CADENCE: Sensor-Aware Glial Calcium Modeling and Control

Calcium Adaptive Dynamics Engine for Neuroscience, Control, and Estimation

CADENCE is a reproducible research prototype for testing how calcium-indicator memory affects latent-state inference and downstream in-silico control. It pairs a labelled synthetic benchmark with a hierarchy-aware secondary analysis of public mouse astrocyte recordings.

Research status: synthetic methods benchmark plus exploratory public-data analysis. CADENCE is not evidence of treatment efficacy, a confirmed astrocyte state mechanism, or a clinical controller.

New to the project? Start with the plain-language tutorial.

Research question

Can a sensor-aware model separate simulated state dynamics from calcium-indicator memory, and which assumptions determine whether model-based control appears to work?

Evidence at a glance

Evidence	Design	Current result	Interpretation
Synthetic state recovery	Four-state simulator; fit on 10 intact traces and score on 20 held-out traces	89.0% offline-smoothed accuracy; 83.4% causal-filter accuracy	The estimator handles indicator memory under this simulator; online inference is harder
Synthetic exit-hazard recovery	Causal states; retrospectively standardized shared exposure decay of 0.90; whole-trace bootstrap	Intact b1 = 0.694, 95% CI [0.558, 0.847]; intact−blocked difference 0.725 [0.580, 0.864]	Recovers a contrast deliberately encoded in simulation; not a biochemical measurement
Multi-seed synthetic sensitivity	Five independently randomized simulator runs; same 30-trace / 20-held-out-trace design; seed-specific trace bootstraps	Causal accuracy 76.0–83.4%; all 5/5 intact−blocked contrast intervals exclude zero	Robustness within this simulator family only; not biological replication
Real astrocyte secondary analysis	Public H1R data; 147 ROIs nested in 13 slices; ROIs aggregated before slice comparison	KO−WT response ΔAUC positive in 5/5 and 6/7 paired slices	Descriptive biological context; animal identifiers are unavailable
Offset-free H1R robustness	Alternative median-response metric on the same recordings, common within-slice MAD scale, and 36-specification grid	Common-scale KO−WT effect positive in 5/5 and 6/7 slices; mean effect positive in all 36 specifications in both cohorts	Direction is not created by the +100,000 offset, but this is not an independent dataset or assay and a stricter per-ROI z check is less consistent
In-silico controller	Intervention and endogenous escape share one simulated coefficient	Control is disabled when that coefficient is near zero	Structural code/model check, not independent biological validation

Real astrocyte analysis

CADENCE analyzes two cohorts from Taylor et al.'s public mouse cortical astrocyte histamine-1-receptor dataset. It follows the deposited preprocessing, computes norepinephrine post-minus-pre response ΔAUC, and averages ROIs within each slice before comparing wild-type and knockout signals.

The paired slice differences point in the same direction in most slices across both deposited cohorts. They are not animal-level estimates: the deposited tables omit animal identifiers, slices and ROIs cannot be treated as independent animals, and the two sensor/protocol cohorts are not pooled. One slice with a markedly different raw fluorescence scale is flagged and retained rather than silently excluded. No p-value is reported.

See the real-data methods, machine-readable summary, and derived-data provenance.

Offset-free robustness check

This is an alternative analysis of the same recordings, not an independent replication, assay, or dataset. It does not reuse contextual ΔF/F₀ or AUC and measures each ROI's raw median post-minus-pre response and divides the paired slice contrast by one robust baseline MAD scale shared by both genotypes. This check is invariant to the arbitrary additive fluorescence offset and to common positive gain.

• NE only: mean KO−WT effect +11.663 baseline-MAD units; 5/5 slices positive.
• NE after low histamine: mean +3.559; 6/7 slices positive.
• Across an audit-defined 36-specification grid, the mean effect stayed positive in every specification in both cohorts.
• A stricter within-ROI noise-standardized check was less uniform: 3/5 and 5/7 slices were positive. That disagreement is retained as a limitation.

The grid was defined after the primary data had been inspected. It is an exploratory multiverse analysis, not a preregistration or confirmatory test.

Synthetic benchmark

Real recordings do not provide frame-level state labels or known parameters, so the synthetic benchmark supplies both. A fixed four-state simulator is passed through GCaMP-like sensor kinetics and noise. The sensor-aware model is then evaluated on traces withheld from fitting.

The shared 0.90 exposure decay was standardized during the repository audit, not prospectively preregistered. It is held fixed across conditions to prevent condition-specific tuning.

The downstream variable L is a causal accumulator of recent inferred high-state occupancy. It is a dwell-history proxy—not measured molecular calcium load. Likewise, four states make a controlled, interpretable benchmark; they have not been established as discrete astrocyte biology.

The causal blocked estimate is −0.030 ([−0.058, −0.005]) even though the oracle-state interval includes zero. This small false-positive null bias from hard causal labels is reported explicitly; the cross-condition contrast—not the blocked point estimate alone—is the synthetic result.

Multi-seed sensitivity

The compact benchmark was independently regenerated for five fixed simulator seeds. Held-out causal accuracy ranged from 76.0% to 83.4%; the causal intact-minus-blocked b1 contrast ranged from +0.637 to +0.891, and all five seed-specific trace-bootstrap intervals excluded zero. The causal blocked slope remained negative across the five runs, consistent with the documented hard-label null bias; the contrast remains the relevant endpoint.

This checks variation across random draws from the same simulator—not an independent dataset, a biological replication, or a generalization guarantee. See the full sensitivity methods and machine-readable results.

Show synthetic benchmark figures

The controller figures test behavior under the simulator's assumed plant. The blocked case is expected algebraically because b1 multiplies both the endogenous exit term and the intervention term. A hard adaptive exposure budget was added after causal validation revealed that the heuristic futility check can miss non-response. Lower cumulative cost is therefore a constraint, not an efficiency discovery; CADENCE can also use a larger peak dose than open loop.

Reproduce

Python 3.11 or newer is recommended.

python -m pip install -r requirements.txt

# Recreate the compact reference benchmark and all versioned analyses
python src/run_all.py --quick --n_traces 30 --fit_traces 10 --control_traces 6

# Longer: repeat the synthetic benchmark across five independent generator seeds
python src/seed_robustness.py

# Run unit and claim-level regression checks
python -m unittest discover -s tests -v
python tests/test_pipeline.py

The versioned compact H1R export is sufficient to reproduce the public-data analysis with Python. Re-exporting it from the original MATLAB tables is optional and documented in docs/real_data.md.

Raw Dryad and DANDI source files are not committed. The compact H1R export keeps the source DOI, Dryad file IDs, and SHA-256 checksums. real_data.py also tests NWB ingestion on a pan-neuronal zebrafish recording from DANDI:001076; that asset is an out-of-domain loader/QC check, not glial validation.

What the repository establishes

• Held-out recovery scoring against labelled synthetic data
• Whole-trace uncertainty for a fixed shared synthetic hazard contrast
• Multi-seed robustness of that synthetic contrast under the same simulator
• Machine-readable wiring from the learned synthetic law into the controller
• Hierarchy-aware descriptive analysis of public astrocyte fluorescence
• Source checksums, deterministic derived data, continuous integration, and explicit limitations

What it does not establish

• That astrocytes occupy the simulator's four discrete states
• That state-derived exposure measures a molecular feedback pathway
• Animal-level H1R effects from the deposited tables
• Controller efficacy, safety, or mechanism in tissue, animals, or people

Next decisive experiment

Use an animal-identified glial calcium dataset under a prospectively frozen analysis plan; compare state-history predictors with predictors computed directly from fluorescence; and evaluate model mismatch across indicators, sampling rates, drift, and heterogeneous traces beyond the five same-simulator seed checks.

Research integrity

This repository is not a competition application. The AAN Neuroscience Research Prize requires the applicant's original research and writing. See AAN prize readiness, research disclosure, references, citation metadata, and the contribution guide.

Author

Joseph David · LinkedIn · GitHub