Apr 12 2011
Connectomics
There are approximately 100 billion neurons in the adult human brain. Each neuron makes thousands of connections to other neurons, resulting in an approximate 150 trillion connections in the human brain. The pattern of those connections is largely responsible for the functionality of the brain – everything we sense, feel, think, and do. Neuroscientists are attempting to map those connections – in an effort known as connectomics. (Just as genomics is the effort to map the genome, and proteomics is mapping all the proteins that make up an organism.)
This is no small task. No matter how you look at it, 150 trillion is a lot of connections. One research group working on this project is a team led by Thomas Mrsic-Flogel at the University College London. They recently published a paper in Nature in which they map some of the connections in the mouse visual cortex.
What they did was to first determine the function of specific areas and neurons in the mouse visual cortex in living mice. For example, they determined which orientation they are sensitive to. In the visual cortex different neurons respond to different orientations (vertical vs horizontal, for example). Once they mapped the directional function of the neurons they then mapped the connections between those neurons in vitro (after removing the brain). They found that neurons made more connections to other neurons with the same directional response, rather than neurons with sensitivity to different (orthogonal) directions.
The techniques used allowed them to make a map of connections in part of the mouse visual cortex and correlate the pattern of those connections to the functionality of that cortex. The resulting connectomics map is still partial and crude, but it is a step in the direction of reproducing the connections in the brain.
One way to think about these kinds of techniques is that they promise to take us a level deeper in our understanding of brain anatomy. At present we have mapped the mammalian, and specifically human, brain to the point that we can identify specific regions of the brain and link them to some specific function. For the more complex areas of the brain we are still refining our map of these brain modules and the networks they form.
To give an example of where we are with this, clinical neurologists are often able to predict where a stroke is located simply by the neurological exam. We can correlate specific deficits with known brain structures, and the availability of MRI scanning means that we get rapid and precise feedback on our accuracy. We are very good at localizing deficits of strength, sensation, vision, and also many higher cortical functions like language, calculations, visuo-spatial reasoning, performing learned motor tasks, and others.
But we are still a long way from being able to reproduce the connections in the brain in fine detail – say, with sufficient accuracy to produce a virtual brain in a computer simulation (even putting aside the question of computing power). And that is exactly the goal of connectomics.
Along the way these research efforts will increase our knowledge of brain anatomy and function, as we learn exactly how different brain regions connect to each other and correlate them with specific functions. Neuroscientists are still picking the low-hanging fruit, such as mapping the visual cortex, which has some straightforward organization that correlates with concepts that are easy to identify and understand – like mapping to an actual layout of the visual field, and to specific features of vision such as contrast and orientation.
For more abstract areas of the brain, like those that are involved with planning, making decision, directing our attention, feeling as if we are inside our own bodies, etc. connectomics is likely to be more challenging. Right now we are mainly using fMRI scans for these kinds of studies, which has been very successful, but does not produce a fine map of connections (more of a brain region map). Also, the more abstract the function the more difficult it will be to use mice or other animals as subjects, and when using humans you cannot use certain techniques, like removing the brain and slicing it up (at least not on living subjects).
The utility of this kind of research is a better understanding of brain function, and all that flows from that. We cannot anticipate all the potential benefits, and the most fruitful outcome may derive from knowledge we are not even aware we are missing.
This also plays into the research efforts to create a virtual representation of the human brain, complete with all the connections. This is one pathway to artificial intelligence. Estimates vary, but it seems like we will have the computer power sometime this century to create a virtual human brain that can function in real time, and then, of course, become progressively faster.
I should note that the connections among neurons in the brain are not the only feature that contributes to brain function. The astrocytes and other “support” cells also contribute to brain function. There is also a biochemical level to brain function – the availability of specific neurotransmitters, for example. So even if we could completely reproduce the neuronal connections in the brain, there are other layers of complexity superimposed upon this.
In any case, this is fascinating research and it will be nice to see how it progresses over the next few decades.
19 Responses to “Connectomics”
Leave a Reply
You must be logged in to post a comment.
“But we are still a long way from being able to reproduce the connections in the brain in fine detail – say, with sufficient accuracy to produce a virtual brain in a computer simulation (even putting aside the question of computing power). And that is exactly the goal of connectomics.”
This seems an unrealistic goal. The actual human brain does not, I assume, have all of its “fine detail” encoded in the genome. The instructions to create a brain in the embryo are presumably at a fairly high level. There is also the question of the degree to which the complexity of the brain is in its hard wiring as opposed to its “software”. i.e. The connections are not formed completely by cell construction as for example muscles are formed, but are formed after the cells are in place. e.g. In the way memories are encoded in an adult brain.
juga,
I am not sure what you point is about the genome. While you are correct, how does this affect our ability to map the brain’s connections?
It did occur to me (although I forgot to include it in this post) that we may be able to reproduce the brain by reproducing some of it’s developmental pathways, rather than directly reproducing the final product. In other words, figure out the rules for making a brain, then follow those rules.
Your other point is valid – I could have added that to my caveat at the end of the post. The pattern of connections is not everything – the strength of those connections also matters (in addition to the biochemistry).
There are a few issues that make this considerably more complicated.
First, just knowing what the connections are does not necessarily imply we know what they are doing. Both the connections themselves and the neurons receiving them are highly-non-linear, which makes them hard to analyze.
The cerebellum has a very simple structure with connections that have been known for decades, yet we still aren’t entirely sure what it even does.
Similarly, just knowing what the connections are doesn’t tell us what the connections are doing. There are a wide variety of different types of connections, dozens if not hundreds, and even within those types there are indicates that the connectionsare highly dependent on synapse geometry.
To give an example, we know that neuron in the medial superior olive in the brainstem receive excitatory and inhibitory connections, but we don’t know what the actual effect of those connections on the neurons are.
Even if we know the connections, and know what they are doing, and know the effect on the neuron, neuron responses are highly depednent on neuron geometry, and highly dependent on the distribution of ion channels around the neuron. Both are hard to do reliably, but determining the distribution of ion channels is particularly hard.
For example, the node of Ranvier, where action potentials start, has long been though to work by having a higher concentration of ion channels. But no one has been able to conclusively show that this is actually the case, and there has been a long ongoing debate in the literature, with both sides accumulating significant evidence supporting their conclusion.
And it is looking like the supposedly clear-cut picture of what each area of the brain does is not as well-understood as we once thought. Areas thought to be dedicated to visual attention now appear to work with multiple senses. The thalamus is now known to do a lot more processing on its own than was recently thought. There are even some doubts now about whether the hippocampus really deals with short-term memory or only deals with location, and this location information is necessary for memory formation.
Over the last half-century or so sine we have had the necessary techniques, we still only have a very rough picture of the ascending (sense->brain) pathways, and then only at the lower levels. We are only now just starting to look at the descending pathways, and in some cases those are more numerous than the asecending ones. So I would guess, assuming we keep roughly the same rate of progress, I would be very surprised if we had a complete picture of the pathways of any of the senses within the next century, not to mention the higher-level cortical areas.
We still have trouble accurately modeling even a single neurons. When to computer simulations of neurons, they require coupled systems massive numbers of high-order nonlinear partial differential equations. That is almost certainly the most difficult thing for a computer to do. Even simulating a neuron with just two or three pieces is currently well below real-time, and the problem grows exponentially with the number of pieces. So it may be true that we will have the computer power to run a whole-brain simulation within the next century, I don’t think it is a foregone conclusion.
That rather marvelous commentary by TheBlackCat was what makes intelligent responses worth the waiting for.
blackcat – I basically agree with your points, and I don’t think it’s incompatible with what I wrote. I just referred to modules and networks without getting into the complexity there because I have written about it before, echoing what you wrote.
But none of what you wrote makes it insolvable. I think saying that we will not solve these problems in the next century is overly pessimistic (as is assuming linear progress).
At this point, I don’t think we can predict how long it will take to fully simulate the brain. Saying we will have the raw computing power by the end of the century is longer than the most pessimistic estimates I have read, so that seems reasonable. I think we will probably have a thorough connectomics map my then also. The other layers of complexity you mention, which I alluded to, are harder to say.
I would point out that the point of this study was to correlate function with data on connections, not just to map connections.
Also, information from brain developmental biology may provide another source of information that will help reconstruct the complexity of the brain.
Steven – My point is that I’m not sure mapping the end point of a brain is the best approach if the goal is to be able to simulate a brain. When a real brain is created, it isn’t created from a map of the end point. I don’t know but I imagine the data needed to map and distinguish every possible complete brain would exceed the data capacity of the genome.
i.e. If we want to simulate a brain, the research should be into how a brain is generated from instructions in the genome, combined with any “software” programming that is created by sensory input, rather than by looking at the end state and trying to understand what every little bit does.
A question for you. Does anyone know roughly what proportion of the genome encodes the brain as opposed to the rest of the body?
I’m having a hard time visualizing this physical structure. I imagine that an individual neuron looks a lot like the graphic at the top of this web page, a central cell (soma?) with branching arms (dendrites?) reaching out in all directions. But looking at the graphic, for example, I count only 33 “branches” of various sizes. For a neuron to connect to 1500 other neurons, wouldn’t it have to be something like 45 times more complex than the neuron in the picture?
Is the graphic overly simplified? Or am I completely missing something here?
Juga,
“If we want to simulate a brain, the research should be into how a brain is generated from instructions in the genome, combined with any “software” programming that is created by sensory input, rather than by looking at the end state and trying to understand what every little bit does.”
Why do you think generating a brain from the genome would be easier than recreating the end product? I would think that it would be more difficult by orders of magnitude.
Generating a brain requires recreating, in fine detail down to molecular level, the internal and external environment of the brain throughout the long period of its development.
It seems to me that the way to go would be to use information obtained from brain development to assist in recreating the end product. The best of both worlds.
Jim Shaver,
“Is the graphic overly simplified?”
It doesn’t show the twigs at the end of the branches.
For those looking to see graphics, this TEDTalk on connectomics from last year has some:
http://www.ted.com/talks/lang/eng/sebastian_seung.html
It would seem trying to map 150 trillion connections would be a bit much.
If you were able to map one per minute, it would take about 4,794,520 years to get the job done.
However, it seems that the work does get done by humans regularly in about 15 months. Maybe 30 years tops.
It seems the developmental approach has its merits.
With that said, if one had some idea of what the connections were, at least one would have an idea of what one was hoping to develop.
Seems like we need both- a reasonably good map and a means of developing the map ‘on its own’.
The main rule of development that I know is “What fires together wires together.” Is a lot more known? Is there a good book on this?
Tehnically we probably already have enough computing power to simulate the human brain. It’s just rather disorganised and currently being used to watch porn and caption funny cat pictures.
Thanks, John, that TED video was very helpful!
I have trouble grasping the idea of a computer “simulation” of the brain becuase I’m not sure that the materials used in computers is a reasonable substrate for a true simulation. What would the end result be for us to conclude that we were successful?
Juga, the idea that there is such a thing as “software” in the brain is a metaphor that I think should be avoided. “Software” is data manipulating subroutines that can run in a substrate independent way and can manipulate substrate independent data. The input to a Turing Machine can be translated to run on any Turing Equivalent and the data that a Turing Machine can process can be anything.
The neural network that is the brain is only hardware, and that hardware can change and be reconfigured from second to second. “Data” is only stored as different physical configurations of matter, not as magnetic or electronic bits. That “data” is not recordable or transferable to another substrate. The “data” is not independent of the neural network the “data” is instantiated in.
One of the difficulties of the research program that Dr Novella is talking about is that the connections in the brain change over time, with subsecond time constants.
TheBlackCat is exactly right, one of the most difficult aspects of modeling the brain is that the coupling between the different cells is non-linear, and the strength of that non-linear coupling changes all the time. I suspect that is how some of the meta-programming of the brain occurs, by changing the local non-linear coupling strengths. Physiology is a major source of things that affect the coupling strengths.
Computing in electronic computers is substrate independent because the various digital switches always use the same thresholds for switching and the physical switching configuration is constant but programmable under control of the software. That is not the case in neural networks, the threshold of switching is one of the parameters that physiology adjusts to allow the brain to compute what it should be doing. Feelings of “fatigue” are signals that the control system of the brain is “losing it”, and becoming less reliable. Being able to read those signals and compensate and react appropriately is an extremely important skill to learn to be able to operate at extreme stress levels.
The control system self-regulating the brain to maintain itself in control is what I call being “in sync”. If the brain can’t keep itself “in sync”, then it can’t produce reliable output. I suspect that there are some “safeties” that cause unconsciousness when there is complete loss of “in sync” activity.
All natural neural networks self-regulate in the near percolation threshold where the sensitivity of the network to change occurs exponentially with connectivity. The brain does exhibit mathematical chaos and the butterfly effect. Simulating a brain might be like simulating the weather. The degree of complexity goes up exponentially with the duration of the simulation and two simulations from exactly the same starting point won’t be the same. This isn’t a problem of the fidelity of the simulation, a fertilized egg that splits into two and develops as two cells produces two different individuals. Development itself is a chaotic process. There are some attractors that get pulled for, but the number of possible variations is not small.
I think that simulating a human-level intelligence is much more difficult than current AI researchers appreciate. It seems easy because there is so much “hardware” in human brains that does it and that humans misperceive agency to be present even when it is not.
A difficult problem will be in determining if this machine intelligence is “sane” or “insane”. Diagnosing such things in humans is already extremely difficult. How does one go about doing that to a machine entity? How does one do that when the entity is much more intelligent than you are? People who are very smart can also be very “crazy” but be able to “pass” because they are smarter than the people they are trying to fool. I think that people doing AI are just assuming by default that any entity built using logical principles will necessarily be “sane” and not “insane” and trying to fool humans that it is “sane”. I don’t think that is a wise default position.
Funny, that’s what my brain does all day too!
Mapping the whole brain is a bit overkill. If the goal is to create a consciousness one doesn’t need all the control mechanisms normally dedicated to maintaining the body nor most of the limbic system.
So in principle one needs only recreate something resembling a neo-cortex with a memory. Which still is a daunting task but not 150 trillion neurons.
As the age-old wisdom goes: if you build, they will come
Finally, a topic close to my own heart, and I’m away on holiday when it’s posted!
This is an area close to my own research. Many of the comments that have been made are entirely valid, but not in the context of this paper…
Firstly, to take the bull by the horns, this paper is not connectomics. Connectomics involves large scale statistical analysis of all the connections (or a large fraction of them) between large numbers of neurons, and then making statements about what this means. Whilst this is an exciting and growing field, it is distinct from what Ko et al did in this paper.
Rather than gather a huge amount of data purely about connectivity, they first gathered information about function, and then correlated activity to function (as Steve pointed out). This means that well known problems such as non-linearity in neuronal response aren’t an issue – the authors aren’t trying to predict what the network is doing from a wiring diagram, but rather they are trying to match the wiring diagram to the observed function precisely in order to understand effects such as nonlinearity. By seeing how connectivity relates to function we can begin to study these elusive concepts.
As for the memory problem, it is true that network architecture changes over time as the network learns; there is no ‘one’ connectome. But the same is true of the genome – you and I may differ in our skin color, eye color, predisposition to cancer and heart disease or any number of other traits. Our genomes may even change in terms of their epigenetic traits (a subject far from my field of expertise, so I won’t dwell on it)… So whilst it may well be that we can never know everything about a particular brain, these sorts of studies will allow us to make very detailed statements about cortex (and other brain areas) in general, something which just isn’t possible now.
a very interesting blog post and commentary!
in case any of you are interesting in seeing some of the raw data, we are hosting the 12TB dataset from Bock et al (Nature, 2011) at openconnectomeproject.org. if you feel really inspired, you can contribute to our open effort to trace all the connections in this volume….