Heterogeneous and higher-order cortical connectivity undergirds efficient, robust and reliable neural codes

Daniela Egas Santander, Christoph Pokorny, András Ecker, Jānis Lazovskis, Matteo Santoro, Jason P. Smith, Kathryn Hess, Ran Levi, Michael ReimannbioRxiv 2024 · DOI

TL;DR

This paper reveals that the brain's complex, non-uniform wiring (higher-order connectivity) isn't just noise; it's a crucial mechanism for balancing efficient information processing with robust reliability. By developing a new metric based on "simplicial complexes," the researchers show that different brain regions specialize: some are optimized for speed and efficiency, while others boost overall network reliability, allowing both to coexist within the same brain.

What Problem Does This Paper Address?

The brain faces a fundamental challenge: how to process information efficiently while also being robust to errors and noise. This is often framed as a "robustness-efficiency tradeoff." Simple models suggest that increasing redundancy (making the code more robust) comes at the cost of efficiency. The authors propose that the brain's intricate, heterogeneous structure, particularly its "higher-order" connectivity (beyond simple pairwise connections), might be the key to overcoming this tradeoff. They hypothesize that different "subpopulations" of neurons within the brain might be wired differently, allowing them to optimize for different aspects of this tradeoff, rather than the whole network being uniformly optimized.

Methods

The core of the paper's methodological innovation lies in the development of a new metric to quantify the complexity of neural connectivity. They use the mathematical theory of simplicial complexes to capture "higher-order" structures – patterns of connectivity involving more than just pairs of neurons. This metric, called simplicial neighborhood complexity, compares the actual connectivity of a neural subpopulation to that of a random network. This allows them to distinguish between "low complexity" and "high complexity" neighborhoods. They then use this metric to analyze real neural data and simulations, correlating connectivity complexity with functional measures like efficiency and reliability.

Datasets

The study utilizes several large-scale, experimentally derived neural datasets:

  • C. elegans connectome: A full brain reconstruction of the nematode worm.
  • Drosophila larva connectome: A reconstruction of the fruit fly larva brain.
  • MICrONS connectome: An electron microscopy reconstruction of a cubic millimeter of mouse primary visual cortex, with co-registered functional data.
  • BBP connectome: A morphologically detailed model of the rat somatosensory cortex, used for simulations.

Key Insights and Ideas

  • Brain wiring is not random: Biological neural networks exhibit significantly more complex higher-order structures (like directed simplices) than random network models.
  • Complexity is key to the tradeoff: The brain's heterogeneity in connectivity complexity is not a bug, but a feature that allows it to navigate the robustness-efficiency tradeoff.
  • Specialized neighborhoods: Low-complexity neighborhoods are efficient for information processing, while high-complexity neighborhoods boost overall network reliability.
  • Coexistence is crucial: These specialized neighborhoods can and do coexist within the same biological neural network, enabling a sophisticated balance of function.

Key Results

  • Efficiency in low complexity: Subpopulations with low simplicial complexity were found to exhibit more efficient neural activity, meaning they could encode information with fewer neurons.
  • Reliability in high complexity: Subpopulations with high simplicial complexity played a crucial role in enhancing the overall reliability of the network, acting as a buffer against noise.
  • Global reliability boost: Manipulating network complexity (e.g., by adding reciprocal connections) globally increased neuronal reliability, demonstrating a system-wide effect.
  • Biological relevance: These findings held true across multiple species (C. elegans, Drosophila, mouse, rat model) and were observed in both experimental data and detailed simulations.

Limitations and Future Work

  • Computational cost: While simplex counting is efficient, it can still be computationally prohibitive for very large and dense networks. Future work might involve modifying these methods or simplifying connectivity networks.
  • Weighted metrics: The current analysis primarily uses unweighted network metrics. Extending the analysis to weighted metrics (considering synaptic strength) could provide deeper insights.
  • In vivo data sparsity: The in vivo data (MICrONS) was sparsely sampled, limiting the depth of some analyses, particularly for efficiency and layer-specific reliability. Future work could benefit from more densely sampled in vivo data with co-registered activity.
  • Noise correlation: The study did not directly investigate the effect of trial-to-trial variability (noise correlation) on network efficiency, which is a promising avenue for future research.
  • Human connectomes: While human connectomes show even more complex architectures, their analysis is limited by scale. Future work could explore these using topological methods like q-analysis.

Summary updated

Abstract

Abstract Simplified models of neural networks have demonstrated the importance of establishing a reasonable tradeoff between memory capacity and fault-tolerance in cortical coding schemes. The intensity of the tradeoff is mediated by the level of neuronal variability. Indeed, increased redundancy in neuronal activity enhances the robustness of the code at the cost of the its efficiency. We hypothesized that the heterogeneous architecture of biological neural networks provides a substrate to regulate this tradeoff, thereby allowing different subpopulations of the same network to optimize for different objectives. To distinguish between subpopulations, we developed a metric based on the mathematical theory of simplicial complexes that captures the complexity of their connectivity, by contrasting its higher-order structure to a random control. To confirm the relevance of our metric we analyzed several openly available connectomes, revealing that they all exhibited wider distributions of simplicial complexity across subpopulations than relevant controls. Using a biologically detailed cortical model and an electron microscopic data set of cortical connectivity with co-registered functional data, we showed that subpopulations with low simplicial complexity exhibit efficient activity. Conversely, subpopulations of high simplicial complexity play a supporting role in boosting the reliability of the network as a whole, softening the robustness-efficiency tradeoff. Crucially, we found that both types of subpopulations can and do coexist within a single connectome in biological neuronal networks, due to the heterogeneity of their connectivity. Our work thus suggests an avenue for resolving seemingly paradoxical previous results that assume homogeneous connectivity.
BibTeX
@misc{santander2024heterogeneous,
  author = {Daniela Egas Santander and Christoph Pokorny and András Ecker and Jānis Lazovskis and Matteo Santoro and Jason P. Smith and Kathryn Hess and Ran Levi and Michael Reimann},
  title = {{Heterogeneous and higher-order cortical connectivity undergirds efficient, robust and reliable neural codes}},
  year = {2024},
  doi = {10.1101/2024.03.15.585196},
  howpublished = {bioRxiv},
}
Figure