DANDI 000971 independent-control pilot v0.1
Status: completed; mixed descriptive result; no method promotion (26 July 2026)
Question
Do experimental signal-only bleaching baselines recover approximately the same slow trend as an independently acquired 405-nm isosbestic channel in real dual-site recordings?
The protocol and exact asset manifest were frozen before aggregate signal execution. The compact, machine-readable result retains every case.
Data and execution
One animal was selected from each of four session families in immutable published
DANDI:000971 version 0.260213.1851. Each NWB contributed DMS and DLS calcium plus
matched isosbestic channels: eight region-level cases and sixteen method-case
comparisons. All four downloads matched their recorded byte sizes and SHA-256
digests. All exposed the expected schema, and every fitted baseline was fully
finite after block averaging to approximately 20 Hz.
The executable runner is
scripts/run_dandi_000971_control_pilot.py.
It caches raw assets outside the repository, rechecks integrity on reuse, and
generates the JSON result from the frozen manifest.
Results
| Method | Cases | Median slow-trend correlation | Median relative RMSE | Correlation gate | RMSE gate |
|---|---|---|---|---|---|
| Double exponential | 8 | 0.259 | 4.62% | Fail | Pass |
| Rate-aware AsLS | 8 | 0.713 | 5.36% | Fail | Pass |
The low RMSE and modest correlation are not contradictory. Fluorescence trends have large positive offsets, so two traces can be close in scale while disagreeing about relatively small within-session changes. The individual correlations were heterogeneous: double exponential ranged from -0.482 to 0.804 and AsLS from 0.278 to 0.845. One double-exponential DLS fit was nearly flat while its comparator rose 21.6%, producing the negative correlation. In FP-RR20, both methods tracked the direction and approximate magnitude of decline in both regions.
The corrected-signal/isobestic correlations also varied substantially. They are reported diagnostically, not treated as a success criterion: residual correlation can reflect motion, shared bleaching, biological contamination, or other common structure, and a low value is not by itself proof of valid correction.
Interpretation and decision
The pilot establishes a reproducible real-NWB path and shows that both methods produce numerically plausible fluorescence-scale baselines. It does not show reliable recovery of the independently measured slow trend across preparations. Neither descriptive method gate passed in full, so SDR-0002 remains unchanged: both signal-only methods stay experimental and outside the recommended typed pipeline.
The comparator is not ground truth. The affine isosbestic fit can be invalid, the 405-nm channel can carry non-bleaching structure, and wavelength-specific bleaching can create genuine disagreement. Consequently this result motivates a larger multiverse across comparator construction and datasets; it does not rank AsLS as scientifically superior merely because its pilot median correlation was higher.
Follow-up
- Reproduce raw-to-archived-dF/F behavior in DANDI:000351, retaining its published processing outputs as a parity comparator while clearly marking the Dandiset as a mutable draft.
- Add explicit comparator variants (robust affine, OLS, and slow-scale choices) under a frozen multiverse, then report whether conclusions survive those choices.
- Refresh and execute the prospectively gated IBL analysis on the 18 newly available animals, independently of this control-free baseline decision.
- Seek a larger immutable raw calcium/isobestic cohort before reconsidering promotion.