Dimorphism in the male model

flymsg dimorphism asks, for each validation case, whether the simulated response runs through sexually dimorphic neurons more than expected. Code: src/flymsg/dimorphism.py. Output: runs/dimorphism.txt.

Method

• Responders: neurons other than the stimulated ones that fire during the 300 ms stimulus in ≥ 2 of 3 seeds (validation protocol, defaults: w_syn 0.192, no adaptive threshold; with the v0.3 defaults the pattern was the same, P1 3.4–8.1×, pIP10 5.8–16.3×).
• Categories (MaleCNS annotations; "potentially …" included): fru/dsx+ (any fruDsx value), male-specific, dimorphic (sexually or potentially dimorphic).
• Null: 1,000 random neuron sets of the same size and superclass mix as the responders, drawn from all non-stimulated neurons. Matching superclass matters: responders of a visual stimulus are mostly visual neurons, and dimorphic neurons are not spread evenly.
• Statistic: share among responders / mean null share, with an empirical one-sided p (+1 smoothing, so the smallest possible p is 0.001). 15 tests (5 cases × 3 categories): with Bonferroni, p ≤ 0.0033 is significant.

Results

Bold: p = 0.001. Counts of responders in the category in brackets.

• Courtship circuits run through the dimorphic network. P1 and pIP10 responses are 4.3–18.6× enriched in fru/dsx+, male-specific and dimorphic neurons, within their superclasses. This is the model's side of the known picture: P1 and pIP10 sit in the fruitless circuit that builds courtship song.
• Escape and feeding avoid it. Looming and sugar responses contain about half the expected share of fru/dsx+ and male-specific neurons.
• The giant fiber's reach depends on the model. With the v0.3 defaults its 29 responders included 5 fru+ wing motor neurons (hg1, hg3, ps1); with the density-scaled weight only 2 neurons respond besides TTMn, so the escape output stays narrow.

Silencing (flymsg dimorphism --silencing --null 1000, runs/silencing-1000.txt)

Every neuron of a category is silenced (it never fires, so it transmits nothing: for the rest of the network this equals zeroing its outgoing synapses, as Shiu et al. silence neurons), except the stimulus and the targets. The change of each target's rate is compared with 1,000 silencings of as many neurons outside the category, with the same superclass mix (defaults, 3 seeds). These are the pre-registered T1 runs (below); the smallest possible p is 0.001 and the threshold is 0.00625.

Confirmatory: does the dimorphic network hold dPR1 back? (one-sided for a rise)

Verdict, by the pre-registered rule (both cases must pass):

• dimorphic: confirmed. Silencing the 948 sexually dimorphic neurons raises dPR1 by 61–64 % after either stimulus, more than any of 1,000 matched random silencings.
• fru/dsx+: not confirmed. The rise is as large (+79 … +100 %), but after P1 eleven of 1,000 random silencings of ~4,900 superclass-matched neurons raised dPR1 as much or more (p 0.011, above the 0.00625 threshold). Silencing a fifth of the CNS moves dPR1 a lot by itself: the null mean is +21 % there. The effect is real but not specific to fru/dsx+.

Descriptive (not part of the 8 confirmatory tests): male-specific neurons (1,332/1,418) change dPR1 by −2 % after P1 and −13 % after pIP10 (p 0.99, 0.015 for a drop); fru/dsx+ silencing lowers pIP10 after P1 by 12 % (p 0.001 for a drop), the one clear fall. Looming escape depends on none of the three categories: silencing any of them moves DNp01 by under 1 % and TTMn by −10 … +18 %, inside the null in every case (p 0.11 … 0.93). The 1,000 draws confirm what the v0.5 run with 100 draws showed.

• Enrichment is not necessity. Courtship responses survive the loss of thousands of dimorphic neurons, because they run mostly through the direct P1 → pIP10 → dPR1 links, which stay intact (stimulus and targets are never silenced).
• The v0.5 run with 100 draws reported the same direction and sizes; what 1,000 draws add is the threshold, and they show that the fru/dsx+ effect does not clear it.
• With the v0.3 defaults (20 null draws) silencing the male-specific neurons lowered dPR1 by ~20 %; with the density-scaled model that effect is gone, so it is not robust.

The predicted loop (flymsg dimorphism --loop --null 1000, runs/loop-silencing.txt)

Pre-registered test T2. Two sets are silenced: dMS9 (2 neurons) and the inhibitory feedback it drives (4 vPR9_a + 4 IN00A038). The null silences as many of the case's own responders, matched on superclass and transmitter sign — a stricter comparison than random neurons, since those are active neurons of the same kind.

Verdict: partly supported. dMS9 passes in both cases: silencing those two neurons raises dPR1 by two thirds, far beyond silencing any 2 matched responders. The named inhibitors do not: after pIP10 silencing them lowers dPR1 (−6 %), and after P1 the rise stays inside the null (p 0.036). The second half of the v0.5 prediction is wrong: dMS9 restrains dPR1, but not through vPR9_a and IN00A038 as stated — or not only through them, since the null shows that silencing 8 active VNC inhibitory neurons can raise dPR1 by up to 99 % on its own.

Descriptive: silencing both sets together gives +61 % (pIP10, p 0.012) and +75 % (P1, p 0.001), close to dMS9 alone, which fits dMS9 being the part that matters. Silencing dMS9 also raises pIP10 itself after P1 (+14 %, p 0.001).

Reading the two results carefully

• The two categories overlap little. Only 318 of the 948 sexually dimorphic neurons are also fru/dsx+, so the confirmed result is not a subset of the one that failed.
• A category's size sets the power of its test. Silencing ~4,900 superclass-matched random neurons raises dPR1 by +21 % on average, with draws up to +90 %: only a very large effect can clear that null, and fru/dsx+'s +79 % does not. The 948-neuron null moves dPR1 by ±10 % at the median, which is why the same size of effect passes there. "Not confirmed" therefore means "not distinguishable from silencing that many neurons", not "no effect".
• T2's inhibitory test is stringent by construction. Its null silences 8 active VNC inhibitory responders, which on their own can raise dPR1 by up to 110 %. What sinks the prediction is not only the p value: after pIP10 silencing vPR9_a + IN00A038 lowers dPR1, the opposite of what a feedback inhibitor should do.
• dPR1 is far from saturation (a 2.2 ms refractory period allows ~450 Hz), so the rises are not a ceiling artefact.

Where the restraint comes from (post-hoc, found after T2, not yet tested)

Silencing dMS9 and looking at what changes gives a longer chain than the one v0.5 proposed. The inhibitory input to dPR1 that collapses is IN03B024 (2 GABAergic VNC neurons, 731 and 802 synapses onto dPR1): from 90 Hz to 29 Hz when dMS9 is silenced. Its drive comes from vMS12_a/b/c (12 cholinergic fru+ neurons, a quarter of its excitatory input), which dMS9 drives with 4,054 / 1,028 / 1,480 synapses and which fall to 0 Hz without it:

dPR1 →822→ dMS9 →6,562→ vMS12_a/b/c →1,458→ IN03B024 →1,533→ dPR1

The v0.5 route (dMS9 →670→ vPR9_a →128→ dPR1) is thin next to this one, which fits T2's rejection. This chain was found by looking at the result, so it is a hypothesis, not a finding, and the numbers above are wiring and one descriptive comparison.

T3 (fixed 2026-09-23, before running it; run 2026-10-11, checkpoint resumed, not replaced; same protocol, 1,000 matched-responder null draws, threshold p ≤ 0.05/3 = 0.0167 for a rise of dPR1 after pIP10 and after P1). Report: runs/t3-loop.txt.

1. Silencing IN03B024 (2 neurons) raises dPR1.
2. Silencing vMS12_a + vMS12_b + vMS12_c (12 neurons) raises dPR1.
3. Epistasis: silencing dMS9 together with IN03B024 raises dPR1 no more than silencing IN03B024 alone (their 95 % bootstrap intervals overlap). If dMS9 acts through another route as well, this fails.

The chain is supported only if 1 and 2 pass in both cases and 3 holds; if 1 passes and 2 fails, the drive to IN03B024 comes from somewhere else; if 1 fails, IN03B024 is not the route either.

Result (2026-10-11). All three hold, so the chain is supported. Silencing IN03B024 (2 neurons) raises dPR1 by 101% after pIP10 (88 → 177 Hz) and by 100% after P1 (97 → 193 Hz), p = 0.001 both, the floor of 1,000 draws. Silencing vMS12_a+b+c (12 neurons) raises it by 67% (88 → 147 Hz) and 78% (97 → 172 Hz), p = 0.001 both. Silencing dMS9 together with IN03B024 (4 neurons) raises it by 100% and 99%. The 95% bootstrap intervals of the per-seed dPR1 rise overlap in both cases: pIP10 [0.893, 1.167] for both silencings, P1 [0.966, 1.036] for both. w_syn stayed 0.192.

Which cell types (--by-type, runs/silencing-by-type-*.txt)

Silencing the fru/dsx+ responders one cell type at a time (3 seeds, no null) points to one loop:

dMS9 (fru_high, sexually dimorphic, cholinergic; "Lillvis 2024: dMS9") receives 822 synapses from dPR1 and gives it only 5. It drives two inhibitory neurons that project back onto dPR1: vPR9_a (670 synapses from dMS9, 128 onto dPR1) and IN00A038 (161, 132). From that wiring v0.5 predicted a disynaptic negative feedback loop, dPR1 → dMS9 → vPR9_a / IN00A038 → dPR1. The pre-registered test T2 (above) keeps the first step and rejects the second: silencing dMS9 raises dPR1 by two thirds, beyond any matched silencing, while silencing vPR9_a and IN00A038 does not (it lowers dPR1 after pIP10). T3, above, supports the longer chain dPR1 → dMS9 → vMS12_a/b/c → IN03B024 → dPR1. The two neurons named in v0.5 are not that route. AVLP717m (fru+) relays part of P1's drive to pIP10. These per-type numbers come from a scan without a null and are indications, unlike T1 and T2.

T4, pre-registered before any new per-type run (fixed 2026-10-11, before running it, and without T3's outcome). T3's sets stay the ones fixed on 2026-09-23. This test does not split them, and it is not part of the T3 command.

The bold claims in the table are dMS9's rise of dPR1 and AVLP717m's drop of pIP10 after P1. dMS9 is T2, already decided. vMS12_a belongs to T3's vMS12 set and does not get a second confirmatory test. vPR9_a sits in T2's rejected inhibitory set; vPR9_c is the other half of that row and is not promoted because its indication goes up. The TN1a range is not a claim about one subtype. Every other fru/dsx+ responder type in runs/silencing-by-type-*.txt stays an indication.

One confirmatory type: AVLP717m (2 neurons). The null is the T2/T3 null: as many of the case's own responders, matched on superclass and transmitter sign, never the stimulus, the targets or the tested neurons. 1,000 draws, 3 seeds, 100 Hz, 300 ms, defaults only (w_syn 0.192, th_jump 0; do not pass --w-syn). The one confirmatory test is a drop of pIP10 after P1, p ≤ 0.05. Both courtship cases are run; the silencing after pIP10, and AVLP717m's effect on dPR1, are reported and do not change the verdict. The pIP10-stimulus case has no other excitatory cb_intrinsic responder, so that null is not defined and the case is reported without a p. A failure is a finding.

bg uv run flymsg dimorphism --loop t4 --null 1000 --checkpoint runs/t4-loop.ckpt.jsonl

Result (2026-10-11). The confirmatory test passes. Silencing AVLP717m (2 neurons) drops pIP10 after P1 by 43% (104 → 60 Hz), p = 0.001, the floor of 1,000 draws. The same silencing drops dPR1 after P1 by 28% (97 → 70 Hz), p = 0.001; that row is reported and does not decide. After the pIP10 stimulus, dPR1 is unchanged (88 → 88 Hz) and the matched null is not defined, so that case has no p. w_syn stayed 0.192. Report: runs/t4-loop.txt.

Pre-registered tests (v0.6, fixed 2026-09-22 before running)

The silencing results above are indications: 100 null draws cannot reach a Bonferroni threshold, the p values tested drops while the claims are rises, and the loop comes from a scan without a null. Two confirmatory tests, with the defaults and the validation protocol (3 seeds, 100 Hz, 300 ms; target rate = best neuron's mean rate), follow. The rules below are fixed before any of these runs; a failure is reported as a finding.

T1. The dimorphic network restrains dPR1 (flymsg dimorphism --silencing --null 1000). Same design as above (silence every neuron of the category except stimulus and targets; null = as many neurons outside the category with the same superclass mix), with 1,000 null draws and a one-sided test for a rise: p = (1 + number of null draws whose rise is ≥ the observed rise) / 1,001. Four confirmatory tests: fru/dsx+ and dimorphic, each after P1 and after pIP10, on dPR1. Male-specific, the other targets and looming escape are reported as descriptive.

T2. The predicted loop dPR1 → dMS9 → vPR9_a / IN00A038 → dPR1 (flymsg dimorphism --loop --null 1000). Two silenced sets: dMS9 (both neurons, cholinergic, ascending) and the inhibitory feedback (all 4 vPR9_a and all 4 IN00A038, GABAergic, VNC intrinsic). The null silences as many neurons drawn from the case's own responders (firing in ≥ 2 of 3 seeds), matched on superclass and transmitter sign, never the stimulus, the targets or the tested neurons. This is stricter than a random draw: it asks whether these neurons matter more than other active neurons of the same kind, not whether silencing active neurons does anything. Four confirmatory tests: each set after pIP10 and after P1, one-sided for a rise of dPR1, as in T1. Silencing both sets together is reported as descriptive.

Decision rules. The 8 confirmatory tests (T1 and T2) share one Bonferroni threshold, p ≤ 0.05 / 8 = 0.00625. T1 is confirmed for a category if both cases pass. The loop (T2) is supported if both sets pass after pIP10, the most direct test; partly supported if only dMS9 passes (dMS9 restrains dPR1 through some route, not necessarily the named inhibitors); not supported otherwise. The P1 results qualify the verdict but do not change it. Known beforehand: in the v0.5 per-type scan, silencing vPR9_a alone after pIP10 lowered dPR1 by 9 %, and IN00A038 was never tested (it is not fru+).

Execution notes (added 2026-09-22, before any confirmatory result)

• Runs stop at the end of the stimulus. The tests use only the target rate in the 0–300 ms window. The Poisson input is drawn for the stimulus only and the dynamics are causal, so stopping each run at 300 ms instead of 600 ms gives bit-identical rates. This was checked on the real data for pIP10, P1 and looming, through sim.run and validate.respond, and it is locked by a unit test. It costs 2–2.7× less.
• Confirmatory jobs run first, so that an interrupted run has the 8 decisions before the descriptive tests.
• Interruptions. A first run, which stored nothing until the end, was stopped after 5 hours because the CPU was thermally capped at 900 MHz and would have needed about 20 hours; its output was lost. The runs now save every null draw (--checkpoint). A resumed run replays the saved picks, each checked against its hash, and simulates only the missing draws.
• Disclosed peeks. End-to-end checks of the code on the real data used 2 and 4 null draws. They showed silencing dMS9 after pIP10 raising dPR1 by about 67 %, and silencing the inhibitory feedback lowering it by about 6 %. No rule was changed after them.

Limits

• Annotations, not measurements: fruDsx and dimorphism are the release's best calls; the "potentially" levels are included.
• Enrichment says where the simulated activity goes. Silencing tests necessity with 1,000 null draws: the T1 null matches superclass, and the T2, T3 and T4 null matches the case's responders, superclass and transmitter sign. The per-type table above was run with no null. T4 is the pre-registered null for the one bold claim that T2 and T3 do not already test.
• Neither null matches proximity to the stimulus.