Mindlessly mapping the brain
Seeking clarity for the connectome
As a kid I loved dot-to-dot puzzles. Briefly. Until sense kicked in and I realised you could just see the picture by looking at the dots (yay, another donkey). But, for a brief while, joining dots with pencil lines and seeing the picture emerge was deeply satisfying to my tiny mind. In the sort of metaphorical leap beloved of over-thinking biographers, I’m going to claim that this love of dot-to-dot puzzles led me inexorably to a life as a neuroscientist.
For you see the human brain is the surely the most complex dot-to-dot puzzle ever conceived. We know there are about 87 billion neurons in one human brain; but we do not how each of them are wired to all the others. Indeed, we don’t even know how one of them is wired to all the others. And sadly evolution did not helpfully label each neuron with a number. The scale of the puzzle is mind-blowing: 87 billion dots, and about 1 trillion lines to draw between them. By knowing the wiring, we may know all the routes by which information flows in, through, and out of the brain. So how do we solve this most magnificent of dot-to-dot puzzles?
Enter connectomics. Connectomics is the automated, mechanised, high-volume tracing of the connections between neurons. Some believe connectomics will kick-start a true understanding of how brains works; others believe it a colossal waste of money and talent. I worry that we’ve not thought hard enough about what questions the resulting wiring diagram — the connectome — will answer.
Industrialised science suits technical challenges. The melding of many minds and much money into a single entity focused on a single, technical goal: land on the moon; smash these particles; make fusion work; map the genome. But industrialised projects in neuroscience are heavily criticised, for we lack scientific insight into how the brain works — and without the science in place, we have not (yet) defined any technical goals. We cannot automate the understanding of memory, nor mechanise a theory of emotion.
Connectomics is seemingly immune to such criticisms. Manually tracing the connections between neurons is an exhausting, difficult, time-consuming process. Neurons are microscopic; the connections between them smaller still. The specialised electron microscopes we need to definitively identify a single connection are highly expensive, need intense training to make sense of the images, and are mind-numbingly slow. Tracing one connection can take hours. A single neuron in the cortex of a mouse receives 10,000 connections. Tracing a small circuit of a few hundred neurons by hand alone would take many lifetimes. It’s a problem crying out for automation.
Heeding this call, projects like MICrON and EyeWire have launched themselves into the fray with gusto. After all connectomics is ripe for industrialisation. It has a defined goal: map the connections between neurons. It is a technical challenge: build better kit and algorithms to find the connections.
The well-defined goal of connectomics is an illusion. In reality, there are many possible goals. And which goal we choose depends on the scientific question we want to answer. Do we want a connectome or the connectome? Do we want the adult bauplan or the developmental arc? Do we want the connectome constant to all of a species, or the variation between them?
Each answers a different set of scientific questions; but which we choose will absorb incredible quantities of time and money. So we must choose with care.
A connectome or the connectome? A connectome is a total reconstruction: every connection between every neuron in a single animal. The connectome is the set of connections that are true to every member of that species (that, likely, differ between the sexes). A connectome means we can answer questions specific to that creature, and hope they generalise to other creatures of the same species. The connectome means we can answer questions general to the species, and hope they apply to individual creatures. They need not, of course.
We have one complete connectome, the 279 neurons of the nematode worm C. Elegans (for the pedants: its hemaphrodite form has 302 neurons, of which 279 form a single connected network). Texts uncountable have discussed this wiring diagram as the epitome of “a” connectome. Strictly speaking this is not true. Heroic as the original 1986 paper was, it missed out some connections; these were completed in 2011. The connectome we have is then actually an amalgam of two different creatures. What will happen if we replicate this connectome? Are there really all the exact same number of connections between the exact same neurons in every C. Elegans? It seems that each of the neurons is genetically specified, and in such a minuscule nervous system it is possible that each and every one of the connections is too. But would you bet your house on it?
The recent reconstruction of a maggot’s sensory circuit might give you pause before taking that bet. Here was a reconstruction of all the inputs and outputs of the 223 “Kenyon cell” neurons in one maggot. Heroic. Yet right from the off we see variation in the same animal, with 110 of these neurons on the left and 113 on the right. We cannot tell if the difference is true of all maggots — if the maggot really has an asymmetry between the left and right halves of its brain — or if this is just natural variation we happen to have seen in this particular maggot’s brain. We can’t tell this until we have reconstructed the sensory circuits of many maggots.
Only a handful of these 223 neurons had the exact same patterns of inputs from other neurons on both sides of the brain. The input from neurons activated by smells was seemingly random. And while this random input could turn out to be the best way to combine information about different smells, we cannot conclude this until we reconstruct multiple circuits and find out if these smell inputs are different between different maggots.
To put it another way, is there any scientific purpose in having just N=1 connectome?
If we want a connectome, then do we want to know the variation within the lifetime of one animal? Almost all animals with a nervous system have a baby stage in which the nervous system differs from the adult. We’ve just met a small number of neurons in a maggot, precursor to the adult fly. The maggot of this species has about 10,000 neurons; the adult fly has about 40,000 neurons. So while the general pattern of connections may be preserved — neuron type A connects to type B but not type C — the numbers of connections patently cannot be.
If we just construct the adult connectome, we can answer questions about the dynamics in the adult. But have no understanding of how the connections came to be, of how what happens to the creature during its development from baby to adult drives differences in wiring. If we just construct the baby connectome, we obviously miss the adult stage. But we cannot construct the baby and adult connectome in the same creature, due to its brain being sliced into tiny bits in the baby stage. There is no such thing as “a” connectome. So we must choose which “a” connectome we want.
If we want the connectome, do we want to know the wiring consistent across a species, or the variation in wiring within a species? The connectome of the cortex in one mouse will differ from the connectome of the cortex in a different mouse. This variation could be due to specific genetic differences, or due to experiences during development. Such variation could reflect that each mouse has fundamentally different thoughts about the world: one is mad for gruyere; the other prefers chocolate wheat hoops. Or such variation could be just inconsequential noise — the two cortices have identical dynamics despite differences in the connections, and both mice equally adore Hob-Nobs. If we cannot tell from just the wiring whether the variation is fundamental or noise, there seems little point in reconstructing only the wires.
Worse, only genetics and development cause variations in the wiring that occur on long enough time-scales that we may catch them in the act of changing. Learning creates variation on short time-scales, by changing the connections between the neurons engaged in what is being learnt. Neuromodulators create variation on even shorter time-scales, by changing the strength of connections between neurons depending on how scared, frightened, hungry, or bored the creature is. If we cannot capture the changes in wiring, there seems little point in reconstructing only the wires.
We want the wiring and the activity of the neurons at the same time. We want the specific connectome of the thing we’re recording activity from right now; both the activity of each neuron and the connections between them. Only then can we answer simple questions about whether variation in wiring matters, or about how learning changes the connections.
How do we make that happen? One way would be to record activity then reconstruct the wiring. This is tedious, even longer than just reconstruction alone. And imagine the screams of frustration echoing from the lab every time the brain that the researcher just spent weeks recording from gets smushed beyond use during the reconstruction process.Ideally we would work out the connections from just the activity itself.
If neuron A is connected to neuron B, then we should be able to see the effect of neuron A’s activity on the activity of neuron B. The problem is that we record just the infrequent electrical pulses, the spikes, that neurons A and B send to other neurons. And each of these spikes is driven by summing up the input from many tens to hundreds of other neurons. So the input from neuron A only has a tiny influence over whether neuron B sends a spike or not.
(This hasn’t stopped people from trying to work out connections by just checking whether one spike from neuron A slightly changes the chances of neuron B then sending a spike. Unfortunately this is almost unworkable without extraordinary durations of recorded activity: the change in chance of sending a spike is tiny, and a spike from neuron A followed by one from neuron B could equally be caused by neuron C having inputs to both of them. And if we have long enough recordings, then we cannot rule out that the connections have changed during the recording.)
What we need are recordings of every bit of electrical activity in the bodies of every neuron. (More precisely: we need to record the voltage in each neuron’s body). Then we could see the spikes, and see everything leading up to them — we could see the small change that a spike from neuron A causes in the electrical activity of neuron B.
There is such a solution. There exist dyes that glow according to the voltage of the neuron. So if we could video lots of neurons that have one of these dyes inside them, we can record all the voltages of all the neurons.
This is not a new solution. We’ve had these dyes since the late 70s, but until recently the dyes we had did not glow strongly enough for us to tell the difference between changes in the voltage and noise, except for the big jumps that are the spikes sent to other neurons. That’s now changed — at least in leeches and flies. And with such recordings of every flicker of voltage, we should in theory be able to tell whenever the spike from one neuron causes a tiny change in another, and say: they are connected. Then we would have the connectome of the thing we were recording from right now.
Is there then any purpose to pursuing the connectome alone?
Only connectomics can answer the fundamental question of where the connections are made on the neuron. And it’s a vital question. For you see, neurons are not dots after all. They are elegant structures, a tiny body from which sprout dendrites — twisted, tentacular outgrowths, stretching to capture the inputs from other neurons.
Where an input from neuron B lands on neuron C can make a dramatic difference to how it contributes to neuron C. A connection falling on the body of neuron C can powerfully affect its activity. A connection falling far away from neuron C’s body, up towards the tips of its dendrites, will have little effect by itself on the activity of neuron C. But a cluster of such connections together on the tip can cause a local pulse of activity that sweeps down the dendrite to the body of neuron C, triggering it in turn to send a spike of activity out to other neurons.
An inhibitory input on a bit of dendrite can prevent the excitatory inputs along the same bit of dendrite from transmitting their signal down to the body of the neuron. Indeed, theorists have shown how different orderings of excitatory and inhibitory connections along a bit of dendrite create different local computations. And shown how the spread of connections, whether spread across many bits of dendrite or clustered on one bit, also create different local computations. Computations such as this AND that; this OR that; this AND NOT that; only this OR only that. Logic, in other words.
Large recordings of the activity of many neurons cannot see these local computations. They record only the activity at the neuron’s body, only what it transmits to other neurons. Perhaps then the true scientific purpose of connectomics is to let us understand the computations that could be done by each individual neuron — computations largely invisible to short recordings of the electrical activity of each neuron’s body. Perhaps connectomics exists so we may find out the computational logic of each brain region. And find out that our cortex contains not just 17 billion neurons, but 17 billion individual computers.