Connectome

by Sebastian Seung

  [>] computer would be necessary: Actually, we could still do it by hand if the chain were perfect. But if there are some “inappropriate” connections, such as synapses directed backward, finding a chain becomes more difficult and requires a computer (Seung 2009). Unscrambling neurons is an example of a problem called “graph layout” by computer scientists.

  [>] resemble blinking lights: These stains fluoresce when illuminated, like a sticker that glows in the dark when illuminated by black light. The amount of fluorescence varies with calcium concentration, which in turn is modulated by spiking.

  [>] might not be able to order: Actually, this outcome could leave ambiguity. Perhaps a sequential ordering exists, but our unscrambling algorithms are too poor to find it. Computer scientists will have to work hard to make sure that their algorithms are good enough to find any ordering if it exists.

  [>] go backward or jump: Even if there turn out to be some “inappropriate” connections that violate the sequential ordering, we could still say that the connectome is an approximation to a synaptic chain. But if there were too many such connections, then we’d have to say that the chain is a bad model and cannot explain why the network generates sequential activity.

  [>] HVC neurons in young males: Jun and Jin 2007; Fiete et al. 2010.

  [>] reconnection also plays a role: This was suggested by Jun and Jin 2007.

  [>] Kevin Briggman: Briggman, Helmstaedter, and Denk 2011.

  [>] Davi Bock: Bock et al. 2011.

  [>] great-great-grandma’s dog: What about grounding the memory of the bird’s song? If we found an entire bird connectome, we could examine the pathways from each HVC neuron to the vocal muscles. These pathways are thought to transform the abstract sequence in HVC into the specific motor commands required to make sounds. (This transformation appears to be learned by practice too.) Analysis of the connections in these pathways might make it possible to decode the movement signaled by each HVC neuron. This method would require that we identify rules of connection for neurons related to motor control, which are analogous to the part–whole rule for perceptual neurons. In general, grounding memories requires that we trace pathways all the way from the center of the brain to the sensory and motor periphery.

  [>] rules of connection: Rules of connection can be mathematically formalized as probabilistic models of graph generation based on latent variables at the nodes of the graph (Seung 2009).

  [>] quite improbable too: Mooney and Prather 2005.

  12. Comparing

  [>] Native American and African myths: Davis 2005.

  [>] bedrock assumption: Even more disconcertingly, identical twins challenge the more sweeping axiom that everything is unique—human, animal, or inanimate object. This axiom underlies the lovely claim that no two snowflakes are alike, and may have been behind the animistic beliefs of primitive societies that all objects have souls. Because of mass production by factories, we have grown blasé about material objects that look almost indistinguishable. Such instances were much rarer in the preindustrial world, so I suspect that twins appeared even more magical to our primitive ancestors than they do now. But such thoughts are less relevant for connectomics than fodder for nanotechnologists who promise to make material objects that are truly identical, down to the placement of individual atoms (see, for example, Drexler 1986).

  [>] deviations in DNA sequence: Machin 2009 discusses both genetic and epigenetic differences between identical twins.

  [>] two complete C. elegans connectomes: As mentioned earlier, the researchers actually pieced together the connectome using images drawn from several worms. The published C. elegans connectome is a mosaic, not a unified representation of an individual worm’s nervous system. So we don’t have even one complete connectome of an individual worm, much less two.

  [>] David Hall and Richard Russell: Hall and Russell 1991.

  [>] purebred dogs and horses: Laboratory animals are generally inbred this way to ensure that they are genetically almost identical, which is supposed to make experiments more repeatable. It’s well-known that inbreeding can increase the likelihood of having two defective copies of a gene, and “recessive” disorders are governed by a “two strikes and you’re out” rule. This is why many dog breeds have genetic disorders and why European royalty suffered from hemophilia. Since inbreeding probably makes laboratory animals “dumber,” research on them might not be applicable to their wild counterparts.

  [>] sophisticated computational methods: The most basic computational problem of genomics is finding a matching or alignment between two DNA sequences. This is solved by fast approximations to dynamic programming, a formalism first developed in the 1940s and 1950s for solving problems with a one-dimensional or tree structure. Solving the analogous matching problem for two connectomes will be an important computational challenge for connectomics, and is much harder than aligning genomes. Determining whether two connectomes are the same is known as the graph isomorphism problem, for which no polynomial time algorithm is known. Determining whether one connectome is part of another is known as the subgraph isomorphism problem, which is NP-complete.

  [>] known in antiquity: Gray and white are not the natural colors of living brain tissue, which is pinkish, but rather the colors of preserved brain tissue.

  [>] is all “wires”: As noted by Kostovic and Rakic 1980, Cajal already observed that there are exceptions to this rule, known as “interstitial neurons.”

  [>] straight out of the base: This mental picture is a bit confusing, because the cell body looks like an arrowhead pointing in the opposite direction of information flow along the axon.

  [>] 150,000 kilometers: This crude estimate assumes that the density of axons throughout the cerebral white matter is the same as in the corpus callosum, or 380,000 axons per square millimeter (Aboitiz et al. 1992). The estimate also makes use of the total volume of white matter, which is 400 cubic centimeters (Rilling and Insel 1999).

  [>] Myelination speeds up: The fat in myelin serves as an insulator that prevents leakage of electrical currents out of the axon. This has the effect of boosting the speed at which electrical signals propagate. Electrical signals travel at top speed in myelinated axons, ten or more times faster than in unmyelinated axons. Myelin sheaths are outgrowths of non-neuronal, or glial, cells. Schwann cells myelinate PNS axons, and oligodendrocytes myelinate CNS axons.

  [>] axon enters and branches: If the axon doesn’t branch in a region, it’s probably passing through without making synapses.

  [>] almost completely unexplored: Historically, the white matter of animal brains has been studied by the method of tracer injection. When certain substances are injected into the brain, they are taken up by neurons at that location and transported along axons to other brain regions. By visualizing the destination of such tracer substances, it is possible to identify the regions connected to the injection site. Data from such experiments was compiled in Felleman and Van Essen 1991 to chart the regional connectome of the monkey brain shown here (Figure 51). The Brain Architecture Project, led by Partha Mitra, is systematically applying tracer injections with the goal of producing a complete map of long-range connections in the rodent brain. But the tracer must be injected while the brain is still alive, as its transport depends on active processes in living neurons. Therefore tracer injection is an invasive technique, and is employed only with animal brains. It does not work at all with postmortem human brains. (Certain lipophilic dyes don’t depend on active transport, but are difficult to use as tracers in postmortem brains because they travel so slowly.) My proposal of serial light microscopy does not require injection of tracers. Instead of staining just a small bundle of axons, all myelinated axons in the white matter are stained and imaged. This method could potentially be applied to a postmortem human brain. Furthermore, its high spatial resolution prevents the ambiguities that plague diffusion MRI and naked-eye dissection. My proposal is an example of dense reconstruction, which extracts a complete map from a single brain, rather than ag
gregating data from many brains.

  [>] Diffusion MRI is an exciting: This method works by measuring the direction dependence of the speed of diffusion of water molecules in the brain. Diffusion along the axis of axons is faster than in the perpendicular direction.

  [>] sparking revisions: Friederici 2009. Friederici 2009.

  [>] complementary methods: We’ve focused on comparing connectomes of different individuals using microscopy. This provides snapshots of connectomes at moments in time. Comparing such snapshots can tell us something about how interventions change the brain. (Recall that Rosenzweig’s experiments on environmental enrichment and Antonini and Stryker’s experiments on monocular deprivation of V1 relied on comparisons between different animals or populations of animals.) But we would also like to compare the connectomes of a single individual at different times. Unfortunately, there is currently no good way of doing this. A noninvasive method like MRI can follow the evolution of a connectome over time but cannot deliver the neuronal resolution of microscopy. There are ways of improving the snapshots of microscopy by highlighting changes to the connectome, however. There now exist staining methods for making recently strengthened synapses visible, as well as methods that do the same for newly created neurons. It’s important to invent ways of labeling synapses that were recently created, as well as locations where synapses were recently eliminated. With such images, one could not only quantify the total amount of synapse creation and elimination but go much further, because every created and eliminated synapse would be seen in the context of an entire network. We would know exactly how synapse creation and elimination changed the organization of connectivity, as opposed to a coarse measure like total number of synapses. This would enable us to detect even subtle connectome changes, as well as figure out whether they are causally related to learning.

  [>] brains of the deceased: I mentioned earlier that the two-photon microscope can be used to observe neurons in living brains. This requires opening or thinning the skull, however. Also, it works only for neurons near the surface of the brain, unless the viewing is done through an optical fiber inserted deep inside, an even more invasive procedure. And it can visualize only neurites that are sparsely labeled.

  [>] present special problems: The brains may not be well preserved after death; they may suffer from other abnormalities that are not relevant to the mental disorder in question, such as injury caused by stroke; and they may have been changed by drugs if the deceased person was treated for the mental disorder.

  [>] into the genomes of animals: Nestler and Hyman 2010. Some mental disorders are associated with deletions of parts of the genome, and researchers can create these deletions in animal genomes also.

  [>] simian immunodeficiency virus: According to one theory, HIV originated when SIV mutated and jumped from monkeys to humans.

  [>] numbers of plaques and tangles: Oddo et al. 2003.

  [>] “unbiased, hypothesis-free manner”: Lander 2011.

  [>] stroke of insight: Other times, it’s the available tools of measurement that motivate the hypothesis. For example, Galton hypothesized that intelligence was related to head size mainly because he was able to measure head size, not because this was a great hypothesis.

  13. Changing

  [>] Der Freischütz: This literally means “The Freeshooter,” but it’s typically translated as “The Marksman.”

  [>] suffering of millions of people: Bosch and Rosich 2008.Bosch and Rosich 2008.

  [>] inspired by Weber’s popular opera: Strebhardt and Ullrich 2008. Ehrlich also invented the idea of receptor molecules.

  [>] last-ditch measure: The current practice of psychosurgery and the history of the “frontal lobotomy,” which earned the Portuguese physician Egas Moniz a Nobel Prize in 1949, are described in Mashour, Walker, and Martuza 2005. While lobotomy could reduce the symptoms of psychosis, it also mentally crippled the patients. It became apparent that the side effects were worse than the disease. Because of psychosurgery’s abuses, many regard the prize to Moniz as an embarrassment to the Nobel committee. However, some historians argue that psychosurgery was justifiable in an age before antipsychotic drugs, when the only alternative was confinement in a mental institution. Much of the infamy of the procedure was due to American physician Walter Freeman, who developed a version of the procedure that he called the “transorbital leucotomy.” In his gruesome technique—nicknamed the “ice pick lobotomy”—a mallet was used to drive a sharp instrument resembling an ice pick past the eye through the eye socket into the brain. Moving the tip back and forth destroyed tissue in the frontal lobe. Freeman’s innovation made the procedure so quick and easy that non-surgeons and even non-physicians could perform it.

  [>] surplus or deficiency of neurotransmitter: Schildkraut 1965.

  [>] effects of fluoxetine on the four R’s: Hajszan, MacLusky, and Leranth 2005 found an increase in dendritic spine density, a sign of synapse creation. Wang et al. 2008 demonstrated increased dendritic growth of newborn neurons. The extensive literature on neuron creation in the hippocampus and its role in depression is reviewed in Sahay and Hen 2008.

  [>] specifically target connectomes: Other treatments for brain disorders involve manipulating neural activity. In electroconvulsive therapy (ECT), shocks administered through scalp electrodes induce epileptic seizures. ECT is far from a magic bullet, as the seizures spread unselectively over the brain, yet for some unknown reason ECT can alleviate symptoms of depression and other mental disorders. Better-targeted electrical stimulation can be performed using electrodes that are surgically implanted inside the brain. Symptoms of Parkinson’s disease, for example, can be relieved by stimulating parts of the basal ganglia. Some researchers are developing even more precise therapies based on optogenetics, the optical stimulation of activity in a single neuron type that has been genetically altered to be sensitive to light. Like altering neurotransmitter levels, manipulating neural activity may sound completely different from promoting connectome change, but it’s not. For example, the seizures of ECT may change the connectome through Hebbian plasticity, and it’s quite possible that such changes are responsible for its therapeutic effects (and for side effects like amnesia).

  [>] supplemented by training regimens: It’s intuitively plausible that combining drugs and “talk therapy” might be more effective than either alone. Evidence supporting this idea for the treatment of depression is presented by Keller et al. 2000.

  [>] alive but damaged: Lipton 1999.

  [>] “gene therapy” for Parkinson’s: Yamada, Mizuno, and Mochizuki 2005; Mochizuki 2009.

  [>] degeneration in neurons: Some researchers report that neurons “die backwards” in many diseases. In other words, degeneration affects the synapses and tips of axons first, and then moves backward along the axon toward the cell body. The collapse of the axon in turn might trigger the neuron to initiate the suicidal mechanisms of programmed cell death. See Coleman 2005; Conforti, Adalbert, and Coleman et al. 2007.

  [>] connections are lost: Selkoe 2002.

  [>] before the first onset: Baum and Walker 1995.

  [>] such as lizards: Lledo, Alonso, and Grubb 2006.

  [>] fingertips grow back: Illingworth 1974.

  [>] Injury naturally activates: Carmichael 2006.

  [>] divert them from: Zhang, Zhang, and Chopp 2005.

  [>] survive in recipients’ brains: Mendez et al. 2008.

  [>] whether the transplants actually: Olanow et al. 2003.

  [>] “reprogrammed” to divide: This is known as a patient-derived induced pluripotent stem cell (iPSC).

  [>] skin cells of Parkinson’s: Soldner et al. 2009.

  [>] Whether created naturally: Zhang, Zhang, and Chopp 2005; Buss 2006; Lledo 2006.

  [>] added by transplantation: Brundin 2000.

  [>] molecules that promote plasticity: Murphy and Corbett 2009.

  [>] grow new axonal branches: Carmichael 2006.

  [>] natural molecular processes: Carmichael 2006. Reweighting might a
lso be important for recovery from stroke, by unmasking previously nonfunctional pathways through strengthening of their synapses. Another type of change can unmask pathways, which should perhaps be included in reweighting. This is a change in the threshold for producing a spike. (In the weighted voting model, the threshold specifies the margin between “yes” and “no” votes required from presynaptic “advisors” for a neuron to fire an action potential.) Lowering thresholds can unmask pathways by making neurons more excitable, that is, less choosy about when to spike. This could be especially important for recovery from stroke, because the death of neurons reduces the number of advisors for the surviving neurons. They may receive less “yes” votes, so they will not spike unless their thresholds are lowered.

  [>] effect on learning and memory: Nehlig 2010.

  [>] deprived of cigarettes: Newhouse, Potter, and Singh 2004.

  [>] nine out of ten: Kola and Landis 2004.

  [>] a billion dollars: Morgan et al. 2011. These estimates are uncertain because such financial information is proprietary. Also, pharmaceutical companies have an interest in overstating their costs, to answer criticisms that they are greedily overcharging for their products.

  [>] swept through the psychiatric hospitals: The serendipitous history of antipsychotic drugs is reviewed in Shen 1999. The first-generation, or “typical,” drugs were created by varying the molecular structure of chlorpromazine. The second-generation, or “atypical,” drugs have more diverse molecular structures.

  [>] first antidepressant medications: Lopez-Munoz and Alamo 2009. Iproniazid was the first of the monoamine oxidase inhibitors, and imipramine kicked off the discovery of many tricyclic antidepressants.

  [>] golden age of the 1950s: Since the 1950s, the only major success story has been fluoxetine, which was discovered by rational means rather than serendipity. From studies of the first antidepressants, scientists had formulated the theory that depression had something to do with the brain system that secretes the neurotransmitter serotonin. In the early 1970s, the company Eli Lilly searched for molecules that acted on the serotonin system but lacked the side effects of the tricyclic antidepressants like imipramine. The search turned up fluoxetine, which was finally approved by the U.S. government in 1987. See Lopez-Munoz and Alamo 2009.