A Very Big Map

Connectomics and Its Limits

October 07, 2015

Imagine a gigantic and faraway planet with 100 billion cities and hundreds of trillions of routes connecting them. Imagine too that this vast planet can be viewed only a few square feet at a time, and only with the most sensitive and expensive imaging technologies. 

Making a map of that planet with all its routes and cities would seem a near-impossible task. Yet it would probably be a walk in the park, compared to the stupendous challenge of mapping the human brain’s trillions of connections—mapping that must be done in three dimensions, at scales too tiny for ordinary microscopes, and in fragile, wet tissue.

Even so, brain-mapping research, also known as connectomics, is currently one of the hottest fields in neuroscience, indeed in all science. Funding for connectomics-related efforts such as the EU’s Human Brain Project and the US-supported Human Connectome Project and BRAIN Initiative amounts to well over $100 million annually.

Much of this research has been aimed at gathering relatively low-resolution structural and functional data on the brain, such as from magnetic resonance imaging. But arguably the most exciting connectomics research now is aimed at producing very high-resolution maps (“reconstructions”) of brain tissue—maps that are detailed enough to reveal the actual synapses, or connection-points between neurons.

Lessons from a “saturated reconstruction”

Jeff Lichtman, a Harvard professor and member of the Dana Alliance for Brain Initiatives, runs one of the leading high-resolution connectomics laboratories. In a landmark paper in Cell on July 30, he and his colleagues described an automated method for “probing the structure of neural tissue at nanometer resolution.”

Down at nanometer scale, photons of visible light heave about like sea waves—they are far too large and bendy to be useful in imaging. So traditional light microscopy is out, and researchers instead must rely on electron microscopy (EM), which images targets by peppering them with tiny subatomic particles—electrons.

EM is much more expensive and difficult to use than light microscopy, so connectomics researchers would stick with the latter, and its lower resolution and wider imaging field, if they could. In fact, connectomics as a whole has relied heavily on light-resolution imagery up to now, believing that it is good enough.

“In the neocortex, for example, there have been very few EM reconstructions so far,” said Lichtman.

Light microscopy works down to a resolution of a few hundred nanometers, corresponding to the shortest wavelengths of visible light. That is sufficient to image the structures of neuronal cell bodies, their output stalks (axons), and their input branches (dendrites). It cannot resolve the fine details that indicate actual synapses, but many scientists have believed, with some experimental support, that synapses’ locations essentially can be inferred—to a degree sufficient for understanding and modeling the brain—wherever axons and dendrites are seen in close proximity.

However, several studies in recent years have provided evidence that imaging axon/dendrite crossings with light microscopy is not enough to infer true connectivity—in other words, that axons tend to seek out specific partner dendrites; they don’t just hook up with those that happen to be nearby. One such study determined this from EM images of the mouse retina; another did it using EM images of the mouse hippocampus, an important memory region.

Lichtman’s team, including first author Narayanan Kasthuri, then a postdoc in the Lichtman lab, focused its study on the mouse cortex, the evolutionarily most advanced part of the brain, and in humans also the largest part. The researchers found enough of the same kind of evidence, in their words, to “refute the idea that physical proximity is sufficient to predict synaptic connectivity.”

In addition to discovering that axons don’t simply make synapses with dendrites at a predictable percentage of physical crossings, they found that individual axons commonly make multiple synapses with a given dendrite—which probably helps explain why some pairs of neurons seem more strongly connected than others, in terms of signal transmission.


Axon (blue) making multiple synapses with the same dendrite (green). Credit: Lichtman
Lab / Harvard University

“They showed that it’s simply not true that the vast majority of connections between neurons are formed by singleton synapses,” said Christof Koch, president and chief scientist at the Allen Institute for Brain Science, which is heavily involved in connectomics research. “The sophisticated simulations of brain activity that we and others are working on are going to have to take this into account.”

Arguably the most impressive outcomes of Lichtman’s six-year study were the technical advances. To an unprecedented degree, Lichtman’s team automated the process of slicing, with a diamond knife, extremely thin (29-nanometer) wafers of mouse cortex, and gently guiding them with a tiny, sticky conveyor belt to the EM chamber, for two-dimensional imaging (at three by three nanometers per pixel).

They also modified existing pattern-recognition software to automatically “annotate” the thousands of slice-images by recognizing and labeling every object of significance in the images. These objects included not only the larger structures such as neuronal bodies, astrocytes and other support cells, neuronal axons, and dendrites, but also the tiny structures on axons and dendrites where synapses form. The software even identified and labeled synapse-related objects inside axons and dendrites, such as clusters of energy-producing mitochondria, and the vesicles that carry neurotransmitter molecules.


Neurons and their dendrites traced back from imaged volume (pink arrow). Credit: Lichtman Lab / Harvard University

This process had to include the “stitching” together of objects from adjacent slice-images, and the tracing of synapses to their host neurons, to form a true three-dimensional picture. All the stitching and tracing, and analysis of axon/dendrite crossings, took up most of the time spent on the project, largely because the pattern-recognition algorithms, advanced as they were, just couldn’t do the job accurately without some human editorial help. This aspect of connectomics projects is “currently the main rate-limiting step,” Kasthuri said.

Neurons and their processes, color-coded to distinguish one from another. Credit: Lichtman Lab / Harvard University

The scale of the problem

Since the end of that project, Lichtman’s lab has improved further the automation of imaging and analysis, and now, Lichtman says, could reconstruct the same-sized volume of brain tissue (0.0000015 cubic millimeters, or roughly one cell-width across) in a matter of months, not years.

Still, that is only a very tiny volume, on the order of a billionth of a whole mouse brain—or less than a trillionth of a human brain.

The enormity of the connectomics challenge is also apparent when one looks at the data storage requirements. The annotated imagery generated in Lichtman’s project required about 3 trillion bytes (terabytes) of storage. According to Kasthuri, mapping a full human brain connectome would require more than 1021 bytes (one zettabyte), which represents a large fraction of the total storage capacity on Earth at present.

Connectomics research clearly is not for the faint of heart.

In fact, far from being worried, researchers in the field seem convinced that the required technologies will keep improving, obstacles will be overcome, and connectome-mapping projects will run ever-faster.

“There’s a lot of people working on the problem,” Kasthuri said. “The algorithms are getting close to as good as people can do.”

“The state of the art is always moving,” said Lichtman.

“More and more the challenge is shifting from the acquiring of the data to the analyzing of the data—developing the machine-learning techniques to align, to stitch, and to annotate,” said Koch.

At the Allen Brain Institute, a connectomics team led by Koch’s colleague Clay Reid is trying to do essentially what Lichtman’s team just did, but on a much larger scale—a dense reconstruction of a full cubic millimeter of human cortex.

“In terms of taking all the image data and annotating it, we’re still about a factor of a thousand away,” Koch said. “But there’s good reason to hope that advances in machine vision and machine learning—and we have an ongoing collaboration with Google on this—will improve these algorithms so that over the next five years, we can speed all this up and acquire a complete cubic millimeter.”

Lichtman noted that his lab also is involved in multiple projects aimed at reconstructing brain volumes of different animals, on scales up to a cubic millimeter. “One of my colleagues is even beginning an effort to do a whole brain of a mouse,” he said.

Lichtman and Koch, like other connectomicists, have been particularly emboldened by the success of a previous grand project in biology, sequencing the human genome. Researchers in that field struggled for years with very slow, labor-intensive and expensive DNA-sequencing methods, but eventually made those methods many times faster and cheaper.

“Now you can drop a tissue sample in the lab and a day later have the complete genome,” said Koch. “I suggest that in another ten years, you’ll be able to put a cubic millimeter of brain tissue in a box, and an hour later get a beautifully reconstructed three-dimensional circuit diagram.”

A promissory note

A full connectome of a significant portion of a mouse brain would undoubtedly be a gold mine of new neuroscientific knowledge. Even Lichtman’s project, limited to a tiny brain volume, yielded unexpected, intriguing findings—such as the observation that energy-producing mitochondria, visible in axons in close proximity to synapses, were for some reason rare in the parts of dendrites (dendritic spines) where synapses occur.

Large connectomes from human brains could also, in principle, clarify the mechanisms of major diseases—particularly developmental diseases such as autism and schizophrenia where connection problems are strongly suspected. One could compare the connectome of a diseased adult brain to that of a healthy adult brain, and perhaps also to a connectome of a still-developing child brain, and see precisely how they differ. “I think that’s the kind of investigation that would allow us to make headway in understanding illnesses that we call connectopathies,” said Kasthuri.

To date, however, connectomics has not progressed far enough to generate findings with direct clinical relevance, much less to enable the advanced brain-function models that would revolutionize both neuroscience and artificial intelligence. “It’s still very much a promissory note,” said Koch.

Beyond the connectome

As likely as it seems that connectomics tech will eventually become fast and cheap enough to deliver a whole human-brain connectome—computer storage capacity having expanded accordingly—just having that connection map per se won’t tell scientists how the brain works. Researchers have known for a while, for example, the 302-neuron connectome of the worm C. elegans—an animal frequently used for molecular biology and genetics studies—but as Koch noted: “We still don’t have a good model for how those 302 neurons work together to produce behavior, so that shows you that a connectome may be necessary but by itself won’t be sufficient.”

What more will be needed? First, functional mapping data showing broadly where and how neurons tend to work together—and many connectomics projects are already concerned with gathering these functional data.

Scientists also will need, at least, more detail on brain cells, particularly neurons. In their recent study, Lichtman’s team labeled neurons in the reconstructed volume merely as “excitatory” or “inhibitory.” But as the Allen Brain Institute and others have shown, there are throughout the human brain hundreds if not thousands of different kinds of neurons, distinguished by their surface receptors, neurotransmitters, other secreted molecules, electrical characteristics, gene expression patterns, and so on.

“None of those data are included in the connectomes yet,” said Lichtman. “The challenge is going to be to superimpose it, so that we know not only the connectivity of nerve cells but also their [molecular] identities.”

Currently scientists don’t have methods for making such detailed molecular characterizations of neurons, or other cells, all at once—they typically have to apply stains or fluorescent beacons one at a time to reveal a given cell marker. “It’s not that I’ll want to know where the cholinergic cells are, for example,” said Lichtman. “I’ll want to know, in one dataset, where every type of cell is. And that will be hard—it will require another kind of industrialization, this time for molecular labeling. But it will happen.”

Connectomics and Immortality

The excitement over connectomics has led some to suggest that it might ultimately offer humans a path to practical immortality.

One idea would be to map a person’s connectome after death, and later simulate it on a computer. This, it is supposed, would enable a person’s “mind” to be brought back to life, and it could then be left disembodied, like an app on a smartphone, or perhaps put in control of a robot.

But this approach seems highly unlikely to succeed as intended. Even if a connectome did encode a dead individual’s personality, “uploading” it to a new medium would merely create a copy, instead of transferring the original.

Koch points out that a computer-simulated mind based on an uploaded connectome also probably wouldn’t be conscious.

“A lot of people seem to take for granted that if you could simulate my brain, for example, by knowing its connectome, you would get consciousness too—you would get me, including all my feelings,” he said.

But they shouldn’t take that for granted, he emphasized, because there is no experimental or even halfway serious theoretical evidence for it. Indeed, the most persuasive theory of consciousness implies that the architecture, the integrated activity, the bodily context, and the causal power of a brain add up to consciousness in a way that a computer simulation could never do.

“An astronomer can simulate a black hole on a supercomputer,” Koch said, “but that simulation won’t warp space time around the supercomputer. Similarly with a brain simulation based on the connectome—if the computer doesn’t have the causal power of the brain, you’re not going to get consciousness.”

Conceivably, knowing a person’s connectome would facilitate repairs, for example, if their brain has been frozen in hopes of a future, high-tech resuscitation. Current frozen-storage techniques—in the commercial/technological realm known as cryonics—are thought to cause heavy damage to brains, presumably including damage to their connectomes. The catch, for now, is that there is no way to map the connectome of a brain without completely destroying that brain.