Connectomics Explained in Six Questions: Third Post, Questions 5-6

By MW Pleijzier

In the previous article of this ‘Connectomics Explained’ series, I described the range of approaches involved in connectomic investigation. Microscale, mesoscale and macroscale connectomics all aim to detail the architecture of information flow at different levels of spatial resolution. Any tool in science will advance understanding through refining quantitative data accumulation, but each tool will have its own limitations in the questions it can attempt to answer. For this last post of this series, I will describe what connectomics has in store for the future of neuroscience.

#5 How will connectomics advance our understanding of brains?

Connectomics will not complete our understanding of how nervous systems operate. What connectomics will provide is anatomical maps of brain networks at various spatial resolutions. This is necessary to see how brain circuits behave the way they do. Connectomics is a neuroanatomical approach to reconcile nervous system structure and function. Since Ramón y Cajal used the Golgi method to stain neurons (thereby rejecting the reticular theory of neural connectivity), neuroscientists have been searching for recurrent themes in nervous systems to help explain their function¹. Neuroanatomists search for conserved patterns of connectivity (circuit motifs) between specific cell types – anatomical building blocks that are elaborated upon in different ways, depending on the sensory modality transduced, internal processing executed or motor action performed. Computational neuroscientists (and electrophysiologists/biophysicists) approach this problem from a different perspective; they seek to uncover a ‘neuronal code’, its context within different brain regions and how that context leads to region-specific neural algorithms, emerging as behaviour². Both seek common elements within nervous systems, much in the same way as DNA was found to be the common blueprint for life on earth.

There are many different forms of circuit motifs: see The Circuit Motif as a Conceptual Tool for Multilevel Neuroscience. From the structural perspective, these common circuit motifs can be found throughout the spatial scales occupied by nervous systems: from intimate nanometre resolution to the millimetre and centimetre projections of neuronal tracts. In other words, they exhibit fractal geometry and are self-similar³. This fractal organisation determines the activity of the respective neurons, resulting in neural signals (action potentials) that are non-linear and chaotic². In a non-linear system, the output is not proportional to the input: it may take five neurons to fire simultaneously to elicit one action potential in a postsynaptic neuron, or conversely, one neuron firing may cause a whole train of action potentials in a postsynaptic neuron. Finding which neurons have these motifs, where they occur within brain-wide networks and which neurons they are connecting are goals of connectomics. This is used to build a framework for understanding the chaotic nature of neural communication. The Hungarian biochemist Albert Szent-Györgyi once said, “if structure does not tell us anything about function, it only means we have not looked at it correctly.”⁴

All circuit gifs were created by the author using Neuronify [5], reproduced with permissions from [5]. See http://ovilab.net/neuronify/. The first circuit layout first shows repeated activation of a ‘neuron’ that projects to two independent neurons; the red neuron is inhibitory and the blue neuron excitatory. This excitatory neuron then projects to another excitatory neuron with a voltmeter attached, so we can see the neuron’s activity. Once the inhibitory neuron is connected to the excitatory neuron in layer-2, this inhibits the layer-2 excitatory neuron. This neuron then does not send any excitatory signals to the layer-3 neuron, resulting in no activity in the voltmeter. The excitation from the layer-1 neuron to the layer-2 inhibitory neuron is fed forward through the circuit. Simple feedforward inhibitory motifs normalise the inputs to a neural population, increasing the range of activities these neurons can respond to. This type of motif can be found in the Drosophila antennal lobe [6] which performs olfactory processing, the Drosophila mushroom body [7] (mediating memory), and the pyramidal cells of the rodent cortex [8].

Increasing the number of connectomes produced would potentially enable us to find out how highly conserved particular motifs are between and within nervous systems, and also their evolutionarily determined, developmental regimens. However, every connectome of every animal does not need to be produced to find every particular circuit motif – every genome and the proteins produced by every organism was not studied in detail before the structure or function of DNA was explained. This would be a huge investment of resources with little promise of return. Instead, the biological sciences use model organisms to study fundamental processes, where the simplest system is used to produce the most comprehensible results. For example, most of our understanding of synapse operations comes from studies of homologous proteins in yeast, which don’t have neurons. The majority of our understanding of neural networks comes from studies on C. elegans, which don’t have brains. The Drosophila connectomics project is the first time a connectome of an adult organism’s complete brain is being traced. As this will be accompanied by many genetic and neurophysiological tools (see previous post) used in Drosophila neurobiology, the task of deciphering behaviours and relating these to particular motifs will advance rapidly. Also, this is the first time we will be able to relate specific microscopic motifs to brain regions known to be implicated in complicated behaviours. And these behaviours are complicated.

This circuit layout shows activation flowing from the layer-1 excitatory neuron, via the layer-2 excitatory neuron, to the layer-3 excitatory neuron. The layer-3 excitatory neuron, attached to the voltmeter, also sends an excitatory projection back to a layer-2 inhibitory neuron which has no efferent connections initially. When this inhibitory neuron is attached to the second-order excitatory neuron, the voltmeter activity changes from repeated excitation to delays between activity spikes. Simple feedback inhibition normalises outputs of a neural population, creating sparse activity in a population of neurons resulting in the control of the maximum output. Simple feedback inhibition can be found in the connections between Kenyon cells and the anterior paired lateral neuron of the Drosophila mushroom body [9].

For example, female Drosophila exhibit mate-choice copying¹². When virgin female flies are shown males copulating with other females, the virgin females are more likely to copulate with those males. Male flies that are sexually naïve will tend to spend more time having interspecific sexual encounters (e.g. with D. simulans) but mature flies spend less time copulating and will tend to copulate only within the melanogaster species, increasing their rate of reproductive success¹³. These behaviours incorporate pheromone processing and courtship behaviour, coinciding with experience-dependent learning mechanisms and the subsequent propagation of the genes which promote these behaviours into future generations. The ability to manipulate specific neuronal types and specific neurons, based on our knowledge of specific neuron connectivity, will advance neuroscience by deconstructing behaviours into particular network activities and properties, particular cell characteristics and genome expression patterns. This will not only further our understanding of natural ethologies (behaviours), but also our understanding of many disease states the humble fruit fly is used to model, including drug addiction¹⁴, Alzheimer’s¹⁵, Parkinson’s¹⁶ and depression¹⁷.

Connectomics does not promise to provide a complete, comprehensive understanding of a brain; whether it is a nematode, fruit-fly, mouse or human. Its goals are to find circuit motifs within brain regions across spatial scales, to understand the network properties of biological networks and evaluate how these mediate information transfer across brain regions. It seeks to both confirm experimentalist findings and inform them of possible new hypotheses to be tested. Similarly, while the Human Genome Project never provided a complete, comprehensive understanding of our DNA or the evolutionary origins of Homo sapiens, it was a benchmark for many experimental and computational studies as well as promoting important advances such as increases in computation power and more diverse statistical methodologies.

Three species of Drosophila interacting at a food source. D. hydei are the larger, dark flies whilst D. melanogaster and D. simulans are the smaller, lighter flies. Males of all three species can be distinguished because they approach different sexes of different species in courtship attempts. Reproduced from [18] under the Creative Commons License (CC-BY 4.0).

#6 What is the future of Connectomics?

The speed of slicing brains and building connectomes is increasing rapidly, across species. Last year, a connectomics facility opened in Suzhou, China¹⁹. The HUST-Suzhou Institute for Brainsmatics will contain 50 automated ultramicrotomes for slicing mice brains and using digital reconstruction to provide connectomes of mice. Although the slices of the images will be ~ 1 mm thick (this resolution pales in comparison to the Drosophila slices of 45 nm thick), this is the first attempt to map the circuits of an entire mammalian brain. Even with the increasingly diverse technical hardware being designed and implemented for slicing structures as large as the mouse brain, the main rate limiting factor in connectomic reconstruction is tracing. As a solution, tracing will become increasingly automated through machine learning algorithms. Ambitious members of the Howard Hughes Medical Institute are developing machine learning algorithms to perform this automated reconstruction²⁰. However, one issue that arises in automated tracing is how can an image of hundreds or thousands of neurons be segmented such that individual neurons can be located – how to take neuron membranes and use these to divide the image into its constituent membranes? This particular challenge was tackled by FusionNet²¹, a convolutional neural network that performs highly efficient image segmentation specifically for connectomics. Once the boundaries between neurons are processed, then they can be reconstructed from image to image.

An example of segmentation that machine learning algorithms need to perform in order to execute automated reconstruction. Reproduced from [21] under the Creative Commons License (CC-BY 4.0).

Another development has been SegEM²², a toolset designed by members of the Max Planck Institute for Brain Research. Whilst SegEM provides semi-automated reconstruction (manual reconstruction is also required), it combines highly advanced image segmentation with the ability to detect synapses. For automated reconstruction, there is a trade-off between the rate of reconstructing a neuron’s morphology and the ability to detect synapses, due to the computational power involved. SegEM managed to perform both of these tasks on the mouse retina and cerebral cortex²². If neurons have already been reconstructed without the annotation of synapses, a new tool called SyConn can produce thoroughly annotated connectivity matrices by automatically identifying neuron features – axons, dendrites, myelin, somata and then classify these into cell types²³. A fascinating irony exists here, of machine learning algorithms being used to determine the connectomics of biological learning machines and biological learning centres, which could then perhaps be used to generate new learning algorithms for the improvement of the very processes used to create them.

Connectomics will also advance through data synthesis from different methods of neurobiological investigation. Recent functional methods have advanced to a stage where the neural activity of an entire Drosophila brain can be imaged whilst the fly is behaving naturally, and also in response to stimuli²⁴. Furthermore, the fruit fly can now be imaged over extended periods of time (>24hrs), giving neurobiologists interested in the development of neural circuits, experience-dependent plasticity (memory) and neural ageing, the opportunity for longitudinal studies²⁵. These methodological improvements will enable us to examine an organ which “becomes an enchanted loom where millions of flashing shuttles weave a dissolving pattern, always a meaningful pattern though never an abiding one; a shifting harmony of subpatterns.”²⁶ Relating structural information to these subpattern harmonies, to define the architecture of information flow from the micro to the macro-scale, is the supreme goal of connectomics so that we can advance our understanding of how brains work.

Maximum projection of pan-neuronal GCaMP6F activity simultaneously with grooming behaviour. The white square corresponds to the presentation of UV light presentations. Reproduced from Movie 2 of reference [24], under the Creative Commons License (CC-BY 4.0).

If connectomics becomes more automated and requires less direct human supervision for the creation of connectomes, then perhaps one day fully automated connectomes could be generated in a timescale of a week (but now I’m dipping into science fiction). Within our lifetimes however, microconnectomics should not focus its technical advances on jumping across the clades of the evolutionary tree so that we eventually create a microconnectome of a (hopefully already post-mortem) human brain. Instead, if increased automated tracing speeds and increased speeds of slicing brain tissue are achieved, then multiple connectomes could be generated for simpler organisms within one or several studies. Increasing the sample size of connectomes used (instead of just n = 1) would enable biologists to uncover how individual genes or functionally grouped genes contribute to pleotropic effects on network connectivity and individual neuron morphology throughout the brain. Importantly, we would be able to examine the fundamental processes and changes in network architecture that underlie long term memory. By comparing brains of conspecifics vs. closely related species within the same genera (e.g. Drosophila melanogaster to Drosophila simulans), we could investigate how natural selection has directed the evolution of network connectivity, giving rise to conspecific or species-specific behaviour. Maybe one day, a neuroscience manuscript won’t even be considered for publication in Neuron or Nature without having a set of ‘control’, ‘test_1’ and ‘test_2’ connectomes included.

Conclusion

The playwright George Bernard Shaw once toasted Albert Einstein by saying “Science is always wrong, it never solves a problem without creating ten more.” Another way of thinking about this is “as the circle of knowledge expands, so does the circumference of darkness surrounding it.” Within our current frame of view, eventually the expanding circle of knowledge approaches the borders. From time to time, we must take a step back and evaluate what we know from afar, before mining at the frontier of what is presently known.

Created by the author.

Connectomics makes us appreciate the complicated and complex interconnectivity of brains. Connectomics, however, recognises both the incremental and fundamental technical developments required for neuroscientific advancement. The advent of connectomics in 1986 brought about a slow-burning revolution in the neuroscientific field. The tools are available for finding particular circuit motifs at the microscopic level. Ever since Ramón y Cajal created the depictions of neural structures for which he is so revered, neuroscientists have been searching for these motifs in order to understand neuronal function. For the first time in history, we are able to investigate these motifs in an entire brain. This investigation will require participation from across the scientific spectrum, a major feature of 21^st century science.

Created by the author.

References

1. Braganza, O. & Beck, H. The Circuit Motif as a Conceptual Tool for Multilevel Neuroscience. Trends Neurosci. 41, 128–136 (2018).
2. Vreeswijk, C. Van & Sompolinsky, H. Chaos in Neuronal Networks with Balanced Activity Excitatory and Inhibitory. Science (80-. ). 274, 1724–1726 (1996).
3. Sporns, O. Small-world connectivity, motif composition, and complexity of fractal neuronal connections. BioSystems 85, 55–64 (2006).
4. Buszaki, G. Rhythms of the Brain. (Oxford University Press, 2011).
5. Dragly, S.-A. et al. Neuronify: An Educational Simulator for Neural Circuits. Eneuro 4, ENEURO.0022-17.2017 (2017).
6. Olsen, S. R., Bhandawat, V. & Wilson, R. I. Divisive normalization in olfactory population codes. Neuron 66, 287–299 (2010).
7. Perisse, E. et al. Aversive Learning and Appetitive Motivation Toggle Feed-Forward Inhibition in the Drosophila Mushroom Body. Neuron 90, 1086–1099 (2016).
8. Pouille, F., Marin-Burgin, A., Adesnik, H., Atallah, B. V. & Scanziani, M. Input normalization by global feedforward inhibition expands cortical dynamic range. Nat. Neurosci. 12, 1577–1585 (2009).
9. Lin, A. C., Bygrave, A. M., De Calignon, A., Lee, T. & Miesenböck, G. Sparse, decorrelated odor coding in the mushroom body enhances learned odor discrimination. Nat. Neurosci. 17, 559–568 (2014).
10. Letzkus, J. J., Wolff, S. B. E. & Lüthi, A. Disinhibition, a Circuit Mechanism for Associative Learning and Memory. Neuron 88, 264–276 (2015).
11. Liu, B. hua et al. Broad inhibition sharpens orientation selectivity by expanding input dynamic range in mouse simple cells. Neuron 71, 542–554 (2011).
12. Dagaeff, A. C. et al. Drosophila mate copying correlates with atmospheric pressure in a speed learning situation. Anim. Behav. 121, 163–174 (2016).
13. Dukas, R. Male fruit flies learn to avoid interspecific courtship. Behav. Ecol. 15, 695–698 (2004).
14. Kaun, K. R., Devineni, A. V. & Heberlein, U. Drosophila melanogaster as a model to study drug addiction. Hum. Genet. 131, 959–975 (2012).
15. Moloney, A., Sattelle, D. B., Lomas, D. A. & Crowther, D. C. Alzheimer’s disease: Insights from Drosophila melanogaster models. Trends Biochem. Sci. 35, 228–235 (2010).
16. Feany, M. B. & Bender, W. W. A Drosophila model of Parkinson’s disease. Nature 404, 394–398 (2000).
17. Ries, A. S., Hermanns, T., Poeck, B. & Strauss, R. Serotonin modulates a depression-like state in Drosophila responsive to lithium treatment. Nat. Commun. 8, 1–11 (2017).
18. Markow, T. A. The secret lives of Drosophila flies. Elife 4, 1–9 (2015).
19. Cyranoski, D. China launches brain-imaging factory. Nature 548, 268–269 (2017).
20. Funke, J. et al. A Deep Structured Learning Approach Towards Automating Connectome Reconstruction from 3D Electron Micrographs. 1–11 (2017).
21. Quan, T. M., Hildebrand, D. G. C. & Jeong, W.-K. FusionNet: A deep fully residual convolutional neural network for image segmentation in connectomics. (2016).
22. Berning, M., Boergens, K. M. & Helmstaedter, M. SegEM: Efficient Image Analysis for High-Resolution Connectomics. Neuron 87, 1193–1206 (2015).
23. Dorkenwald, S. et al. Automated synaptic connectivity inference for volume electron microscopy. Nat. Methods 14, 435–442 (2017).
24. Aimon, S. et al. Fast whole brain imaging in adult Drosophila during response to stimuli and behavior. Doi.Org 33803 (2017). doi:10.1101/033803
25. Huang, C. et al. Long-term optical brain imaging in live adult fruit flies /631/1647/245/2186 /631/1647/328/2236 /631/378/2624 /14/69 /14/35 /14/56 /64/24 article. Nat. Commun. 9, 1–10 (2018).
26. Sherrington, C. Man on his Nature. (Cambridge University Press, 1940).