Particle physicists are used to popping champagne corks when they make discoveries at lilliputian scales, but now it’s neuroscientists’ turn. After 12 years and more than $40 million, an eclectic team of 100 biologists, computer scientists, and neuronal proofreaders announced on Wednesday that they have mapped the “connectome” in the central region of the poppy-seed-sized brain of a fruit fly, working out the precise meanderings of 25,000 neurons and their 20 million connections.
The neural map covers one-third of the fly brain, making it the largest connectome, or wiring diagram, ever worked out; besides its 20 million synapses, the precise characterization of more than 4,000 cell types makes it the most detailed. All told, the feat by researchers at Google and the Janelia Research Campus of the Howard Hughes Medical Institute amounts to a big “told you so” to skeptics who said it couldn’t be done this soon, this inexpensively, or this well.
With the connectome of the entire fly brain expected by 2022, the once-unimaginable rate of progress suggests that a human connectome is not the impossible dream skeptics believe.
“It’s extraordinary, it’s huge, it’s a landmark in neuroscience,” said Clay Reid of the Allen Institute for Brain Science, who is mapping the mouse connectome but was not involved in the fly’s. “Nothing like this has ever happened in the field. It will be completely transformative.”
That’s because although the Drosophila brain has only one-ten-thousandth the volume of the human brain, “it is capable of sophisticated behavior,” said Janelia director Gerry Rubin, who led the project. The connectome for its central region, or “hemibrain,” includes circuits for learning, remembering, navigating, sleeping, and maintaining circadian rhythms. By mining the data, which is publicly available and free to all, scientists will be able to identify the sequence of neurons that fire when a fruit fly detects the irresistible aroma of rotting banana, or follows a garbage truck like a long-lost love, among many other behaviors.
“This tells you the whole chain of connections, including long-distance ones,” Reid said. “Instead of seeing maybe 300 connections” as was possible with cruder connectomes, “you can see millions, allowing you to trace neural circuits at a level that was completely unimaginable.”
That matters for the holy grail of neuroscience: the human connectome. Although a “Human Connectome Project” ran from 2009 to 2014, it was more Google Earth than Google Street View, Reid said, mapping only the brain’s large-scale circuitry. In contrast, the Drosophila connectome shows every country lane, overpass, cloverleaf, and roundabout made by the 25,000 neurons, as the scientists describe in a preliminary paper (more are on the way). That’s what a passionate band of neuroscientists aspire to for the human brain: mapping the meanderings of its 86 billion neurons and their trillions of connections.
To neuroscientists, the appeal of the connectome is like that of the genome for geneticists. The Human Genome Project produced a blueprint of heredity and powered discoveries of the genetic causes of diseases and drugs to treat them. A human connectome could reveal the basic wiring that underlies thinking, remembering, reacting, moving, believing, and feeling. With more and more evidence that disorders such as autism and schizophrenia are caused by brain miswiring rather than, say, imbalances of neurochemicals, many neuroscientists believe mapping the connectome is more important than ever.
“This is parallel to and will likely be as impactful as the genome project,” which also started with non-human animals, said neuroscientist Diane Lipscombe of Brown University, the immediate past president of the Society for Neuroscience. “It’s critical to have a reference that everyone in the community can refer to.” Although with current technology scientists cannot map the connectome in a living brain, they can make electrical and other measurements, she pointed out, which could be compared to a reference connectome for that species.
Based on the Drosophila connectome, Rubin estimates that “you could do a 1,000-fold bigger project” such as a mouse connectome “with only 10 times the money in 10 years” — roughly $500 million (in addition to Janelia’s $40 million, Google contributed funding for the fly project). The Human Genome Project had $3 billion in government funding, and the 2020 budget of the National Institutes of Health tops $41 billion.
With a mouse connectome, which Reid’s team at the Allen Institute is pursuing (immediate goal: the 100,000 neurons and 1 billion synapses in 1 cubic millimeter of cortex), mouse versions of autism and schizophrenia and other brain diseases could reveal much more about the neural basis of such disorders.
As for a human connectome, the possibility of using it to understand human behavior and brain disorders is only the low-hanging fruit. Some scientists have even bolder — and controversial — ideas. If all of a brain’s wiring could be preserved after death, then if researchers can decode the connectome they might be able to read its content, and an individual’s memories would transcend death.
Scaling up from mouse to human would probably require another 1-million-fold improvement in mapping speed, Rubin said, on top of the 1,000-fold improvement he and his team achieved since they began the fly connectome in 2008. “But there has been a 1-trillion-fold increase in the speed of DNA sequencing since I did my Ph.D. thesis in 1973,” Rubin said. (The sequencing of 158 bases of yeast RNA that took him two years can now be done by machines in a millisecond.) “So these big numbers don’t worry me.”
In fact, it was technology that got the fly connectome this far.
Mapping a connectome begins with adding special stains to a brain (or hemibrain) to make neurons and other features stand out, then embedding it in epoxy. Technicians then cut it into slabs 20 microns thick, about the width of an extremely fine human hair.
Then comes a key technological breakthrough that’s the microscopic equivalent of surface mining a seam of coal in Appalachia: focused ion beam milling combined with scanning electron microscopy, or FIB-SEM to its fans. A beam of gallium ions blasts off surface atoms, the microscope takes an image of what’s revealed beneath, the beam mills off another 2 nanometers (thousandths of a micron), atomic layer, over and over. Eventually, the actual chunk of brain is gone. But 26 terabytes of image data (26 million photos) record what was. The entire fly brain would amount to 100 terabytes.
Computers stack the images in the same order as the brain slices. Identifying which little blob in one image belongs to the same neuron as a blob in other images, thereby tracing its path through the brain, has long been the bottleneck for constructing connectomes. Until recently, it was beyond the abilities of computers, and it’s so laborious for humans that Rubin originally estimated the fly connectome would take 250 people working for 20 years to map. The only animal whose connectome has been completed is the tiny roundworm Caenorhabditis elegans, with a piddling 302 neurons making about 7,000 synapses, and that took 15 years for a rough draft (released in 1986) and another 20 for a final.
Enter image segmentation, a technology that enables machines to recognize faces and objects in images, including to interpret medical scans and to make self-driving cars “see.” Looking to put segmentation technology to an acid test, computer scientist Viren Jain of Google began collaborating with the connectome team at Janelia, where he’d worked previously. “We thought we could push the state-of-the-art in image segmentation by working on a difficult scientific problem like connectomics,” he said. Specifically, they could test whether segmentation algorithms could analyze the 50 trillion pixels in the Drosophila connectome dataset well enough to trace neurons from one imaged brain slice to the next.
The Googlers developed a tracing algorithm that detects a bit of neuron passing through one slice and determines where its next piece is in the next slice, and the next one, on and on until the neuron’s winding journey has been mapped as completely as Homer did Odysseus’.
Algorithms aside, humans remained central to the process. The algorithm was trained on neurons that had been fully traced by Janelia researchers, for instance, and dozens of “proofreaders” at Janelia spent two years checking the algorithm’s output, making sure it didn’t mistake one neuron’s branches for another’s. (Humans are more accurate than machines at such things, though much slower.)
Whether or not connectomes can illuminate the human mind, heathy and not, technologists believe even partial connectomes can improve machine intelligence by “reverse engineering” the brain. To that end, U.S. intelligence agencies are pouring $100 million into connectomics, including the Allen Institute mouse project, giving hope to dreamers that if scientists can map the human connectome they might one day be able to simulate a mind in silicon.
Such a prospect is likely decades away, if that. Even a mouse connectome will require algorithms that are 1,000 time faster and better than today’s, Jain said, “which is not a trivial thing to do. But when I started in connectomics in 2005, we couldn’t [trace] a single neuron, let alone 25,000. And now look where we are.”
As a person who watched their Mother live through 6 years with Alzheimer’s, this project makes me hope that my children will have a different experience or will not suffer it themselves. Fantastic work.
I am not sure what all this accumulation of ‘information’ would be useful for.
Brain wiring does not change for no reason at all – they change as a RESULT of our experiences (this is called ‘neuroplasticity’). For example, the following longitudinal study showed that psychological stress RESULTS in changes in the connectome.
Magalhães, R., et al. (2018). The dynamics of stress: a longitudinal MRI study of rat brain structure and connectome. Molecular psychiatry, 23(10), 1998-2006.
There are many other studies like this, including human studies. Also, reductions in stress brings about the reversal of these structural changes – for example, mindfulness practices result in measurable changes in the structure and function of the brain in positive ways (e.g. increases in gray matter and cortical thickness, etc.).
Just because you can’t formulate a use right now doesn’t mean it won’t be useful. We studied coronaviruses decades ago when they were just an interesting dead-end in animal microbiology — and not even a fatal one for the animal. Understanding the connectome of a species would involve understanding a range of “normal,” just as it is “normal” to have your heart on the left although some people do not. Similarly, changes in specific brain-maps as a result of stress (or its reduction) would provide important data about pathway building and neural-net options: Learning how a dog with three legs copes neurologically with that deficiency could teach us how to treat the phantom pain of traumatic amputation on a neurological level, for example.
It’s depressing to see someone cite science as a reason for less science.
Your section regarding image segmentation was so interesting. I can’t even imagine how exciting the end result must have been for all who worked so diligently on this project and how it will eventually impact humanity!
Comments are closed.