Abstract
Sorting action potentials (spikes) from extracellular recordings of large groups of connected neurons is essential to understanding brain function. Simulations with known spike times have driven significant advances in spike sorting, but present models do not account for neuronal heterogeneity and its effect on sorting accuracy. Here, we used a large-scale detailed cortical microcircuit model to simulate recordings, evaluate modern spike sorters, and link their performance to neuronal heterogeneity. We also exposed the network to various stimuli to investigate how sorting errors affect stimulus discrimination. Spike sorters successfully isolated about 10% of neurons within 50 microns of the electrode shank. This undersampling had no impact on stimulus discrimination ability. However, sorting biases related to firing rate, spike extent, synaptic type, and layer reduced its discrimination ability by nearly half. These findings show realistic models are a complementary method to evaluate and improve spike sorting and, hence, improve our understanding of neural activity.
Introduction
In modern neuroscience, "spike sorting" is the foundational computational step of assigning electrical pulses recorded by extracellular probes to individual neurons. To validate these algorithms, researchers typically rely on "ground truth" simulations where the exact timing of every spike is known. However, most existing benchmarks utilize oversimplified models that treat neurons as identical points, ignoring the complex, heterogeneous reality of biological brain tissue. This study addresses this "validation gap" by quantifying the information loss and distortion that occurs when state-of-the-art sorters are applied to a biophysically realistic environment.
Methods and Simulation Framework
The researchers utilized the Blue Brain Project’s large-scale model of a rat primary somatosensory cortex (S1) column. This model is exceptionally detailed, containing 30,190 neurons across six layers, encompassing 60 different morphological types and 11 electrical firing classes.
To simulate extracellular recordings, the team modeled two probe designs—Neuropixels 1.0 and a "dense" probe—using line-source and point-source equations to replicate how electrical currents propagate through tissue. They evaluated the performance of the entire Kilosort suite (versions 1 through 4) and Herdingspikes against this biophysical ground truth.
Datasets and Probe Geometries
The study leveraged a newly generated, publicly available dataset of simulated cortical recordings with full biological heterogeneity. To ensure the simulation's validity, results were benchmarked against real-world recordings from the Marques-Smith and Horvath datasets, as well as the standard but less detailed Buccino synthetic model.
Key Insights and Results
The Bias Toward "Loud" Neurons
Spike sorters exhibited a heavy bias toward "prominent" neurons, particularly Layer 5 pyramidal cells. These neurons possess large spatial footprints and high firing rates, making them significantly easier for algorithms to detect. Conversely, smaller or more sparsely firing interneurons were frequently ignored, leading to a systematic overestimation of mean firing rates in sorted data by a factor of 3 to 4.
Undersampling vs. Selection Bias
A critical finding was that while sorters only isolated approximately 10% of neurons within 50µm of the probe, this undersampling was not the primary driver of information loss. If the sampling were random, decoding performance would remain stable. Instead, the damage stems from selection bias. Because sorters systematically prefer specific cell types and firing patterns, they distort the mathematical structure of population activity (the neural manifold).
Information Loss
This distortion resulted in a nearly 50% reduction in the ability of a linear classifier to distinguish between simulated whisker-deflection stimuli. Furthermore, sorting accuracy was found to be ten times lower in the biophysical model compared to simpler synthetic models, suggesting that current tools struggle with biological complexity far more than previously recognized.
Discussion and Future Directions
While the model provides a powerful "digital twin" for validation, it does not yet replicate slow-wave oscillations (1-4 Hz) typical of anesthetized states, which may further complicate sorting. Future work will focus on using these realistic simulations to train a new generation of spike sorters capable of handling and correcting for the inherent biological biases of extracellular recordings. To assist current researchers, the authors developed a Lasso-regularized logistic regression model that automates the identification of high-quality single units with 90% precision.
Summary updated