Improved estimation and interpretation of correlations in neural circuits

Area of Science:

• Neuroscience
• Computational Neuroscience
• Statistical Modeling

Background:

• Recording neural activity in large, dense populations is crucial for understanding brain function.
• Interpreting correlation matrices from such recordings presents significant statistical challenges.
• Regularization techniques can improve correlation matrix estimation by imposing structure.

Purpose of the Study:

• To identify statistically efficient estimators for neural correlation matrices in dense cortical recordings.
• To test the hypothesis that a 'sparse+latent' model best represents neural dependencies in these datasets.
• To understand how estimator structure informs about dominant neural interaction types.

Main Methods:

• Recorded in vivo calcium signals from nearly all neurons in mouse visual cortex using 3D random-access laser scanning microscopy.
• Applied and cross-validated various regularized covariance matrix estimators, including a proposed 'sparse+latent' model.
• Analyzed the resulting interaction graphs for relationships with physical distance and orientation tuning.

Main Results:

• The 'sparse+latent' covariance matrix estimator consistently outperformed other regularized methods in cross-validation.
• The sparse component revealed an interaction graph where 'excitatory' connections decreased with distance and orientation differences.
• Negative 'inhibitory' interactions showed less selectivity based on these properties.

Conclusions:

• The 'sparse+latent' estimator provides a more physiologically relevant representation of functional connectivity in dense neural recordings.
• This method enhances the statistical efficiency of analyzing large-scale neural population activity.
• The identified interaction patterns offer insights into local cortical circuit organization.