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).
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?
Decreases in brain volume seen via MRI are considered to be signs of decreased synapse density, neuron loss, and cell shrinkage. Tondelli et al. hypothesized that healthy subjects who would develop AD by up to 10 years later would show reduced brain volume in certain brain areas in an MRI at baseline.
Specifically, via voxel-based morphometry, they found greater brain volume in certain regions in individuals who did not develop AD, with the largest differences seen in the medial temporal lobe (hippocampus and amydgala). They also report a significant correlation between the degree of bilaterial temporal lobe atrophy and a measure of cognitive impairment, although it would have been nice to see the actual coefficient.
Perhaps the most interesting part of their study is when the authors used image segmentation of subcortical regions with standardized vertices to compare across subjects. This allowed them to find shape differences in the right hippocampus that could distinguish (> 93% classification accuracy) between subjects who converted to AD from those that remained healthy. This kind of measure could allow for greater insight into the mechanisms of deterioration at play.
Tondelli M, et al. 2011 Structural MRI changes detectable up to ten years before clinical Alzheimer’s disease. Neurobio Aging. PubMed.
Last week the Allen institute released their new atlas with lots of microarray data localized in different regions of human brains. It comes in the form of free software called the “Brain Explorer 2,” which takes awhile to download but seems to be worth it. Here is a screenshot of the user interface:
The two brains correspond to their two donors, showing some of the variability in brain structure. The maps were created using structural MRI and diffusion tensor imaging.
Note that you can choose certain regions of the brain to be displayed or not by clicking on that region in the menu to the right. You can rotate the brain relative to any orientation by clicking and dragging the human head in the upper right corner. If nothing else this makes it a useful tool for learning brain anatomy. Check it out.
Cardona et al describe their research towards the Drosophila connectome here. One advantage of studying Drosophila is its lineage tracts, which are groups of neurons in discrete compartments of the brain that differentiated from the same neuroblast early in development. These allow groups of synapses to be broken up into lineage tracts by light microscopy, and analyzed individually. In the following cross section of one hemisphere, one of these lineage tracts is shown in purple:
They aimed for a resolution of 3-4 nm / pixel, which allows them to image synaptic vesicles with diameters of 30-40 nm. These are indicated by arrows in the EM image below:
For data analysis, they define a “hierarchy” of objects based on their relationship to one another. This hierarchy ranges from “neuropile compartments” to “neuropile tracts” to “lineages” to “neurons” to “primary branches of a neuron” to “secondary branches of a primary branch.” Branches connect to other neurons at synapses, of course.
One advantage of the hierarchical object classification is that it will eventually allow the researchers to upload their data sets to neuron modeling programs such as neuron, to simulate the firing of their neurons and try to understand what kind of output the network might produce.
Next, they classify the neural projections (dendrites + axons) they image into four types:
varicose (“V,” branched, thick are 0.5 – 1.5 μm, thin are 0.15 – 0.4 μm);
globular (“G”, variant of varicose with large swellings > 1.5 μm at points, often form end points of axons);
dendritiform (“D,” highly branched, often change direction, and < 0.2 μm).
Below are generic models for each, followed by representative 3d digital reconstructions of each type, followed by a 3d digital reconstruction that includes all four types of projections:
Although the same types of projections show up across brain regions, their relative number, direction, and the placement of synapses along them differ between regions. These can form network motifs, which is used in this context to indicate a recurring form, shape, or figure in a design.
The most common type of network motif they found they called a “dense overlapping regulatory motif”, in which each axon has multiple different targets and each dendrite has multiple different inputs. Less commonly, they found feed forward motifs, in which one projection (neurite) receives input from a second neurite, and both provide output to a common target. Here are schematics of both:
The dense overlapping motif demonstrates how convoluted the Drosophila neural connections can be. In the following example, one presynaptic neurite (“a”, bright green) contacts 12 postsynaptic “dendritiform” neurites (yellow-brown) at 3 synapses. Concomitantly, 7 other presynaptic elements (transparent green) contact these same postsynaptic dendritiform neurites. The following shows a representative brain section of the motif as well as front and side views of their reconstructed 3d model:
On the other hand, the feed forward motif underscores how neural signals can be amplified and recombine in a hierarchical manner. In this example, a globular neurite (G) is postsynaptic to a varicose neurite segment (V). Concomitantly, a dendritiform neurite (D, orange) is postsynaptic to both the globular neurite (G) and the varicose neurite (V). Here are two representative brain segments of this motif from the ventral nerve cord cube and a lateral view of the 3d reconstruction:
Not only does this paper have good info about Drosophila neural architecture, but it also shows how researchers are still iterating towards that connectome…
Reference: Cardona A, Saalfeld S, Preibisch S, Schmid B, Cheng A, et al. (2010) An Integrated Micro- and Macroarchitectural Analysis of the Drosophila Brain by Computer-Assisted Serial Section Electron Microscopy. PLoS Biol 8(10): e1000502. doi:10.1371/journal.pbio.1000502.
Chklovskii et al recently published a paper in Current Opinion in Neurobiology that discusses the problem of attempting to build a map of neural connections (a “connectome”) in the fruit fly. Here is a sweet image of their broad plan for attacking the problem:
Here’s how I understand how they reconstruct the sections: First the images are mapped globally using a software program like TrakEM2. Then, individual images are overlapped pairwise by different software on the basis of, e.g., image correlation. These overlapping images are transformed or “connected” together. Finally, transformed images are fit into a global coordinate system using the least squares method of linear regression.
This process is almost exactly how one might solve a jigsaw puzzle, if the puzzle was in 3d, there was no guarantee that you had all of the necessary images, and the results would be one of the most important breakthroughs in neuroscience in some time.
Once the images are built into a 3d image, they need to be segmented into biologically interesting portions, such as the axons and the synapses. This involves the choice of various algorithms for boundary detection, which are improving as the world pours more and more research into image recognition in general. Many of these approaches use the Rand index, which measures similarity between adjacent pixels.
Although it is less sexy, researchers still proofread and annotate these connections manually, and there is software designed specifically for that. Perhaps they could jump on the bandwagon of crowdsourcing and formulate this problem as a public and interactive computer game.
The researchers are currently in the processing of reconstructing 250 columns in the medulla neuropile optic lobe that fills a volume of 90 μm × 90 μm × 80 μm. One column of 250 has so far been reconstructed, and they note that it took two “person years” to do so. So let’s hope that we can get some Moore’s law-like advancement going in the area of neural circuit reconstruction!
Chklovskii DB, Vitaladevuni S, Scheffer LK. 2010 Semi-automated reconstruction of neural circuits using electron microscopy. Curr Opin Neurobiol. doi:10.1016/j.conb.2010.08.002
Saalfeld et al have a cool new paper (here) explaining their algorithm for registration (i.e. proper stacking) of serial section transmission electron microscopy images. Microscopy instruments alone are too imprecisefor seamless stitching of the images, leaving some images rotated and out of place. So, computer tech is needed to help for automation.
Their algorithm extracts landmarks (e.g. “blobs”), from all section micrographs (“tiles”), identifies landmark correspondences between tiles, and estimates the tile configuration that minimizes (via the function arg min) the sum of all the squared correspondence displacements between landmarks. In the demonstration below, feature candidates are circled, with size proportional to the feature’s scale. On bottom, two correspondence matches are shown as an example:
They tested their algorithm on 6000+ Drosophila larval brain 60 nm sections imaged at 4.68 nm per pixel resolution. They found that their reconstruction yields continuity of structures such as axon bundles within and across image sections. They conclude that “globally optimal reconstruction of entire brains on TEMlevel will enable registration of 3D light microscopy data ontoelectron microscopy volumes. By that it will be possible toestablish the connection between brain macro (neuronal lineages)and micro (synaptic connectivity) circuitry.” Exciting stuff.
Saalfeld S, et al. 2010 As-rigid-as-possible mosaicking and serial section registration of large ssTEM datasets. Bioinformatics. doi:10.1093/bioinformatics/btq219.
Just et al have an interesting paper (here) using brain images to predict what noun participants were looking at (for 3 s) in their visual field. Importantly, they did not see an actual picture of an object based on the word, although they were instructed to think about what qualities the word connotes. Their model takes into account four factors of a word to predict its brain activation: manipulation, shelter, eating, and word length. Here’s a tantalizing picture of one participant’s expected and actual brain activations upon seeing the words “apartment” and “carrot.”
coronal slice in the parahippocampal area (dark blue ellipse) and precuneus area (light blue ellipse); doi:10.1371/journal.pone.0008622
The real test of any true model is in prediction. For data within one participant, in answering the question “What will the activation patterns be for these two new words, given the relation between word properties and activation patterns for the other 58 words?” the model had a mean accuracy of 0.801. Still within one participant, in answering the question “Which of the 60 words produced this activation pattern, given information from an independent training set?” the model had a mean accuracy of 0.724. These accuracies are far, far above chance.
Even across participants, the model was accurate. In generating predictions concerning two previously unseen words for a previously unseen participant (from training data of the 10 other participants and 58 other words), the model had a mean accuracy of 0.762. Possibly brain imaging needs a Turing test to decide exactly what would be required to say that researchers can “read minds” in an fMRI machine? I would say, > 98% accuracy for what word subjects focus on when they decide to focus on a word, out of all possible dictionary words, would be pretty close to mind reading to me. Or maybe they don’t have to actually guess the word entirely correct but just get its meaning down in terms of various components? All I know is, it will probably be easy to tell when people have the munchies, because there will be a lot of activation in the “eating” factor.
Just MA, Cherkassky VL, Aryal S, Mitchell TM (2010) A Neurosemantic Theory of Concrete Noun Representation Based on the Underlying Brain Codes. PLoS ONE 5(1): e8622. doi:10.1371/journal.pone.0008622
The major input of the basal ganglia, the striatum, contains ~ 6 cell types, primarily: 1) medium spiny neurons ( ~ 96%), 2) Deiter’s neurons (2%), 3) cholinergic interneurons (1%; tonically active but usually stop firing in response to reward). Medium spiny neurons are particularly interesting little buggers.
All medium spiny neurons in the striatum express at least one of DR1 and DR2 receptors, and a minority express both. They receive dopamine-based (“dopaminergic”) input from the substantia nigra that modulates the neurons’ responses to glutamate-based excitatory input from the cortex, hippocampus, amygdala, and etc.
Medium spiny neurons themselves use GABA as their neurotransmitter and thus have an inhibitory effect on the neurons in the globus pallidus that their axons project to.
Dopamine input increases the intrinsic excitability of medium spiny neurons in part by decreasing the inactivation rate of their inhibitory A-type potassium channels. Azdad et al (here) show how rats with pharmacologically-depleted dopamine also have decreased spine density in their medium spiny neurons:
So, this is possibly a homeostatic regulation system whereby dopamine-based changes in the morphology of medium spiny neurons is accommodated for by channel-based increases in excitability.
Identifying this type of neuron in section slices accurately is a major challenge of circuitry analysis. Matamales et al (here) have shown that one way to identify the types of cells in the striatum is by staining their nuclei with a dsDNA intercalating fluorescent molecule (TO-PRO-3). This allows nuclear diameter, nuclear shape, and heterochromatin distribution to be visualized.
As opposed to traditional antibodies specific to one type of molecule (red in the following figure), the nuclear DNA morphology-based system is useful for classifying cells using just one type of marker:
The authors note that “similar nuclear appearances were observed for principal neurons in other brain regions such as pyramidal neurons of the cerebral cortex,” suggesting that their method might be broadly useful in other brain regions, too.
Finally, here is a 3d reconstruction of a medium spiny neuron from a mouse collected using multiphoton imaging:
One day there will probably be a massive database with probabilistic reconstructions of many of these types of neurons arrayed in some sort of physiologically relevant order. How the world will change…
Martone, M. E., Gupta, A., Wong, M., Qian, X., Sosinsky, G., Ludaescher, B., and Ellisman, M. H. A cell centered database for electron tomographic data. J. Struct. Biology 138: 145-155, 2002.
Azdad K, Chàvez M, Bischop PD, Wetzelaer P, Marescau B, et al. (2009) Homeostatic Plasticity of Striatal Neurons Intrinsic Excitability following Dopamine Depletion. PLoS ONE 4(9): e6908. doi:10.1371/journal.pone.0006908
Matamales M, Bertran-Gonzalez J, Salomon L, Degos B, Deniau J-M, et al. (2009) Striatal Medium-Sized Spiny Neurons: Identification by Nuclear Staining and Study of Neuronal Subpopulations in BAC Transgenic Mice. PLoS ONE 4(3): e4770. doi:10.1371/journal.pone.0004770
Variance at the postsynaptic density (PSD) has the potential to account for various differences in human mental characteristics. Hahn et al (here) examined the prefrontal cortex (BA 9) of 12 human brains after death with mean postmortem intervals of 9.6 hours and mean freezing times at – 80°C of 10.2 +/- 1.9 years. They then isolated fractions of the PSDs and visualized them with electron microscopy.
The authors noted that the neuronal ultrastructures were surprisingly well-maintained. Here is a EM image of synaptic membrane fractions, with arrows on right indicating filamentous crossbridges:
Given that these brains were not frozen immediately following death like other animal brains usually are, the sound structure of their PSDs is interesting.
In order to map synaptic connections in large volumes such as the retina, Anderson et al have argued (see here) that molecular profiling needs to be correlated with electron microscope (EM) images. More on their paper later.
For now, here’s an example of how a molecular profiler can be put to good use. Li et al electroporated expression constructs with genes for horseradish peroxidase targeted to the plasma membrane. Horseradish peroxidase emits amplifiable light at a wavelength of 428 nm. These researchers used it as an anatomical label to correlate spatially with the serial section EM of their tadpole neurons.
One of the advantages of horseradish peroxidase is a uniform plasma membrane distribution including mitochondria / vesicles. It also helps identify long axons / dendrites with small diameters. But on the other hand it has to be electroporated while the animal is still alive to have an effect, unlike some other markers.
Here’s a series of EM images of a distal dendritic branch (blue) that synapses with axon terminals (pink) at white arrow heads:
You can see how the dendrite recedes as you look to the right in the series of images, which the researchers can reconstruct in their model of the microcircuit.
In order to identify synapses, these researchers used two main criteria:
1) At least two serial sections in which docked and clustered synaptic vesicles oppose the plasma membrane of the next neuron.
2) The distance from the intracellular portion of the presynaptic membrane to the cytosol of the postsynaptic membrane should be significantly longer at synaptic sites than non-synaptic sites. Consistent with this, the presynaptic membrane and the synaptic cleft should be much thicker in electron dense material at synaptic sites than non-synaptic sites.
The following chart quantifies the average membrane thickness at synaptic and non-synaptic sites, although these differences are actually slightly less distinct in neurons expressing horseradish peroxidase:
Li J, Wang Y, Chiu S and Cline HT (2010) Membrane targeted horseradish peroxidase as a marker for correlative fluorescence and electron microscopy studies. Front. Neural Circuits doi:10.3389/neuro.04.006.2010. Link here.