Opening the black box toward a modular approach to spike sorting

Spike sorting is an algorithmic process that extracts the activity of individual neurons from extracellular electrophysiology recordings. With the ballooning use of high-density probes, such as Neuropixels, this essential processing step is increasingly becoming time-consuming and computationally expensive. Although many software tools have been proposed to address spike sorting, they are usually constructed and benchmarked as monolithic 'black boxes', making it difficult to factor out the effects of individual algorithmic steps on the final outcome, especially when varying datasets and parameters. To address this issue, we developed a modular and common framework to develop, benchmark, and assemble the key computational steps that are used in state-of-the-art spike sorting algorithms. Relying on fast and efficient ground truth generation of biophysically plausible recordings, we show that we are able to individually benchmark and precisely quantify the performance of different steps in a spike sorting pipeline (i.e. peak detection, feature extraction, clustering, and template matching). We then leverage these results to create a modular, component-based spike sorter that can outperform Kilosort4 on dense and large simulated recordings, and produce similar quantitative results on real data. In addition, we find that the major bottleneck of all modern spike sorting pipelines is in the physical motion of probes, regardless of the drift-correction strategy. The component-based spike sorting framework presented here has the potential to foster community engagement in the field by lowering the barrier to contributions and providing a flexible yet powerful framework to construct end-to-end spike sorting solutions.