Large-scale electron microscopy (EM) can provide detailed reconstructions of neuronal networks in the brain (i.e. connectomes). In this project – ConnectomesToANNs – we developed and implementing a set of computational approaches that allow the extraction of rules that explain the wiring properties underlying connectomics data, the transfer of these anatomical principles into the design of ANNs, and the evaluation of how these principles impact performance. Previously, we had developed a model of the dense neuropil structure and connectome of the vibrissa-related part of the rat primary somatosensory cortex – the so called barrel cortex. We had validated the model against all available empirical data and thereby demonstrated that the model mimics prospective large-scale EM reconstruction in both complexity and biological details. We were hence in the position to already start developing our methods for building ANNs from dense EM data, even though such large EM datasets are still in early stages of reconstruction. For this purpose, we utilized recent advances in simulation-based Bayesian inference, and showed that we can now infer which wiring rules are most likely to have generated empirically observed connectomes. In addition to the prototypic connectome data from the barrel cortex model, we demonstrated applicability of our computational approaches to connectome datasets from EM reconstructions of the mouse primary visual cortex and human temporal cortex. Based on the connectome models and datasets, we developed a second approach to transform such dense reconstructions of the neuropil into all structurally possible connectomes, and to determine how likely each connectome occurs for wiring rules that represent different empirically observed synapse formation strategies – e.g. neurons specifically form or avoid connections depending on their cell type, activity, target domain, and combinations thereof. These rule-based probability distributions, which we refer to as statistical ensembles of connectomes, allow revealing which assumptions about synaptic specificity are necessary, sufficient and best-suited to account for the empirically observed connections between any of the neurons in the connectome datasets. Thus, we can derive specificity matrices that generate empirical connectomes directly from the underlying EM data. Our approach thereby reduces the large EM datasets to a set of generating principles which facilitate (1) interpretability of the origins that give rise to the enormous complexity of connectomes, and (2) generalizability across datasets from different brain regions, species and/or individuals. An open-access publications of this work is currently in preparation and will soon be submitted for peer-review. Because we achieved a primary goal of our project – i.e. implementing a set of computational approaches that allow the extraction of rules that explain and hence generate the wiring properties underlying connectomics data – we were able to develop methods for transferring of these anatomical principles into the design of ANNs, and to evaluation of how these principles impact their learning performance. We demonstrated these methods by generating ANNs that capture increasing amounts of features as observed empirically in cortical connectomes – such as sparse connectivity, recurrence, nonrandom topologies, excitation/inhibition – and by testing their performance on a set of machine-learning tasks. An open-access publication of this work is currently in preparation and will soon be submitted for peer-review.
