A Persistent Homology Pipeline for the Analysis of Neural Spike Train Data

In this article, we introduce a Topological Data Analysis (TDA) pipeline for neural spike train data. Understanding how the brain transforms sensory information into perception and behavior requires analyzing coordinated neural population activity. Modern electrophysiology enables simultaneous recording of spike train ensembles, but extracting meaningful information from these datasets remains a central challenge in neuroscience. A fundamental question is how ensembles of neurons discriminate between different stimuli or behavioral states, particularly when individual neurons exhibit weak or no stimulus selectivity, yet their coordinated activity may still contribute to network-level encoding. We describe a TDA framework that identifies stimulus-discriminative structure in spike train ensembles recorded from the mouse insular cortex during presentation of deionized water stimuli at distinct non-nociceptive temperatures. We show that population-level topological signatures effectively differentiate oral thermal stimuli even when individual neurons provide little or no discrimination. These findings demonstrate that ensemble organization can carry perceptually relevant information that standard single-unit analysis may miss. The framework builds on a mathematical representation of spike train ensembles that enables persistent homology to be applied to collections of point processes. At its core is the widely-used Victor-Purpura (VP) distance. Using this metric, we construct persistence-based descriptors that capture multiscale topological features of ensemble geometry. Two key theoretical results support the method: a stability theorem establishing robustness of persistent homology to perturbations in the VP metric parameter, and a probabilistic stability theorem ensuring robustness of topological signatures.

💡 Research Summary

This paper introduces a novel analytical pipeline that applies Topological Data Analysis (TDA), specifically persistent homology, to decipher population-level information from simultaneously recorded neural spike trains. The central challenge addressed is that while individual neurons may show weak or no selectivity to different stimuli, their coordinated ensemble activity might still encode discriminative information. The proposed framework extracts this ensemble-level structure where traditional single-neuron analyses fail.

The methodology is built upon a rigorous mathematical foundation. First, key entities are formally defined: spike trains as finite sets of time points, neurons as functions mapping stimuli to probability distributions over spike trains, and neuronal networks as distributions over ensembles (rasters) of spike trains. To compare spike trains within an ensemble, the biologically-inspired Victor-Purpura (VP) distance is employed. The authors provide an equivalent reformulation of this distance as an optimal partial bijection problem, which is more amenable to subsequent theoretical analysis.

The core of the pipeline treats a single-trial raster (a set of spike trains from multiple neurons) as a finite metric space using the VP distance. Persistent homology is then applied to this space. This process tracks the evolution of topological features (like connected components and loops) across a range of distance scales, resulting in a topological signature (e.g., a persistence barcode) that summarizes the multiscale geometric organization of the neural ensemble in response to a stimulus.

A significant contribution of the work is the provision of theoretical guarantees for the pipeline’s robustness. Two key stability theorems are proven: 1) A Lipschitz stability theorem showing that small changes in the VP distance’s cost parameter q lead to bounded changes in the resulting persistent homology, and 2) A probabilistic stability theorem ensuring the robustness of the topological signatures to variability and finite sampling from the underlying neural response distributions. These theorems provide mathematical confidence in the method’s reliability under realistic experimental conditions.

The utility of the pipeline is demonstrated through both synthetic and biological experiments. Synthetic data experiments showcase a powerful strength: even when individual neurons carry zero discriminative information about two stimuli, the population-level topological structure can enable perfect classification. In biological experiments, the method is applied to spike train ensembles recorded from the mouse insular cortex during the presentation of deionized water at different non-nociceptive temperatures. The results show that classification based on ensemble topological signatures often outperforms analysis based on individual neuron responses, effectively differentiating the thermal stimuli.

In conclusion, this work establishes a general, theoretically-grounded TDA framework for analyzing population activity. It demonstrates that topological signatures can capture perceptually relevant information in neural ensemble dynamics that is inaccessible to standard methods. While focused on neuroscience, the framework is broadly applicable to any data modality consisting of simultaneous point processes, such as gene expression timing or financial trade data, offering a new lens for analyzing temporal patterns in complex systems.