Spike sorting with Kilosort4

Spike sorting, the computational process of identifying the precise firing times of individual neurons from recordings of local electrical fields, is a fundamental yet complex task in neuroscience. The complexity arises from several factors inherent to the experimental setup and the biological signals being measured. Recordings are inherently non-stationary, meaning that the characteristics of the signals can change over time. This can be attributed to various factors, such as electrode drift within the brain tissue. Furthermore, the electrical fields generated by nearby neurons significantly overlap, making it difficult to isolate the contribution of individual neurons. These challenges have motivated the development of sophisticated computational methods to perform accurate spike sorting. This paper focuses on the Kilosort framework, an open-source software package that has been under continuous development to improve the accuracy and efficiency of spike sorting. The paper describes the algorithmic improvements introduced in different versions of Kilosort, culminating in the release of Kilosort4, which incorporates a novel graph-based clustering algorithm leading to substantial performance gains. The improved performance of Kilosort4 is rigorously evaluated using a novel realistic simulation framework that accounts for the non-stationarity and overlapping nature of the signals in real-world recordings. The improved ability to accurately extract neural activity from dense recordings is of significant importance to the neuroscience field, enabling more precise investigations into brain function and network dynamics.

Classical spike-sorting frameworks typically involve a series of sequential operations: preprocessing, spike detection, clustering, and postprocessing. Early approaches often relied on simple clustering methods, such as k-means, and suffered from limitations in handling non-stationary data and resolving overlapping spikes. Modern approaches have incorporated more advanced clustering algorithms, including density-based approaches and agglomerative approaches using bimodality criteria, to improve clustering accuracy. However, these methods often lack the ability to effectively handle overlapping spikes which can contaminate the signals being analyzed. The original Kilosort attempted to address this limitation by innovatively combining the spike detection and clustering steps into a single template learning process, alongside a novel matching pursuit technique for resolving overlapping spikes. This approach has proven effective in improving performance compared to previous methods, but further improvements were sought through ongoing development of the framework.

Kilosort4 builds upon previous versions, inheriting some algorithmic steps such as drift correction and matching pursuit, while introducing key improvements in the clustering algorithm. The core innovation lies in the adoption of a graph-based clustering approach based on modularity optimization, combined with a merging tree strategy. This framework uses a feature extraction pipeline to initially detect spikes and extract relevant features. This involves the use of simple templates to identify initial spike waveforms. Subsequently, a graph-based clustering algorithm is applied to refine the template set and generate refined templates for spike detection and background subtraction. The core clustering algorithm in Kilosort4 is applied twice. The first application occurs during template deconvolution to learn the templates, and the second application is performed on the deconvolved features to determine the final cluster identities. The graph-based approach constructs a graph where nodes represent spikes and edges connect nearest neighbors in Euclidean space. A cost function, derived from graph properties, guides the clustering. The algorithm addresses challenges faced by traditional modularity optimization algorithms, such as Leiden and Louvain, by incorporating a merging tree approach that incorporates domain knowledge (refractory period violations and projection bimodality) to make merge/split decisions. To enhance computational speed, a landmark-based version of graph clustering is used. This involves using nearest neighbors within a subset of all data points, which significantly reduces the computational complexity. The simulation framework employs hybrid and full simulations, generating spike trains and background activity to reflect real-world conditions. Hybrid simulations use real recordings as background and add simulated spikes from ground truth waveforms at vertically offset positions to create a more realistic scenario. Full simulations build upon this by adding simulated 1/f noise and multi-unit activity using units with small spike norms. Realistic drifting simulations were generated using actual drift patterns identified in the IBL dataset, encompassing various drift conditions (no drift, medium drift, high drift, fast drift, and step drift). Waveforms from real recordings exhibiting large drift were utilized to generate the simulation waveforms. The drift was added by applying drift patterns from real experiments to simulated spike trains. The performance of Kilosort4 was compared to several other state-of-the-art spike-sorting algorithms using established metrics such as false positive (FP) and false negative (FN) rates. An ablation study examined the individual contribution of various steps in Kilosort4, quantifying the impact of each component on the overall performance. The software uses Python and the PyTorch library for GPU processing and includes a PyQt GUI for user interaction.

The paper's key findings demonstrate the superior performance of Kilosort4 across a variety of simulated conditions. Nearly all versions of Kilosort outperformed other algorithms, with Kilosort4 achieving the best results in all cases. In simulations without drift, all Kilosort versions except Kilosort1 outperformed competing algorithms, with Kilosort4 showing the highest accuracy. This performance advantage was maintained in full simulations, which included simulated 1/f noise and multi-unit activity. Importantly, the study addressed the limitations of previous biophysical simulations, demonstrating that existing biophysical models produce unrealistic waveforms, leading to inaccurate assessments of spike sorter performance. The simulations with realistic drift showed a consistent performance advantage for Kilosort4, even under challenging conditions such as high drift, fast drift, and step drift. Kilosort4 consistently achieved higher accuracy in unit identification compared to other algorithms, including IronClust which accounts for drift differently. The ablation study revealed that drift correction, deconvolution, and cross-correlogram-based merging/splitting significantly impact the performance of Kilosort4, while nonrigid motion correction and deconvolution for feature extraction played smaller, though still consistent, roles. Importantly, the high accuracy of Kilosort4 was not achieved at the cost of a high false positive unit rate. The analysis shows that Kilosort4 had similar false positive unit rates compared to other algorithms. Finally, Kilosort4 demonstrated minimal dependence on ground-truth firing rates, spike norms, and spatial extents, in contrast to other algorithms which displayed substantial dependence on these parameters. The improved performance of Kilosort4 was consistently observed across different drift conditions and types of recordings, including datasets from other recording probes and configurations.

The findings of this study highlight the significant improvements achieved in spike sorting through the development of Kilosort4. The superior performance of Kilosort4, particularly in handling drift and overlapping spikes, is a major advance in the field. The incorporation of graph-based clustering, combined with the merging tree strategy and domain knowledge, has proven crucial in achieving better accuracy and robustness. The development of a realistic simulation framework, which accurately represents the complexities of real electrophysiological data, allows for robust benchmarking and assessment of spike-sorting algorithms. The findings have implications for various areas of neuroscience research, enabling more accurate and reliable analysis of neural data from high-density recordings. The improved spike sorting performance will facilitate more detailed investigations of neural circuit function, contributing to a deeper understanding of the brain's complex computations. The robustness of Kilosort4 across different drift conditions suggests its broad applicability and potential to standardize neural data analysis across laboratories and experimental setups.

This paper presents Kilosort4, a significant advancement in spike-sorting algorithms. Its graph-based clustering algorithm and merging tree strategy offer superior performance compared to existing methods, especially in handling drift and resolving overlapping spikes. The use of realistic simulations allows for robust benchmarking and highlights the limitations of previous approaches. Kilosort4's adaptability to various recording types and its open-source nature facilitate broader adoption and further development within the neuroscience community. Future work may explore further refinements to the algorithm and its application to increasingly complex recording environments, such as those employing ultra-high-density probes.

While Kilosort4 demonstrates superior performance, certain types of data may require special considerations. Data lacking well-defined geometry or with excessive vertical spacing between electrodes may present challenges for drift correction. Recordings from chronic experiments over multiple days, potentially separated by long intervals, were not explicitly tested but the step drift simulation provides some indication of performance in such scenarios. The reliance on the IBL dataset for simulation parameter generation may introduce a degree of bias, though efforts were made to use a wide range of data from different laboratories and brain areas. Future work should investigate the algorithm's performance on diverse datasets spanning a wider range of recording technologies, experimental conditions, and neural systems.