TL;DR
This paper introduces CILVA, a novel statistical model that disentangles stimulus-evoked neural activity from spontaneous, internal brain activity in calcium imaging data. By simultaneously modeling both, CILVA provides a more accurate, single-trial understanding of how neural populations encode information, overcoming biases introduced by spontaneous fluctuations.
What Problem Does This Paper Address?
Understanding how neural populations encode sensory information is a fundamental question in neuroscience. Calcium imaging, a popular technique for monitoring neural activity, is particularly susceptible to a major challenge: spontaneous activity (SA). SA refers to neural firing that isn't directly triggered by external stimuli. This ongoing, internal brain activity can significantly influence and even mask the responses to stimuli, leading to biased estimates of neural encoding. This is especially problematic in calcium imaging due to the slow kinetics of calcium indicators, which blur the temporal resolution and make it difficult to separate stimulus-evoked events from spontaneous ones on a single-trial basis. Existing methods often struggle to disentangle these two sources of activity, particularly within a single experimental trial.
Methods
The core contribution of this paper is the development of the Calcium Imaging Latent Variable Analysis (CILVA) model. CILVA is a generative statistical model that simultaneously estimates both stimulus-evoked responses and latent sources of spontaneous activity.
Key aspects of the CILVA model include:
- Joint Modeling: Unlike sequential approaches that first estimate evoked activity and then analyze the residuals for spontaneous activity, CILVA models both components concurrently.
- Latent Factors for SA: CILVA posits that spontaneous activity arises from a small number of underlying "latent factors." These factors are not directly measured but are inferred from the data and are assumed to influence subsets of neurons.
- Linear Combination of Sources: The model assumes that the intensity of calcium influx for a neuron at any given time is a linear combination of the stimulus drive (processed through a receptive field) and the influence of these latent SA factors.
- GCaMP Kernel Convolution: The model explicitly incorporates the known kinetics of calcium indicators (like GCaMP6s) by convolving the inferred activity intensity with a difference-of-exponentials kernel.
- Sparsity Prior on Latent Factors: To encourage the model to attribute activity to stimuli when possible, a sparsity-inducing prior is placed on the latent factors, effectively encouraging them to be active only when necessary.
- Model Selection: A procedure is described for selecting the optimal number of latent factors and the sparsity parameter.
The paper also contrasts CILVA with a baseline approach using non-negative least squares (NNLS) and non-negative matrix factorization (NMF) on residual data, highlighting CILVA's advantages in avoiding bias.
Datasets
The authors applied CILVA to two distinct datasets:
- Larval Zebrafish Optic Tectum: Two-photon calcium imaging data from developing zebrafish larvae responding to visual stimuli presented at varying angles. This dataset was collected by the authors.
- Mouse Primary Visual Cortex (V1): Publicly available data from mouse V1, featuring more complex stimuli (drifting gratings with multiple orientations, spatial, and temporal frequencies) presented more rapidly.
The paper also utilized simulated data to validate the model's performance under controlled conditions.
Key Insights and Ideas
- Spontaneous Activity is Structured and Informative: The study demonstrates that SA is not just random noise but often exhibits structured, low-dimensional patterns that can be identified and modeled. This suggests SA may carry meaningful information about internal brain states or computations.
- Decoupling is Crucial for Accurate Encoding Estimates: By explicitly accounting for SA, CILVA provides more refined estimates of stimulus-evoked receptive fields and retinotopic maps, which are less biased by spontaneous fluctuations.
- Latent Factors Target Specific Neuron Groups: The identified latent factors for SA tend to influence distinct, non-overlapping sets of neurons, suggesting specialized internal processes.
- Individual Neurons Vary in Their Drive: The analysis reveals a spectrum of neural responses, with some neurons primarily driven by external stimuli (evoked activity) and others by internal factors (spontaneous activity), and many exhibiting a mix.
Key Results
- Realistic Model Fits: CILVA provided good fits to both zebrafish and mouse visual cortex calcium imaging data, with correlation coefficients between raw traces and model fits significantly higher than those obtained with shuffled data.
- Improved Receptive Field Estimates: In zebrafish, CILVA-estimated tuning curves led to more refined retinotopic maps compared to simple averaging methods.
- Quantification of EA vs. SA Contribution: The "drive ratio" analysis showed that neurons were largely biased towards being either stimulus-driven or spontaneous, rather than responding equally.
- Significant Improvement with SA Modeling: Incorporating SA into the model led to a substantial increase in the quality of model fits, particularly for neurons that were poorly explained by stimulus-evoked activity alone. For example, in the zebrafish data, accounting for SA improved the mean correlation coefficient by 34% on average across 8 fish.
- Identification of Latent SA Patterns: CILVA successfully identified low-dimensional patterns of SA in both datasets, which were often difficult to discern in the raw data.
Limitations and Future Work
- Additive Interaction Assumption: The current model assumes an additive interaction between evoked and spontaneous activity. The authors acknowledge that more complex interactions (e.g., multiplicative effects) might exist and could be incorporated in future work by using nonlinear activation functions.
- Identifiability: Like other factor analysis methods, CILVA faces identifiability challenges, which are addressed through standardization procedures.
- Temporal Dynamics of Latent Factors: The model assumes sparse latent factor activity, with temporal dynamics arising from the convolution with the calcium kernel. While this simplifies the model, explicitly modeling smooth temporal dynamics for latent factors could be explored.
- Private Variability: The model distinguishes between shared SA (driven by latent factors) and private SA. The authors note that their estimate of private variability is an upper bound, as some "private" SA might be shared but unobserved due to limitations in the recorded neuron population.
- Future Directions: The authors suggest exploring multiplicative interactions, incorporating more complex temporal dynamics for latent factors, and applying CILVA to other brain regions and data types.
Summary updated