frthjf/spe1-zhao2026-benchmark

SPE-1 — Paired patch-clamp + Neuropixels recordings (Arrow conversion)

This dataset is an Arrow / 🤗 datasets re-packaging of the SPE-1 ground-truth electrophysiology dataset by Marques-Smith et al. (2018), preprocessed and split into per-cell configurations for use with the spike-localization benchmarks by Zhao et al. (2026).

The original raw recordings (~270 GB of int16 binaries) are not included here, but can be downloaded using the included scripts if needed.

License

The original SPE-1 release is distributed under CC-BY-SA 4.0 (see Licence CC BYSA 4.0.pdf). This Arrow conversion inherits the same license.

Original sources:

• CRCNS: https://crcns.org/data-sets/methods/spe-1
• Google Drive mirror: https://drive.google.com/drive/folders/13GCOuWN4QMW6vQmlNIolUrxPy-4Wv1BC

Citation

If you use this dataset, please cite the original recordings and the benchmark methodology:

• The original publication describing the SPE-1 dataset: Marques-Smith, A., Neto, J.P., Lopes, G., Nogueira, J., Calcaterra, L., Frazão, J., Kim, D., Phillips, M., Dimitriadis, G., Kampff, A.R. (2018). Recording from the same neuron with high-density CMOS probes and patch-clamp: a ground-truth dataset and an experiment in collaboration. bioRxiv 370080; doi: https://doi.org/10.1101/370080
• The SPE-1 dataset itself: André Marques-Smith, Joana P. Neto, Gonçalo Lopes, Joana Nogueira, Lorenza Calcaterra, João Frazão, Danbee Kim, Matthew G. Phillips, George Dimitriadis and Adam R. Kampff (2018); Simultaneous patch-clamp and dense CMOS probe extracellular recordings from the same cortical neuron in anaesthetized rats. CRCNS.org http://dx.doi.org/10.6080/K0J67F4T
• The spike-localization benchmark: Zhao, H., Zhang, X., et al. (2026). Benchmarking spike source localization algorithms in high density probes. https://doi.org/10.1371/journal.pcbi.1014059

Cells included

12 cells from the SPE-1 release are converted:

Cell	Notes
c5	Longest WC-IC recording — used as the whole-cell V(t) reference
c14, c15, c16, c19, c24, c26, c28, c29, c37, c45, c46	The 11 juxtacellular cells selected by Zhao et al. 2026

For each cell, the first 10 s (settling period) are skipped and 300 s of recording are kept.

Repository layout

frthjf/spe1-zhao2026-benchmark
├── README.md                       (this file)
├── chanMap.mat                     2-D probe geometry (384 × 2 µm, x and y)
├── Data Summary.xlsx               per-cell ground-truth electrode (chan_predicted)
├── Data Summary.csv                  (same converted to plain text csv)
├── Recording Catalogue.pdf         original SPE-1 catalogue
├── Licence CC BYSA 4.0.pdf         original SPE-1 license
├── localization_benchmark.json     COM / MT / GC template-level numbers
│                                   produced by scripts/reproduce_zhao2026.py
├── scripts/                        self-contained reproduction pipeline
│   ├── prepare.sh                    download (CRCNS) + run convert_to_arrow
│   ├── convert_to_arrow.py           SPE-1 raw → Arrow conversion
│   └── reproduce_zhao2026.py         COM / MT / GC baselines (Zhao et al. 2026)
└── c{ID}/                          one HF DatasetDict per cell
    ├── dataset_dict.json
    ├── conversion_metadata.json    (see schema below)
    ├── dataset_stats.json          per-channel median / IQR / abs-IQR
    ├── templates.npz               peak-aligned mean template + probe geometry
    ├── patch.npz                   ground-truth patch trace + spike times
    ├── patch_quicklook.png         sanity-check plot of patch.npz (where shipped)
    ├── train/
    │   ├── dataset_info.json
    │   ├── state.json
    │   └── data-*.arrow            ~ 138 MB per 3-s segment
    └── test/
        └── data-*.arrow

Per-cell file schemas

Arrow segments (train/, test/)

Each row = one ~3-second segment of preprocessed Neuropixels recording (bandpass 300-3000 Hz + global common-median reference, gain -> µV applied). 80 % of segments -> train, 20 % -> test (all segments are in order, so concatenating train+test yields original data).

Column	Type	Shape / units	Description
sample_id	string	scalar	Zero-padded NPX sample index of the segment start
cp	list<list<float32>>	[384][T] (T = ~90 000 samples ≈ 3 s)	Per-channel preprocessed voltage traces (µV)
cit	list<list<int32>>	[384][n_spikes_on_channel]	Per-channel unit indices for each ground-truth spike (always 0 since single patched neuron)
ctt	list<list<float64>>	[384][n_spikes_on_channel] (ms)	Spike times within the segment, in milliseconds

Spike events are only assigned to the unit's peak channel (unit_peak_channel); the other 383 channels carry empty cit/ctt lists for that segment. Sampling frequency is 30 kHz.

templates.npz

Mean spike template extracted with the Zhao et al. 2026 method (1 ms before + 2 ms after the patch-clamp peak; raw mean over snippets).

import numpy as np
tpl = np.load("c14/templates.npz")
tpl["templates"]            # float32 [1, 90, 384]    (units x time x channels)
tpl["sampling_frequency"]   # 30000.0
tpl["probe_positions"]      # float32 [384, 2]        (x, y in um)
tpl["nbefore"], tpl["nafter"]  # 30, 60

patch.npz

Authoritative patch-clamp recording for the same window used by the Arrow splits.

patch = np.load("c5/patch.npz", allow_pickle=False)
patch["patch_v"]        # float32 [N]    patch-clamp voltage (mV)
patch["t_patch_s"]      # float64 [N]    timestamps (s, NPX-aligned)
patch["spike_times_s"]  # float64 [K]    GT spike times (s, NPX-aligned)
patch["patch_type"]     # 0-d str        "wc-ic" / "juxta" / etc.
patch["cell_id"]        # 0-d str        e.g. "c14"

conversion_metadata.json

{
  "source_metadata": {
    "mode": "spe1_paired_recording",
    "dataset": "Marques-Smith et al. (2018) \u2014 CRCNS spe-1",
    "cell_id": "c14",
    "skip_seconds": 10.0,
    "duration_s": 300.0,
    "patch_sample_rate_hz": 50023.0,
    "preprocessing": "bandpass_300_3000Hz + global_CMR",
    "spike_source": "wc_spike_samples.npy (authoritative GT)"
  },
  "template_metadata": {
    "extraction_method": "hao2026_raw_mean_template",
    "ms_before": 1.0,
    "ms_after": 2.0,
    "nbefore": 30,
    "nafter": 60,
    "n_spikes_total": 814,
    "n_spikes_used": 814,
    "n_out_of_bounds_discarded": 0
  },
  "sampling_frequency": 30000.0,
  "num_units": 1,
  "num_channels": 384,
  "unit_peak_channel": {
    "0": 161
  },
  "unit_locations_um": {
    "0": [
      27.0,
      1600.0,
      0.0
    ]
  },
  "gt_electrode_chan_idx": 159,
  "bad_channel_ids": [
    36,
    75,
    112,
    151,
    188,
    303
  ],
  "n_spikes_total": 814,
  "n_spikes_used_for_templates": 814,
  "segment_duration_s": 3.0,
  "train_fraction": 0.8
}

unit_locations_um[0] is the (x, y, z=0) µm position of the ground-truth electrode (chan_predicted from Data Summary.xlsx), and serves as the regression target for all localization metrics. Channel indices listed in bad_channel_ids were flagged by SpikeInterface's detect_bad_channels on the bandpass-filtered (pre-CMR) recording and should be masked before running amplitude-weighted localizers (COM / MT / GC).

Quick start

Load one cell

from datasets import load_dataset

ds = load_dataset("path/to/spe1", name="c14")   # -> {'train', 'test'}
row = ds["train"][0]
print(row["sample_id"], len(row["cp"]), len(row["cp"][0]))  # 384 channels x ~90k samples

Or directly from disk after a clone / huggingface-cli download:

from datasets import load_from_disk
dsd = load_from_disk("spe1/c14")

Reproduce Hao et al. 2026 baselines (template-level localization)

The self-contained systems/datasets/spe1/scripts/reproduce_zhao2026.py runs Center-of-Mass, Monopolar-Triangulation and Grid-Convolution directly on the shipped templates.npz files:

python scripts/reproduce_zhao2026.py                    # all 11 Hao cells
python scripts/reproduce_zhao2026.py --cells c14 c45    # subset
python scripts/reproduce_zhao2026.py --with-spikes      # also spike-level Acc/MAE (slower)

Results are printed and written to localization_benchmark.json. Bad channels (bad_channel_ids in conversion_metadata.json) are masked automatically before localization, exactly as in the published baseline.

Reproducing the Arrow conversion from scratch

Only needed if you want to re-derive the Arrow files from the ~270 GB raw binaries (e.g. to change duration_s, segment_duration_s, or the preprocessing pipeline). The dataset ships a self-contained pipeline in scripts/ that depends only on standard PyPI packages (requests, scipy, pandas, openpyxl, spikeinterface, datasets, numpy):

# 1. Register a free CRCNS account at https://crcns.org/register
# 2. From the dataset root:
CRCNS_USERNAME=<user> CRCNS_PASSWORD=<pass> bash scripts/prepare.sh

prepare.sh will:

1. Download chanMap.mat and Data Summary.xlsx from CRCNS into .raw/ (and mirror them at the dataset root so consumers don't need .raw/).
2. Download and extract c{ID}.tar.gz for each requested cell into .raw/Recordings/c{ID}/.
3. Run scripts/convert_to_arrow.py once per cell to produce the per-cell Arrow train//test/ splits + templates.npz + conversion_metadata.json.

Environment overrides (see prepare.sh header for full list):

Variable	Default	Description
SPE1_RAW_DIR	./.raw	Where raw downloads land
SPE1_OUT_DIR	. (dataset root)	Where Arrow datasets are written
SPE1_CELLS	All 12 cells	Space-separated cell IDs
SPE1_DURATION	300	Seconds of recording per cell
PYTHON	python3	Python interpreter

Provenance & preprocessing summary

• Probe: Neuropixels 1.0, 384 channels, 30 kHz, int16 raw → µV (gain 2.34375 µV/bit), 2-D probe geometry from chanMap.mat.
• Filtering: causal bandpass 300–3000 Hz (Butterworth) followed by global common-median reference, applied via SpikeInterface (spre.bandpass_filter + spre.common_reference).
• Bad-channel detection: spre.detect_bad_channels on the bandpass output (before CMR) — channels listed in bad_channel_ids should be masked before running amplitude-weighted localizers.
• Ground-truth spikes: from wc_spike_samples.npy (authoritative GT shipped by SPE-1) when available, otherwise threshold detection on the patch trace (12 x MAD below median, 1 ms refractory).
• Templates: peak-aligned mean over all GT snippets, 1 ms before + 2 ms after peak, no per-spike realignment.
• Splits: contiguous 3-s segments, first 80 % → train, last 20 % → test.

Total size

~145 GB across 12 cells (≈ 13 GB / cell, c37 is ~5 GB and c29 ~12 GB because of shorter usable recording windows).