Systematic errors in connectivity inferred from activity in strongly recurrent networks.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Understanding neural computation and learning necessitates knowledge of neural circuitry.
• Directly measuring neural circuit wiring diagrams is difficult, driving interest in algorithmic estimation from activity recordings.

Purpose of the Study:

• To investigate the accuracy and limitations of algorithmic methods for inferring neural circuit connectivity from multicell activity recordings.
• To identify biases in current inference methods and explore potential solutions.

Main Methods:

• Applied sophisticated inference methods to simulated multicell activity data from neural circuits.
• Analyzed the impact of correlated neuronal activity and model-data mismatch on inference accuracy.
• Investigated the effect of perturbing circuit dynamics on inference bias.

Main Results:

• Even with unlimited data, sophisticated methods exhibit bias, inferring connections between unconnected but highly correlated neurons.
• Inference bias is exacerbated by a mismatch between true network dynamics and the inference model, a common issue in real-world modeling.
• Activity-based connectivity estimates require caution in strongly connected networks due to inherent inference limitations.

Conclusions:

• Causal inference is compromised by high variable correlation, necessitating careful interpretation of activity-based connectivity estimates.
• The study highlights a fundamental challenge in reconstructing neural circuitry from observational data.
• Perturbing circuit dynamics by driving them far out of equilibrium may offer a strategy to mitigate inference bias.