Retinal ganglion cell tracing in Eyewire

In order to make serial section electron microscopy neurite reconstruction truly high-throughput, it will be essential to find a way to automate the image recognition component. Unfortunately, as I’ve written before, it’s quite difficult to segment and recognize patterns in electron microscopy images.

Inspired by other citizen science approaches, Sebastian Seung & co have come up with the possibly ingenious idea of enlisting the help of the everyman in this task. Their website is called Eyewire. It challenges users to reconstruct ganglion cells from electron microscopic images in the retina.

The images are stained in their cell membranes via a dye to create contrast. In theory, this contrast allows machines and humans to distinguish precisely where the neurite travels. In practice, the dye can invade to organelles, creating noise, or it can stain the cell membrane incompletely, creating artifacts.

Or, the machine learning algorithm might just miss it, because of some sort of bias, like missing boundaries that are outside of its field of view. This is where you come in. Your task is to move from slide to slide and pick out the regions that the algorithm misses.

I just opened up the game and in the first section I was assigned, I came upon this error. Here’s the first slide, which, as you can see, is completely filled in within its stained cell membrane boundaries:

And here’s the next image stack up:

As you can see, but for whatever reason the ML algorithm cannot, there is a hole in the second image which should be filled in. Eyewire allows you to do this yourself,

by filling the hole in with the light teal.

Sometimes the missing holes are more consequential. Filling in some holes means that whole undiscovered branches of a neurite can be found.

In a very nice feature, the algorithm automatically propagates your changes to the rest of the image stacks, so that you don’t have to do so manually.

When you have enough people doing this, the results can be pretty interesting. For example, here is the current reconstructed version of cell #6:

How would you go about quantifying the branching neurites of this neuron and what can you learn from its structure about how it works? These are the kinds of questions that we’ll be able to address as we collect more of these.

Sebastian Seung calls the game “meditative.” In the hours I’ve played so far (my account name is porejide), I have found it quite fun when it’s working fast and I can zoom through the stacks.

On the other hand, at times the internet connection at my house couldn’t really keep up, leading to some lag, which caused me to experience a sensation that I would not call meditative. But perhaps that’s just the fault of my internet connection.

One angle that I especially appreciate is the friendly competition between users. After you fill in a set of image stacks, the game rewards you with a number of points that is meant to be proportional to what you accomplished.

I have no small amount of pride in reporting that yesterday I played well enough (and for long enough) to reach #2 in points for the day, with 981 points, although xo3k was way ahead of me with 3450. As I was playing I could see user vienna717 was gaining ground on me quickly, which gave me the competitive juices I needed to go faster.

This is a great infrastructure, and has the potential to get even more fun if they gamify it further. For example, perhaps users could join teams with other people and play for a glory greater than the self.

This all sounds dandy, but what if you don’t care about retinal ganglion cells? Frankly, I don’t care that much myself. To the best of my understanding, the main thrust of the game is not to build the 3d maps of these ganglion cells, although that will be informative.

Rather, the idea is to provide a huge training set for machine learning algorithms, so that they can learn to better incorporate the insights of humans. This will scale much better than having humans do it, and will in theory allow us to reconstruct neural connections on much larger scales.

This, in turn, will allow us to rigorously test some of the most fundamental questions in neuroscience.

There is no guarantee that Seung & co’s approach will actually get us there, and even if it does, it will take a lot of time and effort. In the meantime, I’ll see you on the leaderboard!

Correlating infant brain network topology with neurologic outcome

The ability to use variability in structural connectivity to explain variability in clinical outcomes would be a critical validation of connectomics, and would help push the field forward.

Towards that end, Tymofiyeva et al used DTI to map the structural connectivity of a clinical cohort of six-month old infants with perinatal hypoxic ischemic encephalopathy.

Since there is no well-established atlas for the rapidly changing brain at such an early stage of development, the authors relied on two unbiased parcellation techniques, illustrated here:

colors arbitrarily refer to distinct brain regions; a = partitioned spherically, b = partitioned along linear dimensions; doi:10.1371/journal.pone.0031029

They then used these parcellation schemes to compute adjacency matrices for each baby. Here are representative matrices for each of the above parcellation schemes:

binary adjacency matrices where a connection (denoted in white) were called if there was a fiber tract with end points located in both regions; c = partitioned spherically, d = partitioned along linear dimensions ; doi:10.1371/journal.pone.0031029

The authors then attempted to correlate a neuromotor score of the infants with the brain network’s degree of clustering. For this, they chose to test 1) the average shortest path length between any two nodes, and 2) the average clustering coefficient.

By the authors’ own admission, a larger sample size and a longitudinal design would be ideal to make inferences from this sort of study. In the scatterplots they presented, they found one positive correlation below the arbitrary threshold of p=0.05, but the effect seems to be mediated in no small part by one outlier.

Still, this is an innovative approach and has great potential. One thing that might be interesting would be to take a more unbiased approach to the data analysis. That is, instead of choosing the network summary statistics to test a priori, use some sort of ensemble learning method to let the data tell you how best to predict the neuromotor scores on the basis of the adjacency matrices.


Tymofiyeva O, Hess CP, Ziv E, Tian N, Bonifacio SL, et al. (2012) Towards the “Baby Connectome”: Mapping the Structural Connectivity of the Newborn Brain. PLoS ONE 7(2): e31029. doi:10.1371/journal.pone.0031029

Casanova R, et al. 2012 Combining Graph and Machine Learning Methods to Analyze Differences in Functional Connectivity Across Sex. doi:  10.2174/1874440001206010001.

Visualizing the Golgi-stained mouse brain

Visualizing nervous systems, both the raw images and their reconstructions, is a hot field for a good reason. Once we have these sort of maps we’ll be able to make much more precisely quantitative statements about how the information flow in neuronal networks is constrained.

On this front, in Nov ’11 Chung et al published a paper describing their research into visualizing mouse brain-wide data generated by a knife-edge scanning microscope. The 29 s, soundless video below shows one of their data sets sweeping through each of the imaging planes (sagittal, coronal, and horizontal).

You can view the data set in a web browser here. It is still in “beta” mode and on my browser it is pretty slow, but worth the wait.

At high zoom, the data is fairly precise. In my screenshot below, you can make out individual somata.

Specimen: C57BL/6J Mouse; Stain: Golgi; Dimension: 600 x 375 x 10 (pixels); Current Layers: 2951 - 2960

The Golgi is called a “sparse” stain because it marks only a subset of the neurons, typically ~1%. On the plus side, it is considered to stain neurons randomly, so any conclusions drawn from the connectivity differences between brain regions in this data set should not be systematically biased.

Of course, we’d first have to convert the image stacks to structure calls, which is far from a settled problem.


Chung JR, Sung C, Mayerich D, Kwon J, Miller DE, Huffman T, Keyser J, Abbott LC and Choe Y (2011) Multiscale exploration of mouse brain microstructures using the knife-edge scanning microscope brain atlas. Front. Neuroinform. 5:29. doi: 10.3389/fninf.2011.00029

Harnessing DNA sequencing to understand neuronal network activity

What has been the growth rate of computing power, multi-neuron recording, and DNA sequencing over the past decade? Konrad Kording provides an illuminating chart pertaining to this question:

neurons recorded = the number of neurons that can be recorded from simultaneously; the neuron and computer scales are exponential fits to data; doi:10.1371/journal.pcbi.1002291

Given the above DNA sequencing trends, it’s no surprise that groups in many different fields are developing strategies to turn the problem they are trying to study into a sequencing problem.

See, for example, Jonathan Weissman’s talk on ribosome profiling, which is an elegant way to use DNA sequencing of mRNA molecules tethered to the ribosome as a way to study translation.

In his article, Kording touches on a couple of intriguing sequencing technologies that might help make the “data-out” step of a given neuroscience experiment more high-throughput.

The method for connectomics he describes is particularly fascinating. The idea is to assign neurons a unique DNA barcode that is spread to each of its synaptic partners via a transsynaptic virus, and then sequence the set of barcodes from a given group of cells.

One aspect that I think Kording might have underemphasized is that these technologies would improve greatly if we improved our ability to sequence the DNA of individual neurons.

For example, typical protocols for probing the expression of intermediate early genes rely on harvesting cells from mass culture or coarse brain regions before sequencing. This is powerful, but it would be much more so if we could analyze the distribution of gene expression between cells rather than across them.

Single-cell genomics is advancing, but it is not yet at the point of routine laboratory use for a typical sequencing experiment. And in order to really take advantage of DNA sequencing technology in understanding how networks of neurons work together, it will presumably need to reach that point.


Kording KP (2011) Of Toasters and Molecular Ticker Tapes. PLoS Comput Biol 7(12): e1002291. doi:10.1371/journal.pcbi.1002291

Link to Jonathan Weissman’s 11/16/11 talk.

Oyibo H, et al. 2011 Probing the connectivity of neural circuits at single-neuron resolution using high-throughput DNA sequencing. Presentation at Computational and Systems Neuroscience Meeeting, pdf.

Saha RN, et al. 2011 Rapid activity-induced transcription of arc and other IEGs relies on poised RNA polymerase II. doi: 10.1038/nn.2839.

Kalisky T, et al. 2011 Single-cell genomics. doi:10.1038/nmeth0411-311

Compensatory brain activation in siblings of children with autism spectrum disorders

Humans seem to have developed dedicated systems for detecting the prototypical gait of moving animals. One paradigm for operationalizing this ability is a point light display, which simulates animals moving in the dark with just a few lights on their joints.

We are able to classify these sparse moving points as biological motion and can often even make inferences about the characteristics of the moving agent. See for yourself in this 31 s video:

Previous studies have indicated that toddlers with autism have deficits in perceiving biological motion. This is not surprising, because social information is embedded within the stimuli.

Kaiser et al took this further by using this point light display paradigm and fMRI on 1) children with ASD, 2) siblings of children with ASD, and 3) control children.

They looked for regions differentially activated between biological light displays and scrambled light displays. They then compared the degree of differential neural activity between groups.

Brain regions were classified as having 1) less differential activation in ASD children in biological conditions as compared to siblings and controls (orange below), 2) less differential activation in ASD children and siblings as compared to controls (yellow), 3) enhanced differential activation in siblings (green), or 4) no statistically significant difference in differential activation between groups (uncolored).

top = sagittal slice; middle = coronal; bottom = axial; doi: 10.1073/pnas.1010412107

Their approach helps tease out the neural circuits underlying why some individuals with genetic risk factors don’t develop ASD. The two main brain regions they implicated were the vmPFC (of emotional decision making fame) and the right posterior STS. Could we imagine some study attempting to stimulate these regions in a model of ASD to mimic the development of compensatory mechanisms?

Kaiser M, et al. 2010 Neural signatures of autism. PNAS doi:10.1073/pnas.1010412107.

How to make mathematical sense of connectomics data

…[C]onsider the example … regarding the significant resources and time being put into deciphering the structural connectome of the brain. This massive amount of accumulating data is qualitative, and although everyone agrees it is important and necessary to have it in order to ultimately understand the dynamics of the brain that emerges from the structural substrate represented by the connectome, it is not at all clear at present how to achieve this. Although there have been some initial attempts at using this data in quantitative analyses they are essentially mostly descriptive and offer little insights into how the brain actually works. A reductionist’s approach to studying the brain, no matter how much we learn and how much we know about the parts that make it up at any scale, will by itself never provide an understanding of the dynamics of brain function, which necessarily requires a quantitative, i.e., mathematical and physical, context.

That’s Gabriel Silva, more here, interesting throughout.


Silva GA (2011) The need for the emergence of mathematical neuroscience: beyond computation and simulation. Front. Comput. Neurosci. 5:51. doi: 10.3389/fncom.2011.00051

Structural variability and the usefulness of understanding neural connectivity patterns

Two recent accounts from Jeff Lichtman about the technical progress in neural connectivity research can be found in his interview with Ira Glass, and his article with Winfried Denk. On a more philosophical note, the end of their article notes that:
During the study of the mouse ear muscle described above, it became clear that every instantiation of the wiring diagram was different from every other one. Some will take such variability to mean that nothing can be learned from doing this kind of tedious, data-intensive, and highly expensive work.
It’s not clear if this argument is merely a straw man, but let’s take them on their word that some critics might espouse such a line of reasoning. From the perspective of explanatory power, it is easy to see why this is a flawed argument.
Indeed, if the connectivity patterns were the same between organisms of the same species, it would mean that these connectivity patterns would be unable to explain any differences in their cognition and behavior.

As an analogy, imagine a hypothetical universe in which the DNA of every organism in the same species were exactly the same, and all of the differences between individuals were mediated via epigenetic modifications.

If this were the case, knowledge of an individual’s DNA sequence would have greatly diminished utility. We wouldn’t be able to correlate genetic variability with molecular, cellular, and organismal variability.

To be fair, it is similarly true that if the DNA of every organism were so variable that we could call it totally random, it would also not have any utility in explaining differences between individuals. The same is true for neural connectivity patterns.

So, for both neural connectivity and DNA base pairs, we can loosely think of the relationship between potential explanatory power and structural variability like this:

The shape of this distribution is modeled after the expected surprisal of a coin flip versus the fairness of the coin. That is, I’d expect extreme degrees of variability or non-variability to be especially uninformative.

The great assumption of connectivity research is that the variability patterns will fall in the “sweet spot” of the above distribution. But Lichtman’s point is that this assumption is not just limited to neural connectivity research–it is an overarching theme of biology.


Lichtman and Denk. 2011 The Big and the Small: Challenges of Imaging the Brain’s Circuits. Science DOI: 10.1126/science.1209168 

Link to Lichtman’s NPR interview.

Modeling the development of the Xenopus spinal cord

There are (at least) two major approaches to building structural models of brain regions: 1) to map all of the connections precisely, or 2) to simulate the development of the connections using validated rules.

Borisyuk et al.’s study in modeling the developmental of the spinal cord in a Xenopus tadpole is an example of the latter. Here is their general model of the spinal cord that the researchers simulated the development of:

"A short length of spinal cord shown in section and after cutting down the dorsal midline and opening flat. Rectangles containing neurons represent the two sides (RL and RR), separated by the ventral floor plate (dark gray rectangle, RF). Examples of the cell body positions (ellipses), dendrites (thick lines), and axon projections (thin lines) are illustrated. Neuron types are listed on the right: RB, Rohon Beard sensory neuron; dla, dlc, dorsolateral ascending and dorsolateral commissural sensory interneurons; dIN, cIN, aIN, descending, commissural, and ascending premotor interneurons; mn, motoneurons." doi: 10.3389/fninf.2011.00020

First, the researchers split their model of the spinal cord into subintervals, each containing empty “spaces” that a neuron could “fill.” Then, in each of these subintervals, they added the appropriate number of neurons of each type, randomly shuffling their relative positions.

This “appropriate number” of each type of neuron was given by previous experimental evidence, either via antibody staining, anatomical labeling (i.e., injecting horseradish peroxidase), or intracellular labeling with neurobiotin. The empirical distribution of these neuron types along the longitudinal axis is below:

"Longitudinal distributions of neuron cell body numbers (per 100 μm). The curves show smoothed, theoretical distributions based on current anatomical estimates...." doi: 10.3389/fninf.2011.00020

After assigning the positions of the ~2000 neurons, the researchers sampled from an experimental dendrite length distribution to assign dendrites to each of the neurons.

Finally, the researchers used a model for axon generation and growth, which updates the position and orientation of the growth cone at each time step. They optimized the parameters controlling this updating to fit with experimental data for the axons of different cell types.

One of the interesting analyses they did with their connectome was to quantify the relative frequency of synapses with various connection distances. As you’d expect and can see below, the relative frequency drops off logarithmically with increasing distance.

"The overall distribution of connection distances for synapses between neurons on a single side of the tadpole connectome."

Since their developmental model follows stochastic rules, it might be interesting to simulate many connectomes and get a sense of the variability between simulations. This might help us to answer the question, how robust are different sets of developmental rules?

As computational power and the ease of programming continues to expand, such an extension might become relatively easy.


Borisyuk R, al Azad AK, Conte D, Roberts A and Soffe SR (2011) Modeling the connectome of a simple spinal cord. Front. Neuroinform. 5:20. doi: 10.3389/fninf.2011.00020

Dynamic causal modeling to study beta oscillations in Parkinson’s model rats

New statistical methods for studying neural connectivity show tremendous promise for understanding diseased brain states. An example of such a technique is dynamic causal modeling, which, broadly, attempts to create realistic models of the interaction between neural activity in different brain regions.

Moran et al. used dynamic causal modeling and a rat model of Parkinson’s disease to parse out the differences in effective connectivity that are apparent in circuits implicated in its pathogenesis.

Note that effective connectivity implies some degree of causation of neural activity from one brain region to another, whereas functional connectivity only implies a statistical association of activity between regions.

A key physiological component of Parkinson’s that their model captured was the oscillations at beta frequencies (13-30 Hz) in regions such as the subthalamic nucleus. These oscillations are associated with motor impairments, including bradykinesia and rigidity.

The authors used a biologically-motivated generative Bayesian model to fit simultaneous local field potential recordings from key brain regions. Here are their results for the circuit connection strengths:

left = control animals, right = parkinson's model (6-OHDA-lesioned) rats; numbers are connectivity strengths (averaged across animals), +/- = 95% credible intervals, differences called significant (**) if they pass a 99.99% confidence threshold; black arrows = excitatory connections, grey arrows = inhibitory connections; light grey boxes = GABAergic cells, dark grey boxes = glutamatergic stellate cells, black boxes = glutamatergic projection cells; doi:10.1371/journal.pcbi.1002124

As you can see above, there is consistent increase in connection strength between the cortex and the subthalamic nucleus in Parkinson’s model animals, while there is a consistent decrease in connection strength between the subthalamic nucleus and the external globus pallidus.

Subsequent analyses showed that the connection from the cortex to the subthalamic nucleus had one of the largest contributions to the beta oscillations seen in the Parkinson’s model rats. This result helps to confirm the importance of this connection pathway in Parkinson’s pathophysiology.

The authors speculate that connections in which there are apparent differences between healthy and diseased brains may represent viable targets for therapeutic strategies. If so, this paper would provide an example of how emerging statistical techniques for studying connectivity can offer downstream clinical benefits.


Moran RJ, Mallet N, Litvak V, Dolan RJ, Magill PJ, et al. (2011) Alterations in Brain Connectivity Underlying Beta Oscillations in Parkinsonism. PLoS Comput Biol 7(8): e1002124. doi:10.1371/journal.pcbi.1002124

Connection strength and specificity of medium spiny neurons in the mouse striatum

The strength and specificity with which a neuron forms synapses is a fundamental question in evaluating its function. To address this for striatal medium spiny neurons (MSNs), Chuhma et al used an optogenetic approach reminiscent of the aforementioned Bardi et al study.

First, they bred a mouse that selectively expressed ChR2 in striatal MSNs. Then, they recorded (in brain slices) the responses of target neurons before and just after shining light. If the neuron’s response to the light had an amplitude exceeding the mean + 2 SD of its baseline, then that neuron was counted as a synaptic connection.

For example, of the 22 neurons the researchers tested in the pars reticulata of the substantia nigra, 21 (or 95%) had an amplitude exceeding the baseline-derived cutoff following light exposure. Here’s the overall summary of what they found:

thickness of arrows = relative strength of connections (as determined by the relative inhibitory postsynaptic current), color of circle = % of target neurons with detectable light-evoked responses (black = 0%, white = 100%), MSN = medium spiny neuron, GP = globus pallidus, SN = substantia nigra, dStr = dorsal striatum, FSI = fast-spiking interneuron, TAN = tonically active neurons (large, cholinergic interneurons), pA = pico-amps; doi: 10.1523/ JNEUROSCI.3833-10.2011

As the authors note in the discussion, the magnitude and strength of MSN-MSN connections indicate that these cells likely play a large role in determining striatal output. Their selectivity for particular cell subtypes is also striking. It’d be interesting to see if this was altered at all in a model of Huntington’s disease, which heavily affects striatal MSN processing.


Chuhma N, et al. 2011 Functional Connectome of the Striatal Medium Spiny Neuron. J Neuro, doi: 10.1523/​JNEUROSCI.3833-10.2011.