Summary

The original question was whether some brain areas contain more neurons whose activity is suppressed at movement initiation. Exploration showed that a more defensible confirmatory question was broader and signed: do non-root Cosmos brain regions differ in their average movement-initiation modulation more than expected from within-probe anatomical label exchangeability?

The confirmatory result was positive. In held-out data, the variance across non-root Cosmos region means was larger than every statistic generated by 99,999 within-probe label shuffles. With the pre-specified p-value convention, this gives the minimum attainable one-sided p-value, p = 0.00001.

The strongest positive broad-region mean was in thalamus (TH, mean metric 0.1166), while several regions had means near zero or below zero. This should be read as an event-aligned modulation result, not as a claim about continuous movement versus quiescence states.

Original Question

Which brain areas contain more or fewer neurons with activity suppressed at movement initiation?

The analysis began with this suppression-focused framing. During exploration, two issues became clear:

A stored peak-response feature was not appropriate as the main statistic because it preferentially selected positive transients.
The most stable anatomical result was not a simple Beryl thalamus-versus-other suppression contrast, but a broader difference in signed modulation among Cosmos regions.

The locked confirmatory question therefore became:

Do broad non-root Cosmos brain regions differ in their movement-initiation modulation more than expected from within-probe anatomical label exchangeability?

Biological Context

Movement initiation is a major behavioral transition in the IBL task. Neural activity around firstMovement_times can reflect motor preparation, sensory feedback, action initiation, reward expectation, or other covariates that are aligned to the same trial epoch. Because Neuropixels recordings sample many anatomical locations, a brain-wide map can ask whether this event-aligned modulation is anatomically structured.

The analysis does not try to assign a causal role to any region. It asks a statistical anatomy question: given the units recorded on each probe, and given each probe’s anatomical label composition, are the observed region labels associated with more separation in movement-initiation modulation than would be expected if the labels were exchangeable within probe?

Refinement During Exploration

The unit metric was defined from spike counts around movement initiation:

event anchor: firstMovement_times
baseline window: [-0.2, 0) seconds
post window: [0, 0.2) seconds
baseline rate: baseline spikes divided by valid trials and window width
post rate: post spikes divided by valid trials and window width
primary metric: (post_rate - baseline_rate) / (post_rate + baseline_rate + R)

The regularizer R controls how much low-rate cells can dominate the normalized ratio. Exploration compared multiple values. R = 1 Hz was selected for confirmation because it was a moderate choice: less unstable than R = 0.1, but less suppressive of low-rate units than R = 10.

Exploratory sensitivity of the pre/post normalized movement-initiation metric to time windows and regularization values.

Single-unit examples and split-half checks were used to inspect whether the scalar metric matched interpretable peri-event responses. These checks motivated treating the scalar as an event-aligned modulation index rather than a literal movement-state selectivity measure.

Exploratory label-shuffle analyses supported a Cosmos-scale omnibus statistic. The diagnostic below shows the observed non-root Cosmos structure and the shuffle-calibrated null behavior in the exploration set.

Exploratory observed and shuffled distributions for the R=1 Cosmos label-shuffle analysis.

Locked Confirmatory Plan

Six sessions used for exploration were excluded. All remaining sessions in the local Brain Wide Map ephys dataset were held out for confirmation.

The primary analysis used R = 1 Hz. Eligible primary regions were non-root Cosmos regions with at least 50 held-out units and at least 10 held-out subjects. The Cosmos root label was excluded because it is too broad to be an interpretable anatomical group.

The observed statistic was computed in two steps:

Compute the unit-weighted mean metric within each eligible non-root Cosmos region.
Compute the unweighted sample variance across those region means.

This is an “average first, then variance” statistic. It tests whether broad region-level means are separated, not whether every probe individually shows a regional spread.

The null distribution was generated by permuting anatomical labels within each probe/insertion. Each shuffle kept the unit metric, subject, session, probe, and cluster identity fixed, and preserved the exact label counts within every probe. For each shuffled dataset, the same region means and across-region variance were recomputed.

The planned p-value was one-sided:

p = (1 + number of shuffled statistics >= observed statistic) / (99,999 + 1)

With 99,999 shuffles, the p-value denominator is 100,000 and the minimum possible p-value is 0.00001.

Confirmatory Results

The full confirmatory run computed rates for 73,182 held-out units. The primary Cosmos test included 67,764 units across 10 eligible non-root Cosmos regions.

The primary statistic was larger than every shuffled statistic:

observed non-root Cosmos variance: 0.0014228
shuffle null mean: 0.0006199
shuffle null SD: 0.0000259
shuffles: 99,999
p-value: 0.00001

Confirmatory primary statistic showing the observed non-root Cosmos variance relative to the within-probe label-shuffle null.

The region means showed a clear broad anatomical spread. Thalamus (TH) had the largest positive mean metric (0.1166). Isocortex, central nuclei, midbrain, hindbrain, and hippocampal formation were closer to zero but positive on average. Cerebellum, olfactory areas, hypothalamus, and cortical subplate were near zero or negative on the normalized metric.

The primary result supports the confirmatory claim: non-root Cosmos labels are associated with more across-region separation in event-aligned movement-initiation modulation than expected under within-probe anatomical label exchangeability.

Sensitivity Analyses

The Cosmos variance result was robust to the pre-specified regularizer sensitivities:

Analysis	R	Observed	Null mean	p-value
Cosmos non-root variance	0.1	0.001953	0.000811	0.00001
Cosmos non-root variance	1	0.001423	0.000620	0.00001
Cosmos non-root variance	10	0.000487	0.000235	0.00001

The Beryl-scale sensitivity analyses were also positive:

Analysis	R	Observed	Null mean	p-value
Beryl area-mean variance	1	0.004524	0.002125	0.00001
Beryl thalamus-minus-other contrast	1	0.106600	0.079555	0.00001

The Beryl thalamus contrast was kept as sensitivity evidence rather than the primary claim because the exploratory version of that targeted contrast was not as stable across regularizer choices as the broader Cosmos variance statistic.

Raw delta-rate summaries were used as scale diagnostics, not as primary tests. In the primary Cosmos regions, mean raw post_rate - baseline_rate was positive in all regions, while the normalized metric was near zero or negative in some regions. This difference is not a contradiction: the normalized metric weights the pre/post change by the unit’s overall event-window firing rate and the regularizer, so it captures relative modulation rather than absolute Hz change.

Interpretation

The result indicates that movement-initiation modulation is anatomically structured at a broad Cosmos-region scale. The clearest feature is high positive normalized modulation in thalamus relative to the other broad regions. The result is not driven only by selecting a single thalamic Beryl contrast: the primary statistic was an omnibus variance across non-root Cosmos region means, and it remained significant across the pre-specified regularizer sensitivity analyses.

The p-value is calibrated by the within-probe label-shuffle null. This is important because units are not independent biological replicates in the usual sense. The test asks whether the anatomical labels within the same probe carry more information about modulation than expected after shuffling labels among units recorded by that probe.

Caveats

The movement-initiation metric is event aligned. It does not isolate movement from all correlated task variables, such as sensory changes, action selection, reward expectation, or trial timing.

The null is conditional on the recorded probes, units, metrics, and probe-level anatomical compositions. It does not test whether the set of recorded probes is a random sample of the whole brain.

The statistic is unit-weighted within region and then unweighted across region means. This makes the anatomical means stable within each region while giving each eligible Cosmos region one vote in the final variance statistic. A subject-balanced version would answer a related but not identical question.

The primary metric depends on regularization. The result survived the pre-specified R = 0.1 and R = 10 sensitivity analyses, but the exact rank ordering and metric magnitude should not be overinterpreted as a direct physiological scale.

Anatomical hierarchy matters. The primary claim is at the broad Cosmos scale with root excluded. Finer Beryl results are useful for interpretation and sensitivity, but they should be treated as secondary unless separately locked and confirmed.

Reproducibility

Primary scripts and outputs:

confirmatory-analyses/021_run_confirmatory_analysis.py
artifacts/021_confirmatory_unit_metrics.parquet
confirmatory-analyses/021_confirmatory_permutation_tests.csv
confirmatory-analyses/021_confirmatory_area_means.csv
artifacts/021_confirmatory_permutation_nulls.parquet
confirmatory-analyses/022_graphical_summary.py

Key run settings:

local Brain Wide Map ephys dataset: data/bwm_ephys/1.1.0
primary regularizer: R = 1 Hz
permutation seed: 20260523
permutations: 99,999
permutation block: probe/insertion
p-value denominator: 100,000

Suggested Future AI Instructions

Several reusable workflow lessons from this project are saved in ai_instruction_suggestions.md. The most important are to state the exchangeability block explicitly for label-shuffle tests, separate unit-level sensitivity from valid p-value calibration, and run regularizer sensitivity before locking normalized firing-rate metrics.