Connectome

by Sebastian Seung

  This question could be investigated in a number of interesting ways if the synaptic chain model of HVC is correct. During the off season, does a dormant synaptic chain continue to store the memory of song? When the new neurons enter HVC, do they integrate into the chain? If so, how do they do it? Neural Darwinism predicts that newly created neurons are randomly connected with other neurons. This prediction could be tested empirically by connectomics, with the aid of special stains that mark new neurons.

  Similar questions can be asked about the elimination of neurons. What causes neurons to commit suicide? Is it triggered by elimination of synapses and branches, which in turn happens because the neuron fails to integrate into the chain? This hypothesis could be probed using connectomics, through snapshots of neurons caught during the process of dying. To prepare for the off season, are neurons eliminated in such a way that prevents the chain from breaking?

  Because of technical limitations, neuroscientists have had to settle for counting increases and decreases in the number of neurons. These studies suggest that regeneration is important, but they do not reveal its exact role in memory. To make further progress, it’s crucial to know how new neurons get wired up to the existing organization, and whether the elimination of neurons depends on how they are wired. This kind of information can be provided by connectomics. The function of rewiring could also be studied in HVC by investigating how the growth and retraction of branches of neurons depends on their connections with other neurons.

  I’ve outlined a plan for finding synaptic chains in the HVC connectome and cell assemblies in the CA3 connectome. I’ve called this “reading memories” from a connectome. More precisely, I’ve proposed a way of analyzing connectomes to guess activity patterns that are replayed during recollection of a memory. But let me emphasize: That’s not the same as knowing what the memory means. By analyzing the HVC or CA3 connectomes, we won’t know what the bird’s song sounds like, or what’s in the videos that were previously seen by a human research subject. We might call this the reading of an “ungrounded” memory, one that is divorced from its meaning in the real world.

  I already proposed one way of grounding the memory, which is to measure HVC activity in birds as they sing, or CA3 activity in humans while they describe what they’re experiencing. Then each neuron could be placed in correspondence with a particular movement or reported idea. This sort of approach uses measurements of spiking in a live brain to ground memories read out after the brain is dead. It’s the only approach possible in the near future, as long as we can find only partial connectomes from small chunks of brain.

  In the long term, though, I expect that we will be able to find connectomes of entire dead brains. Then it may become possible to ground memories without measuring spiking in live brains. To do this, we’d have to figure out, for example, whether a CA3 neuron is selectively activated by Jennifer Aniston or some other stimulus. Could this be possible by analyzing the pathways that bring information from the sense organs to the CA3 neuron?

  It might be, if we employ the hypothesized rules of connection for perceptual neurons—for example, “A neuron that detects a whole receives excitatory synapses from neurons that detect its parts.” The Jennifer Aniston neuron might receive inputs from a “blue-eye neuron,” a “blond-hair neuron,” and so on.

  For now, researchers are starting to test this part–whole rule by combining measurements of spiking with connectomics in animals. The first step is to determine the functions of neurons in perception by measuring their spiking in response to various kinds of stimuli, as in the Jennifer Aniston experiment. This is done as described earlier, by staining the neurons so that they blink when active, and observing the neurons through a light microscope. Then researchers image this particular chunk of the brain using an electron microscope to discover how the neurons are connected. Kevin Briggman and Moritz Helmstaedter have accomplished this feat with retinal neurons, working with Winfried Denk. Studies of neurons in the primary visual cortex have been performed by Davi Bock, Clay Reid, and their collaborators. This approach, as it develops, will make it possible to see whether there are in fact connections between neurons that detect parts and wholes.

  In the coming years the part–whole rule of connection will be tested in this way. For the sake of discussion, let’s suppose that the rule is true, and speculate about how we could use it to read connectomes. The driving idea behind the rule is that a neuron stands on the shoulders of other neurons. We could start by applying the rule to the neurons near the bottom of the hierarchy and guess which stimuli they used to detect. These are the neurons just one step away from the sense organs. Then we could move step by step up the hierarchy, each time guessing the stimuli that neurons detect from the part–whole rule. Eventually we might reach the top of the heap—CA3 neurons—and guess which stimuli used to activate them in the live brain. (A neuron that receives connections from neurons that detect floppy ears, sad brown eyes, wagging tail, and loud bark—that’s the neuron that detected your great-great-grandma’s dog. )

  Reading memories from dead human brains might sound cool—you could certainly imagine an entertaining movie being built around this plot device—but it’s too far off to be considered seriously as an important practical application of connectomics. What I’m proposing instead as a basic research challenge is to decode the HVC connectome. It would be a way of improving our understanding of how the brain’s function depends on the connections between its neurons.

  I’ve discussed several ways of analyzing connectomes: carving them into brain regions, carving them into neuron types, and reading memories from them. These approaches may seem quite different, but all can actually be viewed as the formulation of rules of connection governing neurons. Each approach in the list is progressively more accurate at predicting connections, because its rules are based on more specific neuronal properties.

  For example, carving the avian brain into regions would yield coarse rules, such as “If two neurons are in HVC, they are likely to be connected to each other.” It’s certainly true that a connection between two HVC neurons is more likely than a connection between an HVC neuron and, say, a neuron in a visual region called the Wulst, which doesn’t happen at all. Nonetheless, this rule would still be lousy at predicting whether two arbitrary HVC neurons are connected, as this turns out to be quite improbable too.

  To make the rule more accurate, it might help to divide HVC into multiple neuron types. I didn’t mention it before, but our previous discussion was actually specific to just one type of HVC neuron, the one that sends axons (“projects”) to RA. This neuron type is of special interest because it generates the kind of sequential spiking characteristic of a synaptic chain. We could use it to formulate a revised rule: “If two HVC neurons both project to RA, they are likely to be connected to each other.” This more specific rule could well be more accurate.

  Even better would be to make the rule depend on the spike times of the neurons during song: “If two HVC neurons both project to RA, and their spike times during song are one after the other, they are likely to be connected to each other.” If the synaptic chain model is correct, then this rule would be highly accurate at predicting connections.

  If we really want to understand how the brain works, we need this third kind of rule, which depends on functional properties of neurons as determined by measurements of spiking. The coarser rules of connection, which depend on region or neuron type, get us only part of the way there. Knowing the regional connections that lead from HVC to the syrinx tells us why HVC neurons have functions related to song. But that’s not enough for elucidating why different HVC neurons spike at different times during song.

  Likewise, knowing regional rules of connection might tell us why the Jennifer Aniston and Halle Berry neurons do similar things—both are activated by visual stimulation—but no fan would say that they do exactly the same thing. We’d like to know why the Jennifer Aniston neuron responds specifically to Jen and not Halle, and vic
e versa. For this we need something like the part–whole rule of connection, which again depends on the functional properties of neurons.

  In the most general sense, decoding connectomes means reading out the roles played by neurons not only in memories but also in thoughts, feelings, and perceptions. If we can succeed at decoding, we’ll know that we’ve finally found rules of connection precise enough for understanding how the brain works. And then we’ll be ready to return to the question that we started with, the one that motivates this book: Why do brains work differently?

  12. Comparing

  In elementary school my friends and I tried not to gawk at identical-twin classmates, but we couldn’t help staring as we strained to tell them apart. Photos of Siamese twins were even more riveting. We looked at them long and hard while flipping through beat-up copies of the Guinness Book of World Records. Twins just seemed spooky, though we weren’t sure exactly why.

  Native American and African myths are full of stories about twins. The Navajo people trace their ancestry to the goddess Changing Woman. Impregnated by sunbeams, she bore twin sons named Monster Slayer and Born for Water. They grew up in twelve days, traveled to find their father, the Sun, and went on to engage in deadly combat with giants and monsters.

  Many more twins figure in the world’s legends and literature. Fraternal twins have always seemed special, and identical twins perhaps even magical. Why do we feel that way? For one thing, identical twins assault our bedrock assumption that every human being is unique; we’re unsettled by their alikeness. But we’re also fascinated by the slight differences that are visible if we look closely.

  In Greek myths, twins were often the offspring of one mother but two different fathers, one divine and the other mortal, which explained the twins’ different natures and fates. Today we know that we can account for those differences by pointing to the genomes of fraternal twins, who share only half of their genes. Identical twins, however, look almost indistinguishable from each other because of their duplicate genomes. I mentioned this claim about identical twins earlier when discussing the genetics of autism and schizophrenia, but it needs some qualification. Recent genomic studies have demonstrated that tiny deviations in DNA sequence arise during the process of twinning, the divergence of a fertilized egg into two embryos. These deviations might explain why identical twins look slightly different, and perhaps why they don’t think and act in exactly the same way. But genes do not fully explain mental aspects that depend on learning. Even for twins who remain conjoined (the term that has replaced Siamese) instead of being surgically separated, life experiences do not match exactly. Such twins are literally inseparable, but their memories are not identical.

  According to connectionist thinking, identical twins have different memories and minds chiefly because their connectomes differ. Many people have wondered what it would be like to have a twin sibling. Sometimes I fantasize about a mad scientist creating my “connectome twin,” a person with a brain that is wired exactly like mine. Would I be enraptured to meet him? Would my girlfriend grow jealous of our close relationship, complaining about yet another proof of my narcissistic tendencies? I suppose I could confide anything to my twin, who would be guaranteed to understand me. Then again, maybe it would be boring to pour out my problems to someone who thinks in exactly the same ways.

  And what if, after a week of getting to know each other, we were kidnapped by a team of crazed gunmen? Let’s say they decide to shoot one of us and send the body along with the ransom note, as proof of abduction. Should I fear being shot, or should I be altruistic and volunteer to take the bullet? Maybe it doesn’t matter, as all my memories and personality will survive in my twin even if I die, and vice versa. But wait. A week has passed since the mad scientist breathed life into my replica. Our connectomes have been changing since then. They diverged from the first instant after duplication, so our minds are no longer identical.

  Luckily I’ll never be forced to engage in the head-scratching required to solve this distressing philosophical dilemma. We won’t be seeing human connectome twins any time soon. But what about worms? I referred in the Introduction to “the” connectome of C. elegans, implying that any two worms are connectome twins. But is this really true? Certainly the neurons are identical, so we should be able to take two connectomes, match up their neurons one to one, and check to see whether the connections are the same.

  Such a comparison has never been done in its entirety, because it would require two complete C. elegans connectomes, and finding just one was difficult enough. David Hall and Richard Russell took the shortcut of comparing partial connectomes from the tail ends of worms. They didn’t find a perfect match. If two neurons were connected by many synapses in one worm, in all likelihood they were also linked in another worm. But if two neurons were connected by a single synapse in one worm, there might be no synapse at all between them in another.

  What caused these variations? The worms had been highly inbred in the laboratory for many generations, by exaggerating the methods used to create purebred dogs and horses. That made all lab worms genomic twins, but a few differences did remain in their DNA sequences. Could these differences account for connectome variation? Or is such variation a sign that worms learn from experience? Or perhaps the variation is due neither to genes nor to experience, but rather to random sloppiness as the worm’s neurons wire together during development. Any of these explanations could be true, but more research is needed to test them.

  Did connectome variation affect behavior, giving worms distinctive “personalities”? Hall and Russell did not study this question, so we don’t know. Their worms were inbred but otherwise normal. Other researchers have identified genetically defective worms that also behaved abnormally. Finding their connectomes has yet to be done, but after that is accomplished, it should be straightforward to compare the connectomes of abnormal and normal worms if the neurons can be placed in one-to-one correspondence. If there are missing neurons, or additional neurons, then matching the connectomes will be a bit more difficult; still, it should be possible. Research of this type will take off as it becomes easier to find C. elegans connectomes.

  Comparing the connectomes of animals with big brains will be much more challenging. As I mentioned in the Introduction, big brains vary greatly in number of neurons, so there’s no way of placing neurons in one-to-one correspondence. Ideally, we would find some way to match up neurons with similar or analogous connectivity. According to the connectionist mantra, such neurons would also have similar functions, like a Jennifer Aniston neuron in one brain and a Jennifer Aniston neuron in another. The correspondence would not be one to one, as the number of Jennifer Aniston neurons might vary across individuals. (Some people might even lack Jennifer Aniston neurons altogether, having never had the benefit of exposure to her.) This kind of matching would require sophisticated computational methods yet to be developed.

  An alternative approach is to compare connectomes after coarsening them. We could define reduced connectomes for brain regions or neuron types, as described earlier. Since these are expected to exist in all normal individuals, it should always be possible to place them in one-to-one correspondence. Comparing reduced connectomes of big brains would be as simple as comparing worm connectomes.

  Previously I argued that regional or neuron type connectomes would be insufficient for understanding our memories, the most unique aspect of our personal identities. But other distinguishing mental characteristics, such as personality, mathematical ability, and autism, seem more generic than autobiographical memories. These properties of minds might be encoded in reduced connectomes.

  ***

  In principle, we could find reduced connectomes by carving up neuronal connectomes. Even for rodent brains, however, finding an entire neuronal connectome is a long way off. An alternative is to develop shortcut methods that find reduced connectomes directly, without requiring neuronal connectomes. Such methods would be technically easier, as they would not require col
lecting so much image data.

  Some neuroscientists would like to use light microscopy to find connectomes for neuron types—an approach pioneered by Cajal, who concluded that two neuron types were connected when one type extended axons into a region occupied by dendrites of the other type. His approach was piecemeal, but with modern technologies it could be applied systematically. To find a neuron type connectome, though, we would have to combine neurons imaged in many brains, as light microscopy can reveal only a small fraction of a single brain’s neurons. Therefore, this approach might be less useful for finding differences between individual brains.

  Light microscopy could also be used to map regional connectomes. To apply this approach to the cortex, we must map a specific part of the cerebrum that I haven’t discussed yet—the cerebral white matter. Recall that the cerebrum atop the brainstem resembles a fruit on a stalk. The “peel” of the fruit is the cortex, otherwise known as the gray matter. Cutting the fruit open reveals its “flesh,” called the white matter, as shown in Figure 48.

  Figure 48. Gray versus white matter of the cerebrum

  The distinction between gray and white matter was known in antiquity, but their fundamental difference became clear only after the discovery of neurons. The outer gray matter is a mixture of all parts of neurons—cell bodies, dendrites, axons, and synapses—while the white matter contains only axons. In other words, the inner white matter is all “wires.”

  Most white-matter axons come from neurons in the surrounding cerebral cortex. They belong to pyramidal neurons, which constitute about 80 percent of all cortical neurons. Earlier I mentioned that this neuron type has a cell body with a triangular or pyramidal shape, and an axon that travels a long distance from the cell body. Let’s refine the picture here. The apex of the pyramid points toward the exterior of the brain. The axon comes straight out of the base of the pyramid, perpendicular to the cortical sheet, and plunges into the white matter, as Figure 49 shows.