Dimensionality reduction of calcium-imaged neuronal population activity

Tze Hui Koh, William E. Bishop, Takashi Kawashima, Brian B. Jeon, Ranjani Srinivasan, Sandra J. Kuhlman, Misha B. Ahrens, Steven M. Chase, Byron M. Yu

Abstract

Calcium imaging has been widely adopted for its ability to record from large neuronal populations. To summarize the time course of neural activity, dimensionality reduction methods, which have been applied extensively to population spiking activity, may be particularly useful. However, it is unclear if the dimensionality reduction methods applied to spiking activity are appropriate for calcium imaging. We thus carried out a systematic study of design choices based on standard dimensionality reduction methods. We also developed a novel method to perform deconvolution and dimensionality reduction simultaneously (termed CILDS). CILDS most accurately recovered the single-trial, low-dimensional time courses from calcium imaging that would have been recovered from spiking activity. CILDS also outperformed the other methods on calcium imaging recordings from larval zebrafish and mice. More broadly, this study represents a foundation for summarizing calcium imaging recordings of large neuronal populations using dimensionality reduction in diverse experimental settings.

Introduction

In modern neuroscience, researchers often record from hundreds or thousands of neurons simultaneously. To interpret this high-dimensional data, dimensionality reduction (DR) is used to extract latent variables—the low-dimensional "neural trajectories" that represent underlying computations. While DR was originally developed for electrophysiology (spikes), it is increasingly applied to calcium imaging.

Calcium imaging offers scalability but presents a challenge: it measures fluorescence, which rises quickly upon firing but decays slowly over hundreds of milliseconds. This "blurring" effect introduces artificial temporal correlations. Applying standard DR directly to raw fluorescence can lead to models mistaking slow calcium decay for slow neural processing. This paper investigates whether to deconvolve data first or if a joint approach can better recover the true neural signals.

Methods

The study compares three primary approaches to extracting latent variables from calcium imaging data:

  1. LDS on Fluorescence: Applying a standard Linear Dynamical System (LDS) directly to raw fluorescence traces.
  2. deconv-LDS: A two-stage approach where each neuron is independently deconvolved (e.g., using the OASIS algorithm) to estimate spikes, followed by LDS.
  3. CILDS (Calcium Imaging Linear Dynamical System): A novel unified probabilistic framework that performs deconvolution and dimensionality reduction simultaneously.
Comparison of three dimensionality reduction approaches for calcium imaging data.

CILDS assumes observed fluorescence is a noisy, filtered version of a latent state evolving via linear dynamics. By treating the population jointly, CILDS leverages shared activity across neurons to improve deconvolution estimates for individual cells. The researchers also tested CIFA (Calcium Imaging Factor Analysis), a version without temporal dynamics, to demonstrate that a dynamics model is essential to distinguish neural signals from calcium decay.

Results

The performance of these methods was validated using synthetic data with known "ground truth" latents and biological datasets from larval zebrafish and mouse V1.

Performance of CILDS compared to LDS and deconv-LDS across various experimental settings and parameters.

Key Insights

  • The Joint Advantage: CILDS outperforms two-stage methods by using population-level information. Shared context helps the model distinguish between real neural events and noise in individual fluorescence traces.
  • Dynamics are Essential: Without a model of neural state evolution, it is mathematically difficult to separate slow neural processes from slow calcium decay. CIFA failed significantly as neural activity became slower, erroneously attributing neural smoothness to calcium kinetics.
  • The Speed Threshold: CILDS is most beneficial for fast-paced neural activity (tens to hundreds of milliseconds). While simpler methods catch up when neural activity is very slow (seconds), CILDS remains the most robust across timescales.

Performance Metrics

  • Simulation Accuracy: CILDS consistently recovered ground truth latent trajectories with higher $R^2$ values than LDS or deconv-LDS across varying timescales, neuron counts, and noise levels.
  • Predictive Power: In "leave-neuron-out" tests on real data, CILDS predicted held-out neuron activity significantly better than competitors. In zebrafish data, CILDS outperformed deconv-LDS for 85% of neurons.
  • Decoding Accuracy: In mouse V1, latent variables from CILDS allowed a decoder to identify visual stimuli 1.16x more accurately than deconv-LDS and 1.43x more accurately than standard LDS.

Discussion

While CILDS offers superior accuracy, it is more computationally intensive than standard LDS due to complex matrix inversions required at each time point. The current model assumes linear dynamics and a first-order autoregressive process for calcium kinetics. Future work could incorporate non-linear dynamics or higher-order calcium models to capture more complex brain states and higher-speed imaging nuances. Despite these limitations, CILDS provides a robust foundation for summarizing large-scale calcium imaging data.

Summary updated

Abstract

ABSTRACT Calcium imaging has been widely adopted for its ability to record from large neuronal populations. To summarize the time course of neural activity, dimensionality reduction methods, which have been applied extensively to population spiking activity, may be particularly useful. However, it is unclear if the dimensionality reduction methods applied to spiking activity are appropriate for calcium imaging. We thus carried out a systematic study of design choices based on standard dimensionality reduction methods. We also developed a novel method to perform deconvolution and dimensionality reduction simultaneously (termed CILDS). CILDS most accurately recovered the single-trial, low-dimensional time courses from calcium imaging that would have been recovered from spiking activity. CILDS also outperformed the other methods on calcium imaging recordings from larval zebrafish and mice. More broadly, this study represents a foundation for summarizing calcium imaging recordings of large neuronal populations using dimensionality reduction in diverse experimental settings.
BibTeX
@article{koh2023dimensionality,
  author = {Tze Hui Koh and William E. Bishop and Takashi Kawashima and Brian B. Jeon and Ranjani Srinivasan and Sandra J. Kuhlman and Misha B. Ahrens and Steven M. Chase and Byron M. Yu},
  title = {{Dimensionality reduction of calcium-imaged neuronal population activity}},
  journal = {Nat. Comput. Sci.},
  year = {2023},
  doi = {10.1038/s43588-022-00390-2},
}
Figure