Toward large-scale connectome reconstructions
Introduction
A resurgence of interest in high-throughput, high-resolution quantitative neuroanatomy, known as connectomics, has been accompanied by a passionate debate [1, 2, 3, 4, 5]. The proponents of this approach believe that knowing all synaptic connections in the brain will lead to understanding in ways that any lesser detail cannot. The opponents of this approach argue that it is largely a distraction which will not lead to advances in our understanding of brain function. Until recently, the debate about the role of connectomes in neuroscience has been largely theoretical as very few connectomes of biologically interesting circuits have been actually reconstructed.
In the last couple of years, new ‘experimental data points’ have been obtained and can provide experimental grounding to the debate. Thanks to pivotal technological advances, several connectomes of biologically interesting circuits have been reconstructed. The prime example is the connectome module of the fruit fly medulla [6••], which allowed identification of the neurons and the circuit motif involved in motion detection thus setting the stage for finally answering the sixty year old question about the biological implementation of the elementary motion detector (EMD). Others are the connectome of Caenorhabditis elegans male [7•], and the reconstruction of the mouse retina [8••].
In this review, we first revisit the connectomics debate in the light of new experimental data and demonstrate that stereotypical connectomes can indeed provide insight into neural computation when used in combination with other approaches. We further argue that, in less stereotypical connectomes such as those in the mammalian cortex, connectomics can help identify stereotypical features such as circuit motifs. One such famous but still experimentally unproven circuit motif has been proposed by Hubel and Wiesel (HW) to explain the emergence of orientational selectivity in the visual cortex from orientationally nonselective thalamic inputs. As the relevant cortical neurons span the volume larger than has been reconstructed before, we next discuss technological developments necessary to obtain this ‘experimental data point’.
Section snippets
What are connectomes good for?
One obvious use of connectomes is as brain atlases: to identify neuronal pathways, individual neurons, and their upstream and downstream synaptic partners. Such information helps focus investigations using other approaches on relevant targets. The C. elegans connectome has been irreplaceable in guiding the work in the field. For example, the knowledge of connections allowed reverse engineering the circuit for forward and backward movement in the worm by ablation of identified neurons [9]. Of
Dataset acquisition and reconstruction flow
In this survey, we will examine current approaches of reconstructing the connectome and the barriers to use them on a large scale, such as the scale required to test the HW circuit motif.
Because of the nanometer-level precision required to exhaustively map neural connections, we restrict our discussion to connectomes reconstructed from high-resolution EM datasets. The EM reconstruction flow shown at a high level in Figure 3 provides the outline for further discussion in this paper. The first
Shape reconstruction: skeletonization and volumetric
One of the main tasks of reconstruction is extracting the neuron shape from the images. This process can be manual or computer assisted, and can be based either on skeletons or a volumetric representation, Figure 4. In skeletonization, the user manually determines 3D line segments that pass through each branch of the neuron. In volumetric reconstruction, each neuron is defined by all of its voxels, which can be assigned through manual contouring techniques [30] or by starting with an initial
Synapse identification and characterization
A connectome, either sparsely or densely traced, requires identifying synaptic connections between neurons. Previous studies [6••, 7•, 10, 39, 40•] manually annotated presynaptic terminals and postsynaptic densities. However, this approach is not practical for larger connectomes because it takes too much time. For example, extrapolation from the experience in [6••] to the full Drosophila brain tells us that without automation this step alone would take hundreds of person years. Advances in
Managing uncertainty
Because reconstruction errors are inevitable, and since the accuracy required depends on the intended use, characterizing types of errors and their rates is a critical part of any reconstruction. Errors arise from many causes including imaging artifacts, image ambiguity, incorrect segmentation, and human mistakes. The primary means of reducing errors involves duplicate judgment of multiple proofreaders/annotators [6••, 8••]. Even so, the true error rates are hard to quantify since the ‘truth’
The logistics of terabyte-scale reconstruction
The logistics required to reconstruct an organism include sample preparation and imaging, distributed data management and versioning, obtaining and maintaining a suite of automated tools for extracting neurons and identifying synapses, procuring sufficient human annotation effort and a team of experts to review ambiguous portions of the reconstruction, implementing quality control mechanisms throughout the process, providing solutions for mining and analyzing connectome results, and software
Conclusions
Scaling up connectomics requires attention to all parts of the flow from data collection to disbursement. The scale, inherent ambiguities, and complexity of the task will require a much more engineered and adaptable solutions than those employed so far.
First, we need better raw data, including image quality, contrast of relevant features, and field of view. These improvements in sample preparation, staining and imaging will be instrumental in reducing ambiguities that are difficult for machines
References and recommended reading
Papers of particular interest, published within the period of review, have been highlighted as:
• of special interest
•• of outstanding interest
Acknowledgements
We would like to thank the FlyEM team at Janelia Farm Research Campus for providing insights on reconstruction methodology and its many challenges. In particular, we thank Arjun Bharioke, Patricia Rivlin, Toufiq Parag, William Katz, and Donald Olbris for their correspondence. Thanks to Viren Jain for discussion about segmentation strategies. Thanks to Harald Hess, Shan Xu, and Kenneth Hayworth for discussions on advances in imaging. Thanks to Ian Meinertzhagen for discussion on all topics,
References (69)
Reading the book of memory: sparse sampling versus dense mapping of connectomes
Neuron
(2009)- et al.
Sparse reconstruction of brain circuits: or, how to survive without a microscopic connectome
NeuroImage
(2013) - et al.
Ultrastructural analysis of hippocampal neuropil from the connectomics perspective
Neuron
(2010) - et al.
A neuron as a signal processing device
- et al.
Receptive fields, binocular interaction and functional architecture in the cat's visual cortex
J Physiol
(1962) - et al.
Network anatomy and in vivo physiology of visual cortical neurons
Nature
(2011) - et al.
Serial block-face scanning electron microscopy to reconstruct three-dimensional tissue nanostructure
PLoS Biol
(2004) - et al.
Wiring economy and volume exclusion determine neuronal placement in the Drosophila brain
Curr Biol
(2011) - et al.
Different classes of input and output neurons reveal new features in microglomeruli of the adult Drosophila mushroom body calyx
J Comp Neurol
(2012) - et al.
Ome sweet ome: what can the genome tell us about the connectome?
Curr Opin Neurobiol
(2008)