Uncovering Neuronal Circuitry Using Computational Modelling

Post by Amanda McFarlan

What's the science?

With the advancement of technology, it is now possible to acquire recordings from large populations of neurons in anaesthetized or awake, behaving animals. These recordings generate large data sets that may be useful for understanding how neuronal activity is related to behaviour. However, interpreting and analyzing the results from a large data set can be very challenging and requires sophisticated computational methods. This week in Nature Communications, Kobayashi and colleagues investigated the efficacy of their computational model in identifying neural connectivity in both synthetic simulations and biological data sets.  

How did they do it?

The authors estimated neuronal connectivity using a computational model by cross-correlating the spike (or action potential)  times for each neuron pair. Then, they applied a generalized linear model to identify pairs of neurons with small (millisecond) differences in spike timing to determine the pairs that were likely monosynaptically connected. To determine the level of conservatism that optimally balanced the rate of detected false positives and false negatives, the authors applied their model to spike train data obtained from a synthetic neuronal network of 1000 neurons (800 excitatory, 200 inhibitory) for which the connectivity was already known. Next, the authors sampled neurons from the total population to evaluate the efficacy of their model at predicting neural connections. They generated a connection matrix with four quadrants to represent the different combinations of neural connections: inhibitory-excitatory, excitatory-excitatory, excitatory-inhibitory and inhibitory-inhibitory. Then, they applied their model to spike train data from hippocampal neurons in the CA1 region of a rat to assess its efficacy in estimating neuronal connectivity when applied to a real, biological data set. Finally, the authors established the validity of their model by comparing it to two other commonly used methods, the cross-correlation method and the jittering method, while using both synthetic and biological data.

What did they find?

The authors showed that their simulation of a neuronal network produced results that were consistent with balanced state network models. They determined that balancing the rate of false positives and false negatives was optimally achieved using a significance level of α=0.001 to determine neuronal connections. Compared to excitatory neurons, inhibitory neurons tended to have higher firing rates and exhibit more regular spiking. Additionally, the likelihood of identifying neuronal connections was positively correlated with the duration of the recording, suggesting that it may be possible to correctly identify a larger number of connections between excitatory and inhibitory neurons by extending the length of a recording. Next, the authors showed that applying their computational model to biological data in the rat hippocampus yielded similar results as applying it to synthetic data, suggesting that their model is useful for estimating neuronal connections in real-life examples. Finally, they revealed that their computational model, with only 13 false connections, was superior to the cross-correlation and jittering methods which had 88 and 27 false connections, respectively.


What's the impact?

This is the first study to show a new computational model that combines the generalized linear model and cross-correlations to reliably identify neuronal connections in complex neural networks. The authors have demonstrated that their model is superior to other commonly used methods and is also effective when applied to both synthetic and biological data. This computational model may be instrumental for mapping neural connections in large data sets from extracellular recordings in animal subjects and may provide insight into patterns of neuronal connectivity throughout the brain.  


Kobayashi et al. Reconstructing neuronal circuitry from parallel spike trains. Nature Communications (2019). Access the original scientific publication here.

Illustration by Kai Shinomoto

Illustration by Kai Shinomoto