Thanks for testing this out @fleabrained.bsky.social!

Please feel free to open issues on the GitHub repo with suggestions on how to improve any aspect of the pipeline :)

6/8 Benchmarking confirms Kilosort4 > Kilosort2.5 in accuracy, precision, and recall—especially for low-SNR units—while recovering more complete spike trains.

5/8 We also introduce a benchmarking pipeline that generates hybrid ground-truth data with drift-aware template injection, enabling rigorous, scalable comparison of sorting and preprocessing algorithms.

4/8 The pipeline parallelizes across probes/shanks/recordings, cutting processing time dramatically. A 6-probe Neuropixels 2.0 Quad Base experiment drops from >7 days to ~10-15 hours — 15-20x faster, at lower cost.

3/8 The workflow includes: preprocessing (phase shift, filtering, denoising, drift estimation), spike sorting (Kilosort4/2.5 or SpyKING CIRCUS2 - and more options to come!), postprocessing, automated curation (quality metrics + UnitRefine), QC visualizations, and NWB export.

2/8 We introduce an end-to-end, modular, scalable spike sorting pipeline built with Nextflow, SpikeInterface, and containerized environments. It runs reproducibly on workstations, HPC clusters, or the cloud.