Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Integration of Calcium Imaging Traces via Deep Generative Modeling

Berta Ros, Mireia Olives-Verger, Caterina Fuses, Josep M Canals, Jordi Soriano, Jordi Abante

TL;DR

This work tackles batch effects and the need for single-neuron representations from calcium imaging traces by introducing supervised variational approaches. The authors develop a supervised variational autoencoder (SVAE) and a Gaussian Process–augmented SVAE (GPVAE) that incorporate batch labels and temporal structure to learn compact, biologically meaningful latent representations from calcium traces without spike inference. Across synthetic and experimental datasets, these models outperform Bayesian factor analysis and standard VAEs in preserving underlying neuronal dynamics while mitigating session-to-session variability, enabling robust visualization, clustering, and interpretation of single-neuron dynamics and revealing species- and condition-specific firing patterns. The framework holds promise for multispecies integration and multimodal single-cell analyses, offering a practical tool for analyzing heterogeneous neuronal activity at single-neuron resolution.

Abstract

Calcium imaging allows for the parallel measurement of large neuronal populations in a spatially resolved and minimally invasive manner, and has become a gold-standard for neuronal functionality. While deep generative models have been successfully applied to study the activity of neuronal ensembles, their potential for learning single-neuron representations from calcium imaging fluorescence traces remains largely unexplored, and batch effects remain an important hurdle. To address this, we explore supervised variational autoencoder architectures that learn compact representations of individual neurons from fluorescent traces without relying on spike inference algorithms. We find that this approach outperforms state-of-the-art models, preserving biological variability while mitigating batch effects. Across simulated and experimental datasets, this framework enables robust visualization, clustering, and interpretation of single-neuron dynamics.

Integration of Calcium Imaging Traces via Deep Generative Modeling

TL;DR

This work tackles batch effects and the need for single-neuron representations from calcium imaging traces by introducing supervised variational approaches. The authors develop a supervised variational autoencoder (SVAE) and a Gaussian Process–augmented SVAE (GPVAE) that incorporate batch labels and temporal structure to learn compact, biologically meaningful latent representations from calcium traces without spike inference. Across synthetic and experimental datasets, these models outperform Bayesian factor analysis and standard VAEs in preserving underlying neuronal dynamics while mitigating session-to-session variability, enabling robust visualization, clustering, and interpretation of single-neuron dynamics and revealing species- and condition-specific firing patterns. The framework holds promise for multispecies integration and multimodal single-cell analyses, offering a practical tool for analyzing heterogeneous neuronal activity at single-neuron resolution.

Abstract

Calcium imaging allows for the parallel measurement of large neuronal populations in a spatially resolved and minimally invasive manner, and has become a gold-standard for neuronal functionality. While deep generative models have been successfully applied to study the activity of neuronal ensembles, their potential for learning single-neuron representations from calcium imaging fluorescence traces remains largely unexplored, and batch effects remain an important hurdle. To address this, we explore supervised variational autoencoder architectures that learn compact representations of individual neurons from fluorescent traces without relying on spike inference algorithms. We find that this approach outperforms state-of-the-art models, preserving biological variability while mitigating batch effects. Across simulated and experimental datasets, this framework enables robust visualization, clustering, and interpretation of single-neuron dynamics.
Paper Structure (13 sections, 9 equations, 2 figures, 1 table)

This paper contains 13 sections, 9 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Synthetic data
  4. Experimental data
  5. Models
  6. Bayesian Factor Analysis
  7. Variational Autoencoder
  8. Supervised Generative Model
  9. Gaussian Process Supervised Generative Model
  10. Evaluation Metrics
  11. Results
  12. Discussion
  13. Conclusion

Figures (2)

  • Figure 1: Simulation results. (A) Examples of simulated traces, consisting of two groups, with three replicates each, with varying proportions of neuronal dynamic behaviors (RSor LTS) shown in different colors. (B) Silhouette over firing labels vs. kBET score over batch labels for noise with $\sigma^2 = 1.0$, showing that the supervised models reduce the impact of the technical variability in the embedding, resulting in smaller values of kBET. (C) UMAP of latent representation for BFA (left) and SVAE (right) model for $K = 4$, showing the reduction of batch effect in the latent space with a supervised model. (D) kBET score over batch labels for each model over different levels of noise, showing that supervised models are more robust to noisy data.
  • Figure 2: UMAP of experimental dataset obtained with SVAE for $K = 4$, showing the separation between mouse and rat traces and integration between experimental batches.