spikebias

Measure the biases of spike sorting on dense extracellular recordings, such as Neuropixels, against the ground truth of a large-scale, detailed biophysical model of the rat cortex.

spikebias works on SpikeInterface extractors, so it applies to any sorter and any recording with a ground truth. It also streams the simulated recordings, ground truths and sortings of the paper from the DANDI Archive (dandiset 001250).

What it measures

Install

pip install spikebias

Python 3.10 or later and SpikeInterface 0.101 or later.

Quick start

import spikebias as sb
from spikebias import datasets

# stream the ground truth and the Kilosort 4 sorting of the spontaneous Neuropixels simulation
ground_truth = sb.first_minutes(datasets.load_ground_truth("npx_spont"), minutes=10)
sorting = datasets.load_sorting("npx_spont", sorter="ks4")      # first 10 minutes

# sorting accuracy of each ground truth neuron
accuracy = sb.sorting_accuracy(ground_truth, sorting)
print(sb.count_well_sorted(accuracy), "of", len(accuracy), "neurons sorted with an accuracy >= 0.8")

# spike assignment bias of each sorted single-unit
biases = sb.classify_sorted_units(ground_truth, sorting, duration=600)
print(sb.bias_proportions(biases))

# yield per recording site, and selection biases by synapse class, layer and electrical type
recording = datasets.load_recording("npx_spont")                # streamed: only what is read is downloaded
print(sb.yield_per_site(sorting, recording))
print(sb.selection_biases(ground_truth, sorting))

Your own data work the same way: pass a SpikeInterface ground truth sorting and a sorting of the same period. The sampling and accuracy-model functions use the neurons' properties (layer, synapse_class, etype, x, y, z), which the biophysical datasets provide.

Model of the sorting accuracy

recording = sb.first_minutes(datasets.load_recording("npx_spont", local_folder="data"), minutes=10)
features = sb.neuron_features(ground_truth, recording, duration=600, n_jobs=8)
data = features.join(sb.sorting_accuracy(ground_truth, sorting)).dropna()   # drop neurons without spikes

model = sb.AccuracyModel(transform="rank")
r2 = sb.cross_validate(model, data.drop(columns="accuracy"), data["accuracy"], n_splits=100)
print(sb.summarize(r2))                       # McFadden pseudo R-squared on held-out neurons
print(model.fit(data.drop(columns="accuracy"), data["accuracy"]).weights)

neuron_features reads the spikes of every neuron from the recording, so download it first for large datasets (local_folder); the Neuropixels recordings are about 100 GB.

Definitions

• Hit: a sorted spike within delta_time = 1.3 ms of a ground truth spike.
• Agreement score of a ground truth neuron and a sorted unit: hits / (hits + misses + false positives).
• Sorting accuracy of a ground truth neuron: its best agreement score with a sorted (single-)unit; it is well sorted at 0.8 or more.
• Chance score of a pair: the agreement score expected from two independent Poisson spike trains with the units' firing rates.
• Bias classes of a sorted unit: false positive if no agreement score is above chance; otherwise good (a score of at least 0.8), overmerger (more than one neuron matched above chance), oversplitter (a matched neuron is also matched above chance by another sorted unit), or a mixed combination.

Datasets

datasets.DATASETS lists the datasets of dandiset 001250, and datasets.list_assets(name) their files.

Sorters: ks4, ks3, ks2.5, ks2, ks (Kilosort 4 to 1) and hs (Herdingspikes), on the first 10 minutes (minutes=10) or, for some datasets, the full recording (minutes=None).

Citation

If you use spikebias or the dataset, please cite:

• Laquitaine S, Imbeni M, Tharayil J, Isbister JB, Reimann MW. Spike sorting biases and information loss in a detailed cortical model. bioRxiv (2024). https://doi.org/10.1101/2024.12.04.626805
• The dataset: DANDI Archive, dandiset 001250. https://dandiarchive.org/dandiset/001250

The code that reproduces the paper's figures is at https://github.com/spikebias/spikebias-paper. For the curation of sorted single-units, see spikecurate.

License

GPL-3.0. Author: Steeve Laquitaine (laquitainesteeve@gmail.com).