TL;DR
This paper introduces a novel method to directly estimate signal and noise correlations in neuronal populations from two-photon calcium imaging data. By bypassing the challenging step of spike deconvolution and explicitly modeling the fluorescence observation process, their approach offers a more robust and accurate way to understand neural circuit dynamics. This is crucial for deciphering how neural populations encode information, especially in complex biological systems.
What Problem Does This Paper Address?
Understanding how groups of neurons work together is fundamental to neuroscience. This collective activity, known as neuronal correlations, can be broadly categorized into "signal correlations" (responses time-locked to a stimulus) and "noise correlations" (trial-to-trial variability not tied to the stimulus). Two-photon calcium imaging is a powerful tool for observing these populations, but it presents significant challenges. The fluorescence signals are inherently blurred and noisy, making it difficult to accurately infer the underlying spiking activity. Existing methods often rely on deconvolution to estimate spikes, which is prone to errors and can propagate biases. Even if spikes are perfectly recovered, inferring network features from binary spike trains is complex due to non-linear relationships. This paper tackles the problem of directly and accurately estimating signal and noise correlations from raw calcium imaging data, avoiding the pitfalls of intermediate spike deconvolution.
Methods
The authors propose a Bayesian framework that explicitly models the two-photon fluorescence observation process. Instead of deconvolving spikes, their method directly links the observed fluorescence traces to underlying signal and noise correlations through a multi-tier model. Key to their approach is the integration of techniques from point processes and state-space modeling. They employ variational inference, a method for approximating complex probability distributions, combined with Pólya-Gamma augmentation and compressible state-space estimation. This allows them to efficiently estimate the signal and noise correlations without requiring spike deconvolution. The method provides theoretical guarantees on the performance of the estimator, including bounds on bias and variance.
Datasets
The authors evaluated their method using both simulated data generated according to their proposed model and experimentally recorded two-photon calcium imaging data from the mouse auditory cortex. The real data studies included:
- Mouse auditory cortex under random tone presentation.
- Mouse auditory cortex under spontaneous vs. stimulus-driven activity.
- Spatial analysis of signal and noise correlations in the mouse auditory cortex.
Key Insights and Ideas
- Bypassing Deconvolution is Key: The most significant insight is that accurate correlation estimation doesn't necessitate perfect spike deconvolution. Directly modeling the fluorescence signal and its relationship to correlations simplifies the problem and avoids error propagation.
- Unified Framework for Signal and Noise: The method jointly estimates both signal and noise correlations, allowing for a more comprehensive understanding of functional connectivity and how stimulus-dependent and independent information is processed.
- Robustness to Model Mismatches: The proposed method demonstrates remarkable robustness to deviations from its underlying modeling assumptions, such as non-linear stimulus integration or variations in calcium decay models, which is crucial for real-world biological data.
- Theoretical Guarantees Matter: Providing theoretical bounds on bias and variance offers confidence in the estimator's performance, especially in scenarios with limited trials or noisy data.
Key Results
- Superior Accuracy: In simulation studies, the proposed method consistently outperformed existing methods (Pearson correlations, two-stage Pearson, and two-stage GPFA) in estimating ground truth signal and noise correlations, achieving lower Normalized Mean Squared Error (NMSE) and reduced "leakage" between signal and noise estimates.
- Invariant Noise Correlations: For real data, the method showed that noise correlations remained invariant across spontaneous and stimulus-driven conditions, supporting hypotheses about distinct functional networks.
- Distinct Spatial Structures: The analysis of spatial correlations in the mouse auditory cortex revealed novel and distinct spatial trends for signal and noise correlations, differing from previous findings that suggested higher correlation between signal and noise.
- Scalability and Efficiency: The method's low-complexity variational updates allow it to scale favorably to larger neuronal populations and is applicable to single-session analyses with limited trials and duration.
Limitations and Future Work
- Spiking Rate Assumption: The generative model initially assumes at most one spiking event per time frame, which might not be optimal for bursting activity. However, the inference framework relaxes this by allowing larger spike magnitudes.
- Low Spiking Rate/High Noise: In regimes of extremely low spiking rates and high observation noise, the method's performance can degrade to that of existing Pearson correlation methods.
- Future Work: The authors suggest extensions to incorporate other correlation notions (e.g., connected correlation function), account for non-Gaussian structures, and extend to time-varying correlations or non-linear models using neural networks. They also emphasize the utility of their framework for constructing confidence intervals due to access to approximate posterior densities.
Summary updated