README.md

metadata

license: cc-by-nc-4.0
pretty_name: AxoBench v2
task_categories:
  - time-series-forecasting
  - tabular-regression
tags:
  - neuroscience
  - computational-neuroscience
  - neuron-simulation
  - single-cell
  - benchmark
  - dendrites
  - interventions

AxoBench v2

AxoBench is a benchmark for data-driven surrogates of layer-5 pyramidal neurons: given six seconds of synaptic input events, predict the somatic voltage and spikes of a detailed NEURON model of the cell. Version 2 covers 15 cells, adds paired interventions at three doses each, and stores the somatic voltage at 4 kHz in addition to the 1 ms targets.

Variants

Each variant is a self-contained dataset root in its own top-level folder.

Folder	Status	Traces per training cell	Size
minimal/	released (tag minimal-v2.0)	2,000 train, 250 val, 250 test	14.6 GB
full/	released (tag full-v2.0)	8,000 train, 250 val, 250 test	35.6 GB

full/ contains every minimal/ shard unchanged, plus 6,000 more training traces per training cell. Its val/, test/ and interventions/ are the same files as in minimal/, so scores on the two variants are directly comparable. The extra shards were generated with the same simulator, gains and input statistics, and take shard indices from 20,000 up, so no index or seed of minimal/ is reused.

Download with the axobench package (axobench download-data --variant full, or axobench.hub.download_dataset), or fetch single folders with huggingface_hub.snapshot_download(..., allow_patterns=["full/val/*"]).

Cells

• 14 Allen Cell Types all-active models of mouse layer-5 cells with intact apical dendrites (Gouwens et al. 2018), each with its fitted biophysics and a per-cell excitatory gain calibrated to a common firing rate.
• The NeuronIO layer-5b cell (Hay et al. 2011, neuronio_hay_cell1) as the anchor to earlier work.

Three Allen cells (354190013, 480353286, 488683425) are held out: they appear only in test (zero-shot track). morphologies.json lists every cell with its source and held-out flag; morphologies/ holds the Allen SWC files.

Layout of minimal/ and full/

Folder	Content
train/<cell>/	minimal/: 6,000 shards, 24,000 traces (12 training cells x 2,000); full/: 24,000 shards, 96,000 traces (12 x 8,000)
val/<cell>/	756 shards, 3,000 traces (12 x 250)
test/<cell>/	945 shards, 3,750 traces (15 x 250)
interventions/<condition>_<dose>/	15 folders, 600 pairs each (40 per cell), 9,000 pairs in total
morphologies/, morphologies.json	cell index and reconstructions
dataset_package_manifest.json	counts per split, cell and condition; generator settings and gains
checksums.sha256	SHA-256 of every file in the variant

Interventions: event dropout, excitatory dropout and inhibitory dropout at 10, 35 and 70%; site silencing at 10, 25 and 50%; temporal jitter of 1, 2 and 8 ms.

Shard schema

Each .npz shard has a JSON sidecar with the generator settings. Each split and intervention folder has a manifest.json that lists its shards (paths relative to the folder; split shards sit in one subfolder per cell) and sample ids.

• inputs: (n, 6000, 1278) int8, input events per 1 ms bin on 639 excitatory and 639 inhibitory synapse sites.
• targets: (n, 6000, 2) float32 at 1 ms: spike indicator (0/1) and the NeuronIO-style somatic target, (min(V, -55 mV) + 67.7 mV) / 10.
• soma_mv: (n, 6000) raw somatic voltage (mV) at 1 ms.
• soma_mv_hr: (n, 24000) raw somatic voltage (mV) at 4 kHz.
• Intervention shards hold the same arrays with baseline_ and intervention_ prefixes, plus changed_event_mask.

Withheld data

A confirmation split (private-test, 125 traces per cell) is not published. Predictions on it are scored on request, rate-limited.

License and attribution

The data are released under CC BY-NC 4.0. The Allen cells are simulated on reconstructions and models from the Allen Cell Types Database, whose Terms of Use allow derived works for research and other noncommercial purposes only. Commercial use needs written permission from the Allen Institute.

If you use these data, please cite:

• Allen Institute for Brain Science (2015). Allen Cell Types Database [dataset]. Available from celltypes.brain-map.org.
• Gouwens, N. W., et al. (2018). Systematic generation of biophysically detailed models for diverse cortical neuron types. Nature Communications, 9, 710. https://doi.org/10.1038/s41467-017-02718-3
• Hay, E., Hill, S., Schürmann, F., Markram, H., & Segev, I. (2011). Models of neocortical layer 5b pyramidal cells capturing a wide range of dendritic and perisomatic active properties. PLoS Computational Biology, 7(7), e1002107. https://doi.org/10.1371/journal.pcbi.1002107
• Beniaguev, D., Segev, I., & London, M. (2021). Single cortical neurons as deep artificial neural networks. Neuron, 109(17), 2727–2739.e3. https://doi.org/10.1016/j.neuron.2021.07.002 (input statistics and generator design)