into systems that carry information from the senses to the cortex and from the cortex to the muscles. Following two pathways — touch in and movement out — introduces most of the anatomy needed in the rest of the book (Figure 1.12).

Touch: from skin to cortex

Touch the tip of your finger. Receptors in the skin excite the peripheral branch of a pseudo-unipolar sensory neuron whose cell body lies in a dorsal root ganglion next to the spinal cord. The neuron’s central branch enters the spinal cord through a dorsal root and divides. Local branches end in the spinal gray matter, where they feed reflex circuits; the main branch, without making a synapse, ascends on the same side, in the dorsal columns of the white matter, all the way to the medulla. There it excites neurons of the dorsal column nuclei, whose axons cross the midline and ascend through the brain stem as the medial lemniscus to the thalamus. Thalamic neurons, in turn, project to the primary somatosensory cortex in the postcentral gyrus.

Three features of this pathway recur throughout the nervous system:

• It is a chain of relays, and the information is transformed at each synapse, not merely passed along. In Kandel’s phrase, sensation is an abstraction, not a replication, of reality.
• It is mapped. Neighboring points on the skin are represented by neighboring neurons at every stage, so each relay holds a map of the body surface (it is somatotopic). In the cortex, the map is distorted in proportion to sensory acuity: the fingers and lips occupy far more cortex than the back, a distortion depicted by Wilder Penfield’s homunculus. The map is not fixed: in violinists, the area devoted to the fingers that press the strings is enlarged.
• It crosses the midline — in the medulla, for this pathway — which is why the left hemisphere feels the right hand. Not every pathway crosses completely: only about half the fibers of each optic nerve cross, hearing reaches both hemispheres, and smell stays mostly on the same side.

Different kinds of bodily sensation travel in separate, parallel pathways. Pain and temperature, for example, ascend in the anterolateral system, which crosses the midline within the spinal cord rather than in the medulla.

The thalamus, gateway to the cortex

Except for smell, all sensory information reaches the cortex through the thalamus, a cluster of as many as fifty nuclei. Each sense has its own relay nuclei — the lateral geniculate nucleus for vision, the medial geniculate nucleus for hearing, and the ventral posterior nuclei for touch — while others relay the output of the cerebellum and basal ganglia to the motor cortex, form loops with emotional and association areas, or regulate arousal. The thalamus is not a passive relay. In the lateral geniculate nucleus, for example, axons returning from the cortex make more synapses on the relay neurons than the retina does, although the retina still drives them. And a thin shell of inhibitory neurons surrounding the thalamus, the reticular nucleus, samples the traffic between thalamus and cortex and inhibits the relay cells, gating what gets through — a mechanism implicated in attention and sleep.

The cortical sheet

Most of the cortex is neocortex, which has six layers, numbered I at the surface to VI at the bottom. The layers differ in their cell types and connections, in a pattern repeated across areas (Figure 1.13): input from the thalamus arrives mainly in layer IV; layers II and III connect to other cortical areas; layer V sends the cortex’s output to structures below it — the basal ganglia, brain stem, and spinal cord; and layer VI projects back to the thalamus. Layer I contains mostly dendrites and axons. Perpendicular to the layers, neurons are organized in columns, each about a third of a millimeter wide, that share response properties: in the somatosensory cortex, the neurons of a column respond to the same kind of stimulus at the same place on the body.

The relative thickness of the layers varies from area to area, and the variation follows function. Layer IV, the input layer, is so thick in the primary visual cortex that it is divided into sublayers, and it is essentially absent from the primary motor cortex, whose output layer V is prominent instead. In 1909 Korbinian Brodmann used this variation — the cytoarchitecture of the cortex — to divide it into about fifty areas, whose numbers are still in use: the primary visual cortex is area 17 and the primary motor cortex area 4. Functionally, the areas form a hierarchy. Primary sensory areas feed unimodal association areas, which elaborate a single sense, and these feed multimodal association areas, in the prefrontal, parietal–temporal, and limbic cortex, which combine the senses and support planning, language, and memory. Connections usually run in both directions, and the layers tell them apart: forward (feed-forward) projections leave mainly from layer III and arrive in layer IV of the next area, while backward (feedback) projections leave from layers V and VI and end outside layer IV.

Movement: from cortex to muscle

The command for a voluntary movement leaves the primary motor cortex, in the precentral gyrus, along the axons of large pyramidal neurons in layer V. They descend through the internal capsule, the base of the midbrain, and the pons; gather into the pyramids on the front of the medulla; and most of them cross the midline at the pyramidal decussation, where the medulla meets the spinal cord. They continue down the corticospinal tract, about a million axons strong, to interneurons and to the motor neurons of the ventral horn. Because the pathway crosses, the left motor cortex moves the right side of the body. Two great loops refine the commands: the cerebellum compares intended and actual movements, and the basal ganglia help select which actions to perform. Both act largely through the thalamus, back onto the cortex.

Modulatory systems

Alongside these point-to-point pathways, the brain has broadcast systems (Figure 1.14). Small clusters of neurons in the brain stem and the base of the forebrain send highly branched axons over vast territories of the brain and release neuromodulators, which change how their targets respond rather than carrying specific messages:

• Dopamine neurons, in the substantia nigra and the ventral tegmental area of the midbrain, project to the striatum (the input stage of the basal ganglia) and to the frontal cortex. They are involved in movement, motivation, and learning from reward. Rats allowed to stimulate these reward circuits by pressing a lever prefer it to food, water, or sex.
• Norepinephrine neurons, in the locus coeruleus of the pons, project to almost the entire CNS and regulate arousal, vigilance, and the response to novelty and stress.
• Serotonin neurons, in the raphe nuclei along the midline of the brain stem, also project almost everywhere and influence mood, sleep, and impulse control.
• Acetylcholine neurons, in the basal forebrain (and in the brain stem), project to the cortex and hippocampus and regulate attention, arousal, and learning.

These systems contain few neurons — a human locus coeruleus has only tens of thousands — but they influence motivation, emotion, attention, and memory throughout the brain, and many psychiatric drugs act on them.

The peripheral nervous system

Outside the CNS, the somatic division of the PNS comprises the sensory neurons that serve the skin, muscles, and joints and the axons of the motor neurons that command the skeletal muscles. The autonomic division carries sensation from the internal organs and controls them, along with the blood vessels and glands. Its sympathetic system prepares the body for action (the fight-or-flight response), its parasympathetic system for rest and digestion, and its enteric system — a network of hundreds of millions of neurons in the wall of the gut — runs digestion largely on its own.

Memory: the hippocampal loop

Memory shows that the brain’s systems are not only sensory and motor. In 1953 a young man known for decades only as H.M. — Henry Molaison — had much of the hippocampus and the surrounding medial temporal lobe removed from both sides of his brain to treat severe epilepsy. His seizures improved, but he lost the ability to form new memories of people, places, and events: he could hold a conversation but forgot it minutes later. Yet his memories from long before the operation were largely spared, and so were his perception, his intelligence, and his ability to learn new motor skills. The hippocampal system is needed to form lasting explicit memories — the kind we can recall consciously — but not to store them indefinitely; they are ultimately stored elsewhere, in the cortex, by mechanisms still unknown. Implicit memories, such as skills and habits, depend on other structures, including the basal ganglia and the cerebellum.

The hippocampal formation, in the medial temporal lobe, sits at the top of the cortical hierarchy. It comprises the entorhinal cortex, the dentate gyrus, the hippocampus proper, and the subiculum. The entorhinal cortex is its entry point, collecting information from the highest-level, multimodal association areas. Inside, connections run mostly in one direction, around a loop (Figure 1.15): from the entorhinal cortex to the dentate gyrus, then to the CA3 and CA1 fields of the hippocampus proper, then to the subiculum, and back to the entorhinal cortex, which returns the result to the neocortex. One stage breaks the pattern: CA3 neurons are richly connected to each other. A network that excites itself this way can store patterns of activity and recall an entire pattern from a fragment, which makes CA3 a candidate for the brain’s associative memory.

The loop also shows that circuits are not all alike: their structure suits their function. Sensory and motor systems are hierarchical and reciprocal, with each stage sending connections back to the one before it, but the hippocampal system is mostly serial and one-way. That makes it fragile in a particular way: damage to a single stage can break the whole chain. A patient known as R.B., who lost only the neurons of one field, CA1, after an interruption of blood flow to the brain, suffered a profound loss of memory.

1.9Computation

Classical artificial intelligence tried to reproduce intelligence with serial, symbolic algorithms. It worked well for some problems, such as chess, and poorly for problems that brains find easy, such as recognizing faces and understanding speech. Neural network models took the opposite approach: many simple units, connected in feed-forward and recurrent architectures, computing in parallel. Such networks share two properties with real circuits. Their power comes less from the complexity of their units than from the number of units and the pattern of their connections. And, like the brain, which stays busy during sleep, recurrent networks can generate activity of their own, without any input. A chapter new to the sixth edition of Kandel’s book asks the question from the biological side: what computations do real circuits perform, and how?

Recording and controlling circuits

For decades, neurophysiologists recorded from one neuron at a time. Today, silicon probes with hundreds of recording sites capture the spikes of hundreds of neurons at once,¹⁵15 Jun JJ et al. (2017). Fully integrated silicon probes for high-density recording of neural activity. Nature 551:232–236. and microscopes image the activity of thousands of neurons through genetically encoded indicators that glow when calcium enters an active cell — more slowly than electrodes, but following the same identified cells for weeks. Just as important, neurons can now be controlled. Optogenetics introduces genes for light-sensitive proteins into chosen types of neurons — channels such as channelrhodopsin to excite them, pumps such as halorhodopsin to silence them — so that light switches them on or off within milliseconds; chemogenetics does the same, more slowly, with engineered receptors that respond only to a designer drug. Recording reveals what a circuit represents; control reveals whether that representation causes behavior. Making sense of the data is a computational problem in its own right: statistical methods extract spikes from noisy signals, dimensionality reduction finds the few patterns that dominate the activity of many neurons, and models range from detailed simulations of single cells to abstract descriptions of whole populations.

Encoding and decoding

A neuron encodes a variable if its firing depends on it. Much of what we know comes from measuring tuning curves: firing rate as a function of some stimulus or movement variable. In the 1960s David Hubel and Torsten Wiesel found that neurons in the primary visual cortex respond best to edges of a particular orientation, each neuron preferring its own; neurons in the motor cortex similarly prefer particular directions of arm movement. Codes can also use time: the precise timing of spikes, and their order across a population, can carry information that a firing rate averages away.

Decoding runs the other way: from the activity of a population, infer the variable. It is not only something experimenters do — every brain area must, in effect, decode the activity of its inputs. The hippocampus provides the clearest example. In 1971 John O’Keefe discovered that some hippocampal neurons fire only when an animal is in a particular part of its environment — the cell’s place field — as if each marked a location on a map. Together, place cells cover the whole environment, and from the activity of about a hundred of them it is possible to track where a rat is, moment by moment, to within a few centimeters (Figure 1.16).¹⁶16 Wilson MA, McNaughton BL (1993). Dynamics of the hippocampal ensemble code for space. Science 261:1055–1058. Single cells are noisy, but the population is redundant, and a good decoder weights each cell by how reliable it is. Place cells have since been found in bats, monkeys, and people. The entorhinal cortex adds grid cells, each of which fires at many locations arranged in a triangular lattice, giving the map a metric.¹⁷17 Hafting T, Fyhn M, Molden S, Moser M-B, Moser EI (2005). Microstructure of a spatial map in the entorhinal cortex. Nature 436:801–806. O’Keefe and May-Britt and Edvard Moser shared the 2014 Nobel Prize for these discoveries, which made concrete Edward Tolman’s 1948 idea of a cognitive map.

Decoding also reveals what the brain does offline. During pauses and sleep, hippocampal populations emit brief, high-frequency bursts called sharp-wave ripples, lasting 50 to 500 milliseconds, and within them place cells fire in sequences that replay paths the animal has taken — compressed 10- to 20-fold in time, so that a ripple of a fifth of a second can replay several seconds of running. Replay sometimes runs in reverse, and sometimes traces paths the animal has never taken.¹⁸18 Gupta AS, van der Meer MAA, Touretzky DS, Redish AD (2010). Hippocampal replay is not a simple function of experience. Neuron 65:695–705. It is thought to help consolidate memories and plan routes.

Feed-forward hierarchies

In the visual system, information flows from the retina through the thalamus to the primary visual cortex (V1), and then along the ventral stream (Section 1.4), from the occipital lobe to the underside of the temporal lobe — through V2, V4, and the inferotemporal cortex (IT) (Figure 1.17). Along the way, receptive fields — the part of the visual field to which a neuron responds — grow larger, and preferences grow more complex: edges in V1, contours and simple shapes in V2 and V4, and in IT whole objects and categories such as faces, recognized across changes in position, size, and viewing angle. Each stage builds on the features computed by the one before.

Deep convolutional networks share the design ideas Kandel identifies in this hierarchy: receptive fields that tile the image like a map (the biological counterpart of convolution), that grow from stage to stage, and that become more selective for features and more invariant to position and size. The resemblance turned out to run deeper than design. Networks trained only to classify images develop stages that resemble the visual hierarchy, their later layers predict the responses of IT neurons better than models built by hand,¹⁹19 Yamins DLK et al. (2014). Performance-optimized hierarchical models predict neural responses in higher visual cortex. PNAS 111:8619–8624. and on hard images their errors partly match those of people and monkeys. The differences matter just as much. The visual cortex has massive feedback and lateral connections, learns from very few labeled examples, runs on a fraction of the brain’s 20 watts, and is not fooled by the tiny, carefully chosen changes to an image that fool many networks.

Expansion: the cerebellum

The cerebellum uses a different motif (Figure 1.18). Its input arrives through some 200 million mossy fibers, which excite about 50 billion granule cells, the most numerous neurons in the brain. Each granule cell receives only about four mossy fibers, a different combination for each cell and often from different sources — one signaling, say, the sight of a moving ball and another the flexing of the wrist — so each responds to a particular conjunction of inputs, a mixed representation. The combinations multiply quickly: 100 input channels can be paired in 4,950 ways and grouped in threes in more than 150,000. The granule layer thus recodes the input into a much larger and sparser representation: an expansion. Its wiring appears to be largely random — convenient, since a specific pattern would be hard to specify genetically.

The granule cells’ axons, the parallel fibers, run along the cerebellar cortex and cross the flat dendritic trees of the Purkinje cells, each of which collects input from more than a hundred thousand of them. Each Purkinje cell also receives one very powerful input from a single climbing fiber, which comes from the inferior olive in the brain stem and signals errors or unexpected events. When a climbing fiber fires, the parallel-fiber synapses that were active just before it are weakened, a change called cerebellar long-term depression.

David Marr (1969) and James Albus (1971) recognized this as a learning machine.²⁰20 Marr D (1969). A theory of cerebellar cortex. Journal of Physiology 202:437–470. Albus JS (1971). A theory of cerebellar function. Mathematical Biosciences 10:25–61. The expansion makes many different input patterns separable by a simple, linear readout, and the climbing fiber acts as a teacher that adjusts the readout’s weights. This is how you recalibrate a movement after errors — for example, how your eye movements adapt when you start wearing new glasses.

Eyeblink conditioning shows the mechanism at work. A tone is repeatedly followed by a puff of air to the eye, and the animal learns to close the eye just before the puff arrives. The tone reaches the Purkinje cells through mossy fibers and granule cells, the puff through the climbing fiber. Pairing them weakens the parallel-fiber synapses that were active just before the puff, which carves a well-timed pause into the Purkinje cells’ steady, inhibitory firing; the pause releases neurons in the deep cerebellar nuclei, which drive the eyelid closed. The timing is thought to come from granule cells that respond at different delays after the tone, giving the readout a set of time-shifted signals to choose from. In machine-learning terms the cerebellum learns from errors — supervised learning, though with a crude teacher: in eyeblink conditioning the climbing fiber reports only that something aversive happened, which is why Kandel’s chapter calls it a form of reinforcement learning.

Recurrence: holding and integrating

Feed-forward circuits respond to their input and fall silent when it stops. Many functions need the opposite: activity that outlasts its input. When you move your eyes to a new target, premotor neurons in the brain stem and the superior colliculus send the eye’s motor neurons a brief burst of spikes that encodes the velocity of the movement. But to hold the eyes still at their new position, the motor neurons must keep firing at a new, steady rate for as long as you look there. A circuit in the brain stem, the neural integrator, converts one signal into the other: it integrates velocity into position, in the mathematical sense (Figure 1.19). The leading explanation is recurrent connections: the integrator’s neurons excite one another, so their activity sustains itself after the burst ends. Persistence survives blocking the ion channels that might let single neurons sustain their own firing, which points to the network, though its exact wiring is not settled. Integration is a general-purpose operation. Similar circuits could accumulate evidence over time — as when you pick out a familiar voice in a noisy room — and are thought to produce the persistent activity that holds information in prefrontal cortex for seconds, during working memory.

Learning rules

How do circuits come to be wired correctly? Two kinds of learning rule are well established.

Hebbian plasticity. In 1949 Donald Hebb proposed that when one neuron repeatedly helps to fire another, the connection between them strengthens — often summarized as “cells that fire together wire together.” In its simplest form, the change in a synaptic weight w is proportional to the product of presynaptic activity x and postsynaptic activity y: Δw = η x y, where η is a small learning rate. The physiological counterpart was found in 1973: long-term potentiation, a lasting strengthening of synapses after strong, coincident activity on both sides.²³23 Bliss TVP, Lømo T (1973). Long-lasting potentiation of synaptic transmission in the dentate area of the anaesthetized rabbit following stimulation of the perforant path. Journal of Physiology 232:331–356.

Pure Hebbian growth is unstable — a strengthened synapse makes the postsynaptic cell fire more, which strengthens the synapse further — so neurons also regulate their total input, a stabilizing process called homeostatic plasticity that has been observed experimentally. With normalization, the rule does something remarkable: the neuron’s weights converge to the dominant pattern in its input, the direction along which the input varies most (Figure 1.20). A Hebbian neuron with normalization extracts the first principal component of its inputs.²⁴24 Oja E (1982). Simplified neuron model as a principal component analyzer. Journal of Mathematical Biology 15:267–273. In a few lines of code:

# Oja's rule: Hebbian growth plus a decay term that keeps |w| near 1.
w = rng.normal(size=d)
for x in inputs:               # one input vector at a time (zero mean)
    y = w @ x                  # postsynaptic activity
    w += eta * y * (x - y * w) # strengthen co-active inputs; normalize
# w converges to the first principal component of the inputs

Hebbian learning is unsupervised: it finds structure in the input without any teacher.

Error-driven plasticity. In the cerebellum, as we saw, a teaching signal adjusts the synapses that were active just before an error: supervised learning. A third kind, learning from reward, uses the neuromodulator dopamine. Dopamine neurons fire when an outcome is better than expected and pause when it is worse: they signal a reward prediction error, the same quantity that drives temporal-difference learning in reinforcement learning (Section 1.12).

1.10Genes

The nervous system’s design starts with genes. Genes do not encode behaviors; they encode RNAs and proteins, and these build neurons, wire them into circuits, and run their signaling. This section covers how genes work and how we know they matter; the next shows what they do for behavior.

Heritability

The oldest evidence that genes influence behavior comes from families. In 1875 Francis Galton, who popularized the phrase “nature and nurture,” became the first to use twins to weigh one against the other. Identical (monozygotic) twins develop from a single fertilized egg and share all their genes; fraternal (dizygotic) twins share about half, like any siblings. If identical twins resemble each other more than fraternal twins do, genes are the likely reason, since both kinds of twins usually share their environment. Identical twins separated early in life and raised in different families make the point vividly: as adults they still show remarkable similarities in personality. A meta-analysis of fifty years of twin studies, covering 17,804 traits, found that on average about half of the variation in human traits is attributable to genetic differences.²⁶26 Polderman TJC et al. (2015). Meta-analysis of the heritability of human traits based on fifty years of twin studies. Nature Genetics 47:702–709. The other half shows that the environment matters about as much.

Psychiatric disorders follow the same logic (Figure 1.21). The lifetime risk of schizophrenia, about 1% in the general population, rises with genetic relatedness to an affected person, to about 48% for an identical twin.²⁷27 Gottesman II (1991). Schizophrenia Genesis: The Origins of Madness. W. H. Freeman. That number cuts both ways: genes matter enormously, yet about half of the identical twins of people with schizophrenia never develop it. The pattern also rules out a single cause. Parents, siblings, children, and fraternal twins all share half their genes with an affected person, yet their risks range from 6% to 17% — a spread that one gene could not produce, but many genes acting together with the environment could.

How genes work

A gene is a segment of DNA, a double-stranded molecule whose strands are chains of four nucleotides — adenine (A), thymine (T), guanine (G), and cytosine (C) — that pair across the strands, A with T and G with C. Because each strand is a template for the other, DNA can be copied faithfully when a cell divides. The human genome has about 3 billion base pairs and about 20,000 genes that encode proteins; the protein-coding sequences make up less than 2% of the genome. The rest includes the regulatory sequences that control when and where genes are used, tens of thousands of genes for RNAs that are never translated into protein, the introns described below, and long stretches of repeated DNA whose function, if any, is unclear.

A gene has two kinds of parts (Figure 1.22). Its regulatory regions — a promoter next to the gene and enhancers that can be far away — bind transcription factors, proteins that switch transcription on or off. Its transcribed region alternates exons, which are kept, with introns, which are removed. Expression takes three steps: the gene is transcribed into a single-stranded copy made of RNA; the introns are spliced out, leaving a messenger RNA (mRNA) that leaves the nucleus; and the mRNA is translated by ribosomes into a protein. Because splicing can join exons in different combinations, one gene can produce several proteins.

The brain uses more of the genome than any other organ, and different types of neurons express different combinations of genes. That is how a fixed set of about 20,000 genes can produce thousands of neuron types: the variety comes from combinatorial regulation, not from one gene per type. Gene expression is also dynamic. Experience and neural activity switch genes on and off, which is how long-term memories are stabilized (Section 1.7).

Chromosomes, genotype, and phenotype

Genes are arranged along chromosomes. Humans have 23 pairs: 22 pairs of autosomes and a pair of sex chromosomes (XX in females, XY in males). We are diploid: every cell except eggs and sperm carries two copies of each autosome, one from each parent. The position of a gene on its chromosome is its locus, and the different versions of a gene that occur in a population are its alleles. Mitochondria, the cell’s power plants, carry a small genome of their own, inherited from the mother.

It is essential to distinguish an individual’s genotype, its genetic makeup, from its phenotype, its observable traits. The mapping between them is rarely one to one. A variant may have a large effect, a small one, or none, depending on other genes and on the environment; most traits depend on many genes; and the same genotype can produce different phenotypes. Even eye color, often taught as a simple trait, depends on several genes, most strongly on variants in and near OCA2, which control how much pigment the iris makes.

Genes vary between individuals. Most variants are single-nucleotide polymorphisms (SNPs) — two people differ at about one position in a thousand — but stretches of DNA can also be deleted or duplicated (copy-number variants), and everyone carries some 70 to 90 new, de novo point mutations not present in either parent — on average about one of them changes a protein — with more in the children of older fathers. A recessive mutation shows its effect only when both copies of a gene are affected; a dominant one needs only one. Many disorders of brain development arise from haploinsufficiency: one working copy of the gene is not enough. As a rule, common variants have small effects and rare ones large effects, though there are exceptions: a common variant of the APOE gene raises the risk of late-onset Alzheimer’s disease about fourfold.

Conserved from worms to humans

Genes are conserved through evolution. About 80% of mouse genes have a single counterpart, an ortholog, in the human genome;²⁸28 Mouse Genome Sequencing Consortium (2002). Initial sequencing and comparative analysis of the mouse genome. Nature 420:520–562. more than half of human genes resemble genes of worms or flies; and where the human and chimpanzee genomes can be aligned, they differ by only about 1%.²⁹29 Chimpanzee Sequencing and Analysis Consortium (2005). Initial sequence of the chimpanzee genome and comparison with the human genome. Nature 437:69–87. Yet humans and chimpanzees are profoundly different. Most of the difference lies not in new genes but in how shared genes are regulated: when, where, and how much they are expressed. The basic molecular machinery of neurons — for development, signaling, and gene regulation — was already present in the common ancestor of worms, flies, mice, and humans, and genes important for the human brain tend to be among the most conserved. That is why simpler animals can reveal how genes shape behavior.

Two strategies are used. Forward genetics starts with behavior: mutate many animals at random, find individuals that behave abnormally, and identify the responsible gene. Reverse genetics starts with a gene: delete or alter it — in specific cell types and at specific times, with tools such as Cre/loxP recombination and CRISPR gene editing — and observe the consequences.

1.11Behavior

Single genes rarely explain a behavior. In animals, though, some genes have strikingly specific effects, and they illuminate how genes act. Kandel’s examples span timekeeping, foraging, social bonds, and human cognition.

A genetic clock

Many behaviors — sleep and waking, activity, body temperature, hormone release — follow a daily, or circadian, rhythm. The rhythm is not simply a response to daylight: animals kept in constant darkness keep cycling, with a period close to but not exactly 24 hours (about 24.2 hours in humans). An internal clock generates the rhythm, and light keeps the clock synchronized with the sun.

In 1971 Ronald Konopka and Seymour Benzer screened mutant fruit flies for abnormal daily rhythms and found three mutations in a single gene, which they named period: one shortened the cycle to about 19 hours, one lengthened it to about 28, and one abolished it.³¹31 Konopka RJ, Benzer S (1971). Clock mutants of Drosophila melanogaster. PNAS 68:2112–2116. Until then, many doubted that there could be true “behavior genes” — genes not needed for an animal’s basic physiology — and the period mutants were otherwise healthy. A single gene could change the speed of a behavioral clock, so its product had to be part of the timekeeper itself, not merely a component the clock needed to run. In 1994 Joseph Takahashi’s group found a comparable mutation in mice, in a gene they named Clock: mice with one mutant copy run slow, and mice with two lose their rhythm altogether in constant darkness. A group of genes, not a single one, keeps time, and the same genes do it in flies and mice.

The clock is a transcriptional oscillator: a molecular negative-feedback loop (Figure 1.23). Two transcription factors, CLOCK and CYCLE (called BMAL1 in mammals), switch on the period and timeless genes. Their proteins, PER and TIM, accumulate in the cytoplasm over hours, pair up, and enter the nucleus, where they shut off CLOCK and CYCLE — and therefore their own production. As PER and TIM are degraded, the inhibition lifts and the cycle starts again. Most of the delay comes after the proteins are made: enzymes called kinases tag PER with phosphate groups that mark it for rapid destruction, so it accumulates only when enough TIM is present to bind and protect it. These delays stretch each turn of the loop to about a day. The clock’s transcription factors also switch on output genes that carry the time signal to the rest of the body — in flies, for example, the gene for a neuropeptide, PDF, that sets the daily rhythm of locomotor activity. Light resets the clock by triggering the destruction of TIM.

A similar loop, built from related genes, runs in mammals, where a master clock in the hypothalamus receives light information directly from the retina. In people with familial advanced sleep phase syndrome, a mutation in a PER gene shifts the whole cycle earlier: they fall asleep around 7:30 p.m. and wake around 4:30 a.m.³²32 Toh KL et al. (2001). An hPer2 phosphorylation site mutation in familial advanced sleep phase syndrome. Science 291:1040–1043. Jeffrey Hall, Michael Rosbash, and Michael Young received the 2017 Nobel Prize for working out the mechanism.

Exploring or exploiting

Fruit-fly larvae come in two natural varieties. Rovers move a lot while feeding and travel between patches of food; sitters move little and stay on one patch. Marla Sokolowski traced the difference to the foraging gene, the first gene found to underlie a natural variation in behavior rather than one created in the laboratory. It encodes a protein kinase (PKG) activated by the intracellular messenger cGMP — one of a family of enzymes that are especially important for turning short-term signals into long-term changes in neurons — and rovers have more PKG activity than sitters.³³33 Osborne KA et al. (1997). Natural behavior polymorphism due to a cGMP-dependent protein kinase of Drosophila. Science 277:834–836. Both variants persist in wild populations because each wins under different conditions — in crowded environments rovers do better, and in sparse ones sitters exploit their food more thoroughly. Reinforcement-learning practitioners will recognize the trade-off between exploration and exploitation, here with a single gene setting the default policy. The same gene shapes the division of labor in honeybees, where young bees work inside the hive as nurses and older bees forage outside. Foragers express more of it, and boosting PKG activity in young bees makes them start foraging early.³⁴34 Ben-Shahar Y, Robichon A, Sokolowski MB, Robinson GE (2002). Influence of gene action across different time scales on behavior. Science 296:741–744.

Social behavior

Social behaviors vary enormously between species, yet they have large innate components. In the roundworm C. elegans, some strains feed alone and others feed in groups. The difference maps to a single amino acid in npr-1, a receptor for a neuropeptide — one of a large class of signaling molecules that neurons release to coordinate the activity of whole networks; in mammals, neuropeptides have been implicated in feeding, sleep, pain, and much else.³⁵35 de Bono M, Bargmann CI (1998). Natural variation in a neuropeptide Y receptor homolog modifies social behavior and food response in C. elegans. Cell 94:679–689. Two of them, oxytocin and vasopressin, regulate social bonds, oxytocin mainly in females and vasopressin mainly in males. Prairie voles form lifelong pair bonds, and the males help raise the young; closely related meadow and montane voles do neither. The species differ in where their brains express receptors for these neuropeptides, notably in reward circuits, and adding vasopressin receptors to a reward area in the brains of male meadow voles is enough to make them prefer a single partner.³⁶36 Lim MM et al. (2004). Enhanced partner preference in a promiscuous species by manipulating the expression of a single gene. Nature 429:754–757. Whether these molecules play a similar role in human attachment is not known.

Genes and the human brain

In humans, rare genetic syndromes open windows onto the biology of cognition and social behavior:

• Phenylketonuria (PKU) is caused by recessive mutations in the gene for phenylalanine hydroxylase, the enzyme that converts the amino acid phenylalanine into tyrosine. Untreated, it causes severe intellectual disability; a diet low in phenylalanine, begun in infancy, largely prevents the damage. The phenotype comes from genotype and environment together: the same mutation is devastating on one diet and largely harmless on another.
• Fragile X syndrome, the most common inherited cause of intellectual disability, and Rett syndrome are caused by mutations in genes (FMR1 and MECP2) that regulate the expression of many other genes in neurons. Both genes lie on the X chromosome, which is why fragile X affects boys more severely, and why Rett syndrome is seen almost only in girls: the mutations are often lethal to male embryos, which have a single X.
• Williams syndrome, caused by the loss of one copy of about 27 genes on chromosome 7, produces unusually sociable, talkative people with poor visuospatial skills, intellectual disability, and high anxiety. Gene dosage cuts both ways: an extra copy of the same region raises the risk of autism.
• The 22q11.2 deletion syndrome raises the risk of schizophrenia about twenty-five-fold, to roughly one in four.
• Mutations in FOXP2, a transcription factor, disrupt the fine control of mouth and face movements needed for speech, and language with it. The human form of the protein differs slightly from that of other primates, which made the gene a candidate for one of the changes that made speech possible — an idea still debated.

Common psychiatric disorders are different: they are polygenic. Autism spectrum disorder, which affects 2 to 3% of people, is highly heritable. Part of the risk comes from rare de novo mutations and copy-number variants of large effect — sequencing parents and their children has implicated more than a hundred genes — and part from many common variants of small effect. The genes fall into two broad groups: genes for synaptic proteins, such as the neurexins and neuroligins, which bind each other across the synaptic cleft and help hold synapses together and set their strength; and genes that regulate the expression of other genes. The same rare variants often raise the risk of several conditions — autism, schizophrenia, bipolar disorder, intellectual disability — a one-to-many pattern that cuts across psychiatric diagnoses.

Schizophrenia shows how such risk is being tracked down. Early studies tested candidate genes, chosen by hypothesis, in small samples, and their findings mostly failed to replicate.³⁷37 Among the most studied candidates were neuregulin 1 and DISC1, which affect the migration of neurons and the formation of synapses. Genome-wide studies have not confirmed them as risk genes, although the broader idea they supported — that schizophrenia begins with disturbed brain development — has held. Farrell MS et al. (2015). Evaluating historical candidate genes for schizophrenia. Molecular Psychiatry 20:555–562. Genome-wide association studies scan the whole genome without a hypothesis, correct for the vast number of comparisons, and need enormous samples; one of nearly 40,000 patients and more than 110,000 controls found 108 risk loci.³⁸38 Schizophrenia Working Group of the Psychiatric Genomics Consortium (2014). Biological insights from 108 schizophrenia-associated genetic loci. Nature 511:421–427. The strongest signal lies in the major histocompatibility complex, a cluster of immune genes, and part of it maps to C4, a gene of the complement cascade, which tags synapses for elimination. Variants that raise the expression of its C4A form increase risk, suggesting that excessive pruning of synapses during adolescence contributes to the disease.³⁹39 Sekar A et al. (2016). Schizophrenia risk from complex variation of complement component 4. Nature 530:177–183.

Each common variant has only a tiny effect: even the highest-risk form of C4 raises a person’s lifetime risk from about 1% to about 1.3%, whereas having an affected parent or sibling raises it about tenfold. But the effects add up. Summed into a polygenic risk score, they separate people sharply: in the largest study, people in the top tenth of scores had 8 to 20 times the risk of those in the bottom tenth, depending on the sample. Even here genes act alongside the environment: risk is also raised by factors such as famine or infection during pregnancy and heavy cannabis use in adolescence. In the end, the behavioral differences between people emerge from the interplay of genes, environment, chance, and individual choice.

1.12Imaging

Most of what we know about circuits comes from animals, where electrodes, microscopes, and genetic tools can reach individual neurons. To watch the healthy human brain at work, neuroscience relies mostly on functional magnetic resonance imaging (fMRI), which measures changes in blood oxygenation throughout the brain while people perceive, remember, and decide. Kandel calls it the dominant approach in humans for measuring biological processes and linking them to behavior.

What fMRI measures

MRI exploits the magnetism of hydrogen nuclei, which are abundant in the water of the body. In a strong magnetic field — 3 tesla, about 60,000 times Earth’s — a tiny excess of the nuclei, about ten in a million, lines up with the field, and they precess around it at a frequency proportional to its strength, the Larmor frequency: about 128 MHz at 3 tesla. A radio-frequency pulse at that frequency tips them out of alignment, and as they precess in step they induce an oscillating current in the scanner’s receiver coil. The signal fades as the nuclei fall out of step with one another — a decay measured by the time constant T2, or T2* when local distortions of the field are included — and as they relax back into alignment (T1). Small gradients added to the main field make the frequency depend on position, which is how the signal is mapped to locations in the brain. Fast pulse sequences, such as echo-planar imaging, acquire a slice in under a tenth of a second and the whole brain every second or two.

The functional part comes from blood. Hemoglobin that has released its oxygen is paramagnetic: it distorts the local magnetic field, so nearby nuclei fall out of step faster and the signal dims. When neurons in a region become active, blood flow to the region increases more than their oxygen consumption does, so the blood becomes more oxygenated and the signal brightens. This blood-oxygen-level-dependent (BOLD) signal, discovered by Seiji Ogawa in 1990,⁴⁰40 Ogawa S, Lee TM, Kay AR, Tank DW (1990). Brain magnetic resonance imaging with contrast dependent on blood oxygenation. PNAS 87:9868–9872. tracks neural activity indirectly, through a chain of events called neurovascular coupling whose mechanisms, including the role of astrocytes, are still being worked out. It is slow: after a brief burst of neural activity, the BOLD response begins to rise after a second or two, peaks at about five seconds, and returns to baseline after 12 to 15 seconds (Figure 1.24). And it is coarse: each measurement element, or voxel, is typically 2–3 mm across and contains hundreds of thousands to millions of neurons.⁴¹41 Logothetis NK (2008). What we can do and what we cannot do with fMRI. Nature 453:869–878. What exactly the signal reflects is still debated. It seems to follow the synaptic input and local processing in a region more closely than its output spikes; it cannot tell excitation from inhibition; and blood flow can even rise without extra firing, as when an expected visual stimulus fails to appear.

Analyzing fMRI data

Raw fMRI data first go through preprocessing: correcting for head motion and for the different acquisition times of the slices — up to a second or two apart within one image of the brain — filtering out slow drifts, smoothing, and aligning each brain to an anatomical template so that people can be compared. Three families of analysis then answer different questions:

• Univariate activation asks where: which voxels respond more in one condition than in a matched control? Each voxel’s time series is modeled as the task’s timing convolved with the hemodynamic response, using a general linear model, and the difference between conditions is tested across people. This is how the fusiform face area, on the underside of the temporal lobe, was found: it responds much more strongly to faces than to other objects.⁴²42 Kanwisher N, McDermott J, Chun MM (1997). The fusiform face area: a module in human extrastriate cortex specialized for face perception. Journal of Neuroscience 17:4302–4311. Variants replace the on–off timing of the task with a quantity from a model, such as memory load or a reward prediction error, or exploit adaptation — the weaker response to a repeated stimulus — to probe what a region is tuned to.
• Multivariate pattern analysis asks what: is there information about a stimulus or task in the pattern of activity across many voxels? It comes in two flavors. Classification trains a classifier on part of the data and tests it on the rest; categories such as faces, houses, and chairs can be decoded from the ventral temporal cortex, even from regions that do not respond most strongly to them.⁴³43 Haxby JV et al. (2001). Distributed and overlapping representations of faces and objects in ventral temporal cortex. Science 293:2425–2430. Representational similarity analysis asks how the information is organized, by comparing which stimuli evoke similar activity patterns — across regions, people, species, and computational models such as the layers of a deep network.⁴⁴44 Kriegeskorte N, Mur M, Bandettini P (2008). Representational similarity analysis — connecting the branches of systems neuroscience. Frontiers in Systems Neuroscience 2:4.
• Functional connectivity asks which regions work together, by correlating their activity over time. Two regions can correlate simply because both respond to the same stimulus, so responses evoked by the task are removed first. Even at rest, the brain is organized into large networks of regions whose activity fluctuates together; one of them, the default mode network, is most active when a person is not engaged in an external task.⁴⁵45 Raichle ME et al. (2001). A default mode of brain function. PNAS 98:676–682.

What fMRI has taught us

fMRI has shaped neuroscience in three ways. It has guided research in animals: after face-selective regions were found in humans, fMRI in monkeys revealed similar “face patches,” and electrodes placed in them found that nearly all of their neurons respond selectively to faces.⁴⁶46 Tsao DY, Freiwald WA, Tootell RBH, Livingstone MS (2006). A cortical region consisting entirely of face-selective cells. Science 311:670–674.

It has challenged theories from psychology, often by accident, because it records the whole brain at once. Activity in prefrontal and parietal cortex while an experience is encoded predicts whether it will later be remembered, so memory formation involves more than the hippocampus.⁴⁷47 Wagner AD et al. (1998). Building memories: remembering and forgetting of verbal experiences as predicted by brain activity. Science 281:1188–1191. Brewer JB et al. (1998). Making memories: brain activity that predicts how well visual experience will be remembered. Science 281:1185–1187. And the hippocampus, thought to serve only memories that can be consciously recalled, takes part — interacting, and even competing, with the striatum — in trial-and-error learning tasks long considered unconscious.⁴⁸48 Poldrack RA et al. (2001). Interactive memory systems in the human brain. Nature 414:546–550. Together these findings loosened the link between the hippocampus and conscious awareness, and brought human memory research closer to animal studies.

And it has tested computational models: the human striatum responds to rewards like a reward prediction error, the quantity that dopamine neurons signal in monkeys and that drives temporal-difference learning.⁴⁹49 O’Doherty JP, Dayan P, Friston K, Critchley H, Dolan RJ (2003). Temporal difference models and reward-related learning in the human brain. Neuron 38:329–337. Deep networks trained on vision likewise predict activity along the human ventral stream, layer by layer.

Reading fMRI carefully

fMRI results are easy to overinterpret. Three cautions:

• Correlation, not causation. fMRI shows that a region’s activity varies with a task, not that the task needs the region. Lesions, brain stimulation, and causal manipulations in animals supply that evidence; and since every method has limits — recordings from neurons are correlational too — conclusions are strongest when several methods converge.
• Reverse inference. Most studies reason forward, from a process they manipulate to the regions it engages. Reasoning backward — “the amygdala was active, so the person felt fear” — is valid only to the extent that the region is selective for that process, and most regions are not: one region serves many processes, and one process engages many regions.⁵⁰50 Poldrack RA (2006). Can cognitive processes be inferred from neuroimaging data? Trends in Cognitive Sciences 10:59–63. Databases that pool thousands of studies, such as Neurosynth, now estimate how selective each region is.⁵¹51 Yarkoni T, Poldrack RA, Nichols TE, Van Essen DC, Wager TD (2011). Large-scale automated synthesis of human functional neuroimaging data. Nature Methods 8:665–670.
• Statistics at scale. A colored brain map is a map of statistics, not of activity: the colored voxels are those that passed a threshold. A brain image has 100,000 or more voxels, so without correction for multiple comparisons some will pass by chance. In a famous demonstration, an uncorrected analysis found “activity” in a dead salmon that was shown photographs of people.⁵²52 Bennett CM, Baird AA, Miller MB, Wolford GL (2010). Neural correlates of interspecies perspective taking in the post-mortem Atlantic salmon: an argument for proper multiple comparisons correction. Journal of Serendipitous and Unexpected Results 1:1–5. Selecting voxels with one analysis and then measuring the same data inflates effects in the same way.

Used carefully, fMRI can do remarkable things. In a study published in 2006, a patient diagnosed as being in a vegetative state was asked, in the scanner, to imagine playing tennis and then to imagine walking through her home. Her brain produced the same distinct patterns of activity as healthy volunteers’, suggesting that she understood the instructions and was following them.⁵³53 Owen AM et al. (2006). Detecting awareness in the vegetative state. Science 313:1402.

No single method sees everything (Figure 1.25). Electrodes resolve individual spikes but sample a tiny fraction of the neurons; calcium imaging sees thousands of identified cells but mostly in animals and near the surface of the brain; EEG and MEG, which record electric and magnetic fields outside the head, resolve milliseconds but blur centimeters; fMRI covers the whole human brain at the scale of millimeters, but only second by second. Much of modern neuroscience consists of combining them.

1.13Recap

This chapter assembled a system overview of the brain: seven major parts with specialized functions, built from neurons and glia, signaling with graded potentials and all-or-none spikes, wired by a handful of motifs into reflexes, sensory and motor pathways, modulatory broadcast systems, and memory loops; computing with population codes, hierarchies, expansions, and recurrent dynamics; learning through Hebbian, error-driven, and reward-driven plasticity; built and tuned by genes in constant interplay with the environment; and observable in humans, slowly and indirectly, with fMRI. Nine principles summarize it (Table 1.3).

For engineers, four themes are worth carrying into the chapters ahead:

• Organization beats component speed. The brain’s parts are slow, noisy, and frugal; its strength is the way they are wired — with specificity, in deep, parallel, recurrent architectures.
• Learning is plural. Unsupervised, supervised, and reinforcement learning coexist in different structures, and neuromodulators set the mode in which circuits operate and learn.
• Much is innate. Genes specify the architecture, the initial wiring, and the learning rules — strong priors that experience then tunes.
• Every method sees part of the system. Claims about the brain are only as good as the measurement behind them.

The next chapter, Neurons, opens the cell: how membranes, ion channels, and pumps produce the resting potential, and how the action potential is generated and propagated.

Further reading

• Kandel ER, Koester JD, Mack SH, Siegelbaum SA, eds. (2021). Principles of Neural Science, 6th ed. McGraw Hill. Part I covers everything in this chapter in depth.
• Luo L (2020). Principles of Neurobiology, 2nd ed. Garland Science. A modern, experiment-driven textbook with exceptionally clear figures.
• Sterling P, Laughlin S (2015). Principles of Neural Design. MIT Press. The brain as an engineered system, organized around the costs of energy and information.
• Dayan P, Abbott LF (2001). Theoretical Neuroscience. MIT Press. The standard mathematical introduction to neural coding, dynamics, and learning.
• Lindsay G (2021). Models of the Mind. Bloomsbury. How mathematics and computation have shaped neuroscience, for a general audience.
• Kandel ER (2006). In Search of Memory. W. W. Norton. A memoir of the science behind many of this chapter’s principles.