Connectome

by Sebastian Seung

  Figure 49. Collateral and main branches of a pyramidal neuron’s axon

  As the axon dives down, it sends out side branches, called “collaterals,” which are for making synapses onto nearby neurons. But the main branch of the axon finally leaves the gray matter and enters the white matter to start its journey to other regions. In each of its destination regions, it forks out many branches to make connections with neurons there.

  Some axons don’t travel very far, reentering the gray matter close to where they started. But most axons of pyramidal neurons project to other regions in the cortex, some going as far as the other side of the brain. Some white-matter axons—a small minority—connect the cortex with other structures in the brain, such as the cerebellum, the brainstem, or even the spinal cord. These axons make up less than one-tenth of the white matter. The cortex is highly self-centered, primarily “talking” with itself rather than the outside world.

  Here’s another way to think about it: If the axons and dendrites in the gray matter are like local streets, the axons of the white matter are like the superhighways of the brain. They are relatively wide and unbranched, and also extremely long. In fact, the total length of these axons is roughly 150,000 kilometers, over a quarter of the distance from the Earth to the Moon. And herein lies the challenge: Finding the regional connectome requires tracing the journey of every axon in the white matter.

  It seems like an impossible task, but it could be done by slicing and imaging all of the white matter and using computers to follow the path traveled by each axon in the images. The start and end points of every path would define a connection between two locations in the cortex. Is this approach too difficult to be practical? After all, the cerebral white matter is comparable in volume to the gray matter, and we are still struggling to reconstruct one cubic millimeter of that. Given this, it might seem outlandish to propose reconstructing hundreds of cubic centimeters of white matter. My proposal seems less crazy once you know that white-matter axons are visible at a lower resolution.

  To understand why, take a look at the cross-sectional image shown in Figure 50. As axons exit the gray matter, most of them undergo an important transformation—they become ensheathed by other cells that wrap around them repeatedly. Thus the brain not only wires itself up but also, amazingly, manages to wrap sheets of insulation around its “wires.” The sheets are made of a substance called myelin, which is composed mostly of fat molecules. It’s those molecules that make white matter look white. (The epithet “fathead” may sound derogatory, but it’s actually accurate for everyone.) Myelination speeds up the propagation of spikes, which is important for transmitting signals quickly in large brains. Diseases of myelination, such as multiple sclerosis, have catastrophic effects on brain function.

  Figure 50. Cross-section of myelinated axon

  The myelinated axons of the white matter are much thicker (typically 1 micrometer) than the mostly unmyelinated axons of the gray matter. Furthermore, if we only care about finding regional connections, there’s no need to see synapses. If an axon enters and branches in a region of the gray matter, we can be almost certain that it makes synapses there, so tracing the “wires” of the white matter is enough for finding the regional connectome. If we restrict ourselves to myelinated axons, we could accomplish the job with serial light microscopy, which is similar to serial electron microscopy but employs thicker slices and produces images with lower resolution.

  Of course, mapping white-matter axons is still a daunting technical challenge for a brain of human size. Studying white matter in smaller brains, such as those of rodents and nonhuman primates, is a good starting place. We can check the results by comparing them with those from older techniques for studying white-matter pathways in animals. These techniques were used to find connections between the visual areas of the monkey cortex, as shown in Figure 51. (The areas, but not their connections, were shown earlier.) Since the older techniques are not applicable to human brains, our own white matter has gone almost completely unexplored.

  Figure 51. Connections between visual areas of the rhesus monkey cortex (see Figure 39)

  The Human Connectome Project is already trying to find a map like the one in Figure 51 for the human brain using diffusion MRI (dMRI) rather than microscopy. Diffusion MRI is different from MRI, which is used to find the sizes of brain regions, or fMRI, which is used to measure their activations. Unfortunately, dMRI is subject to the same basic limitation as other forms of MRI: poor spatial resolution. MRI typically yields millimeter-scale resolution, which is not enough for seeing single neurons or axons. Given its poor resolution, how can dMRI hope to trace the wires in the white matter?

  It turns out that white matter has an interesting feature that makes its structure simpler than that of gray matter. Have you ever forgotten to stir the spaghetti after dropping it into boiling water? You discover your mistake a few minutes later, when you see that some of the strands have stuck to each other to form bundles. This culinary embarrassment resembles white matter; gray matter is more like a bowl of fully entangled spaghetti.

  When axons bundle like unstirred spaghetti, they form a “fiber tract” or a “white-matter pathway.” The bundles are similar to nerves, except that they run within the brain. Why do axons bundle? Well, why do so many people follow the same dirt paths through lawns? First, they are shortcuts, more efficient than the paved walkways installed by landscape designers. Second, there is a “follow the leader” effect—once a few trailblazers have worn down the grass a bit, everyone else follows them, trampling it down completely. Similarly, axons take efficient paths through the white matter, assuming that it evolved to achieve wiring economy. Since an efficient solution is often unique, we’d expect axons sharing the same origin and destination to take the same path. Also, it’s known that the first axons to grow during brain development often blaze the trail, providing chemical cues for other axons to follow.

  Fiber tracts may be thick, even though a single axon is microscopically thin. The largest is the famous corpus callosum, the huge collection of axons that travel between the left and right hemispheres. Neuroanatomists in the nineteenth century discovered other large tracts through naked-eye dissection of the brain. Diffusion MRI is an exciting advance, because it’s a way of tracing white-matter pathways in the living brain. It computes an arrow at every location that indicates the orientation of the axons there. By connecting these arrows, it’s possible to trace the paths of axonal bundles. In one notable success, dMRI has uncovered white-matter pathways connecting Broca’s and Wernicke’s regions, other than the classical one in the arcuate fasciculus. As I mentioned earlier, such discoveries are sparking revisions of the Broca–Wernicke model of language.

  Such stories are encouraging, but dMRI also has limitations. Because of its poor spatial resolution, dMRI has difficulty following thin fiber tracts. And even thick tracts can be problematic if they intersect and their individual axons become intermingled. Think of this crossing as a chaotic traffic intersection packed with pedestrians, bicyclists, animals, and cars—you have to watch carefully to see whether any particular traveler goes straight or turns. Similarly, once axons enter the region where two bundles intersect, it’s difficult to see, using dMRI, where they end up. The only foolproof way of mapping the white matter is to use a method that can trace individual axons, like the one I’ve proposed here.

  Mapping regional connectomes is already problematic with dMRI; the method is even more ill-suited for neuronal and neuron type connectomes. Of course, dMRI has the important advantage that it can be performed on a living brain. At the very least it will detect gross connectopathies, like a missing corpus callosum. Since dMRI can be used quickly and conveniently to study many living brains, it will find correlations between mental disorders and brain connectivity. But these correlations might remain weak, just like the earlier phrenological ones.

  MRI experts are continuing to improve resolution, but the rate of improvement is not that fast, and there is a long
way to go. Roughly speaking, the current resolution of dMRI is a thousand times worse than light microscopy, which in turn is a thousand times worse than electron microscopy. Inventors might create better noninvasive imaging methods than MRI. But let’s not forget that seeing through the skull into the interior of a living brain is fundamentally more challenging than chopping up a dead brain and examining the pieces with a microscope. Microscopy already delivers the resolution we need to find connectomes; we just have to scale it up to handle larger volumes. In contrast, MRI requires breakthroughs far more fundamental. For the foreseeable future, then, microscopy and MRI will remain complementary methods.

  To find connectopathies, we will use the methods I outlined above to map reduced connectomes of abnormal and normal brains, and compare them. Some differences may be detectable by dMRI, but subtle ones will require microscopy. We will also compare neuronal connectomes of small chunks of brain using electron microscopy. The use of microscopy poses difficulties, as it must be carried out on the brains of the deceased. People do bequeath their brains to science—there is a long tradition of such generosity—but even if we have postmortem brains, many of them present special problems.

  One alternative is to search for connectopathies in the brains of animals. Such research will also be important for developing therapies, which are often tested first on animals and only later on humans. The legendary French microbiologist Louis Pasteur produced the first vaccine for rabies by growing the virus in rabbits and then weakening it. The vaccine was tested on dogs before its dramatic first human trial on a nine-year-old boy who had been bitten by a rabid dog.

  Studying human mental disorders with animals is no easy task. The rabies virus leads to the same disease, whether it infects rabbits, dogs, or humans. But is there such a thing as an autistic or schizophrenic animal? It’s not clear whether such animals occur naturally, but researchers are now attempting to create them using the methods of genetic engineering. Researchers insert the faulty genes associated with autism and schizophrenia into the genomes of animals, usually mice, with the expectation of giving them analogous disorders. Ideally, such animals would serve as “models” for human disorders, approximations to the real thing.

  But this strategy, a variation on Pasteur’s, sometimes fails even for infectious diseases. Human immunodeficiency virus (HIV), which causes AIDS in humans, fails to infect many primates, making it difficult to test HIV vaccines. In monkeys, AIDS is caused by simian immunodeficiency virus (SIV), which is related to HIV but not identical. The lack of a good animal model for human AIDS has slowed down research on finding a cure. Likewise, inserting faulty human genes into animals might not give them autism and schizophrenia. Some analogous but different genetic defect might be necessary.

  Because of these uncertainties, the problem of validating animal models for mental disorders has risen to the fore. It’s not clear what criteria should be used. Some emphasize similarity of symptoms, but even for infectious diseases this criterion doesn’t always work. Sometimes the same microbe can infect both animals and humans but produces very different symptoms. An animal might tolerate infection with little adverse effect at all. And if human genes for autism or schizophrenia turned out to produce very different symptoms in mice, it wouldn’t necessarily mean that the mouse models were useless. (Some might argue that it’s pointless to compare symptoms, as mental disorders involve behaviors that seem uniquely human.)

  An alternative criterion is similarity of neuropathologies, already being applied to evaluate mouse models of neurodegenerative disorders such as Alzheimer’s disease (AD). In humans, AD is accompanied by abnormal buildup of plaques and tangles in the brain. Normal mice do not develop AD, but researchers have genetically engineered several mouse models that do. Their brains generate large numbers of plaques and tangles. Researchers are still arguing about whether any of these models are good enough for studying AD. But at least they have a target: a clear and consistent neuropathology to emulate.

  Along these lines, similarity of connectopathies might be a good criterion for animal models of disorders like autism and schizophrenia. Of course, for this to work we would have to identify connectopathies in animal models, as well as analogous ones in patients afflicted by autism and schizophrenia.

  You may have noticed that the plan for comparing connectomes sounds very different from the plan for decoding them. The connectionist theory of memory proposes particular hypotheses—the cell assembly and the synaptic chain—that can be tested using connectomics. In contrast, the connectopathy idea is open-ended. Without specific hypotheses, wouldn’t searching for connectopathies be a wild-goose chase?

  One of the leaders of the Human Genome Project, Eric Lander, has summed up the decade since its completion in this way: “The greatest impact of genomics has been the ability to investigate biological phenomena in a comprehensive, unbiased, hypothesis-free manner.” It doesn’t sound like what we were taught about the scientific method in school, where we learned that science proceeds in three steps: (1) Formulate a hypothesis. (2) Make a prediction based on the hypothesis. (3) Perform an experiment to test the prediction.

  Sometimes that procedure works. But for every success story, there are many more stories of failure caused by choosing the wrong hypothesis to investigate. It can take a lot of time and effort to test a hypothesis, which might turn out to be wrong or—even worse—simply irrelevant. In the latter case, it would lead to research that ends up being a complete waste of time. Unfortunately, there’s no well-defined recipe for formulating hypotheses, beyond a stroke of insight or inspiration.

  We do have an alternative to “hypothesis-driven,” or deductive, research—the “data-driven,” or inductive, approach. It too has three steps: (1) Collect a vast amount of data. (2) Analyze the data to detect patterns. (3) Use these patterns to formulate hypotheses.

  Some scientists gravitate to one approach over the other, because it fits their personal style. But the two approaches are not really in opposition. The data-driven approach should be viewed as a way of generating hypotheses that are more likely to be worth exploring than ones based purely on intuition. It can be followed by hypothesis-driven research.

  If we have the right technologies, we’ll be in a position to apply this approach to mental disorders. Connectomics will provide more and more accurate and complete information about neural connectivity. With so much data available, we’ll no longer have to search for our keys under the lamppost. Once we identify connectopathies, these will suggest good hypotheses about the causes of mental disorders that are worth exploring further.

  To resort to another metaphor, searching for the causes of mental disorders is like looking for a needle in a haystack because the brain is so complex. How to succeed? One way is to start from a good hypothesis about the location of the needle. Then you need search only a small part of the haystack. This will work if you are lucky or smart enough to have a good hypothesis. Another way is to build a machine that rapidly sifts through all the material in the haystack. You are guaranteed to find the needle with this technology, even if you’re not lucky or smart. This is analogous to the connectomic approach.

  To understand why minds differ, we have to see better how brains differ. That’s why comparing connectomes is so crucial. Uncovering just any kind of difference won’t be sufficient, however, since many differences could end up being uninteresting. We’ll have to narrow in on the important ones, those that are strongly correlated with mental properties. These are the differences that will finally give connectionism more explanatory power than phrenology. They will accurately predict mental disorders for individuals, as well as faithfully estimate the intellectual abilities of normal people. (For connectomes obtained using microscopy on dead brains, the test would actually involve “postdiction,” guessing the mental disorders or abilities of the deceased from their brains.)

  Identifying connectopathies will be an important step toward understanding certain mental disorders. But understanding
goes only so far. Ideally we will capitalize on it by developing better treatments for these maladies, or even cures. In the next chapter I’ll envision how this will be done.

  13. Changing

  In 1821 the composer Carl Maria von Weber premiered his opera Der Freischütz. To marry Agathe, the hero, Max, must impress her father by prevailing in a shooting contest. Driven to desperation by fear of losing his love, he sells his soul to the devil for seven magic bullets, which are guaranteed to hit their mark. Max not only wins the hand of Agathe but manages to evade the devil, and the opera ends happily.

  In 1940 Warner Bros. released Dr. Ehrlich’s Magic Bullet, which dramatized the life of the German physician and scientist Paul Ehrlich. After sharing a 1908 Nobel Prize for his discoveries about the immune system, Ehrlich didn’t rest on his laurels. His institute discovered the first antisyphilis drugs, relieving the suffering of millions of people. By creating the first man-made drugs for any disease, Ehrlich effectively invented the entire pharmaceutical industry. He was guided by his theory of the “magic bullet,” the name of which may have been inspired by Weber’s popular opera. Ehrlich first imagined—and then discovered—chemicals that killed bacteria but spared other cells, like a magic bullet that unerringly flew to its target.