Constraining computational models using electron microscopy wiring diagrams
Introduction
Theoretical models in neuroscience often make assumptions about synaptic connectivity that lead to predictions about neural activity and behavior. These assumptions range from specifying the parameters of a statistical distribution that characterizes the wiring of many neurons (for example, populations of pyramidal cells in a volume of neocortex [1]) to specifying the properties of connections among individual neurons with prescribed functions (for example, specific connections among motion-selective retinal neurons [2••]). Except for systems with sufficiently few neurons, these assumptions are often informed by incomplete knowledge of connectivity obtained from electrophysiological measurements of a subset of connections.
Electron microscopy (EM) reconstruction techniques promise a more complete picture of neuronal interconnectivity, obtained by tracing the processes and identifying the synaptic connections of all neurons in an imaged volume of brain tissue [3]. An EM wiring diagram has existed for the nematode C. elegans since the mid-1980s [4], and for the larva of the ascidian Ciona intestinalis since 2016 [5]. Efforts are underway to map the full nervous systems of the adult [6] and larval [7, 8, 9] Drosophila melanogaster fruit fly, larval zebrafish [11], volumes of rodent brains including retinal [12] and cortical areas [13, 14, 15], and other systems [16,17].
These efforts suggest the possibility of inferring the connectivity of models directly from EM data, rather than assuming it. Such an approach may lead to better models whose activity and interactions can be more readily compared to experiments. We survey recent studies that build models based on synaptic wiring diagrams, highlighting results that have been obtained and the assumptions that are required to build the models. We argue that new quantitative techniques must be developed to exploit EM data as a meaningful constraint in models with many uncertain parameters.
Section snippets
EM wiring diagrams and the information they provide
EM reconstructions of neuronal connectivity are based on images obtained from thin sections of a volume of brain tissue [3]. These images are analyzed to identify structures of interest, typically 3-d reconstructions of neurons, their processes, and their synaptic connections. While this is often done manually, advances in automated image segmentation methods are accelerating the speed at which analyses can be performed [18]. EM reconstructions are effective at identifying neurons and the
Graph theoretic approaches
One approach to this challenge is to develop methods that draw conclusions based only on the graph of connections defined by a synaptic wiring diagram and independent of unknown quantities. Methods have been developed to extract structure such as the presence of distinct cell types from combined connectivity and anatomical data [29••], or to infer latent variables that characterize neurons and their connectivity based on a graph of connections [30]. Studies of C. elegans have quantified the
Models constrained by function
Among the first results provided by the wiring diagram of C. elegans was a characterization of its circuitry for detecting touch [37]. Systems close to the periphery are attractive targets for modeling based on EM wiring diagrams. For many of these systems, inputs and outputs can often easily be identified and relatively complete circuits can be reconstructed, leading to readily-formed hypotheses about function that can be tested even if certain system parameters are unknown.
A major target of
Constrained optimization of neural network models
For many complex neural circuits, it may not be possible to manually infer or tune unknown model parameters in order to produce a desired function, even with knowledge of the circuit's connectivity. In recent years, artificial neural networks (ANNs), optimized using stochastic gradient descent, have proven effective at performing tasks such as object classification and at predicting neural responses in higher visual cortical areas [55]. Can such optimization approaches benefit from knowledge of
Incorporating other sources of knowledge
Constraints in addition to connectivity and optimization for task performance are likely necessary to achieve a good match between model and recordings. Many of the above studies relied on knowledge of neuronal response properties or neurotransmitter identities, which can be used to infer the signs of weights corresponding to excitatory or inhibitory synapses. Future work should also aim to infer, for distinct neurotransmitter types, distinct mappings from synapse count or size to effective
Partial wiring diagrams
The strength of EM reconstructions comes from the comprehensiveness of connectivity data that they provide, but so far the only “complete” connectomes to have been annotated are those of C. elegans [4] and Ciona intestinalis [5], and only recently at the level of detail needed to assess variability across individuals or sexes [59]. While a relatively complete wiring diagram of the Drosophila adult and larva will likely be available soon, for larger organisms such completeness is still far away.
Conclusions
Future theoretical work focused on EM datasets should attempt to develop techniques for identifying features of interest in connectivity graphs independent of assumptions on neural dynamics. Techniques to automatically cluster neuronal types [29••], identify latent connectivity structure [30], and visualize wiring diagrams are needed to facilitate the discovery of connectivity patterns that suggest further experimental or modeling study. Such techniques must be robust to reconstruction errors
Conflicts of interest
The authors report no conflicts of interest.
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
The authors thank L. F. Abbott for comments on the manuscript. Support was provided by the Simons Collaboration on the Global Brain, the Gatsby Charitable Foundation, NSF award DBI-1707398, and the Burroughs Wellcome Foundation (A.L.-K.) and the Howard Hughes Medical Institute (S.C.T.).
References (61)
Space-time wiring specificity supports direction selectivity in the retina
Nature
(2014)The structure of the nervous system of the nematode Caenorhabditis elegans
Phil Trans R Soc Lond B
(1986)Anatomy and function of an excitatory network in the visual cortex
Nature
(2016)System-wide rewiring underlies behavioral differences in predatory and bacterial-feeding nematodes
Cell
(2013)Cellular-resolution connectomics: challenges of dense neural circuit reconstruction
Nature Methods
(2013)The neural circuit for touch sensitivity in Caenorhabditis elegans
J Neurosci
(1985)The comprehensive connectome of a neural substrate for ‘ON’ motion detection in Drosophila
eLife
(2017)Dense EM-based reconstruction of the interglomerular projectome in the zebrafish olfactory bulb
Nature Neurosci
(2016)Whole-animal connectomes of both Caenorhabditis elegans sexes
Nature
(2019)Is cortical connectivity optimized for storing information?
Nat Neurosci
(2016)