2024 Theses Doctoral

Localizing spike sources for improved registration, spike-sorting, and decoding in large-scale Neuropixels recordings

Boussard, Julien

Neuroscience, the study of the brain and nervous system in humans and animals, relies more and more on data acquisition and analysis. Since the first recording of a neuron by Hubel and Wiesel in 1957, using a tungsten electrode probe, neuroscientists have developed a multitude of devices to record neuronal signals at various spans and spatiotemporal resolutions, leading to a rapid, exponential increase in the number of recordable neurons. Methods such as calcium-imaging, multi-electrode arrays (MEAs), or optogenetics are now at the core of research aiming at understanding the processes underlying cognition and the encoding of information by neural populations, improving brain-machine interfaces, and potentially discovering new cures for neurological diseases.

Combining multiple advances in probe design, Neuropixels probes, introduced in 2017, have 384 recording channels arranged in a narrow and elongated shape that allows in vivo recordings of a large number of neurons from multiple brain regions, in arbitrary brain locations, with high spatiotemporal resolution, and in many different species of unrestrained animals. Thanks to these advantages as well as their low cost, this technology has been widely adopted by numerous labs that design experiments to record data from various brain regions and species. Multiple laboratories, and in particular the International Brain Laboratory, a group of 22 neuroscience labs, or the Allen Institute, have now released thousands of Neuropixels recordings from multiple brain regions and species.

These large-scale experiments, which allow neuroscientists to examine the coordinated action of large neuronal populations in superficial and deep structures of the brain, present a fantastic opportunity for studying global brain dynamics. However, Neuropixels probes produce large volumes of high-dimensional data, and extracting information from these recordings is challenging. The main challenge is spike-sorting, i.e. detecting and assigning spikes to individual neurons. This step is critical to many downstream tasks, such as cell type classification or decoding. Unfortunately, spike sorting algorithms are inacurate and do not generalize well to different brain regions or animals, often requiring manual supervision which makes this process expensive and inefficient. The analysis of large-scale Neuropixels recordings thus requires accurate, robust, modular, and scalable spike sorting algorithms that generalize well across multiple species and brain regions, different existing probes, and even new probe designs.

During my thesis, we developed methods for improving spike-sorting in Neuropixels recordings. We specifically tackled the problem of probe motion. Due to their elongated shapes, Neuropixels probes move relative to the brain. The neurons’ spike shapes thus change over time as the probe drifts, making it hard to cluster them properly. Inspired by image registration, we developed a decentralized registration method for Neuropixels recordings to estimate the movement of the probe relative to the brain, by treating the distribution of spike amplitudes as an image. We then developed a localization method to infer the three-dimensional position of the detected spikes relative to the probe using a simple model for the propagation of the electrical field generated by a neuron in the brain. We showed how these locations can be leveraged to improve registration. I then contributed to extending the above methods to developing Dredge, a registration method that shows good performance across a variety of data modalities i.e. different Neuropixels probes, different species, and different frequency bands.

We then built a spike sorter, DARTSort, that improved upon existing spike sorters by explicitly modeling the unit’s spike shape variability as a function of probe motion, rather than interpolating the data to correct for drift. Moreover, we aimed at building a modular and interpretable spike sorter, allowing each of its components to be easily isolated, that generalizes well to a variety of probe designs. DARTSort was used to sort Ultra-High Density Neuropixels recordings. These very dense, newly designed probes allow for improved spike detection, yield, and cell-type classification, at the expanse of a shorter recording span. Finally, the localization and registration methods were utilized for building a state-of-the-art decoding model, taking as input the spike density along the probe, acting as an efficient and uncertainty-aware proxy for spike sorting.