Sebastian Seung on connectomics

You can find a video of Seung’s Davos ‘10 talk here. He talks about neuroscience from ~ 2:00 to ~ 5:00. The “powerful microscope” he discusses is the transmission electron microscope. In my mind he is on track in discussing the wide potential of connectomics. But, one quibble, which Seung surely knows quite well:

Although we do have the entire 302-neuron connectome of the nematode C. elegans, from what I can tell scientists still aren’t quite able to model all of its properties on a computer. In fact, it’s difficult for neuroscientists to even model the 11-neuron gastric mill network of the lobster, dependent as it is on intrinsic oscillator neurons and other non-linear activity.

So, we do no doubt need better ML algorithms to “read” and make sense of the TEM data. But, we also need tons more physio data (i.e., here), to construct and validate working input-output models of all important neural and glial cell classes, before we will be able to understand what Seung calls the “essence that makes us uniquely human.”

How many cell types are there in the brain? Masland noted in 2004 that there are ~ 60 cell types in the retina, and reports one guess that there are ~1000 cell types in the cortex. These are currently classified by electrophysio profile, morphology (dendritic arborization / cell size / position), and type of neurotransmitter. But eventually all putative cell types should be validated by their distinct genetics, perhaps determined via the Gal4/UAS system of transgenes to drive the cell-specific expression of some fluorescent protein.

The connectivity between our brain cells matters, but that’s not all that matters. Not all brain cells are made alike.

References

Meinerzthagen IA. 2010 The organisation of invertebrate brains: Cells, synapses and circuits. Acta Zoologica 91: 64-71 . DOI: 10.1111/j.1463-6395.2009.00425.x

Masland RH. 2004 Neuronal cell types. Current Bio 14: 497-500. doi:10.1016/j.cub.2004.06.035

Combining x-ray tomography with serial section transmission electron microscopy

Haberthur et al wanted to image gold particles (200 and 700 nm) in rat lungs. In order to do so they performed vascular perfusion on the tissue samples and shaped them with a watchmakers lathe. They then did x-ray tomographic microscopy at a wavelength of 11.5 kilo electron volts, yielding the pixel resolution of 350 nm by 350 nm by 350 nm. This allows for volumetric analysis so that a computer program can reconstruct a 3D image of the tissue sample.

But since the 200 nm gold particles are smaller than the resolution of x-ray tomographic microscopy, the researchers then used transmission electron microscopy to determine the precise location of these molecules. In order to do so, they chopped up their perfused tissue samples with serial sectioning. After correcting for rotation (which tissue samples are liable to do in the microtome), they found a high degree of correlation between their recounstructed image stacks and the real slices. They were also able to track the 200 nm gold particles over a series of TEM images, as shown via the arrow in these two slices 80 nm apart:

Haberthur et al 2009

One might imagine neuro investigations using tomographic microscopy to build 3D models of a given tissue region and then confirming the precise location of individual structures (like synaptic ribbons) within the 3D space via electron microscopy.

Reference

Haberthur D, et al. 2009 Multimodal imaging for the detection of sub-micron particles in the gas-exchange region of the mammalian lung. Journal of Physics: Conference Series 186. doi:10.1088/1742-6596/186/1/012040. Link.

In vitro neuroscience competition to test input-output models?

Bosco Ho describes the most important findings from the past decade in computational structural biology and imparts this fascinating endeavor of which I was quite unaware:

Every 2 years, a whole bunch of computational structural biology labs effectively shut down for business, and throw every man, woman and workstation together to attempt to crack a set of problems. This same set of problems is simultaneously being attempted in labs all around the world, as researchers race against a clock to predict the 3 dimensional atomic structure of protein sequences published at the CASP protein-folding competition website.

We often say that science is a competition but is is astonishing to me how the computational structural biology community has embraced formally organized competitions such as CASP. Here, we have pure naked competition, complete with a scoring system, judges, and rankings that determine winners and losers. It has all the drama that you’d expect from a reality TV show: recriminations, anger and tears. And it has taken the field of protein folding much farther than anyone would have imagined 10 years ago…

The field of protein folding had been drifting along in some kind of crappy fitness valley and it wasn’t until CASP came along, that we could even define what protein folding was, in a concrete definitive way. In terms of protein folding, the targets of CASP could be used to define a good fitness function, which was enough to spur the field to scramble out of the valley and up a fitness peak.

Here’s a similar competition that one might design for in vitro anatomical neuroscience. Have some central organization, analogous to the CASP, decide on a set of cultured neurons to analyze.* The org would then perform a few tests on the cultured neurons to determine a few variables, including chemical analysis to determine the types of neurotransmitters used, western blotting / optical density readings to determine the relevant protein densities at each synapse, some sort of microscopy from various angles to determine cell shape and type, etc. The org would then perform various electrical inputs to each of the neurons at controlled time sequences and measure the outputs, but crucially, would not make this output data public. Competitors would be given the anatomical data and the magnitude of the electrical inputs and attempt to predict the output and activity of each neuron on this basis. This could help test and refine neuron models. What say ye? Doable?

* Probably start small in terms of numbers and in terms of neural complexity. They wouldn’t want too many endogenous oscillator neurons in the first few iterations!

Question #12: Two types of central pattern generators

Central pattern generators (CPGs) are networks of neurons that endogenously produce rhythmic output, typically used in motor control. Scott Hooper’s accepted definition of CPGs is that they must have rhythms in which: “(1) two or more processes that interact such that each process sequentially increases and decreases, and (2) that, as a result of this interaction, the system repeatedly returns to its starting condition.” Developmentally, indications are that CPG properties are innately established, and that new motor patterns are acquired by the animal but old capabilities are not lost, such that the CPG becomes increasingly multifunctional. There are two major types of CPGs:

1) CPGs driven by an endogenous oscillator neuron, whose oscillations are induced via interactions of its own membrane currents. An example of this is the pyloric network of crustaceans, driven by an endogenous oscillator called the anterior burster neuron. If you isolate the endongenous oscillator neuron in vitro, it will still produce its typical oscillatory output as driven by its own membrane currents.

2) CPGs driven by the activity of a large network. For example, consider the leech heartbeat rhythm generator. In this CPG, motorneurons have two possible states on each side of the heart, one of which beats the heart from back to front, and one of which causes the heart to beat in unison. Around every 20 beats the two sides of the heart switch states. Two of the neurons of the network possess both a hyperpolarization–activated inward current and a low-threshold persistent Na current. Neurons 3 and 4 are reciprocally inhibited by neurons 1 and 2. The result of this inhibition is that neurons 3 and 4 fire in antiphase, to allow only one of each state to be activated at a given time. This second type of CPG is more common.

CPGs are believed to underly many human functions. These include locomotion (with CPGs found in the thoracolumbar segments of the spinal cord), swallowing (with independent CPGs in the brainstem controlling oral, pharyngeal, and esophageal phases, and requring sensory feedback), respiration, ejaculation (neurons called the spinal generator for ejaculation in the spinal cord are capable of self-sustained rhythmic output to relevant motoneurons), and scratching.

Inspired by CalTech’s Question #12 for cognitive scientists: “What is a central pattern generator (CPG)?”

References

Lang IM. 2009 Brain stem control of the phases of swallowing. Dysphagia DOI: 10.1007/s00455-009-9211-6.

Hooper SL. 1999 Central Pattern Generators. Embryonic ELS. Pdf here.

Guertin PA, et al. 2009 Key central pattern generators of the spinal cord. Journal of Neuroscience Research DOI: 10.1002/jnr.22067

Buono PL, et al. 2002 A mathematical model of motorneuron dynamics in the heartbeat of the leech. Physica D: Nonlinear Phenomena doi:10.1016/j.physd.2003.08.003

Question #11: Where does the plasticity of the vestibular ocular reflex occur?

One of the main roles of the semicircular canal system is to maintain the location of eye focus despite any head movements. The three pairs eye muscles are matched by the three planes of the semicircular canals, and each plane of the canals have direct control over one of these muscle pairs–either the medial and lateral rectus, or the superior and inferior rectus, or the inferior and superior oblique. This control must occur quickly in order to be effective. Eye movements lag head ones by only 10 milliseconds, and the pathway from semicircular canal to eye muscle contains only three feedforward neurons, making this vestibular-ocular reflex (VOR) one of the fastest in the human body.

Researchers in this field define “gain” as magnitude of the eye movement velocity divided by the magnitude of the head movement velocity during head turns in darkness. In typical primates the gain should be 1.0, but the reflex shows adaptation via motor learning if the image tends to have additional movement in one direction or the other. For example, if head movements lead to image movement in the opposite direction, the reflex will adapt to increase the gain and keep pace with the image. If the image movement is in the same direction as the head movement, the gain will decrease to compensate. So the VOR reflex in primates is, in a sense, a model system for plasticity, and in particular for simple motor learning. The debate over the past quarter century has focused on which of the following hypotheses best explains how this motor learning occurs:

1) The cerebellum stores the memory for the VOR adaptation. In this model, head movements lead to visually-dependent climbing fiber activity and semicircular canal-dependent parallel fiber activity. The climbing fibers report error between eye velocity and target velocity, leading to a change in the postysnaptic activity of Purkinje cell synapses and altering their synaptic weights with the parallel fibers. This change in synapse strength encodes the altered motor memory. Evidence: Lesions to the cerebellar flocculus eliminate VOR adaptation; activity of floccular Purkinje cells is highly correlated to adaptation of the VOR in monkeys (see here); and injecting 10 μM hemoglobin (to absorb N2O, diminishing cerebellar synaptic plasticity) into the subdural space over the flocculus on the same side as the observed eye eliminates VOR adaptation in monkeys. Moreover, when Nagao and Kitazawa (2003) injected the depressant lidocaine into the floculli, their monkey’s immediately reversed the VOR adaptation they had learned during 2 hours of visual–vestibular training.* These results indicate that the flocculus is at least necessary for short term cerebellar memory.

2) The role of cerebellum is to compute signal guiding induction of plasticity, but not to store the motor memory. In this model, Purkinje cells in the cerebellum convey the instructive error signal to the vestibular nucleus of the pons and medulla. So, the synaptic changes necessary for memory encoding occur between the axons aferrent to and neurons in the vestibular nucleus. Evidence: It’s possible that the correlations of floccular Purkinje cells and adaptation of the VOR in monkeys (from above) can be better explained by altered input to the cerebellum from mossy fibers that relays an efference copy of adaptation stored in the vestibular nucleus. To test this, researchers isolate the input from vestibular pathways to cerebellar Purkinje cells, possibly by using VOR cancellation. They find that the pattern of Purkinje cell sensitivity is opposite to that required by the VOR adaptation, meaning that VOR plasticity could not be dependent upon the Purkinje cells. Moreover, lesions to the cerebellum do not completely eliminate the memory of VOR adaptation. Finally, Porill and Dean’s (2007) computer sim that includes a realistic 100 ms delay before a report of the retinal error to the cerebellum prevents cerebellum-based learning above frequencies of 2.5 Hz, even though VOR adaptation can occur at ranges up to 25 Hz. In order to account for the biological VOR adaptation capability, extracerebellelar plasticity must be postulated. This second theory now seems to be the consensus view.

For more recent work on the role of the cerebellum and brainstem in the VOR, see here and here. The general lesson I draw from this research is that models based heavily on anatomy, like the classical Marr-Albus-Ito theory (#1) can look very appealing but must be heavily validated just like any other theory before they can be accepted.

* During the adaptation phase the monkeys in the lidocaine group had an increase in gain from 0.76 +/- 0.05 to 0.95 +/- 0.05, but this gain decreased to 0.76 +/- 0.08 only 10 minutes after injection of lidocaine. As opposed to the control solution in which the adaptation remained steady for ~ 1 hour of darkness.

References

Tabata K, et al. 2002 Computational Study on Monkey VOR Adaptation and Smooth Pursuit Based on the Parallel Control-Pathway Theory. J Neurophysiol. Link.

Nagao S, et al. 2003 Effects of reversible shutdown of the monkey flocculus on the retention of adaptation of the horizontal vestibulo-ocular reflex. Neuroscience doi:10.1016/S0306-4522(02)00991-0 .

Nagao S, et al. 1991 Subdural application of hemoglobin to the cerebellum blocks vestibuloocular reflex adaptation. Neuroreport. PubMed.

Porrill J, et al. 2007 Cerebellar Motor Learning: When Is Cortical Plasticity Not Enough? PLOS Comp Bio. Link.

Boyden ES, et al. 2004 Cerebellum-dependent learning: the role of multiple plasticity mechanisms. Annu. Rev. Neuroscience. doi: 10.1146/annurev.neuro.27.070203.144238

De Schutter E, et al. 1996 The cerebellum: cortical processing and theory. Current Opinion in Neurobio. Link.

Question #10: Three findings from the crustacean stomatogastric nervous system

The crustacean stomatogastric nervous system has some particularly attractive features as a model system for neuroscience. Both their 14-cell pyloric network and 11-cell gastric mill network are anatomically separated from the rest of the nervous system and produce distinct motor patterns, allowing researchers to study the properties of each network individually. Here are three of the major findings from this research:

1) Many if not most neurons in the network have intrinsic modes of firing, including endogenous repetitive bursting, rhythmic oscillatory capability, bistability, post-inhibitory rebound, and spike adaptation. Real-life neurons do not just integrate and fire! These different modes of activity can be modulated by the opening / closing or activation / deactivation of various types of ion channels (i.e., types of Na, Ca, K, Cl). In order to model the neural networks accurately with respect to their known biological activity, researchers must incorporate these complex intrinsic properties of the component neurons. Knowledge of the synaptic connections alone is not enough.

2) The neural networks are subject to extensive biological modulation via injection of hormones, stimulation of input nerves, or some other manipulation. For example, the  electrophysiological properties of a given neuron can be changed, inducing an intrinsic mode of firing, or the synaptic connections between neurons can be made stronger or weaker. Enough of these changes can produce a distinct motor output–there is no evolutionary need for a separate neural network for each motor output.

3) Individual networks that appear independent will interact with one another in non-trivial ways given the appropriate environmental cues. It is useful to consider each network as a unit central pattern generator segment that can aggregate with others to form more complex behaviors. This is true of other animals too. For example, Cramer and Keller (2006) found that microstimulation of the vibrissae representation in the motor cortex of rats in frequencies of 50–90 Hz led to evoked whisking frequencies of 5 to 15 Hz, suggesting that the stimulation activated a subcortical central pattern generator instead of allowing for direct motor control.

Research into simple neural networks such as the STNS and C. elegans is indicative of  the progress in neurobio generally: Until we can accurately model these systems, there is little reason to suspect that we will be able to accurately model systems with even more neurons and connections.

Inspired by CalTech’s Question #10 for cognitive scientists: “Describe several main findings resulting from the study of the crustacean stomatogastric nervous system and their implications for the study and understanding of local circuit function in larger, more complex systems.”

Reference

Harris-Warrick RM, et al. 1992 Dynamic Biological Networks: The Stomatogastric Nervous System. Parts available on Google Books here.

Cramer NP, Keller A. 2006 Cortical Control of a Whisking Central Pattern Generator. J Neurophysiol doi:10.1152/jn.00071.2006 .

Correlations Between Neurogenesis and Behavior

Jason Snyder has compiled a very interesting list of 224 studies that have attempted to correlate neurogenesis with behavior while controlling for outside variables. Check out the full list here. Taking the author’s assertions at face value, it appears that 14/47 (29.8%) studies that correlated neurogenesis with depression / anxiety found a significant association, 0/12 studies that correlated neurogenesis with locomotion found a significant association, and  37/71 (52.1%) of studies that correlated neurogenesis with memory found a significant association, in some cases via reduced hippocampal dependence. So based on his cursory lit review the most consistent correlation between adult neurogenesis and behavior in adults can be found in differences in memory.

Question #9: Variation of neural processing in sensory systems

The fact that so many sensory systems eventually project to the cerebral cortex indicates that there are likely similarities in their processing. Evolutionarily, it’s plausible that most neural sensory systems diverged from one common system that involved early analogue of the cortex.

Within mammals, there is a large amount of variation in the number of cortical areas, with most of these subdivisions in all mammals involved in sensory-perceptual tasks. Additionally, some sensory regions of the neocortex have been well preserved throughout the mammalian lineage, including visual cortices V1 / V2, somatosensory cortices S1 / S2, and possibly the primary auditory cortex A1. In order to add additional regions from this baseline amount, evolution might have either selected for rapid duplication of one existing region, or the gradual differentiation of one region into two.

In the visual pathway, the different brain regions contain neurons with different cell morphologies that are arranged in hierarchical fashion. Elston et al gathered data from layer III pyramidal neurons of V1, V2, the medial temporal lobe, V4, the inferior temporal cortex, and finally the superior temporal polysensory of the macaque. When they arranged the data from the pyramidal cells in order they found a beautiful pattern in the basal dendritic field areas, spine densities, and somal areas. I’ve reproduced part a of their figure 2 below but check out the open-access pdf of their paper for all of the data:

This data shows that there is a hierarchical organization of pyramidal cell morphology in the visual system. If there is indeed a common form of neural processing between the various types of sensory systems, we should expect a similar hierarchical organization of cells in non-visual sensory systems as well.

Inspired by Inspired by CalTech’s Question #9 for cognitive scientists: “In mammals, somatosensory, visual, auditory, and olfactory sensory systems all project to the cerebral cortex. To what extent does this imply some common form of neural processing? Justify your answer by referring to and comparing specific details of cortical anatomy and physiology.”

Reference

Kaas JH. 1989 The evolution of complex sensory systems in mammals. Pdf.

Elston GN, Tweedale R, Rosa MG. 1999 Cortical integration in the visual system of the macaque monkey: large-scale morphological differences in the pyramidal neurons in the occipital, parietal and temporal lobes. Proc Biol Sci. 1999 July 7; 266(1426): 1367–1374. PubMed link here.

Question #8: Explaining the role of gamma oscillations

Lakatos et al have noted that “the role of neuronal oscillations in brain operations has been debated since the discovery of the electroencephalogram.” So this is no small cookie. But the current consensus seems to be that gamma waves with frequencies between 25 and 100 Hz are necessary for sensory processing.

Here’s the theory: Activated groups of neurons have the tendency to oscillate in coherent fashion, affecting the output of the given group and its sensitivity to input. Thus, two groups of neurons will be able to communicate much more effectively if their oscillations are phase-locked. Conversely, neural groups that don’t have this frequency oscillation synchrony will have a much reduced capability to communicate. So, incoming sensory info from the currently attended stimulus will have an advantage during recieving in upstream cortical regions. Also, a “broadcasting center” in the thalamus could distribute the selected rhythm to appropriate cortical regions and prime them to preferentially recieve certain frequencies of sensory input.

The theory implies that the spike-traveling time from sending to recieiving group must be timed correctly and have high fidelity. This is true of afferent thalamocortical axons. Spikes in thalamic neurons arrives in cortical cells in between 1-4 ms and peaks at 2 ms. Indeed, Salami et al (2003) found that the conduction velocity along axons on the thalamocortical tract is 10 times faster than other afferents in mice, due to its selective myelination. So this particular tract must be selected from development to be able to have a low latency interaction with the cortex.

The empirical work I’ve seen backs this theory up. For example, Dockstader et al (2010) used MRI and MEG on healthy participants while administering electrical current stimuli for 0.2 ms just above motor threshold. The participants attended either to the electrical stimuli or a distracting video. Selective attention to the electrical stimuli significantly increased activation in the early phase-locked contralateral primary somatosensory cortex gamma response, starting at 20 ms post-stimuli presentation. This shows that selective attention is likely mediated in neural circuits via gamma oscillations.

Note: Very high oscillations (100-500 Hz) are associated with epilepsy, and those between 250 and 500 Hz are often identified via EEG near the onset of focal seizures. The consensus on the role for lower Hz delta-range oscillations is much less distinct, but it may also be involved in early sensory selection. It is of course possible and likely that there are multiple roles for the gamma waves, and that sensory integration is only one of them.

Inspired by CalTech’s Question #8 for cognitive scientists: “What do you know about high-frequency oscillations (20-50 Hz) in invertebrates or vertebrates? What causes them?”

References

Salami M, et al. 2003 Change of conduction velocity by regional myelination yields constant latency irrespective of distance between thalamus and cortex. doi: 10.1073/pnas.0937380100 .

Fries P. 2005 A mechanism for cognitive dynamics: neuronal communication through neuronal coherence. doi:10.1016/j.tics.2005.08.011 .

Dockstader C, et al. 2010 Cortical dynamics of selective attention to somatosensory events. doi:10.1016/j.neuroimage.2009.09.035 .

Jirsch JD. 2006 High-frequency oscillations during human focal seizures. doi:10.1093/brain/awl085.

Question #7: Describing primary sensory pathways in humans

Visual: Path goes, photoreceptor cells in retina –> ganglion cells — (via optic nerve) –> optic chiasm (crossing of axons) –> lateral geniculate nucleus (mostly) + superior colliculi (can mediate saccades) + pretectal area (pupillary light reflex: if light shined in one eye, both pupils constrict). Then from lateral geniculate nucleus — (via optic radiations) –> V1 –> V2 –> parietal visual cortical areas (moving objects around in your head) + temporal visual areas (complex perception of patterns and forms). See here.

Auditory: Path starts in the hair cells of the cochlea, specifically the center axis called the spiral ganglion — (via auditory nerve) –> cochlear nucleus –> location of sound detection: (ventral cochlear nucleus –> superior olive of medulla) + quality of sound: (dorsal cochlear nucleus: frequency differences) — (via lateral lemniscus fiber tract) –> inferior colliculus –> auditory nucleus of the sensory thalamus (aka medial geniculate nucleus) –> primary auditory cortex of temporal lobes. See here.

Olfactory: Path is from olfactory receptors of roof of the nasal cavity — (via axons of receptors projecting as first cranial nerve) –> olfactory bulb — (via axons of mitral cells projecting as olfactory tract) –> olfactory cortex — (via mediodorsal nucleus of the thalamus) –> insular cortex (taste integrates with smell to produce flavor) + orbitofrontal cortex (odor-taste association learning at single neuron level). Notice that olfaction is the one sensation that gives info directly to the cortex from receptors without first passing through the thalamus. See here.

However, there are many multisensory connections in the cerebral cortex, superior colliculus, and thalamus, such that our sensations can feedback and “correct” one another well before conscious awareness. For example, the premotor cortex achieves multisensory integration by converging visual, auditory, and somatosensory inputs, and has large amounts of overlap with axons of various sensory systems sending projections to other cortical regions. These integration mechanisms vary by behavioral task as well type of sensory input and context must be taken into account.

Inspired by CalTech’s Question #7 for cognitive scientists: “Describe the main pathways between sensory receptors and cortex (including intra-cortical circuits) for mammalian vision, hearing and olfaction.”

References

Critchley HD, Rolls ET. 1996 Olfactory neuronal responses in the primate orbitofrontal cortex: analysis in an olfactory discrimination task. Abstract.

Cappe C, et al. 2009 Multisensory anatomical pathways. doi:10.1016/j.heares.2009.04.017.

The Washington University School of Medicine Neuroscience Tutorial, here.