Efficient "Shotgun" Inference of Neural Connectivity from Highly Sub-sampled Activity Data.

Area of Science:

• Statistical Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Inferring neuronal network connectivity is crucial but challenging due to the 'common input' problem.
• Existing methods are limited by observing only small network fractions, leading to biased connectivity estimates.
• Common input from unobserved neurons confounds direct causal connections between observed neurons.

Purpose of the Study:

• To propose and validate a new experimental design and computational method for accurate neuronal network inference.
• To address the limitations of current techniques in distinguishing true connectivity from spurious correlations.
• To enable the study of large-scale neuronal networks with limited simultaneous observation capabilities.

Main Methods:

• Introduced a 'shotgun' experimental design: serial, brief observation of multiple neuronal sub-networks.
• Developed a scalable Bayesian inference method using a generalized linear model for spiking recurrent neural networks.
• Employed an approximate expected log-likelihood-based approach for network inference from partial observations.

Main Results:

• The shotgun design effectively mitigates biases caused by common input effects.
• Simulations demonstrate accurate and rapid estimation of large neuronal networks (thousands of neurons).
• The proposed method achieves significant speed-ups compared to previous approaches.

Conclusions:

• The shotgun experimental design combined with the Bayesian inference method is a powerful tool for neuronal network connectivity.
• This approach overcomes the common input problem, enabling more reliable inference from limited observational data.
• The method facilitates the study of complex, large-scale neural systems with unprecedented efficiency.