Early lock-in of structured and specialised information flows during neural development

David Shorten, Viola Priesemann, Michael Wibral, Joseph T. LizierbioRxiv 2021 · DOI

TL;DR

This study investigates how information processing emerges in developing neural networks. Using a novel continuous-time transfer entropy estimator on cultured neurons, researchers found that information flow dramatically increases and its spatial structure becomes "locked-in" early in development, with neurons adopting specialized roles during bursts that are also established early.

What Problem Does This Paper Address?

Understanding how complex computational abilities arise from the intricate wiring and dynamic interactions of neurons during development is a fundamental question in neuroscience. While we know a lot about information processing in mature brains, how this capacity emerges from nascent neural networks has been largely unexplored. This gap exists partly because traditional methods for analyzing information flow in neural data struggle with the precise timing and complex dynamics of developing systems. Specifically, existing tools often fail to capture information processing across multiple timescales simultaneously, a crucial aspect of neural activity.

Methods

The core of this research lies in applying a continuous-time transfer entropy (TE) estimator to analyze information flow in developing neural cultures. Unlike older discrete-time methods that require arbitrary time binning and can miss subtle temporal dynamics, this new estimator is designed to work directly with event-based data like spike trains. It can capture information flow across both short and long timescales simultaneously, is more data-efficient, and is statistically consistent. The researchers used this estimator to analyze multi-unit activity (MUA) from electrodes in cultured rat cortical neurons at various developmental stages. They also analyzed the dynamics during "population bursts" – synchronized periods of high neural activity – to understand how neurons adopt specialized roles. To explore the generality of their findings, they also applied the same analysis to a simulated network of Izhikevich neurons developing under a Spike-Timing-Dependent Plasticity (STDP) learning rule.

Datasets

The study utilized an open, freely available dataset of recordings from developing dissociated cortical rat neuron cultures. These recordings were collected by Wagenaar et al. [21] and are described in detail in their publication [24]. The dataset includes overnight recordings from multiple cultures at various days in vitro (DIV), ranging from early development to more mature stages. No new datasets were introduced.

Key Insights and Ideas

  • Information Flow "Locks-In" Early: Perhaps the most striking finding is that the spatial structure of information flow in these developing networks becomes remarkably stable early on. This suggests that the fundamental architecture of information processing is established relatively soon after the network begins to form, and then reinforced over time.
  • Specialized Roles During Bursts: During periods of intense, synchronized activity (population bursts), neurons don't just fire randomly. Instead, they appear to adopt distinct computational roles: some act as information "transmitters," others as "receivers," and some as "mediators" connecting the two.
  • Roles Also Lock-In Early: Crucially, these specialized roles during bursts also seem to be determined early in development, mirroring the lock-in of the overall information flow structure. This implies that a neuron's function within these dynamic events is set early and then maintained.
  • Model Mimics Biological Development: The fact that a simulated network developing under a biologically plausible learning rule (STDP) exhibits similar phenomena (increased information flow, early lock-in of structure, and specialized roles) suggests that these developmental patterns might be a general principle of neural network self-organization.

Key Results

  • Dramatic Increase in Information Flow: The total amount of information flowing across the neural cultures increases by orders of magnitude from early to later developmental stages. For example, mean TE values were near zero in early recordings but exceeded 0.049 nats.s⁻¹ in recordings beyond 20 DIV, and 0.15 nats.s⁻¹ beyond 28 DIV.
  • Spatial Structure Stability: Spearman correlations between TE values of source-target pairs across different recording days were positive and statistically significant in 14 out of 16 early-late pairs studied. This indicates a strong tendency for information flow patterns to persist.
  • Burst Position and Role Correlation: Later-bursting neurons showed a tendency for higher incoming information flow (suggesting receiver roles), while earlier-bursting neurons showed higher outgoing flow (suggesting transmitter roles). This correlation was statistically significant in many cases, particularly in later developmental stages.
  • Early Lock-in of Roles: The proportion of outgoing to total burst-local TE, indicative of a transmitter role, showed positive correlations between early and later DIV, suggesting that a neuron's role as a transmitter or receiver is established early.

Limitations and Future Work

  • Multi-Unit Activity (MUA): The study analyzed MUA from electrodes rather than individual sorted neurons. This means the analysis is at a slightly coarser spatial scale than possible with spike sorting, and it limits the ability to track specific neurons across recordings. Future work could involve spike sorting to gain finer-grained insights.
  • Model vs. Biological Data: While the model replicated key phenomena, there were differences in the magnitude of TE values, attributed to the model's higher spike rate. Further investigation is needed to understand these discrepancies and the stronger lock-in of outward TE in biological data versus inward TE in the model.
  • Higher-Order TE: The study primarily used bivariate TE. The authors suggest that future work could explore higher-order multivariate TE measures to more fully characterize complex computational roles and their emergence.
  • Log-Normal Distribution Fit: While previous studies often found log-normal distributions for TE values, this study's use of a continuous-time estimator showed deviations from both normal and log-normal distributions, suggesting the new estimator captures a more nuanced picture of information flow. Further investigation into these distribution differences is warranted.

Summary updated

Abstract

The brains of many organisms are capable of complicated distributed computation underpinned by a highly advanced information processing capacity. Although substantial progress has been made towards characterising the information flow component of this capacity in mature brains, there is a distinct lack of work characterising its emergence during neural development. This lack of progress has been largely driven by the lack of effective estimators of information processing operations for the spiking data available for developing neural networks. Here, we leverage recent advances in this estimation task in order to quantify the changes in information flow during development. We do so by studying the changes in the intrinsic dynamics of the spontaneous activity of developing dissociated neural cell cultures. We find that the quantity of information flowing across these networks undergoes a dramatic increase across development. Moreover, the spatial structure of these flows is locked-in during early development, after which there is a substantial temporal correlation in the information flows across recording days. We analyse the flow of information during the crucial periods of population bursts. We find that, during these bursts, nodes undertake specialised computational roles as either transmitters, mediators or receivers of information, with these roles tending to align with their spike ordering — either early, mid or late in the bursts. Further, we find that the specialised computational roles occupied by nodes during bursts tend to be locked-in early. Finally, we briefly compare these results to information flows in a model network developing according to an STDP learning rule from a state of independent firing to synchronous bursting. The phenomena of large increases in information flow, early lock-in of information flow spatial structure and computational roles based on burst position were also observed in this model, hinting at the broader generality of these phenomena. AUTHOR SUMMARY This paper studies the development of computation in biological systems by analysing changes in the flow of information in developing neural cell cultures. Although there have been a number of previous studies of information flows in neural cell cultures, this work represents the first study which compares information flows in the intrinsic dynamics across development time. Moreover, we make use of a recently proposed continuous-time transfer entropy estimator for spike trains, which, in comparison to the discrete-time estimator used previously, is able to capture important effects occurring on both small and large timescales simultaneously. We find that information flows begin to emerge after 5-10 days of activity, and crucially, the spatial structure of information flows remains significantly temporally correlated over the first month of recording. Furthermore, the magnitude of information flows across the culture are strongly related to burst position, and the roles of regions as information flow sources, sinks and mediators are found to remain consistent across development. Finally, we confirm that these early lock-ins also occur in a simple model network developing under an STDP update rule, suggesting a plausible mechanism undergirding this phenomenon.
BibTeX
@misc{shorten2021early,
  author = {David Shorten and Viola Priesemann and Michael Wibral and Joseph T. Lizier},
  title = {{Early lock-in of structured and specialised information flows during neural development}},
  year = {2021},
  doi = {10.1101/2021.06.29.450432},
  howpublished = {bioRxiv},
}
Figure