A unified Granger-based framework links stimuli, neurons, and behavior from two-photon calcium imaging

# What this paper does

# Why this matters

Two-photon calcium imaging records large neuronal ensembles but provides indirect, noisy, and slow proxies of spiking. Traditional work often treats encoding, readout, and functional connectivity separately, which obscures how sensory inputs are transformed into behavior by distributed populations. A single statistical framework that recovers directional interactions across these elements enables more coherent inferences about information flow and functional roles within neural populations.

# Methods in plain terms

• The framework models latent spiking and calcium dynamics using state-space representations that account for the slow, noisy nature of calcium imaging.
• Variational inference is used to fit the model and estimate latent variables and parameters efficiently across many neurons.
• Point-process modeling is integrated to link latent neural activity to discrete behavioral events and to formalize neuron-to-behavior Granger effects.
• Directional influences are formalized with Granger causality: a component improves prediction of a target's future when included in the predictor set.

# Validation and data

# Practical takeaways for researchers

• Use when you have population calcium imaging and want directional, temporally grounded inferences linking stimuli, neurons, and behavior within one framework.
• The pipeline addresses common calcium-imaging issues: indirect observations, noise, and slow temporal dynamics, by explicitly modeling latent spiking and calcium processes.
• The G-taxonomy provides a means to label neurons by whether their stimulus encoding contributes to behavioral readout, enabling population-level summaries beyond single-cell tuning curves.
• Implementations and data are publicly available (authors provided a DOI for MATLAB code), which supports reproducibility and adoption.

# Findings specific to the mouse auditory cortex data

The framework separated neurons into distinct functional categories based on sensori-behavioral relevance. It also revealed task-related reconfigurations of directed functional connectivity tied to behavioral accuracy, suggesting that network-level directional interactions differ between correct and incorrect trials.

# What to consider before applying this pipeline

• The approach is estimation- and model-driven: performance depends on model specification, inclusion of relevant predictors, and quality of imaging data.
• Computational components (state-space and variational inference) require appropriate choices of priors and initialization for stable fitting across large populations.

# Where to get the code and data

The authors made data and MATLAB implementations available via a DOI link included in the paper.