henhaobio
Sign in

bioRxivNeurosciencePreprint

YASS: Yet Another Spike Sorter

JinHyung LeeiD, David CarlsoniD, Hooshmand Shokri, Weichi YaoiD, Georges A. GoetziD, Espen HageniD, Eleanor Batty, E. J. ChichilniskyiD, Gaute T. EinevolliD, Liam PaninskiiD

PDFDOI
In-depth full-text digest available · Jump to digest ↓

Abstract

Spike sorting is a critical first step in extracting neural signals from large-scale electrophysiological data. This manuscript describes an efficient, reliable pipeline for spike sorting on dense multi-electrode arrays (MEAs), where neural signals appear across many electrodes and spike sorting currently represents a major computational bottleneck. We present several new techniques that make dense MEA spike sorting more robust and scalable. Our pipeline is based on an efficient multi-stage \"triage-then-cluster-then-pursuit\" approach that initially extracts only clean, high-quality waveforms from the electrophysiological time series by temporarily skipping noisy or \"collided\" events (representing two neurons firing synchronously). This is accomplished by developing a neural network detection method followed by efficient outlier triaging. The clean waveforms are then used to infer the set of neural spike waveform templates through nonparametric Bayesian clustering. Our clustering approach adapts a \"coreset\" approach for data reduction and uses efficient inference methods in a Dirichlet process mixture model framework to dramatically improve the scalability and reliability of the entire pipeline. The \"triaged\" waveforms are then finally recovered with matching-pursuit deconvolution techniques. The proposed methods improve on the state-of-the-art in terms of accuracy and stability on both real and biophysically-realistic simulated MEA data. Furthermore, the proposed pipeline is efficient, learning templates and clustering much faster than real-time for a [~=] 500-electrode dataset, using primarily a single CPU core.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (19 Jun 2017) · CC BY 4.0

A dense multi-electrode array spike sorter that triages collisions, clusters clean waveforms with a coreset Dirichlet process mixture, then deconvolves, giving more stable clusters and faster template learning than KiloSort.

The paper names spike sorting, including manual cleanup, as the main computational bottleneck for dense multi-electrode arrays (MEAs), which already reach the order of 10^4 electrodes with 10^6 targeted. Prior pipelines either pass collisions and false positives from threshold detection into clustering or, like Kilosort, run matching pursuit directly on the time series; YASS instead sets collided and noisy events aside, clusters only clean waveforms, and recovers the rest by deconvolution.

Key results

  1. SuggestedA convolutional detection network trained on curated or simulated spikes rejects collisions and noise events much better than hand-designed detectors. It also aligns spikes more precisely, so clusters in PCA space look more Gaussian.

    On simulated 30-electrode low signal-to-noise ratio (SNR) ViSAPy data, false positives and collisions fell by 90% at a matched true-positive rate; Section 3.2 reports an order-of-magnitude reduction in false positives. ROC curves against SpikeDetekt are shown for clusters with SNR 1.4 and 2.3. k-NN triage then removed a further 70% while discarding a small share of real detections. · 2.1 Overview; 3.2 Synthetic Datasets; C.5 Empirical Performance and Scalability; Figure legends

  2. SuggestedOn biophysically simulated recordings, YASS outperformed KiloSort, Spyking Circus and MountainSort in the number of ground-truth units recovered above accuracy thresholds, especially at low SNR. It also produced fewer low-stability clusters.

    Used 30-channel ViSAPy datasets in high-collision and low-SNR settings, with 16 ground-truth units. Inferred clusters were matched to the truth with the Hungarian algorithm. Stability came from Poisson-perturbation resampling with rate parameter 0.25. Differences are shown only as curves (Figure 2); the text gives no summary statistics. · 3 Performance Comparison; 3.2 Synthetic Datasets; Figure legends

  3. SuggestedOn real primate retina data, YASS beat KiloSort, Spyking Circus and MountainSort against a curated gold standard, most clearly in cluster stability. KiloSort ranked second overall, and YASS also returned fewer total clusters.

    Used four sets of 49 contiguous electrodes from a 30-min, 512-channel ex vivo recording. The gold standard was hand-validated on the full array: 355 reference neurons of known types plus 31 consistent with unreported types. Results are plotted (Figure 3, Figure S1) without summary numbers. No method received manual post-processing. · 3.1 Competing Algorithms; 3.3 Real Datasets; B.1 Results on Three Additional Real…; H Constructing a proxy to ground truth for a…

  4. SuggestedOn the 512-channel, 30-min retina recording, YASS learned templates and clusters several-fold faster than Kilosort while running mostly on a single CPU core.

    KiloSort in full GPU mode took about 30 minutes. Spyking Circus took over a day on a 6-core machine. MountainSort was not timed; linear extrapolation gives approximately 10 hours. Only YASS detection used a GPU (Nvidia Titan X), and templates were estimated from one-fifth of the data. Per-step YASS times are in Table 1. · 3 Performance Comparison; 3.3 Real Datasets; Footnotes; Tables

  5. SuggestedTemplate estimation tolerates subsampling: templates from a random fraction of the recording, summarised by a coreset that enforces coverage, reach near-full accuracy. Coreset and clustering costs grow roughly linearly with recording length.

    On two 49-channel, 30-min retina recordings, 5-10 minutes of data matched full-recording performance, and accuracy loss became insignificant above 20% of the length. In a four-cluster Gaussian toy example, the authors' coreset covered clusters that an existing coreset method missed. The authors also state that 100k points reduce to a ≃10k-point coreset with trivial accuracy loss. · 2.1 Overview; B.2 Accuracy with Respect to Data Length; E Additional details on the Coreset Construction; Figure legends

Methods, data and reproducibility

A convolutional network trained on prior curated sorts detects spikes. Collisions and nearest-neighbour outliers are triaged. Clean waveforms are reduced to a recursive k-means++ coreset and clustered per primary channel with a split-merge variational Dirichlet process Gaussian mixture model (DP-GMM). Merged templates drive Kilosort's matching-pursuit deconvolution, which recovers the triaged spikes.

  • ViSAPy biophysics-based simulated extracellular recordings, 30 channels, at several noise levels and collision rates (high-collision and low-SNR settings reported) · n = multiple datasets; 16 ground-truth units
  • Ex vivo extracellular recording of peripheral primate retina on a high-density 512-channel array under spatiotemporal white-noise stimulation, with a hand-curated gold-standard sort · primate (species not stated) · n = 30 min; four 49-electrode subsets; 355 (+31) reference neurons
  • Two 49-channel, 30-min retina recordings used to test template estimation from random data subsets · primate (species not stated) · n = 2 recordings
  • Simulated Gaussian mixture data used to illustrate coreset coverage · n = 4 clusters
Methods & tools
YASS, ViSAPy, KiloSort, Spyking Circus, MountainSort, SpikeDetekt, K-means++, Hungarian algorithm, Tarjan's algorithm, Adam optimizer, Ng-Jordan-Weiss spectral clustering, Butterworth bandpass filter
Reagents & environment
Competing sorters are identified only by download date: KiloSort 30 Jan 2017, Spyking Circus 26 May 2017, MountainSort 7 Jun 2017. No YASS version or dependency list is given.
Assay conditions & compute
CPU timings used one core of a 6-core i7 workstation (32GB RAM); GPU timings used an Nvidia Titan X in that workstation. Detection took 6 s for a 5-min, 49-channel recording; network training took 2 min for the 512-channel data.
Reproducibility gaps
The text gives no code repository, software version or data accession. It does not say whether the primate retina recording or the gold-standard sort is available.
Limitations (6)
  • The detection network depends on curated training data from earlier sorts on the same recording setup, which the authors note may not be practical everywhere. · stated by the authors
  • Previous sorts used as training labels may contain missed or poorly aligned spikes; the authors propose noise-augmented synthetic spikes from templates to compensate. · stated by the authors
  • The real-data gold standard is a partial proxy: it is not claimed to capture every spike and probably lacks small or collided spikes. · stated by the authors
  • Waveform drift was not handled: nonstationarity was reported as minimal in the retina recording (data not shown), and template drift and refractory violations are left to future work. · stated by the authors
  • Accuracy was assessed on one primate retina recording (49-electrode subsets) and 30-channel simulations, with results given only as plots. JRClust was not compared, and several supporting claims cite results that are not shown. · noted in this summary
  • MountainSort's runtime is extrapolated, not measured. YASS's speed partly comes from estimating templates on one-fifth of the recording, and the text does not say whether competitors ran on comparable subsets. · noted in this summary

Read this if you sort spikes from dense multi-electrode arrays and want a modular pipeline that handles collisions separately from clustering, or a coreset-based Dirichlet process mixture for scalable clustering.

Citations
58

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISLee '20Hilgen '17Buccino '25Rossant '15Jun '17Yger '16
89% semantic matchNeuroscience
JinHyung Lee, Catalin Mitelut, Hooshmand Shokri+14 more·bioRxiv·Mar 18, 2020· 63 citations