2 of 11

Objective & approach

Goal

• Such a tool can be used in real time localization of the probes in the brain.

• Measure how an IBL-trained brain-region decoder generalizes to a different lab's rig (Allen / AIND).

Three data sources tied together

• No datasets exist on DANDI which have raw AP and LF data for mice.

• Raw AP+LF — streamed from AIND public S3 (WavPack-Zarr), not downloaded

• Region labels — DANDI 001637 electrodes table (Allen acronym per channel)

• Training — ea_active vintage 2026_W26 (383k channels / 998 IBL probes)

Thanks to Carter for helping with access of the Data

Method

• Full feature set: 49 features (LF, CSD, AP, spike-detection)

• 5 × 5 s snippets per probe, aggregated + denoised

• XGBoost model trained on IBL data

Two metrics reported:

• all-channel (headline, incl. out-of-brain) vs in-brain (real regions only)

ephys-atlas 2026_W26 Cosmos classifier · Allen (AIND) NP1 · 129 probes / 23 sessions

3 of 11

Predictive Procesing Allen Open Scope

Dataset: OpenScope Community Predictive Processing (DANDI 001637), Neuropixels, mouse sensory mismatch

4 of 11

Repository architecture — modular feature calculators

One compute workflow; each data source is a small plug-in — stream or read locally, same pipeline

Class hierarchy (src/ephysatlas/feature_calculators/)

BaseFeatureCalculator

abstract source API + shared compute_snippet()

│

├─ SpikeGlxLikeFeatureCalculator (spikeglx.Reader)

│ ├─ SpikeGLXFileFeatureCalculator local SpikeGLX

│ └─ IBLPIDFeatureCalculator IBL via ONE

│

└─ SpikeInterfaceFeatureCalculator (SI BaseRecording)

├─ NwbFeatureCalculator NWB: local·URL·DANDI

└─ AllenAindFeatureCalculator AIND zarr on S3 ★

Point it at your data — same call, any source

local file NwbFeatureCalculator.from_local(ap_path, lf_path)

stream URL NwbFeatureCalculator.from_url(ap_url, lf_url)

DANDI asset NwbFeatureCalculator.from_dandi("000409",

ap_filepath=…, lf_filepath=…)

IBL probe IBLPIDFeatureCalculator(pid, one=one)

AIND S3 ★ AllenAindFeatureCalculator(ap_url, lf_url)

then, identical for every source:

calc.compute_snippet(

SnippetWindow(t_start, duration_ap, duration_lf),

FeatureComputationOptions(

features_to_compute=["lf","csd","ap","waveforms"]))

Add a source = implement one open-recording hook; geometry, snippet slicing, channel metadata & feature computation are inherited. Acquisition (local / URL / DANDI, HDF5 / Zarr) is composed via NwbSource, not subclassed.

ephys-atlas 2026_W26 Cosmos classifier · Allen (AIND) NP1 · 129 probes / 23 sessions

5 of 11

Result: transfer ≈ IBL generalization

Crossing the IBL→Allen lab gap costs almost nothing beyond normal probe-to-probe variability

Takeaways

• Accuracy grouped by probe (mean ± sd over 129 probes)

• Allen all-channel 0.58 ± 0.13 = IBL across-probe (grouped-by-PID) CV 0.58

• In-brain 0.67 ± 0.15

• Chance = 0.08 (1 of 12 regions)

ephys-atlas 2026_W26 Cosmos classifier · Allen (AIND) NP1 · 129 probes / 23 sessions

6 of 11

How accuracy varies across probes & sessions

Median per-probe 0.57 all-channel / 0.68 in-brain; best probes exceed 0.80

ephys-atlas 2026_W26 Cosmos classifier · Allen (AIND) NP1 · 129 probes / 23 sessions

7 of 11

Example predictions

8 of 11

Example predictions

9 of 11

Example predictions - Bad predictions

10 of 11

Which regions transfer — and where errors go

Hippocampus & cortex transfer strongly; thalamus & olfactory collapse into cortex/HPF

ephys-atlas 2026_W26 Cosmos classifier · Allen (AIND) NP1 · 129 probes / 23 sessions

11 of 11

Key findings & limitations

Findings

• Small domain gap — IBL→Allen transfer matches held-out-probe IBL CV (both 0.58)

• HPF 0.91 and Isocortex 0.74 transfer strongly (robust laminar / theta-ripple signatures)

• Full feature set + matched denoising closes the gap that limited the LFP-only attempt

Things to do

• Thalamus (0.18) & olfactory (0.00) poorly detected — main headroom

• Weak out-of-brain (void) detection drags the all-channel number down

• Check with different probe geometries, and with more powerful models.

• Try with more powerful models.

• Create a plugin for acquisition softwares like SpikeGLX or Openephys.

ephys-atlas 2026_W26 Cosmos classifier · Allen (AIND) NP1 · 129 probes / 23 sessions