Transient Pauses and Regular High-Rate Units Across IBL Brain Regions

Interim exploratory report

1 Original Question

The user asked:

Find all brain regions where there are neurons of high baseline firing rate that show transient pauses in activity. SNr is a known motivating example, but the goal is to search across all recorded regions.

The project was run on the local IBL Brain Wide Map ephys dataset, data/bwm_ephys/1.1.0, using good-unit metadata, task trials, and local spike shards. The analysis is still exploratory. It has not yet reached a locked confirmatory analysis.

2 Introduction

High-rate neurons in basal ganglia, cerebellar, midbrain, and brainstem circuits can encode task events by brief silences as well as by firing-rate increases. Substantia nigra pars reticulata (SNr) was the motivating positive-control region: many SNr neurons fire tonically and can show sharp stimulus- or movement-related suppressions. The scientific challenge is that a “pause” is not a single obvious mathematical object. A low-spike-count window can mean a genuine event-locked suppression, a spontaneous long inter-spike interval in an otherwise regular cell, a slow state transition, nonstationarity, or a spike-sorting artifact.

The analysis therefore followed an exploration-first workflow. Early analyses used SNr as a positive-control region to develop source-level diagnostics and candidate metrics. Later analyses scaled one family of metrics across all local BWM good units and mapped regional patterns of pausing and regularity.

Two concepts are separated throughout:

• Pausy high-rate units: cells with more near-silent short windows than expected under a local-rate reference.
• Regular high-rate units: cells whose ISI structure or low-window count is more regular than a local homogeneous-Poisson expectation.

These are not opposites. A region can contain both strongly regular tonic cells and a pausy upper tail. SNr is a clear example of that heterogeneity.

3 Refinement Of The Question

The initial wording, “transient pauses,” was too broad. The first SNr-only branch tried to detect multi-second recovered pauses using binned and smoothed session-wide rate traces. That branch produced plausible examples but also many slow state-like or nonstationary events, and the user clarified that the intended phenotype was closer to fast, roughly 100 ms pauses.

Early SNr diagnostic separating candidate recovered pauses from slower state-like low-rate epochs. This branch was useful, but the target was later refined toward faster pauses.

The active question became:

Which regions contain high-rate neurons with brief, roughly 100 ms pauses or unusually regular spike trains, and what do example spike trains look like?

The metric development split into three related tracks:

1. Event-locked fast pauses in SNr, especially around stimulus onset and movement onset.
2. Event-free pause-like intervals detected from the spike train alone.
3. Whole-database region summaries using local-rate observed/expected counts of near-silent windows and ISI regularity.

4 Data And Scope

The local bwm_ephys dataset contains good units, session/probe metadata, task trials, event-response features, and per-insertion spike shards. The whole-database feature store built during this project contains 75,395 good units. The region-level summaries used a high-rate analysis subset:

• firing rate >= 10 Hz;
• at least 1000 loaded spikes;
• finite ISI CV2 and finite 110 ms low-window statistic;
• region summaries required enough unit and insertion coverage, with exact thresholds varying by figure.

The primary scaled feature was:

observed_expected_ratio_w110ms =
  observed number of 110 ms scan windows with <= 1 spike
  / expected number under a local homogeneous-Poisson rate estimate

Positive log2 values indicate more near-silent windows than the local-rate reference predicts. Negative values indicate fewer near-silent windows, consistent with regular tonic firing. This is an operational exploratory metric, not yet a locked confirmatory phenotype.

5 Event-Locked SNr Positive-Control Analysis

SNr provided a positive-control population for fast event-locked pauses. In local bwm_ephys 1.1.0 there were 149 SNr good units across 19 insertions, 19 sessions, and 17 subjects. Among high-rate SNr units, the event-locked exploratory metric compared pre-event baseline firing to the lowest 100 ms mean rate in an event-specific search window.

The exploratory candidate rule was:

• baseline >= 10 Hz;
• pause ratio <= 0.6;
• drop >= 8 Hz;
• paired one-sided Wilcoxon p <= 0.01.

Under this rule, 11 of 89 high-rate SNr units were stimulus-onset candidate fast-pause units, and 10 of 89 were first-movement candidate fast-pause units.

SNr event-locked fast-pause summary. Candidate units show sharp suppressions around stimulus onset or movement onset.

Example SNr event-locked fast-pause PSTHs.

6 Event-Free Pause Metric Development

The event-free analysis asked whether pause-like windows can be detected without using task events. SNr unit 128 in NYU-45 was used as a positive-control cell. A naive low-count sliding-window detector was too permissive as an event detector, and narrow ISI-based pause calls missed many visually obvious stimulus-locked pauses.

The most important lesson came from comparing ISI definitions against a fixed event-locked pause target. In unit 128, the 30-100 ms post-stimulus window had zero spikes on 182 of 189 stimulus trials. Yet broad 100-180 ms long-ISI calls recalled only about 54 percent of those fixed-window pauses, and stricter ISI definitions recalled less.

False-negative diagnostic for ISI-based event-free pause calls against fixed stimulus-window pauses in the SNr positive-control cell.

This led to a practical split:

• Event-locked pauses should be measured with direct event-aligned low-spike-count windows.
• Event-free pausing should probably be summarized as a cell-level excess of pause-like windows relative to a local-rate null, not as a perfect event detector.

The leading event-level descriptive detector became sliding_silent_110ms_le1: scan the spike train in 110 ms windows every 10 ms, mark windows with <=1 spike, and merge overlapping positive windows.

Pause-aligned rasters and PSTHs for sliding_silent_110ms_le1 detections. Stimulus-associated detected pauses are a subset of all detected pauses.

For four stimulus-pause examples, ungated recall of fixed zero-spike stimulus pauses was high in three cells and lower in one:

subject	cluster	firing rate (Hz)	ungated fixed-zero recall	p<=0.05 local-rate-gated recall
NYU-45	128	64.7	0.956	0.956
DY_018	0	43.2	0.992	0.632
NYU-45	137	33.2	1.000	0.000
NYU-45	109	53.5	0.869	0.774

The local-rate gate illustrates a caveat: a homogeneous-Poisson p<=0.05 threshold corresponds to about 43.1 Hz for a 110 ms <=1-spike window. It is useful as a graded confidence or eligibility idea, but too strict as a universal hard gate because it excludes visually clear pauses in cells with firing rates below that threshold.

7 Whole-Database Region-Level Exploration

The whole-database feature store computed observed/expected low-window ratios at 50, 80, 110, and 140 ms, plus ISI CV/CV2 summaries for all 75,395 local good units. Region-level plots used distributional summaries rather than medians alone, because pausing and regularity often live in tails.

Region-level scatter showing pausy upper tails and regular lower tails.

Ranked region summaries for pausy and regular tails.

Top regions by pausy upper-tail score at 110 ms included FN, MV, PFL, GPe, SNr, BLA, ENTm, ANcr2, SUV, IP, DN, and SIM. SNr ranked among the strongest pausy regions, with p90 log2(observed/expected) = 1.69 and 15.8 percent of eligible SNr units having observed/expected > 2.

region	units	insertions	subjects	median rate Hz	pausy score	frac ratio > 2	regular score	median CV2
FN	38	3	3	47.9	3.31	0.158	0.845	0.513
MV	362	22	18	29.1	1.80	0.166	0.446	0.511
PFL	47	6	5	20.0	1.74	0.170	0.553	0.609
GPe	207	13	9	37.0	1.71	0.145	0.700	0.508
SNr	101	16	14	49.0	1.69	0.158	3.44	0.467
BLA	25	9	9	24.7	1.68	0.240	0.645	0.637
ENTm	76	11	11	16.2	1.37	0.145	0.063	0.842
ANcr2	231	16	9	23.5	1.37	0.113	0.255	0.691

Top regular regions by low-window lower-tail score included SNc, SNr, LH, FL, MA, PAG, PPT, PB, NPC, LHA, APN, and dhc. SNc was the strongest regular region: regular score = 4.69, with 48.0 percent of eligible SNc units having observed/expected < 0.5.

region	units	insertions	subjects	median rate Hz	pausy score	regular score	frac ratio < 0.5	median CV2
SNc	25	3	3	35.6	0.289	4.69	0.480	0.459
SNr	101	16	14	49.0	1.69	3.44	0.267	0.467
LH	26	8	7	32.0	0.037	2.21	0.231	0.586
FL	20	3	3	32.1	0.808	2.18	0.200	0.532
MA	25	5	5	24.4	0.415	2.12	0.200	0.649
PAG	243	39	26	20.7	0.359	1.93	0.160	0.530
PPT	24	11	10	27.7	0.514	1.83	0.208	0.644
PB	109	24	19	23.4	0.642	1.68	0.110	0.597

The Swanson flatmap made the regional pattern easier to inspect anatomically. The top panels show pausy metrics. The bottom panels were revised to plot inverse-CV2 regularity, so the most regular regions appear bright.

Swanson flatmap of pausy and regularity features. Colored text labels use each panel’s colormap.

8 Non-SNr Pausy Examples: MV

MV was inspected as the first non-SNr region with a high pausy score. The top MV cells by 110 ms observed/expected low-window ratio showed genuine discrete pauses in high-rate spike trains, not just bookkeeping artifacts. The examples split into at least two forms:

• rare interruptions in very regular high-rate firing with little simple task-event alignment;
• more numerous pauses in cells with clear post-event suppressions, especially around feedback.

Example MV pause-like cells with pause-centered spike snippets and task-event rasters.

For selected MV examples, feedback-window enrichment was often strongest:

subject	cluster	event	pause events	associated pauses	fold over occupancy
UCLA033	889	feedback	98	8	2.18
DY_020	362	feedback	2474	334	4.86
UCLA033	634	feedback	530	77	3.88
CSH_ZAD_022	216	feedback	927	131	7.26

This suggests that MV pausing is real, but may often be feedback/outcome/state related rather than stimulus-onset locked like the SNr positive-control examples.

A key caveat is that very regular high-rate cells can have very large observed/expected ratios from modest absolute numbers of detected pauses, because the expected count of <=1-spike 110 ms windows under a local homogeneous-Poisson reference can be extremely small. Region and cell summaries should therefore report both ratio statistics and absolute pause counts or event rates.

9 Regular But Non-Pausy Population Examples

The user then asked to inspect regular but non-pausy regions. The first response used event-averaged PSTH heatmaps, but the intended view was a literal simultaneous-neuron raster. The corrected figures show continuous 12 s windows from the same insertion, with rows as simultaneously recorded neurons, ticks as spikes, colored vertical lines for task events, and 50 ms population spike-rate traces below.

Continuous simultaneous-unit raster for an SNc regular/non-pausy example.

Continuous simultaneous-unit raster for a CLI regular/non-pausy example.

These figures emphasize that “regular” means low near-silent-window excess or low CV2, not absence of task modulation or population structure. Event-averaged diagnostics for SNc, CLI, V, and PAG showed population-mean normalized rates that could dip to roughly half of session-rate baseline around some events. Continuous rasters are more appropriate for visualizing moment-to-moment simultaneously recorded activity.

10 Locked Analysis Plan

No confirmatory analysis has been locked or run yet. The strongest current candidate plan is:

1. Treat event-locked and event-free pauses as separate phenotypes.
2. For event-locked SNr-like pauses, use fixed event-aligned low-count windows with pre/post activity support, not ISI-only event-free definitions.
3. For event-free pausing, use cell-level excess of near-silent 110 ms windows over a local-rate reference, reporting both observed/expected ratio and absolute event counts/rates.
4. Define high-rate eligibility before confirmation, likely using firing rate >= 10 Hz for broad region maps and stricter local-rate or support diagnostics for individual pause-event claims.
5. Choose exploration and confirmation sets by insertion or subject before formal region-level testing.
6. Test region-level prevalence or tail enrichment using insertion/subject as the independent replicate, not individual spikes or simultaneously recorded units.

The current report should therefore be read as an exploratory synthesis, not as a final inferential result.

11 Confirmatory Results

No confirmatory results are available yet. All reported regional rankings, examples, and thresholds are exploratory and were developed after inspecting the data. They are useful for choosing hypotheses and metrics, but they should not be used as p-values or final claims.

12 Post-Hoc And Diagnostic Analyses

The project produced several diagnostic branches that should inform the next phase:

• Multi-second pause detectors were informative but did not match the user’s intended fast-pause phenotype.
• ISI pause definitions are useful for event-free cell-level structure but miss many event-locked fixed-window pauses.
• Sliding 110 ms <=1-spike events recover many stimulus pauses, but local-rate hard gates can exclude visually clear lower-rate examples.
• Region maps show both pausy and regular tails; SNr is heterogeneous and appears in both.
• MV examples show genuine non-SNr pauses, often feedback-associated.
• Regular non-pausy examples can still be strongly task-modulated.

13 Discussion

The main scientific result so far is not a locked list of regions, but a clarified phenotype landscape. Brief silences in high-rate units can arise in several forms:

• event-locked pauses, as in the SNr positive-control examples;
• event-free or weakly task-aligned pause-like intervals;
• feedback/outcome-associated pauses in MV examples;
• regular high-rate tonic firing with fewer near-silent windows than expected under a local-rate Poisson model;
• mixtures of pausy and regular cells within the same region.

The current whole-database metric, log2 observed/expected <=1-spike windows at 110 ms, is useful because it captures both positive pausy tails and negative regular tails. It is also risky if overinterpreted alone. A large ratio can reflect a small number of pauses in a highly regular cell; conversely, a region with strong task modulation can be non-pausy by this metric if it does not produce many near-silent windows.

The most promising next step is to lock a narrower confirmatory question. One candidate is:

In held-out insertions or subjects, are high-rate SNr and MV units enriched for 110 ms near-silent windows relative to other high-rate units after accounting for firing rate and insertion coverage?

Another is event-locked:

In held-out SNr insertions, what fraction of high-rate units show fixed-window stimulus- or movement-locked pauses with pre/post support?

Those questions need a declared replicate definition, region inclusion thresholds, and a locked metric before testing.

13.1 Caveats

• The local-rate Poisson reference is a useful operational baseline, not a full generative model of spike trains.
• Firing-rate nonstationarity, behavioral state, and spike-sorting stability can all create apparent pauses or apparent regularity.
• Single example cells or single insertions are illustrative, not region-wide evidence.
• Region labels use local atlas assignments and should be treated carefully for small or boundary-adjacent nuclei.
• The current whole-database feature store uses good units in local bwm_ephys; any future reproduction should record the dataset version and local data root.

14 AI Instruction Lessons

A standalone file of suggested instruction updates is saved in this report directory:

instruction_suggestions.md

The most important lessons are:

• keep event-locked and event-free pauses separate;
• require raw spike-time diagnostics for pause claims;
• do not call outside-task detections false positives without a true ground truth or appropriate null;
• report absolute event counts alongside observed/expected ratios;
• clarify whether the user wants simultaneous raw rasters or time-averaged PSTHs.

15 References

International Brain Laboratory. Standardized and reproducible measurement of decision-making in mice. eLife 10, e63711 (2021). https://doi.org/10.7554/eLife.63711

International Brain Laboratory, Angelaki, D., Benson, B. et al. A brain-wide map of neural activity during complex behaviour. Nature 645, 177-191 (2025). https://doi.org/10.1038/s41586-025-09235-0

Siegle, J. H., Jia, X., Durand, S. et al. Survey of spiking in the mouse visual system reveals functional hierarchy. Nature 592, 86-92 (2021). https://doi.org/10.1038/s41586-020-03171-x