Inference of synaptic connectivity and external variability in neural microcircuits

Cody Baker, Emmanouil Froudarakis, Dimitri Yatsenko, Andreas S. Tolias, Robert Rosenbaum

TL;DR

This paper tackles the fundamental challenge of inferring the precise wiring (synaptic connectivity) of neural circuits from noisy recordings, a task complicated by unmeasured external inputs. By combining mathematical analysis, simulations, and real neural data, the authors show that while direct inference is tricky, the "precision matrix" of neural activity offers a surprisingly robust proxy for synaptic connections, especially when considering spatial variability.

What Problem Does This Paper Address?

Understanding how neurons are connected is crucial for deciphering brain function. However, inferring this "synaptic connectivity" from recordings is incredibly difficult. The main culprit is the "common input problem": external signals, not part of the circuit being studied, influence all neurons simultaneously, making it hard to distinguish their effect from direct connections between neurons. Many existing methods for measuring "functional connectivity" (how neurons' activities are related) often assume a direct link to synaptic connectivity, but this relationship is rarely proven, especially with real-world data where the true wiring is unknown. This paper aims to bridge this gap by investigating when and how functional connectivity, synaptic connectivity, and external input variability can be disentangled.

Methods

The authors employ a multi-pronged approach:

  • Mathematical Analysis: They derive analytical expressions relating synaptic connectivity to observable statistics of neural activity, particularly the precision matrix (the inverse of the covariance matrix). This involves developing theoretical frameworks for different network models, from simple linear systems to more complex spiking neuron models.
  • Neuronal Network Simulations: They use in silico models to generate ground truth data, allowing them to test their analytical findings and evaluate the accuracy of different inference methods under controlled conditions. They explore various network structures and noise levels.
  • Calcium Imaging Data: They validate their theoretical and simulation-based findings using real in vivo data from mouse visual cortex, demonstrating the practical applicability of their methods.

A key methodological contribution is the focus on the precision matrix as a more informative measure than simple covariance for inferring connectivity, especially when accounting for external input. They also explore how incorporating additional information, like cell type labels or spatial distance between neurons, can significantly improve inference accuracy.

Datasets

The paper utilizes:

  • Simulated Datasets: Generated from various neuronal network models (Ornstein-Uhlenbeck processes, Integrate-and-Fire models, Adaptive Exponential Integrate-and-Fire models) with controlled parameters for synaptic connectivity, external input, and noise.
  • In Vivo Calcium Imaging Data: From mouse primary visual cortex (V1) of PV-Cre/Ai9 mice, recorded using two-photon microscopy. This dataset allows for the evaluation of their methods on real neural activity.

Key Insights and Ideas

  • Precision Matrix is Key: The precision matrix, not just simple correlations, holds significant information about synaptic connectivity, even in the presence of correlated external input.
  • Common Input is a Major Hurdle, but Not Insurmountable: While correlated external input makes direct inference difficult, the precision matrix can still provide a close approximation to synaptic connectivity in practice.
  • Context Matters: Incorporating additional information like cell type (excitatory/inhibitory) or spatial distance between neurons dramatically improves the accuracy of inferring connectivity. This highlights the importance of leveraging biological constraints.
  • Trade-offs in Model Complexity: More biologically realistic models (e.g., spiking neurons) introduce complexities that can reduce inference accuracy compared to simpler linear models, but analytical frameworks can still offer valuable insights.
  • In Silico vs. In Vivo: While simulations are essential for understanding theoretical relationships, bridging the gap to real-world data requires careful consideration of experimental limitations (e.g., temporal resolution, sampling size).

Key Results

  • Improved Inference with Precision: The precision matrix, particularly when considering the spatial structure of neuronal variability, is shown to be closely related to synaptic connectivity.
  • Cell Type Labels Boost Accuracy: Using cell type information (e.g., excitatory vs. inhibitory) significantly improves the area under the ROC curve (AUROC) for connectivity inference, reaching near-perfect classification in some cases.
  • Spatial Distance Helps: Incorporating pairwise distance between neurons in spatial networks also enhances inference quality, demonstrating the value of geometric information.
  • Real Data Validation: Analysis of in vivo calcium imaging data suggests that external input to inhibitory neurons might be more exclusive than previously assumed, deviating from simpler models.
  • Quantitative Performance: AUROC values for distinguishing connected vs. unconnected pairs can range from around 0.6 in basic scenarios to near 1 when leveraging cell type or spatial information.

Limitations and Future Work

  • Idealized Models: The analytical derivations often rely on simplified network models (e.g., linear dynamics, Gaussian noise) which may not fully capture the complexity of real neural circuits.
  • Statistical Estimation Challenges: Accurately estimating the precision matrix from real data, especially with limited recording durations or noisy signals, remains a significant challenge. The paper acknowledges that their analysis provides an upper bound on performance and that regularization techniques are crucial in practice.
  • Subsampling Effects: The impact of subsampling (i.e., not recording from the entire network) on inference accuracy, particularly when external correlations arise from unobserved recurrent interactions, requires further investigation.
  • Non-linearities and Spiking Dynamics: While the paper extends to spiking neuron models, fully characterizing inference in these complex, non-linear systems remains an active area.
  • Future Directions: The authors suggest exploring more realistic network structures (e.g., multi-population models), investigating the effects of subsampling, and developing more robust statistical estimation methods for precision. They also highlight the potential for perturbational methods to infer connectivity, though their scalability is a concern.

Summary updated

Abstract

Abstract A major goal in neuroscience is to estimate neural connectivity from large scale extracellular recordings of neural activity in vivo . This is challenging in part because any such activity is modulated by the unmeasured external synaptic input to the network, known as the common input problem. Many different measures of functional connectivity have been proposed in the literature, but their direct relationship to synaptic connectivity is often assumed or ignored. For in vivo data, measurements of this relationship would require a knowledge of ground truth connectivity, which is nearly always unavailable. Instead, many studies use in silico simulations as benchmarks for investigation, but such approaches necessarily rely upon a variety of simplifying assumptions about the simulated network and can depend on numerous simulation parameters. We combine neuronal network simulations, mathematical analysis, and calcium imaging data to address the question of when and how functional connectivity, synaptic connectivity, and latent external input variability can be untangled. We show numerically and analytically that, even though the precision matrix of recorded spiking activity does not uniquely determine synaptic connectivity, it is often closely related to synaptic connectivity in practice under various network models. This relation becomes more pronounced when the spatial structure of neuronal variability is considered jointly with precision.
BibTeX
@article{baker2020inference,
  author = {Cody Baker and Emmanouil Froudarakis and Dimitri Yatsenko and Andreas S. Tolias and Robert Rosenbaum},
  title = {{Inference of synaptic connectivity and external variability in neural microcircuits}},
  journal = {J. Comput. Neurosci.},
  year = {2020},
  doi = {10.1007/s10827-020-00739-4},
}

Loading related content...

Figure