[Kaggle (Chicago data center)][Open Connectome (Duke, North Carolina)][Causality Workbench (ETH, Zurich)][NYU (New-York)]
We simulated data with a realistic simulator of real neurons and a model of calcium fluorescence recording, providing data closely resembling real recordings of cultured neurons, while providing unequivocal ground truth of synaptic connections. The data were generated using a simulator that was extensively studied and validated (Stetter, 2012). The dynamic behavior of the neurons was adjusted to reproduce collective properties of real networks of cultured neurons. The model also simulates limitations and defects of the imaging technology (calcium fluorescence): limited time resolution (not allowing to separate individual spikes) and light scattering artifacts (by which the activity of given neuron influences the measurements of nearby neurons).
Reconstructing networks from artificially generated data can be thought of as a futile mathematical exercise. This is why we moved away from reconstructing data generated by simple models such as Bayesian networks or Structural Equation Models making over-simplifying assumptions of linearity and Gaussianity. The data simulator that we are using integrates realistic scenarios at three levels (Stetter, 2012):
• Network structure: At the network architecture level we use topologies as close as possible to natural topologies. We use two strategies: (1) topological models: connectivity models taking into account neuron proximity and enforcing a realistic degree of node clustering; (2) re-simulation: use of a network inferred from real biological data using a baseline computational method.
• Neuron models: We use leaky integrate-and-fire models of spiking neurons, as implemented by the NEST simulator (Neural Simulation Technology). The dynamic regimes reproduce faithfully experimentally observed neural recordings. In this first challenge, we consider only excitatory synapses, as an experimentally meaningful simplification, and we further restrict ourselves to constant value synapses.
• Fluorescence model: We simulate the calcium fluorescence time series taking into account time averaging and light scattering effects. The signals generated closely resemble real data.
Data production and challenge workflow. (a) Data prodiction. (b) Challenge workflow.