Abstract
Inferring functional connectivity from neuronal spike trains is a fundamental challenge in neuroscience. Traditional methods often struggle with discretization bias, high computational costs, and sensitivity to hyperparameters. We present ACE (Analysis of Connectivity in spiking Events), a novel statistical framework that models independent firing behavior to detect functional connections. By utilizing a derived null distribution and a $\chi^2$ test statistic, ACE provides a computationally efficient and parameter-robust solution for connectivity inference in both synthetic and realistic biological datasets.
Introduction
Understanding the communication pathways between neurons requires inferring connectivity from discrete spike events. However, synaptic transmission involves inherent delays, making the detection of these correlations non-trivial. Existing methodologies, such as Transfer Entropy or Inverse Covariance, face significant hurdles:
- Discretization Bias: Fixed-width time binning can obscure fine-grained temporal patterns.
- Computational Intensity: Many algorithms scale poorly, hindering real-time or large-scale network analysis.
- Hyperparameter Sensitivity: The need to pre-define search windows for delays often leads to missed connections if the window is poorly specified.
ACE addresses these issues by treating connectivity inference as a statistical hypothesis test against a model of independence.
Methods
The ACE framework operates through a multi-step pipeline designed for efficiency and robustness.
1. Single Neuron Modeling
ACE characterizes individual neurons using two primary parameters: the refractory period ($RP$) and the firing rate ($\lambda$). This allows for a baseline representation of intrinsic firing behavior.
2. The Null Model and Quantile Binning
The core of ACE is the derivation of a probability density function ($pdf$) describing the distribution of delays between two neurons under the assumption of independence. Instead of using arbitrary time bins, ACE employs the null model to create bins based on quantiles.
3. Statistical Testing
By comparing observed spike distributions against the uniform null model using a Pearson’s $\chi^2$ test, ACE identifies significant deviations that suggest functional connectivity. A high $\chi^2$ score indicates that the target neuron's firing is likely influenced by the source neuron.
Datasets and Evaluation
ACE was evaluated using:
- Synthetic Data: Controlled simulations with varying neuron counts (50–200), synaptic latencies, and noise levels.
- Realistic Simulated Data (RAT): Benchmarks from the CRCNS repository simulating the hippocampal CA1 subfield, providing known ground truth under biological conditions.
Key Results
ACE demonstrated superior performance across multiple metrics:
- Accuracy: On synthetic datasets, ACE achieved an AUPRC of 0.8626, significantly outperforming Inverse Covariance (~0.20) and Higher Order Transfer Entropy (HOTE, ~0.78).
- Robustness: On the RAT dataset, ACE maintained an AUROC of 0.9312, whereas HOTE20 dropped to 0.3042.
- Efficiency: ACE is the fastest algorithm tested, with computational complexity scaling linearly with the number of spikes ($M$) and neurons ($N$), enabling potential online processing.
Discussion and Limitations
While ACE is highly effective, it has limitations. It may struggle with "over-firing" scenarios where a source neuron fires faster than the synaptic delay to the target. Additionally, as a pairwise method, it does not inherently distinguish between direct (A $\rightarrow$ B) and transitive (A $ ightarrow$ C $ ightarrow$ B) connections. Future work will focus on post-processing filters to mitigate these transitive effects and further refine network mapping.
Summary updated