Summary

SynapseGenerator is a software package used to predict connectivity at the synapse level based on statistical properties of all neuron populations that co-exist in the given brain region of interest. For each population, the spatial distribution of the dendritic and axonal arborizations needs to be either specified as a probability density function (e.g. a mixture of Gaussians), or can be estimated from samples of reconstructed dendritic/axonal arborizations.

The five main processing steps are illustrated in the figure above:

1. Parameters (lpar): Define the statistical parameters for each neuron population and setup a discrete space to place neurons in. The space can be 1, 2 or 3 dimensional; the discrete grid is chosen fine enough to have at most one synapse per voxel.
2. Gaussian parameters (Yinfo): Parse the input parameters and convert the specified densities to a Gaussian Mixture model, which defines the target density Y. The presynaptic and postsynaptic neurons are specified seperately, each can consist of several neuron populations.
3. Target Density (Y): Compute a discrete spatial sampling of the Gaussian Mixture model. Y is a (sparse) two dimensional matrix in which each row represents a voxel, containing the synapse densities of all contributing neurons.
4. Synaptic Elements (X): Place synapses in target space such that:

• A preset overall synaptic density is reached.
• Synapses are placed where dendritic and axonal densities are highest.
• No more than one synapse is placed in a single discrete voxel.

Several methods to optimize these criteria can be used:

1. Random draws, followed by pruning;
2. The Lasso method (under development), a linear least squares algorithm that uses a constraint that parameters cannot exceed a given value (here: at most one synapse per voxel);
3. The Ising model (under development), which computes an optimal synapse configuration by minimizing an energy function.

The resulting matrix X has the same size as Y but is much sparser since it contains only ones (potential synapse) and zeros (no synapse).

5. Return results:

• Connectivity (conMat): Each synapse is assigned to a pre- and postsynaptic neuron;
• Minimum Spanning Tree: The entire set of synapses for a given neuron is used to reconstruct its arborization, using the minimal spanning tree algorithm, as implemented in the TREES toolbox;
• Standard set of visualizations, to analyse the consecutive steps of the pipeline.

SynapseGenerator is implemented in Matlab; its core elements are currently ported to Python.

Contents

This second iteration of the online documentation consists of the following parts:

• A worked out example of long range, zeroth order connectivity;
• A worked out example of thalamocortical projections to spiny stellates, limited to only a few neurons;
• A worked out example of thalamocortical projections to spiny stellates, limited to a small piece of cortex;
• A description of the lpar input parameter structure;
• A listing of all functions in the SynGenPipeline package with descriptions derived rom AutoDoc comments placed in the code.

The examples are designed such that they run smoothly on normal desktop computers with 8GB memory. With the forthcoming port to Python, we will develop a parallel computing strategy to keep requirements low.

List of contents

• Example 1: Long range, zeroth order pyramidal density model
• Example 2: Spiny stellates in barrel cortex, but only a subset
• Example 3: Spiny stellates in barrel cortex, but partially generated
• Specification of the input parameter structure “lpar”
• Function list

Indices and tables

• Index
• Module Index
• Search Page