All projects
NeuroDataReHack 2026
Benchmarking the ephys-atlas brain-region classifier on Allen (AIND) Neuropixels data
Project Description
- The goal of the project was to develop a FeatureCalculator module for NWB datasets, and use that to test how a region classification model built on IBL data transfers to the AIND Neuropixels 1 dataset.
Dandiset(s) Used
- DANDI:001637 (AIND ecephys): contains CCF acronyms and coordinates for per-channel brain-region labeling. https://dandiarchive.org/dandiset/001637/draft
- Used WavPack-compressed Zarr files to stream raw AP and LF signals from the aind-open-data S3 bucket, cross-referenced with allen_info.tsv.
- **DANDI 001637** (AIND `ecephys`), per-channel Allen brain-region labels (CCF acronyms +
coordinates): https://dandiarchive.org/dandiset/001637/draft
- Raw AP/LF signal streamed from the **AIND public S3 bucket** `aind-open-data`
(WavPack-compressed Zarr), matched to the DANDI files listed in `allen_info.tsv`.
Objectives and Approach
- Stream raw AP + LF per probe directly from AIND S3 (no downloads) and compute the full ephys-atlas feature set (LF, CSD, AP, spike-waveforms) on 5 × 5 s snippets, aggregated and denoised with the library’s own pipeline, the same pipeline that built the training vintage.
- Attach per-channel Cosmos region labels from the matching DANDI 001637 NWB electrodes table (joined on the channel index within each probe).
- Retrain an XGBoost Cosmos classifier on the latest ea_active vintage (`2026_W26`), predict the Allen channels, and report accuracy grouped by probe id.
Progress
- Implemented `AllenAindFeatureCalculator`, a small plug-in to the modular feature-calculator class hierarchy that streams AIND WavPack-Zarr from S3.
- Processed 129 probes (≈49k channels); Scraped the information about the CCFv3 brain region from DANDI dataset.
- Applied the pretrained model on Allen dataset - Allen transfer = 0.58 ± 0.13 all-channel, matching the model’s own IBL grouped-by-PID cross-validation (0.58)
- Per-region: hippocampus 0.91 and cortex 0.74 transfer strongly; thalamus 0.18 and olfactory 0.00 are the weak spots (mostly collapsed into cortex/HPF)
Next Steps
Test with recording of different probes.
Improve out of brain detection, and detection of some regions like Olfactory areas.
Slides -